摘要:

Analyst, August 1996, Vol. 121 (993-1001) 993 Deconvolution and Spectral Clean-up of Two-component Mixtures by Factor Analysis of Gas Chromatograph ic-Mass Spectrometric Data Peter Hindmarch, Cevdet Demir and Richard G. Brereton* School of Chemistry, University of Bristol, Cantock's Close, Bristol, UK BS8 ITS A method is proposed for the deconvolution of GC-MS data. It is applied to two datasets, both of mixtures of the closely eluting and spectroscopically similar compounds salbutamol and clenbuterol. In one set the components are well resolved, whereas in the other, one peak is broadened so that there is complete overlap. After visually looking at the data using principal components plots and baseline correction, the first step is to choose masses characteristic of each component in the dataset.These are then used to estimate elution profiles. Forty masses, ranked in order of significance according to the value of variance/mean in the dataset are then included, to provide better iterated mass spectra and improved elution profiles. Finally, all masses are included. The method is shown to work well even when there is no selectivity in the chromatographic direct ion. Keywords: Factor analysis; gas chromatography-mass spectrometry; deconvolution; clenbuterol; salbutamol 1 Introduction Chemometrics has a long history of use in diode-array high- performance liquid chromatography (DAD-HPLC) for the resolution of mixtures,l-3 and simple approaches such as evolutionary factor analysis3-8 (EFA) have been incorporated into a variety of instrumental software.Most such methods rely on determining so-called composition 1 or selective regions in the chromatogram and using these to guess at the spectra of each pure component, then performing some form of factor trans- formation or rotation. The normal aim in HPLC is the determination of the number of components in a mixture and the spectra and quantities of each component. Factor analysis and chemometric techniques are much less well established in GC-MS. Most methods rely on determining pure masses rather than chromatographic points in time.Y-'3 Each pure mass can then be traced through the chromatogram, giving an estimated elution profile for each component. A major difference between GC-MS and DAD-HPLC is that the former technique results in sparse matrices in which most masses are insignificant or not detected.Hence using the full mass spectral data matrix for factor analysis does not result in good reconstructions, as the data are dominated by zero or baseline readings. In DAD-HPLC, normally all wavelengths have some significance. The potential for factor analysis in GC-MS is large. Particular emphasis can be placed on spectral clean-up. Library searching algorithms often require high-quality reconstructions of spectra. Simply tracing a single diagnostic mass may result in * To whom correspondence should be addressed deconvolution, but not in determining good spectra of pure components. Other problems with GC-MS are that peak shapes are far more commonly distorted, often exhibiting tailing or fronting.Although it is normally possible, experimentally, to tune instruments, this procedure may take several hours, and often with spectra obtained during routine monitoring of a large number of samples, often containing unknowns, there are problems with peak shapes. This can result in the loss of composition 1 or selective regions, meaning that the traditional approach of plotting out a cross-section at a point on the side of each peak in a two-component cluster will not yield the mass spectra of the pure components. Some methods for spectral clean-up such as maximum entropyi4~15 have been proposed, but these normally ignore the second dimension, and act on one- dimensional data. In modern machines, where chromatographic and mass spectral information are acquired simultaneously, it is sensible to take into account information from both dimensions and use chemometric techniques.In this paper, we propose a strategy for the deconvolution of GC-MS data by factor analysis. As an example, two closely eluting and structurally similar compounds, clenbuterol and salbutamol, were chosen. Two datasets are acquired. In set 1, the two components are fairly well resolved, whereas in set 2 the second peak shape is distorted owing to less optimal in- strumental conditions such that there are no composition 1 regions for the fastest eluting component, there being total overlap of the broad peak with the sharp peak. The data were chosen as an example of a method which can, of course, be extended to many other situations. 2 Experimental 2.1 Preparation of Mixture and Standard Solutions Anhydrous quinine was obtained from Fluka (Gillingham, Dorset, UK). Salbutamol, clenbuterol and other chemicals were purchased from Sigma (Poole, Dorset, UK).Trimethylsilyl (TMS) derivatives were prepared by adding N,O-bis- (trimethylsily1)trifluoroacetamide (BSTFA) to the standard samples. Samples were heated at 80 "C for 1 h. Reagents were removed under nitrogen (40 "C). The derivatized samples were re-dissolved in toluene-MSTFA (99 : 1, v/v). Two sets of stock solutions of standards (20 mg ml-1) were prepared separately in this mixture of solvents. Set 1 consisted of 120 ng pl-1 of salbutamol, 120 ng pl-1 of clenbuterol and 80 ng pl-1 of quinine, and set 2 of 40 ng pl-' of salbutamol, 120 ng p1-i of clenbuterol and 80 ng pl-l of quinine.Standard solutions of salbutamol and clenbuterol were prepared in the same manner as set 1. 2.2 GC-MS Conditions GC-MS scans were obtained for the two mixture sets and also for solutions of clenbuterol and salbutamol individually. GC-994 Analyst, August 1996, Vol. 121 MS was performed with a Fisons MD 800 mass spectrometer using the splitless injection technique. Mass spectra were recorded at 70 eV. A fused-silica capillary column (BPXS; 30 m X 0.32 mm id; 0.25 Fm film thickness; SGE, North Melbourne, Australia) was used with helium as carrier gas at 8 psi. The oven temperature was programmed as follows: initial temperature, 90 "C for set 2 and 100 "C for set 1 and the standards; initial hold, 2 min; ramp rate, 20 "C min-1; final temperature, 320 "C; and final hold, 5 min.The transfer line temperature was 280 "C. Full-scan electron impact mass spectra were acquired by scanning the 50-500 u range using an electron energy of 70 eV. Injection volumes were 1 PI. Mass spectra were obtained at a scan rate of 200 min-1 for the well conditioned sets and 100 min-1 for the ill conditioned set 2. 2.3 Software The software for this work was written on a 486 PC using Microsoft Visual Basic for Windows 3 .O Professional Edition. Numerically intensive routines were written as a Dynamic Link Library in C using Microsoft Visual C++ for Windows 1.0. Validation of numerical methods was done by using Microsoft Excel 4.0 spreadsheets. Methods such as principal components analysis (PCA) were independently written in Excel and numerical results, identical with those reported in this paper, were obtained.Data were acquired from the Fisons MD800 mass spec- trometer and decoded using a file format provided by Fisons. 3 Methods 3.1 Baseline Correction Before performing any analysis, it is useful to perform baseline correction. The first step is to select regions of the chromato- gram which represent only noise. The determination of these regions can be assisted by plotting the logarithm of the sum of squares (LSS) of the raw MS data at each elution point i: Plotted graphically, the logarithmic sum of squares values give a very sensitive indicator of the noise regions. Once the noise regions have been identified and selected, the next step is to calculate the mean intensity for each mass number over the selected baseline regions.These means are subtracted from the raw data to give the baseline-corrected data set, bX. A matrix is then formed from the baseline-corrected data and the significant masses, chosen as described below. 3.2 Pre-processing Before performing any data analysis, various approaches to pre- processing the data can be used to obtain more meaningful results and diagrams. 3.2.1 Significant musses Unlike DAD-HPLC, GC-MS results in sparse data matrices, i.e., the intensity at many mass numbers is zero, or close to zero, owing to noise. Only a few mass fragments are significant for data analysis and all others contain no useful information. Poor results are obtained if all masses are used. For example, if a data matrix is used consisting of all ions obtained between mlz = 1 and 1000 and SO data points in time, of the 50 000 elements of the matrix only a few per cent are useful.Including non- essential mass numbers causes problems because ions due to noise are included and the algorithms are slower and possibly less accurate. A second consideration is that low mass numbers have little discriminating power as they represent volatiles, e.g., atmo- spheric species, low-mass hydrocarbon ions and solvent peaks. Hence a lower cut-off of mlz = 50 or 100 is established. The M most significant masses are then selected. The choice of M can be determined visually by performing PCA on an increasing number of masses until there is no significant improvement in the scores plot. This change in appearance of the scores plots can be numerically quantified using Procustes analysis.16 Too few mass number will not give a sufficient representation of the data, but too many mass numbers may introduce unwanted noise.The significance of each mass number is determined by calculating the ratio of the variance to the mean of the intensities as follows: where I = c !xi,; / I i=l 1; is the variance/mean ratio and bEj is the mean of the intensities of mass number j and I is the number of data points in the chromatographic direction. The M mass numbers with the highest ratio can then be selected. The reduced baseline corrected matrix b y , is formed from the corresponding columns of bX. 3.2.2 Stundurdizution Another transformation involves standardizing the data.For raw data, mass numbers with intense peaks dominate the analysis. However, these may not necessarily be the most diagnostic masses. A solution is to transform each element of the data matrix bX i.e., bjl, to a standardized value, bx$ according to the equation ( 3 ) where bzj is the mean of each column of the matrix, i.e., the mean intensity of each mass number across the region of interest. The advantage of this is that the low-intensity masses which are diagnostic of components become more significant, as shown below. 3.3 Principal Components Analysis The next step is to determine how many peaks are in a cluster, where they elute and the characteristic ions representing each component, PCA3,I7 is used for the first stage of exploratory data analysis (EDA). Strictly, conventional statistics defines PCA as singular value decomposition (SVD) on a mean centred data matrix, looking at random variation about a mean.However, in most chemometrics, variation above a baseline is more relevant so decomposition is performed on uncentred data. The terminology PCA is usually employed in the chemometrics literature but it is important to distinguish between the methods.Analyst, August 1556, Vol. 121 995 The effect of centring is discussed elsewhere.1.10 In this paper, PCA is used on uncentred data and decomposed using the NIPALS algorithm. If the elution profiles of each component are characterized by an I X K matrix C , where K is the number of components and I is the number of points in time, and the spectra of the components by S, a K X J matrix where J is the number of mass numbers, then (4) where X is the observed chromatogram obtained by GC-MS and E is the error between the actual and observed data.PCA is used to compute the scores T and loadings P so that J,JX = I,KTK.JP + R ( 5 ) where k is the number of principal components used to describe the system and R is the residual error between the actual and modelled data. The scores and loadings of each PC represented by T and P, respectively, can then be displayed graphically. A principal components scores plot displays the data in a two-dimensional graph which depends upon the characteristics of the data, which are often called latent projection graphs.18 Ideally, each pure component is represented by a linear segment. The shape of the graph depends on how closely the components elute, how similar their spectra are and their relative concentrations.Loadings plots show the relationships between the variables, i.e., mass numbers in GC-MS 3.4 Deconvolution algorithm Methods for the factor analysis of chromatographic data tend to be based upon determining regions in the elution direction where only one component is eluting/absorbing. These ‘compo- sition 1 ’ regions can be used to make an estimate of the spectral data and further regression gives the estimated elution profiles. In GC-MS, peaks are often overlapped so these methods cannot be used. This paper proposes a method which selects diagnostic mass numbers for each component to give the first estimate of the individual components so overcoming this problem.3.4.1 Determination of diagnostic mass numbers The first step in mass-based deconvolution is to determine which mass numbers are most characteristic of each component in the system. Finding the masses provides first estimates of elution profiles for each component. Several approaches to this mass selection are compared in this paper. ( i ) Purity index. The use of purity indices was originally proposed by Windig.19.20 The method is analogous to finding the variance/mean formula for determining the most significant masses, but is modified to account for the effects of noise. The first purity index of a mass number m, is defined as follows : where is the intensity at scan number i, mass number m in the reduced baseline corrected dataset and b?, is the mean intensity of mass number m.The parameter h is a weighting constant to reduce the significance of low-intensity peaks whose variability might simply be due to noise, giving a mean intensity close to zero. In this paper, h is set to be equal to 1% of the maximum value of by,. The first diagnostic mass, z l , is the mass number with the highest first purity index. The second purity index for each mass number is calculated by the use of the correlation coefficient between the elution vectors at each mass bym with the vector at the first diagnostic mass, byz,. This is given by (7) where r(by,,by,,) is the Pearson correlation between the two vectors. In general terms, the correlation between vectors l,Ma and 1.Mb, each with M elements, is defined as M c ( a m - Z)(h, - b) (8) m=I r(a, b) = where a, and b, are the elements and a and 8 are the means of the elements of a and b, respectively.The second diagnostic mass, z2, is given by the mass number with the highest second purity index. Further purity indices can be extracted by matrix determinant based methods.20 (ii) Variance plots. The diagnostic masses for each compo- nent can often be seen by looking at loadings plots. The remaining methods of variance plots and diagnostic loadings attempt to extract these masses algorithmically. The variance plot method2I uses the density of points in the PC 2 versus PC 1 loadings plots to determine the masses associated with each component, and selects the most diagnostic for each. The PC plot is divided into equal angular segments, which in this case were sct to be 10 ’.Fig. 1 ( a ) shows a set of simulated loadings with the 10 O segments. The distances from the origin of all mass numbers in the loadings plot are summed for each segment. This is the total sum of squares vd for each segment d and is given by m E d (9) where rn E d signifies that the sum of squares is performed on all mass numbers rn which fall in segment d. Segments for which vd is greater than 10% of the maximum value of vd are identified and grouped together into clusters. These clusters of segments 1 and 2, and 5 and 6, are marked on Fig. l(a). The number of clusters found corresponds to the number of pure components detected. The mass number in each cluster with the maximum value of pl,,2 + p2,,?2 is chosen as the most diagnostic mass number for that component.Unlike the purity index method, this algorithm needs no prior knowledge of the number of components in the system. (iii) Diagnostic loading. This final method uses trigonometry to determine the diagnostic masses from the loadings plots. The distance, u,, of each mass number from the origin is calculated: (10) 2 2 u, = + P2,m The mass number with the maximum value of u,, is taken as being the first diagnostic mass z I . The angle between the vectors from the origin to z l and from the origin to each other mass number is calculated using the dot product: (1 1)996 Anulyst, August 1996, Vol. 121 where wzl is the vector from the origin to z1 and w, is the vector from the origin to loadings rn. Fig. l(b) shows this angle drawn between the two most diagnostic masses.Finally, the measure of the significance of the distance, 1, , of each mass number from the first diagnostic mass is calculated by multiplying the angle calculated above with the distance from the origin of that point: I, = 0,U, (12) The mass number with the maximum value of 1, is taken to be the second diagnostic mass z2. The method may be extended to determine the diagnostic masses for further components by calculating the significance of the separation of mass numbers, other than 2 1 , from z2. 3.4.2 Estimation of elution projiles and spectra The initial estimate of the concentration profiles l t k is taken to be the columns of hY representing each diagnostic mass, which is equivalent to performing selective ion monitoring (SIM) for ecch mass.By combining the vectors ' t k , a concentration matrix 1C is formed. This matrix can be used with the reduced baselin: corrected matrix hY to give an estimate of the mass spectra, ' S , based on the M significant masses. Since cannot be solved for 1s directly as it is not a square matrix, and hence has no inverse, the pseudo-inverse can be used as follows, 1,dY = I , K 1 e * K , M l S (13) Fig. 1 (a) Variance diagram of simulated loadings in using 10" divisions. Diagnostic masses, Z , and 2 2 are shown. (b) Loadings plot showing angle between diagnostic masses used in diagnostic loading method. 15 = (l*le)-l.l&.by (14) 2 e = bX. li&lglgl)-l 2 3 = ( 2 p 2 e ) - 1.2e.q (16) This estimate is then used to give an improved estimate of the elution profile by using the full baseline corrected matrix, hX: (15) A final regression using 22 and the full baseline-corrected matrix bX gives full estimated mass spectra: Further iterations are possible but it was found that no further improvement is gained.4 Results 4.1 Datasets The total ion current (TIC) chromatogram of set 1 is given in Fig. 2(a). The digitization in this case is very good, but typically routine GC-MS is less resolved. This is often due to tailing or fronting. The TIC for set 2 is shown in Fig. 2 (b). In this case, clenbuterol, the slower eluting peak, appears to have a broad peak shape. Data analysis is performed on both sets of data, although detailed illustration is performed using only the second dataset. 4.2 Preliminary Analysis Baseline regions were determined by sum of squares plots around the peaks of interest.This plot is given in Fig. 3. The areas of low sum of squares are identified as 868 to 875 and 906 to 910. Baseline correction is performed using the means method as described above. The 40 most significant masses between rnlz = 100 and 400 for set 2 were selected using the variance/mean criterion described above. The low-mass cut-off was chosen to remove low-mass volatile and highly intense solvent peaks from the analysis. These masses are given in Table 1. PCA was performed on an uncentred matrix formed by the sequential, unstandardized, baseline corrected data of set 2 Salbutamol Clenbuterol 1 $ UL , , , , , . , , ,JI Q) c 1250 1350 1450 1550 1650 1750 1850 1950 2050 - .- 840 940 1040 1140 1240 1340 Scan number Fig.2 Total ion current chromatograms for (a) set 1 and (b) set 2 with salbutamol and clenbuterol peaks enlarged.Analyst, August 1996, Vol. 121 997 1 ,. 860 880 900 920 Scan number Fig. 3 Expanded TIC chromatogram for set 2 showing a logarithmic sum of squares plot. Table 1 Forty most significant masses, mlz = 369, 370, 262, etc., for set 2, selected by the variance/mean criterion 369 187 166 132 370 174 265 116 262 212 119 117 371 245 193 190 147 189 294 220 264 176 246 277 192 188 133 175 243 263 350 372 173 281 177 168 207 191 205 260 O264 "243 between scan numbers 876 and 904 with these 40 significant masses. The score and loadings plots are given in Fig. 4(a). Several important features can be noted.First, it can be seen that the mass numbers characteristic of each component appear in a similar position in the loadings plots as the purest times in the scores plots. This relationship changes dramatically when the data are standardized, as can be seen in Fig. 4(h). The mass numbers lie approximately on the circumference of a circle. If two standardized PCs perfectly described a system, then the loadings would fall perfectly on a circle; however, the presence of interferents and noise may cause deviations from perfect behaviour. By varying the position of the low-mass number cut- off, this circular fit can be improved as low-mass volatiles, which have a higher degree of noise, are removed from the calculation. In the standardized loadings plots it can also been seen that the significant masses for each component are at the end of the arc (e.g., m/z = 369 for salbutamol and m/z 262 and 264 for clenbuterol) whilst masses common to both components lie in the centre.4.3 Selection of Diagnostic Masses The selection of diagnostic masses was repeated on the same subset of baseline corrected, unstandardized data using the thrce methods described in Section 3.2.2. The loadings plot con- cerned is given in Fig. 4(a). As can be seen, the loadings clearly divide into two components, one with mlz = 369 being the most prominent and the second having m/z = 262 and 264 close together . "1 73 (I. Fig. 4 Scores and loadings plots for Set 2: (a) unstandardized data and (b) standardized data, 40 significant masses.998 Analyst, August 1996, Vol.121 The diagnostic masses found for sets 1 and 2, determined by the methods described above, are summarized in Table 2. For the unstandardized data for set 2, the purity index method gave diagnostic masses of mlz = 369 and 173. The table of purity indices for each component in set 2 is given in Table 3. The variance plot method identifies two components, the first having a diagnostic mass of mlz = 369 and the other mlz = 262. The same masses were obtained by using the diagnostic loading method. For set 1, a diagnostic mass of mlz = 369 is predicted for salbutamol by all three methods. The variance and diagnostic loading methods estimate a mass of rnlz = 261 for clenbuterol whereas the purity index predicts mlz = 242. Diagnostic masses can also be extracted using standardized data using the variance and diagnostic loadings methods.The purity index is not suitable for data which have been standardized as this pre-processing results in each column of the data matrix having a standard deviation of one and a mean of zero, so eqn. (6) would become a division by zero function in all cases. For set I , the variance plot for standardized data identifies a diagnostic mass for clenbuterol of rnlz = 264, which is the same as for the non-standardized data. The diagnostic mass for salbutamol, however, is identified as mlz = 207. From the significant masses listed in Table 1, this mass is the tenth most significant of all masses, which suggests that the variance plot may not be as suited to standardized data as it is to unstandardized data.An explanation for this is that standardi- zation causes all loadings to be projected approximately on to the circumference of a circle as described above. The variance plot then extracts from each angular segment the loading with the maximum sum of squares or distance from the origin. However, owing to deviation from a perfect circumference and equal variances due to noise and interferents, the loading selected will not necessarily be the most diagnostic for that component. This problem is not encountered when using unstandardized data, as the loading with the greatest distance for the origin in each segment is clearly the most diagnostic. A similarly poor estimate is found for the standardized variance plot for set 2. The diagnostic loading method performs slightly better on standardized data as eqn.(1 2) includes a contribution from the Table 2 Diagnostic masses determined for salbutamol (Sal) and clenbuterol (Clen) using the methods described with non-standardized and standardized data for each set Set 1 Set 2 Non-Stand. Stand. Non-Stand. Stand. Sal Clen Sal Clen Sal Clen Sal Clen ~ ~ _ _ _ _ _ _ _ Purity index 369 242 - - 369 173 - - Variance plot 369 261 207 264 369 262 191 173 Diagnosticloading 369 264 369 212 269 262 369 173 Table 3 Best purity indices for each component in set 2 Salbutamol Clenbuterol rnlz Purity index ml: Purity index 369 3.012494 173 2.530117 147 2.951753 164 1.212402 370 2.942495 166 1.184312 371 2.818719 138 1.155183 207 2.67993 264 1.115124 angle of the loading, so loadings, towards the ends of the arc are preferred.4.4 Deconvolution Deconvolution was performed on the two baseline-corrected datasets, sets 1 and 2. Set 2 was pre-treated as in Section 4.2. Set 1 was baseline corrected by the means method using noise regions of sequential points 1245 to 1255 and 1309 to 1317. Deconvolution was performed using the diagnostic masses of mlz = 369 and 264 for set 1 and mlz = 369 and 262 for set 2 for the two components in each set. In the estimate of set 1, two well resolved profiles were observed. In set 2, however, there is a total overlap of the peaks, meaning that methods based on finding 'composition 1 ' regions in the chromatographic dimen- sion for the first peak, salbutamol, would not have been successful. The peak shape of the second component, clen- buterol, is less well defined, possibly owing to the lower digitization of this sample.These two profiles are then used to calculate estimated mass spectra. for the two components in each set, based on the 40 most significant masses determined by the variance/mean criterion. Regression on to the whole datasets gave final estimates of the profiles and are given in Fig. 5(a) and (b). The higher djgitization used for set 1 is apparent as the profiles produced for set 1 are much smoother than those for set 2. Finally, predicted spectra for each component based on all mass numbers between mlz = SO and 400 are given for set 2 in Fig. 6 and for set 1 in Fig. 7. Below rnlz = 100 the fragmentation patterns are similar for each case and so provide little information as to the ability of the method to deconvolute the spectra.Similarly, above rnlz = 300 clenbuterol has no ions whereas salbutamol has an intense cluster at rnlz = 369. More interesting are the results of deconvolution between rnlz = 200 and 300 which contain common ions in addition to the less intense characteristic peaks for clenbuterol. For each mass spectrum, the region mlz = 200-300 is shown zoomed. Key results to note are that in the estimates for I ' J' i,, \, '\ 1267 1294 676 Scan Number 904 Fig. 5 Final estimated chromatograms for (a) set 1 and (h) set 2.Analyst, August 1996, Vol. 121 999 ( i i ) 1M =.. v) a, a, c .- c .- .- c - a, [r: 0 k36 p" p" x v) a, Y .- Y .- B6 50 1W 200 mlz Fig. 6 Mass spectra of (u) salbutamol and (h) clenbuterol showing (i) estimate after deconvolution of set 2, (ii) estimate at apparent composition 1 point in set 2 and (iii) spectrum of the pure salbutamol or clenbuterol standard.In each case the region mlz = 200-300 is enlarged.1000 Analyst, August 1996, VoE. 121 salbutamol in set 1 there is little, or zero, intensity at mlz = 264 which is characteristic of clenbuterol. Fig. 6 includes spectra of the pure salbutamol and clenbuterol standards. Visually, there is a very good correspondence between the pure and estimated spectra of both salbutamol and clen bu terol. Closer inspection of the zoomed regions also confirms the quality of the estimate. For example, in the spectrum of salbutamol for set 2, the cluster of ions at, mlz = 262-268 is very well predicted .It is interesting that the mass number 262 is present in the pure standard even though is scores highly as a diagnostic mass for clenbuterol. This is an important result as it demonstrates that the method can extract ions common to both components in their correct ratios. The relative intensities for the ions in the clusters at mlz = 203-209, 218-222, 276-283 and 294-296 are also very well predicted. Similarly good results can be seen by comparing the predictions of clenbuterol in each set. Conventional methods for extracting mass spectra of in- dividual components from overlapping GC-MS peaks rely on taking a cross-section through the peak where it was assumed that the only one component was eluting i.e., composition 1. Fig. 6 includes these cross-sections for set 2 taken at the leading edge of the first peak and the tailing edge of the second peak.The cross-section for the second component, clenbuterol, is reasonable but for salbutamol it is very poor with ions at mlz = 262 (clenbuterol) and mlz = 369 (salbutamol) detected. Clearly, from the profiles determined above this is due to the fact that the salbutamol peak is totally overlapped by the clenbuterol peak. Normally, however, there is no prior knowledge of peak shapes. The components in set 1 are well resolved so conventional methods would have been sufficient to deconvolute the peaks. It has been shown that the deconvolution methodology presented here produces better predicted spectra than con- ventional methods for cases where a pure composition region can be found.In other cases where there is total overlap of peaks, as is the situation for salbutamol in set 2, the methods presented here produce equally good estimates. 5 Conclusions It has been demonstrated that chemometric methods work well both for deconvolution and spectral clean-up, even in situations where there are no selective regions. Crucial to the methods is the choice of the first significant mass, which, if badly chosen, can result in poor final results. It is particularly important that there is high correspondence between the reconstructed mass spectra for both components and the mass spectra of the pure compounds, suggesting very faithful reconstruction. The method proposed in this paper involves repeating calculations using increasing numbers of masses in three stages. The method can be extended to clusters of more than two components.This depends partly on determining how many peaks are present in a cluster and also on identifying diagnostic masses for each component. If one component is omitted then the information from it will be distributed amongst the other components. With the advent of modestly priced benchtop GC- MS instrumentation, it is likely that most mass spectra will be acquired using this method rather than straight mass spectrome- try, so information from the second, chromatographic, dimen- sion will normally be available. The authors are grateful to P. D'Arcy of Fisons Scientific Equipment Division for providing the LAB-BASE data format which made this work possible and to EPSRC for financial support.Appendix List of Notations Used Throughout this paper matrices are given in upper case bold and italic characters, e.g., S , vectors are given in lower case bold italic characters, e.g., w,, and scalars are giv_en in roman. Estimated variables are denoted by a 'hat', e.g., ' C . Dimensions of matrices are given as left-hand-side subscripts, e.g., 20,1c$ is a matrix of 20 rows by 10 columns. General I Number of scans 1 Individual scan number J j Individual mass number M m Individual significant mass Number of mass numbers between upper and lower limits Number of significant masses selected w.. . mo 4m 100 2cn rnlz 5D Fig. 7 Final estimates of (a) salbutamol and (h) clenbuterol in set I with the region mlz = 20&300 enlarged.Analyst, August 1996, Vol.121 1001 Diagnostic mass selection qnl. 1 First purity index for mass number m Yn1,2 Second purity index for mass number m h Noise weighting constant for purity index r Pearson correlation coefficient d Individual segment in variance plot L'd Sum of squares for segment d U , ? Distance of each mass number from origin on 1, Measure of significance of mass number m from wn, Vector from origin to point m on loadings plot zh Diagnostic mass number for component k loadings plot the first diagnostic mass GC-MS data matrix in region of interest Intensity at scan number i, mass number j in X Baseline-corrected GC-MS data matrix in region Baseline-corrected matrix reduced number of Intensity at scan number i and significant mass Scores Score at time i and component k Loadings Loading for component k at mass m Total number of components Individual component Elution profile for component k Mass spectrum for component k Matrix of all elution profiles Matrix of spectra of interest masses number m D_econvolution results 'C 19 Matrix of first estimates of concentration profile Matrix of first estimates of mass spectra 2e 2s Matrix of second estimates of concentration Matrix of second estimates of mass spectra profile References 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 Brereton, R.G., Analyst, 1995, 120, 2313. Windig, W., Chemom. Intell. Lub. Syst., 1992, 16, I . Mainowski, E., Factor Analysis in Chemistry, Wiley, New York, 2nd edn., 1991. Maeder, M., and Zilian, A., Chemom. Intell. Lub. Syst., 1988, 3, 205. Liang, Y.-Z., Brereton, R. G., Kvalheim, 0. M., and Kahmani, A., Analyst, 1993, 118, 779. Keller, H. R., and Massart, D. L., Chemom. Intell. Lab. Syst., 1992, 12, 209. Brereton, R. G., and Elbergali, A. K., J . Chemom., 1995, 8, 423 Brereton, R. G., Gurden, S . P., and Groves, J. A., Chemom. Intell. Lab. Svst., 199.5, 27, 73. Brakstad, F., Chemom. Intell. Lab. Syst., 1995, 29, 157. Lee, T. A., Headley, L. M., and Hardy, J . K., Anal. Chem., 1991,63 357. Davis, J . E., Shepard, A., Stanford, N., and Rogers, L. B., Anal. Chem., 1974, 46, 821. Rozett, R. W., and Petersen, E. M., Anal. Chem., 1975, 47, 1301. Rozett, R. W., and Petersen, E. M., Anal. Chem., 1975, 47, 2377. Sibisi, S., Skilling J., Brereton, R. G., Laue, E. D., and Staunton, J., Nature (London), 1984; 311, 748. Ferrige, A. C., Seddon., M. J.. Skilling, J., Ordsmith, N., Rapid Commun. Muss Spectrom., 1992, 6, 765. Demir, C., Hindmarch, P., and Brereton, R. G.. Analyst, 1996, in the press. Wold, S., Esbensen, K., and Geladi, P., Chemom. Intell. Lab. Sjist., 1987, 2, 37. Kvalheim, 0. M., Clzemonz. Intell. Lab. Syst.,1987, 2, 283. Windig, W., Chemoni. Intell. Lab. Syst., 1994, 23, 71. Windig, W., Anal. Chem., 1991, 63, 1425. Brereton, R. G., Chemometrirs: Applications of A4athematir.s and Statistics to Luhorutory Systems, Ellis Horwood, Chichester, 1993. Paper 6/01 502 F Receiijed March 4, I996 Accepted May 7,1996

ISSN:0003-2654    

DOI:10.1039/AN9962100993

出版商:RSC    

年代:1996

数据来源: RSC

摘要:

Analyst, August 1996, Vol. 121 (1003-1008) 1003 End-point Determination On-line and Reaction Co-ordinate Modelling of Homogeneous and Heterogeneous Reactions in Principal Component Space Using Periodic Near-infrared Monitoring Timothy Norris and Paul K. Aldridge Pfzer Central Research Laboratories, Groton, CT 06340, USA Advancement of organic homogeneous and heterogeneous reactions to their steady state end-points has been determined directly using near-infrared spectroscopy and fibre optic probes. The reaction times can be obtained without specifically measuring decay of individual starting materials or formation of products. The technique is passive and can be generally applied to most organic reactions in the laboratory or chemical plant. This report uses the formation of zopolrestat ethyl ester and its saponification to zopolrestat sodium salt, as typical examples of a heterogeneous and a homogeneous reaction matrix, respectively.Data has been collected directly from full size chemical reactors. The steady state is detected when the change in the spectra does not change significantly with time. The end-point can be confirmed computationally shortly after the data has been collected using open ended models. These can be constructed directly from the spectral data using commercially available software. The end-point can be confirmed with multivariate calibration methods using dendrograms derived from hierarchical cluster analysis or scores plots obtained from principal component analysis. The reaction co-ordinate can be modelled in principal component space using the locus of the spectral scores plot.Keywords: End-point determination; reaction co-ordinate modelling; principal component analysis; near-infrared monitoring; kinetics; hierarchical cluster analysis; on-line analysis; non-destructive testing Introduction The understanding of how a reaction proceeds to completion is at the core of chemistry. We have given much thought over the last five years as to practical and theoretical methods that might be used to determine easily the true rate of advancement of a reaction and thus determine the time to reach the end-point. We wanted to develop a general and preferably simple methodology that could be widely applied to many organic reactions. In addition we wanted to use a passive technique that did not involve sampling or perturbation of the reaction matrix.As we had a self imposed requirement of ease of use and general applicability, we did not want to relate our models to closed calibrations based on changes of individual components of the reacting system such as concentration of a specific starting material or product molecule because these determinations are elaborate and generally cannot be obtained without sampling and perturbing the reaction matrix. During the course of this work the approach we have taken has been to model the reaction co-ordinate in principle component space using multivariate calibrations that provide open ended empirical models, which can be used to determine the reaction end-points of both heterogeneous and homogeneous organic reactions. Recent literature covering multivariate analysis in process analysis' and use of near-infrared (NIR) to optimize plant operations in a general way2 have been published, but formal general methods of combining the mathematical approaches and spectroscopy are not yet common, particularly in the study of organic chemical reactions.This is partly due to the fact that diverse disciplines, mathematical modelling, organic chemistry, software science, spectroscopy and reaction kinetics have to come together with a unified objective. The desire to make measurements directly from the reaction matrix in both homogeneous and heterogeneous reactions led us to choose the long wavelength NIR region of the spectrum at A l F J Fig. 1 Schematic diagram of the experimental apparatus using large scale chemical reactors.A, agitator shaft; F, fibre optical bundle; S, NIR spectrometer; J, vessel heating jacket; CF, communication fibre; L, quartz lens; PC, remote computer. Fig. 2 Schematic diagram for NIR probe configuration to collect spectra from ( a ) a homogeneous reaction; and (b) a heterogeneous reaction. X, crystal particle in slurry; M, stainless steel mirror; L, NIR light path; Q, quartz lens; R, light receiving fibres; T, light transmitting fibres.1004 Analyst, August 1996, Vol. 121 1100-2500 nm. This region contains information from combi- nation, first and second overtone bands of molecular vibrations. Fibre optical bundles were used to transmit and receive NIR radiation to and from the reaction matrix to the spectrometer.In practice, we demonstrated the spectral collection procedure in the laboratory and in full scale chemical plant reactors. As a result, equipment design considerations limited our practical operation to the region 1100-1800 nm for heterogeneous reactions and 1 100-2 100 nm for homogeneous reactions. Theory The measurement of true rate of a chemical reaction3 as defined in terms of its advancement E is given by eqn. (1). aE at True rate = - If aE/dt represents the true rate in terms of E for the general chemical reaction given by eqn. ( 2 ) , it can be determined that when the reaction advances by an infinitesimal amount the change in the amount of starting material A, is reduced by -vAdE and the amount of product Q is increased by +vQd5 where Y is the number of moles.Similar increases in product and decreases in starting material apply to the other reacting molecules of the reaction. vAA + vBB+ ....... + v ~ Q + vRR+ ....... (2) At equilibrium ( g ) L T YQpQ + YRpR + ....... - v A ~ A - Y B ~ B - ....... = 0 where G is the Gibb’s free energy, and p the chemical potential 1.3, I 1 I S PF E CTT aI 2 0 4 a I100 1200 1300 1400 1500 1600 1700 1800 P d o Goo l&O 1400 lk0 1600 1;oo ,I, Wavelengthhm Fig. 3(a) Typical raw NIR spectrum obtained from Reaction (I); and (b) selected raw NIR spectra (spectrum number shown at right) typical of characteristic stages of Reaction (I). S, solution stage; PF, initial product formation; E, end-point; CT, crystal form transformation in slurry; I, initial crystal slurry; PI, reaction progression to solution stage; P2, reaction progression to end-point; P3 and P4, progression during crystal form transformation.Analyst, August 1996, VoE.121 1005 In other words when the reaction reaches its free energy minimum its advancement ceases. This is equally true of an irreversible reaction. These changes in free energy and hence reaction advancement can be tracked by NIR monitoring of the changes in concentrations of the species in the mixtures. The NIR region of the electromagnetic spectrum can be used to obtain spectral information by passing a beam of radiation via a fibre optical bundle directly into the reaction mixture without perturbation of the reaction matrix. As a reaction advances along its reaction co-ordinate to its equilibrium state or free energy minimum for an irreversible reaction, changes that occur in the reaction matrix due to bond forming and bond breaking processes will be reflected in changes observed in the NIR spectra obtained periodically by a suitable spectrometer.(The energy of the NIR spectral region between 1100 and 2500 nm contains absorption peaks that originate from the first and second overtones and combination bands of predominantly C- H, N-H, and 0-H molecular vibrations.) When spectral data can be obtained directly in this way at the equilibrium point the following conditions will hold: alNIRl = 0. at Assuming the molecular changes in question give rise to NIR absorbances. At the approach to equilibrium or free energy minimum in the case of an irreversible reaction when the m - 1-th and m-th spectra are separated by a time period At the following condition will hold: I NIR I = I alaz .........an I where a, is the spectral amplitude at wavelength n. As a result at the equilibrium end-point of the reaction, NIR spectra taken periodically in the way described will super- impose, or very nearly so. Multivariate calculations with commercially available software that provide open ended models can be used to provide computational confirmation of d the end-point. PCA scores plots can be used to mimic the macroscopic reaction co-ordinate as the reaction moves to its steady state energy minimum. Experimental Heterogeneous Reaction Study The heterogeneous system we studied [Reaction (I)] the coupling reaction4 between ethyl 3-cyanomethyl-4-oxo-3H- phthalazin- 1 -yl acetate, 1, and 2-amino-4-(trifluoromethyl)ben- zenethiol hydrochloride, 2, in refluxing ethanol at = 80°C to yield ethyl 4-0x0-3-{ [5-(trifluoromethyl)benzothiazol-2-yl]- methyl] -3H-phthalazin- 1-yl acetate, 3, the ethyl ester of zopolrestat.* Homogeneous Reaction Study The homogeneous system studied, [Reaction (II)] was the hydrolysis of the ethyl ester of zopolrestat, 3, to zopolrestat sodium salt, 4, in aqueous ethanol in the presence of sodium hydroxide at = 40°C.Apparatus The reactions above were studied in the laboratory using standard glassware modified to accept the NIR probe and at large scale using 4000 and 8000 1 glass-lined steel reaction COzCzH5 cF3 C02C2H5 HCI C2HS0H $N N@+ (I) k 2 - S - -80 "C dy \ N-CN +yJCF3 HS 0 0 NH4CI 1 2 3 -- NaOH 3 a * Zopolrestat is the WHO approved name for ethyl 4-0~0-3-{[5-(trifluo- romethyl)benzothiazol-2-yl]methyl} -3H-phthalazin- 1 -yl acetic acid, an aldose re- ductase inhibitor.- x 10- I 1 I I 4 2 0 e -2 s: z a, C 4 -6 -8 Wave I e ngt h/n m Fig. 4 Typical second derivative NIR spectrum obtained from Reaction (I). 00I006 Analyst, August 1996, Vol. I21 vessels. The general equipment set-up for chemical reactors is shown schematically in Fig. I , The reaction vessel and agitator were made of glass-lined steel and was equipped with a jacket for heating and cooling. The reflux condenser system and other ancillary equipment are not shown. The optical fibre bundles used to transmit and receive NIR radiation consisted of at least 210 fibres for each channel.The lens interface between the reaction media and the fibres was quartz and the metal baffle and probe tip were made of acid resistant stainless steel. The NIR spectrometers used were either a NIRS ystems model 6500 or model 5000 process instrument. The spectrometer was connected through a modem and optical communication fibre to a remote computer. The homogeneous reaction was monitored using a probe configuration shown in Fig. 2(a). Typically a light pathlength of 2.0 mm was used (lens-mirror gap 1.0 mm). The probes and spectrometers (NIRSystems) were purchased from Perstorp Analytical (Silver Spring, MD, USA). The heteroge- neous reaction was a thick crystalline slurry, and was monitored using an indeterminate NIR light pathlength in which light transmitted from the probe quartz interface was reflected back into the probe from the crystal surfaces.Heterogeneous media were monitored using the probe configuration in Fig. 2(b), because this media was a thick crystal slurry near the reaction end-point, and though the pathlength was indeterminate, good data could be collected. The alternative configuration Fig. 2(a) became clogged with crystal fragments between the mirror and the quartz interface. The practical maximum length of the fibre bundle for collecting spectra was 3 m. In the case of the homogeneous reaction useful data was collected between 1100 and 2100 nm, and for the heterogeneous reaction between 1100 and 1800 nm. Longer wavelength data had poor S/N character- istics due to attenuation in the fibre optic.The shorter range for the heterogeneous reaction is due to reduced reflector efficiency of the crystals compared to the stainless steel reflectors. 2.5 I 1 I 1 I I I I I 8 2 m .f! cn IqOO 1200 1300 1400 1500 1600 1700 1800 1900 2000 2 -0 Wavelengthhm S E 30 Fig. 5 Typical raw NIR spectrum obtained from Reaction (11) 2. S, start (initial spectrum); P, reaction progression to end-point; E, end-point region (44 superimposed spectra). -3 6x 10 I 1 I I 1 1 I I I I I iqoo 1200 1300 1400 1500 1600 1700 1800 1900 2000 2100 Wavelengthhm Fig. 6 Typical second derivative NIR spectrum obtained from Reaction (11).Analyst, August 1996, Vol. 121 1007 Results and Discussion Typical raw spectral data between 1100 and 1800 nm, collected from a reactor in which Reaction (I) was run is shown in Figs.3(a) and (b). Transformation to the second derivative to reduce baseline shift is shown in Fig. 4. Second derivative data calculated by the Savitzky Golay polynomial filter method was used to compute the models for the heterogeneous reaction. The spectrum of the reaction matrix was measured automatically at a fixed time interval of about 6 min from the commencement of heating to reflux and throughout the reflux period until start of the cool down operation. Similar data from the homogeneous system, Reaction (111, is shown in Figs. 5 (raw spectra) and 6 (second derivative spectra). NIR spectra from the heteroge- neous matrix, Reaction (I), contain absorbances in the regions 1150-1200, 1350-1450 and 1650-1750 nm that correspond to aromatic and aliphatic C-H bonds, due to the second overtones, the first overtones of combination bands, and first overtones, respectively.The region from 1650 to 1750 nm also contains information from the first S-H overtones. S-H bonds and aromatic C-H bonds are involved in the thiazole ring formation reaction and changes in these regions will be reflected in the advancement of the reaction as it progresses towards its steady state end-point. Fig. 3(b) shows key spectra that correspond to Similarity 0 0.8 0.6 0.4 0.2 0.0 E key stages of the reaction as it advances to its steady state. Reaction (11) is a saponification carried out in aqueous ethanol and NIR absorbances in the region 1300-1600 nm contain information connected with 0-H hydrogen bonding that changes significantly during the progress of the reaction.Fig. 5 shows the progression of the reaction to end-point. At the end- point the spectra superimpose to give a characteristically clear, steady state end-point, typical of a homogeneous reaction. These spectra were subjected to hierarchical cluster analysis (HCA) using commercially available Pirouette software6 to produce the dendrograms shown in Figs. 7(a) and (b), respectively. Dendrograms give a measure of similarity of spectra as the reaction proceeds to its steady state end-point. The distance metric used to construct the dendrogram is eqn. (3) where A4 = 2, when the order in M = 2 this corresponds to the Euclidean distance. Principal Component Space ' . CT P4 "': Fig. 8 Macroscopic reaction co-ordinate described in 3D principle component space for Reaction (I) using by scores plots obtained by PCA. I, Initial crystal slurry; PI, reaction progression to solution stage; S, solution stage; Pz, reaction progression to end-point; CT, crystal form transforma- tion; P3 and P4, progression during crystal form transformation.Spectra corresponding to those shown in Fig. 3(6) are numbered. Principal Component Space f . I Fig. 7 Typical Dendrogram obtained from (a) Reaction (I) second derivative spectral data; and (h) Reaction (11) raw spectral data. P, Direction of reaction progression; E end-point region. Fig. 9 Macroscopic reaction co-ordinate described in three dimensional principle component space for Reaction (11) using scores plots obtained by PCA.I, Initial reaction matrix; PI, P2 and P3, reaction progression to end- point; E, end-point.I008 Analyst, August 1996, Vol. 121 nr dab = c [(&, - Xh])‘1l’M (3) J = 1 The method used to establish the linking between spectra was the single link given by eqn. (4). dAB*C = 0.5dAc + 0.5dBc - 0.5 I dAc - dBc I (4) As Fig. 7(a) shows, the dendrogram in region E (spectra 42-50) indicates that the spectral changes in Reaction (I) are small after about 250 min which corresponds to spectra 42. (Spectra were collected automatically, in this case the period between spectral scans was =: 6 min). This gives an accurate estimate (+I 2 min) of when the reaction has reached a steady state for the prevailing reaction conditions. Correspondingly, Fig. 7(b), for Reaction (11) shows that changes are small, starting in region E after about 168 min (spectrum 28).In the case of the homogeneous Reaction (11) it is not necessary to derivatize the raw spectral data in order to obtain satisfactory reaction end-point results and the beginning of the end-point region is more sharply defined. It is interesting to compare the differences between the heteroge- neous, Reaction (I) and homogeneous, Reaction (11). The agitated crystal slurry obtained at the end-point is a two phase system of product crystals suspended in a solvent solution, that also contains unreacted starting materials and side products. Because the slurry is randomized only by mechanical agitation and measured with a probe configuration that uses an indetermi- nate pathlength, there is more noise or variation in the spectra collected near the end-point, than with a homogeneous reaction solution.Nevertheless, the reaction end-point can be clearly determined for the conditions of the reaction using the dendrogram. Other mathematical methods such as PCA also show a more indistinct, but definable end-point for the heterogeneous, Reaction (I), than for the homogeneous case. We have found this to be quite general in other systems studied. Principal component analysis was carried out using Pirouette software and figures generated using a combination of Pirouette and Matlab.7 PCA in Pirouette is done using the NIPALS (non- linear iterative partial least squares) appr0ach.8.~ This produces a subset of the first k principal components without needing to compute all possible factors.The locus of the scores plot in three dimensions of principal component space are illustrated in Fig. 8, for Reaction (I) and Fig. 9, for Reaction (11). Calculation of eigenvalues against each principal component, shows that most of the variance can be explained by three principal components in the case of heterogeneous Reaction (I) and two principal components for homogeneous, Reaction (11). The scores plot, Fig. 8, for Reaction (I) provides a model of the macroscopic reaction co- ordinate described in three dimensions of principal component space. When the reaction reaches its end-point the scores plot forms a cluster because the variation in the spectra between scans reaches a minimum. In the case of heterogeneous reactions the cluster is not as tight as that observed in a homogeneous reaction (see Fig.9) because of small but significant variations in the sample interrogated and the sample path due to random fluctuations in the agitated crystal slurry (see above). Penrose l o tackles the problems of adequately describing the real world in mathematical terms and these concepts are applicable to models and mimics of processes like organic chemical reactions. His general approach led us to consider ways of modelling the progress of a reaction to its end-point at its free energy minimum. Experience shows that at the macroscopic level chemical reactions have predictable out- comes, and intuitively, it seems that any macroscopic reaction co-ordinate mimic or model that plots the advancement of a reaction should exist in a space of two to five dimensions.The reasoning behind this is as follows. When components of a chemical reaction are mixed in the same molar proportions under the same conditions of temperature and pressure, the reaction proceeds to a definable product mixture. Repeat experiments give the same result. The bond forming and bond breaking processes that occur at the microscopic level are reflected at the macroscopic level by monitoring changes in the electromagnetic spectrum of the reaction matrix that occur with time. It is proposed that the locus of the scores plot in principal component space mimics the macroscopic reaction co-ordinate and provides a good model of the rate of advancement of the reaction as it progresses to its free energy minimum.In practice for the reactions we have studied, it can usually be viewed satisfactorily in two or three dimensions of principal component space if the reaction changes are monitored under isothermal conditions. This work illustrates that the NIR region of the electromagnetic spectrum can be used to define such a model, provided direct and non-perturbing periodic observations of the changes that occur in the chemical reaction matrix are made. Certainly, these scores plots provide a useful visualization of the advancement of a reaction to its end-point at the steady state. They can be used to predict the reaction end-point without prior knowledge of the reaction mechanism or detailed calibration of spectra with decay of a reactant or growth of a product. Conclusions We have successfully demonstrated that this experimental methodology can accurately confirm organic chemical reaction end-points in both heterogeneous and homogeneous reaction matrices directly without removing samples from the reactor.The technique has been applied to reactions in laboratory glassware from 500 ml to plant reactors with a capacity of 8000 1. The computed results are available shortly after the end-point event is reached and the speed is only limited by equipment computational speed. Reaction conditions examined for the first time can be evaluated with a single process run and the time to reach the steady state for the chosen process conditions are easily confirmed with a second experiment run under the same conditions. This method of monitoring organic chemical reactions has been used to ensure consistency in batch operations of computer controlled chemical plants. This method has proved very useful for quickly identifying a reaction time with a specific set of conditions. When developing a process for commercialization this saves a lot of time and significantly cuts down on the number of experiments. References 1 2 3 4 5 6 7 8 9 10 Kourti, T., and MacGregor, J. F., Chemorn. Intell. Lab. Syst., 1995, 28(1), 3. Espinosa, A., Lambert, D., and Valleur, M., Hydrocarbon Process., Int. Ed., 1995, 74(2), 86, 91. Atkins, P. W., Physical Chemistry, Oxford University Press, Oxford, 1978, p. 255, 853. Mylari, B. L., Larson, E. R., Beyer, T. A., Zembrowski, W. J., Aldinger, C. E., Dee, M. F., Siegel, T. W., and Singleton, D. H., J . Med. Chem., 1991, 34, 108. Savitzky, A,, Golay, M. J. E., Anal. Chem., 1964, 36, 1627. Pirouette, Multivariate Data Analysis for IBM PC Systems, version 1.2, Infometrix Inc., Seattle, Washington, USA. Matlab, High Performance Numeric Computation and Visualization Software, version 4.2C. 1, The Math Works Inc., Natick, Massachu- setts, USA. Wold, H., Estimation of Principal Components and Related Models by Iterative Least Squares in Multivariate Analysis, Proc. Inter- national Symposium, June 1965, ed. Krishnaiah, P. R., Academic Press, New York, 1966, p. 391. Geladi, P., Anal. Chim. Actu, 1986, 185, 1. Penrose, R., The Emperor’s New Mind, Penguin Books. USA Inc., 1991, New York, p. 176. Puper 6101 61 I A Accepted April 30, I996

ISSN:0003-2654    

DOI:10.1039/AN9962101003

出版商:RSC    

年代:1996

数据来源: RSC

摘要:

Analyst, August 1996, Vol. 121 (1009-1013) 1009 Analysis of Ageing and Typification of Vintage Ports by Partial Least Squares and Soft Independent Modelling Class Analogy M. Cruz Ortiza, Luis A. Sarabiab, Charles Symingtonc, Fernando Santamariac and Montserrat Iniguezc (I Department of Analytical Chemistry, Faculty of Science and Food Technology and Chemistry, University of Burgos, Pza. Misael Bariuelos SIN, 09001 Burgos, Spain 1’ Department of Mathematics and Computation, Faculty of Science and Food Technology and Chemistry, University of Burgos, Pza. Misael Bariuelos SIN, 09001 Burgos, Spain c Gobierno de la Rioja, Consejeria de Agricultura y Alimentacibn, Estacibn Enolbgica de Haro, C.lBret6n de Los Herreros 4, Haro, La Rioja, Spain Vintage port is a wine produced in very limited amounts in the Portuguese region of the Douro Valley.It ages very slowly in bottles and few analytical data are available. Thus, the composition of 24 wines from the cellars of DOW’S and Graham’s,* corresponding to 12 different years between 1963 and 1990, was analysed. Forty-one analytical parameters determined on each sample were used to predict the year using a partial least squares (PLS) regression and to identify the cellar using a soft independent modelling class analogy (SIMCA) model. Both problems fall naturally within the field of food chemometrics and present the statistical difficulty of having fewer objects than variables. It is, therefore, impossible to apply normal multiple regression and discriminant analysis techniques. The PLS model requires two latent variables and explains 97.4% of the variance of the age variable and 90.2% of the cross-validated variance.As for the characterization of both brands, a SIMCA model with two components for each model correctly classified 22 of the 24 samples; it gives a sensitivity of 83.3% and a specificity of 91.7% for DOW’S vintage ports, whereas for Graham’s vintage ports the sensitivity and specificity are 91.7 and loo%, respectively. Keywords: Pattern recognition; soft independent modelling class analogy; regression; partial least squares; CieLab; vintage port; wine ageing; wine colour Introduction Oporto wine originates in the Portuguese Douro Valley and represents 65% of the total production in the valley. Its production begins with a fermentation period (averaging 36 h) to extract the maximum colourant.Once the sugar content has reached 12.54% m/m, the wine is removed from the fermenta- tion vats. It is then mixed with liquor until it reaches an alcohol content of 20% which inhibits further fermentation and characterizes the wines of Oporto. The wines of Oporto mature in large barrels or in bottles. In the first case, after a period of between three and 40 years they are stabilized and bottled for immediate consumption. These are the mellow Oporto wines. Those Oporto wines matured in bottles are the vintage ports. These wines age for two years in large barrels, are bottled without filtering or stabilizing and continue to age in the bottle. * Both belong to the firm of Silva & Cosens, Oporto, Portugal.Vintage ports are from exceptional years. Furthermore, for an ageing period of at least 18 months the wine is evaluated to see if it is of sufficient quality to resist a long maturation in the bottle. The declaration of a vintage year is extremely important for a cellar and requires the approval of the ‘Instituto de Vinho do Porto’. Generally, less than 3% of the harvest reaches this quality. Vintage ports are mixed wines coming from various ‘Quintas’ (small properties). Wines known as ‘Single Quinta vintage’ are made in the same way as the vintage ports but come from only one Quinta and are produced in years that are not vintage years. Their quality is notably high. Data and Software The vintage ports studied here belong to two brands: Dow’s and Graham’s.* They correspond to the years 1963, 1966, 1970, 1975, 1977, 1982, 1983, 1985, 1986, 1988 and 1990, of which 1982, 1986, 1988 and 1990 are not vintage years and the wines are those of the ‘Quinta de Bomfin’ and ‘Quinta dos Malvedos’.In a vintage year the Quinta de Bomfin dedicates its production to the Dow vintage and the Quinta dos Malvedos to the Graham vintage. There are, therefore, 24 samples covering a period of 27 years and two types (categories): the Dow vintage together with the single Bomfin vintages and those of Graham together with the single Malvedos vintages. Ref. 1 provides a global qualification of the year and of each brand. The analytical determinations carried out on these samples were performed in the Oenological station of Haro following official methods.The values obtained can be seen in the corresponding report.2 following 41 variables were used in this work: Nine routine parameters: alcoholic grade, residual sugars, volatile acidity, total acidity, pH, total SO%, density, dry extract and ashes. Two polyalcohols: glycerol and butan-2-01. Seven major volatile components: ethanal, ethyl acetate, methanol, isobutyl alcohol, pentyl alcohols, ethyl lactate and ethyl succinate. Five organic acids: lactic acid, tartaric acid, malic acid, succinic acid and citric acid. Seven metallic ions, viz., Na, K, Ca, Mg, Fe, Cu and Zn. Eleven parameters related to the phenolic compounds and the colour: absorbance at 420 nm (D0420) and at 520 nm (D0520), Sudraud’s index (IC), tone (TON), Garoglio’s ageing index (IE), tannins, anthocyanins, and1010 Analyst, August 1996, Vol.121 the following four CieLab3.4 parameters, tone (H*), chroma (C*), clarity (L*) and saturation (S). Technological advances in the elaboration of wine have been important in the years covered by this study. Furthermore, the fortification of the partially fermented grape juice with grape liquor has introduced a factor that was not controlled until 1983 when the composition of the liquor was regulated. Clearly some variables are influenced by the elaboration (e.g., volatile acidity, glycerol content, or that of ethyl acetate, among others) and others by the type of liquor (e.g., pentyl alcohols content). All the calculations were made with the PARVUSS package on a Tandon 33/486 computer in the Faculty of Science and Food Technology and Chemistry of the University of Burgos.Results and Discussion Ageing of Vintage Port Analysis of ageing was carried out by means of a regression of the age of the vintage, expressed in years, on the 41 chemical variables. All the variables were autoscaled to avoid any effect caused by the different scales of measurement used to express the parameters and age. As the number of predictor variables is greater than the number of objects it is not possible to use a multivariate regression; there exists colinearity as the samples belong to a sub-space of dimension 24. Even by restricting analysis of this sub-space, the ratio of the greatest to the least eigenvalue of the matrix of correlations is 482, greater than 10, which indicates problems of colinearity in the sub-space.6 Furthermore, high levels of correlation exist among some predictor variables due to their own definition.For example, the Sudraud and Garoglio indices and the absorbances describe aspects of the evolution of the anthocyanins and show high levels of correlation between each other. The same situation occurs with the CieLab parameters. The coefficient of linear correlation is in many instances above 0.95. The partial least squares (PLS) regression which is widely used in chemometrics9-'1 and has been completely justified from a theoretical point of view,12-l4 is adequate for the construction of a model given the characteristics of the data. This is a biased regression technique on latent variables aimed at achieving the highest prediction capacity.In order to decide the number of latent variables and thus the bias, full cross- validation was used. l S , l h More specifically, this was carried out by means of three deletion sets such that in each instance the age of the eight wines (four years and two brands in each) must be predicted using the model constructed with the 16 remaining wines belonging to different years. Table I shows the years corresponding to each deletion set while Table 2 shows the explained and cross-validated variance; it is clear that only two latent variables need to be considered. It can also be deduced that, with this model, a large part of the variability present in the predictor variables (74.06%) is not related to the age. It would, therefore, seem appropriate to select only those predictor variables whose variability is related to the response.The modelling capacity was taken as a criterion, i.e., the proportion of the variance of each variable used by the two latent variables. Twenty-two variables must be eliminated if the threshold value is 25%. A new full cross-validated PLS with the same deletion sets provides the evolution of variances shown in Table 3. According to these results two latent variables must be taken. For this model the residual standard deviation is 2.36 in fitting and 2.67 in prediction. Fig. 1 shows the age calculated using this PLS model versus the real age of the wines studied. Table 4 shows the elements of the model to be analysed: the loading of each parameter in each latent variable and their modelling power.The loadings define the participation of each parameter measured in the latent variable. In this instance the first latent variable is related to the age and grows with it. In contrast, the second latent variable tends to equalize, on average, both commercial brands. This, together with the modelling power, allows one to analyse the influence of each variable in the model. According to Table 4, variables 1-9 have the least modelling power, between 32 and 49%. Their moderate contribution in explaining the age of the vintage is due not only to the chemical evolution of the wine but also to concomitant factors. The 30 years which the samples in this study cover have witnessed great improvements in the fermentation process: (i) The grape is in less contact with the air.Previously the grape remained in low wide wine presses for 1 or 2 d. (ii) The introduction of coating equipment. (iii) Decanting without aeration. (iv) The replace- ment of sodium metabisulfite by SO2 gas. (v) Control of the quality of the liquor used in fortifying the wine. The increase with age of the volatile acidity, the ethyl acetate content, the pentyl alcohols and the sodium content could be due to the age or to the above-related factors. The formation of potassium hydrogentartrate deposits in the bottles explains the decrease in tartaric acid. If one adds to this the decrease in citric acid as the age increases, then decreases in total acidity can be justified. Finally, the alcoholic grade and the glycerol content increase together.Looking at the second latent variable it is observed that the volatile acidity (variable 2) and the citric acid content (variable 8) do not contribute to the differentiation of both brands in relation to age. On the other hand, the alcohol grade, and the glycerol and ethyl acetate content are the parameters that reveal the differences in the ageing process between the two brands. These differences may be due to specific fermentation processes maintained for years by each brand. Variables 10-19 in Table 4 are those that contribute the greatest modelling power, above 70%, in explaining the age of the vintage. These are the parameters related to the polyphenol content, except for the tannin content. It is known that the ageing of a wine is based on the evolution of the polyphenols. In relation to this fact, the Sudraud, Garoglio and other indices, based on the evolution of colour, have been defined.Another reasonable alternative is to use the evolution of colour with CieLab parameters directly. The PLS model allows comparison between the behaviour of both types of parameters. It was discovered that the parameters D0420, D0520, IC, anthocyanins content, C* and S decrease with the age of the wine because their loadings in the first latent variable have a negative sign. They also show differences between one brand Table 1 Deletion sets used to cross-validate the PLS model. In each year there are two vintages Year Deletion set 1963 1966 1970 1975 1977 1980 1982 1983 1985 1986 1988 1990 1st * * * * 2nd * * * * 3rd * * * *Analyst, August 1996, Vol.121 25 101 1 + + - + Table 2 PLS model for age with 41 predictor variables. Evolution of the variance Latent variable Variance 1st 2nd 3rd 4th 5th Explained 86.61 94.61 96.98 98.11 98.57 Cross-validated 82.17 88.65 88.39 87.61 86.83 Of predictor block 19.72 25.94 28.88 31.31 35.01 Table 3 PLS model for age with 19 predictor variables that have a modelling power greater than 25%. Evolution of the variance Latent variable Variance 1st 2nd 3rd 4th 5th Explained 83.96 92.37 94.85 96.41 96.37 Cross-validated 79.52 90.24 87.75 84.42 84.07 Of predictor block 42.77 52.04 53.94 53.91 57.68 j0l 15 U c Q 5 0 ':I : 0 t + + + i + I t + * + + + I * + + + 5 10 15 20 25 Agelyears Fig. 1 actual age. Computed age of vintage ports by means of PLS model versus Table 4 PLS model for age with 19 predictor variables Loadings Number 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 Variable Grade Volatile acidity Total acidity G 1 y c e r o 1 Ethyl acetate Pentyl alcohols Tartaric acid Citric acid Na DO420 DO520 IC Tone IE Anthocyanins H* C* L* S ~ 1st loading variable 0.17 0.21 -0.22 0.15 0.2 1 0.20 -0.22 -0.22 0.20 -0.24 -0.26 -0.25 0.26 0.26 0.26 0.26 -0.25 -0.25 -0.25 2nd loading variable 0.34 0.09 0.42 0.3 1 0.26 0.18 0.08 0.18 0.28 0.24 0.25 0.05 0.05 0.24 0.04 0.22 -0.20 0.27 -0.18 Modelling power 0.36 0.36 0.42 0.41 0.49 0.38 0.32 0.37 0.32 0.7 1 0.90 0.87 0.72 0.74 0.83 0.70 0.84 0.88 0.88 and another since the loading in the second latent variable is above 0.2 in all of them.The variable L* increases with age (they are clearer wines) although this increase is different depending on the brand.It should be pointed out that the differences between the two brands are systematic. This is shown in the structure of the loadings and ceases to be apparent in the four oldest years, i.e., in the vintages of 1975, 1970, 1966 and 1963. On the other hand, the TON, IE and H* indices increase as the vintage ages without there being differences between the brands. Finally, it can be confirmed that the CieLab parameters model the ageing of the wine in a similar way to the Sudraud and Garoglio indices, but are less sensitive to possible errors because they are based on a spectrum and not on the absorbance at two or three wavelengths. Typification of the Dow (Bomfin) and Graham (Malvedos) Brands The results obtained by modelling the ageing process suggest that the difference between the two brands is systematic and it is, therefore, possible to differentiate them with the analytical parameters already described.The possibility that the parame- ters that do not vary with age may be important when discriminating the origin of the vintages should be con- sidered. Since the number of variables is greater than the number of objects and high degrees of correlation exist, the soft independ- ent modelling class analogy (SIMCA) procedure was used. This is a highly efficient method in chem0metrics~~-~9 and more specifically in food chemometrics.20 The data are grouped into two classes: (1) Category 1, Dow vintages and the Bonfim single quinta (2) Category 2, Graham vintages and the dos Malvedos In order to decide the number of principal components of each category, the double cross-validation method was applied, with the data autoscaled separately in each category.A total of 20% of the individual values from the raw data was used as the deletion set: therefore, the prediction error sum of squares (PRESS) is made with five deletion sets. Table 5 shows the PRESS values obtained and the value of the residual standard error (RSE) as a function of the number of components considered. It is clear that in each class two principal components must be considered. Since the number of objects is small, the Wold enlarged range was used. Fur- thermore, the weighted augmented SIMCA distances was considered to model smoothly the lateral sides of the SIMCA box in this problem.Under these conditions one obtains a model for the Dow vintages with a sensitivity of 83.3% (two vintages remain outside the model) and a specificity of 91.6% (one Graham vintage was accepted by the model for Dow vintages). The SIMCA model for the Graham vintages has a sensitivity of 83.3% and specificity of 100% (no Dow vintage was accepted by this model). With the same deletion sets used in the PLS regression (Table l), the cross-validated percentage of correct classification was calculated, obtaining 10 out of 12 correct classifications in each category. The correct classifications in fitting are 11 out of 12 in each class. The Coomans21 diagram (Fig. 2) shows that both brands remain sufficiently separate even when the samples cover a period in which great technical changes have taken place in the production of wines.It was noted that two vintages remained outside the model of each category and that the Graham class had better specificity. vintage. single quinta vintage.1012 Analyst, August 1996, Vol. I21 Each of the analysed parameters either contributes to create the SIMCA model of each brand or distinguishes one model from the other. It is also possible that they perform both or neither of these two functions. The modelling power of one variable is its contribution to the SIMCA model as the fraction of the standard deviation it brings to the two principal components. On the other hand, the category residual standard deviation is 0.72 for the Dow vintages and 0.67 for the Graham vintages.Consequently, the contribution of a variable to the corresponding SIMCA model is considered to be significant if its modelling power is above 0.28 and 0.33, respectively. In Table 6 the values of modelling power above this threshold are underlined. Of the 19 variables chosen for the PLS model (see Table 3), all of them except for alcoholic grade, pentyl alcohols and sodium content are significant in modelling both categories. This is logical; even when the age variable does not appear as such it is present in the structure of each category through its predictors, particularly anthocyanins and colour indices. Table 6 indicates those variables with specific modelling power to the Graham vintages. These are total SO*, density, dry extract, ashes, butan-2-01, pentyl alcohols and sodium which show greater variability in the samples given by these parameters in relation to that found in the model of the Dow vintages.The opposite situation is that of ethanal, ethyl succinate, lactic acid and Mg whose variability is significant in the Dow samples but not in those of Graham. It is observed that the malic acid content has a modelling power close to the threshold, slightly below that of lactic acid. The modelling power of these variables is attributable to the malolactic fermentation in bottles, which was confirmed in four of the samples, whereas this only occurred in one of the Graham vintages. The tannin Table 5 Double cross-validation: prediction error sum of squares (PRESS) and residual standard error (RSE) values as a function of the number of components in each brand Dow's vintage Graham's vintage Component PRESS RSE Ratio PRESS RSE Ratio 1 350.13 461.99 0.76 324.55 461.99 0.70 2 242.06 258.07 0.94 210.32 244.82 0.86 3 204.99 183.88 1.12 240.80 161.20 1.49 4 174.57 137.71 1.27 162.29 120.62 1.34 5 137.75 105.43 1.31 110.80 86.72 1.28 a a D O D D n o $ al a 0 Q 0 0 0 Q '," "Q Q G .1' 95% Distance from class Dow's vintages Fig. 2 Coomans diagram with the squared distance SIMCA for the Dow's vintages, D, and Graham's vintages, G. Continuous lines are the critical SIMCA distance for each category. content has great modelling power in both categories. However, its variability has no connection with the age of the wine (it does not contribute to formation of the PLS model).The variables with the greatest discriminating power are tannins and tartaric acid as well as butan-2-01, succinic acid, citric acid and potassium, i.e., the difference in average values between the two brands is greatest as regards these para- meters. Conclusion Using a PLS regression it is possible to model the age of Dow and Graham vintages from 1963 to 1990. The variables responsible are those related to colour. The CieLab parameters contribute to defining the age in a similar way to the Sudraud and Garoglio indices. The differences between the two brands can be modelled with sufficient sensitivity and specificity using a SIMCA model with two principal components for each brand. The tannin content is the parameter with the greatest discriminating power, and characteristics relating to each brand appear, which can be Table 6 SIMCA model with two principal components in each category.Those modelling power values greater than the threshold of the category are underlined, as is a discriminant power greater than 2 Number 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 Variable Grade Residual sugar Volatile acidity Total acidity Total SO2 Density Dry extract Ashes Glycerol Butan-2-01 Ethanal Ethyl acetate Methanol Isobutyl alcohol Pentyl alcohols Ethyl lactate Ethyl succinate Lactic acid Tartaric acid Malic acid Succinic acid Citric acid Na K Ca Mg Fe c u Zn DO420 DO520 IC TON IE Tannins Anthocyanins H* C* L* S PH Modelling power Category 1 Category 2 0.14 0.00 0.34 0.38 0.18 0.17 0.17 0.27 0.00 0.51 0.15 0.40 0.37 0.00 0.10 0.12 0.10 0.59 0.28 0.41 0.27 0.09 0.56 0.20 0.00 0.26 0.46 0.12 0.07 0.00 0.59 0.77 0.74 0.62 0.64 0.62 0.76 0.62 0.73 0.80 0.74 0.32 0.21 0.38 0.37 0.27 0.41 0.53 0.59 0.53 0.44 0.38 0.03 0.73 0.29 0.04 0.58 0.00 0.13 0.05 0.47 0.11 0.32 0.52 0.38 0.00 0.21 0.00 0.13 0.04 0.00 0.78 0.88 0.88 0.74 0.73 g.5J 0.72 0.x6 0.90 0.89 Discriminant power 1.06 1.48 1.07 1.04 1.30 1.60 1.48 1.59 0.91 1.44 1.61 1.33 1.55 1.12 2.48 0.95 1.39 1.64 1.80 3.42 1.44 1.10 __ 2.84 1.80 1.82 1.19 1.11 1.20 1.38 1.62 1.60 1.06 1.14 5.41 1.30 1.15 1.68 1.86 1.54Analyst, August 1996, Vol.121 1013 correlated with the fermentation process and malolactic fer- mentation. References Suckling, J., Vintage Port, Wine Spectator Press, San Francisco, CA, 1990.Symington, C., and Santamaria, F., in Magister de Viticultura y Enologia, Escuela Politkcnica de Logroiio, Universidad de Zaragoza, 1993. IRANOR, Colorimetric Quantities, UNE 72-03 1-83, Madrid, 1983. IRANOR, Numerical Specification of the Psychophysical Colour of Luminous Stimulus, UNE 72-03 1-84, Madrid, 1984. Forina, M., Leardi, R., Armanino, C., and Lanteri, S., PARVUS: an Extendable Package of' Programs for Data Exploration, Classifica- tion and Correlation, Release 1.0, Elsevier, Amsterdam, 1988. Hines, W. W., and Montgomery, C., Probability and Statistics in Engineering and Management Science, Wiley, New York, 2nd edn., 1980. Wold, H., in Multivariute Andysis, ed. Krishnaiah, P. R., Academic Press, New York, 1966. Wold, H., in Perspectives in Probability and Statistics, Papers in Honour o j M. S . Bartlett, ed. Gani, J., Academic Press, London, 1975. Sharaf, M. A., Ilman, D. L., and Kowalski, B. R., Chemometrics, Wiley, New York, 1986. 10 1 1 12 13 14 15 16 17 18 19 20 21 Martens, H., and Naes, T., Multivariate Calibration, Wiley, New York, 2nd edn., 1988. Brereton, R. G., Chemometrics. Applications of Mathematics and Statistics to Laboratory SystemJ, Ellis Horwood, Chichester, 1990. Lorber, A., Wangen, L. E., and Kowalski, B. R., J . Chemorn., 1987, 1, 19. Frank, I. E., and Friedman, J. H., Technometrics, 1993, 35, 109. Stone, M., and Books, R. J., J . R. Stat. SOC., Ser. B , 1990, 52, 237. Martens, H., and Naes, T., in Near-infrared Technology in the Agricultural and Food Industries, eds. Williams, P., and Norris, K., American Cereal Association, St. Paul, MN, 1987, pp. 57-87. Lanteri, S., Chemom. Intell. Lab. Syst., 1992, 15, 159. Wold, S., J . Pattern Recognition, 1976, 8, 127. Stone, M., and Jonathan, P., J . Chernom., 1994, 8, 1. Kvalheim, 0. M., and Karstang, T. V., in Multivariate Pattern Recognition in Chemometrics, Illustrated by Case Studies, ed. Brereton, R. G., Elsevier Science Publishers, Amsterdam, 1992. Forina, M., Lanteri, S., and Armanino, C., in Chernometrzcs and Species Identification, Topics in Current Chemistry, Springer-Verlag, Berlin, 1987, vol. 141. Coomans, D., Pharm. Thesis, Vrije Universiteit, Brussels, 1982. Paper 6lOO115G Received January 5 , I996 Accepted April 23, 1996

ISSN:0003-2654    

DOI:10.1039/AN9962101009

出版商:RSC    

年代:1996

数据来源: RSC

摘要:

Analyst, August 1996, Vol. 121 (1015-1018) i Import data 1015 I I Generate deconvoluting I I I Generate simulation or import data? Deconvolution of Analytical Peaks by Means of the Fast Hartley Transform I ! A. Economou, P. R. Fielden* and A. J. Packham Department of Instrumentation and Analytical Science, UMIST, PO Box 88, Manchester, UK M60 lQD 4 I peak I I Generate simulation I A ’ & I I I A peak is one of the commonest and more desirable forms of analytical response. In multicomponent systems, where each individual component might give rise to one or more peaks, overlap between adjacent peaks often takes place. Depending on the degree of overlap, the extraction of qualitative and quantitative data can be difficult, and in extreme cases, impossible. In this paper, a method for deconvoluting overlapping peaks, based on the implementation of the fast Hartley transform (FHT), is described.The FHT is superior to the traditionally used fast Fourier transform in terms of speed and memory requirements. The program for deconvolution was developed in an object-oriented, icon-based software development tool (LabVIEW for Windows). The suggested method was applied to electroanalytical data involving overlapping peaks. Keywords: Deconvolution; Hartley transform; Idturnmetry Introduction In analytical chemistry, a bell-shaped profile (peak) is the commonest type of response. In real samples a multitude of peaks of different heights can be produced in a single experiment. Furthermore, some of the peaks may overlap, partially or completely, due to similarities in the physical and chemical properties of the relevant constituents in the sample.This overlap may cause difficulties in the precise quantification of the components, since there will be no stable baseline to act as a reference point for the measurement of peak height. In more * To whom correspondence should be addressed. Generation and manipulation of analytical signal severe situations, overlapping peaks may merge completely, appearing as a single peak; in this case, not only is quantitative evaluation of the data impossible, but also qualitative analysis is complicated, possibly leading to false interpretation of the results. Traditionally, deconvolution procedures based on the appli- cation of fast Fourier transform (FFT) strategies, have been shown to be applicable in resolving electrochemical,14 spectro- scopic5-* and chromatographic’ overlapping peaks.According to this scheme, the coefficients in the FIT spectrum of the analytical signal are divided by the coefficients in the FFT spectrum of a judiciously selected peak-shaped function. The resulting frequency-domain signal is then subjected to an inverse FFT (IFFT) which produces the deconvoluted analytical signal. However, when the sequences to be processed consist of real-valued data points (like the data captured digitally in the course of an analytical experiment), the FFT produces complex- valued sequences in which half of the information is redundant. This happens because, due to symmetry relations, only n/2 harmonics are necessary to describe completely in the fre- quency domain a signal that is comprised of n real values.’ 1,12 The fast Hartley transform (FHT), Hi, of a real-valued digital signal, Yk, is obtained through the discrete implementation of the Hartley integral.’ I.12. H, = 1 [ ~ ~ J Y , [ c o s [ y j +sin [T)j (1) for i = 0, 1, 2, ..., n - 1. The resulting spectrum is a mapping of n discrete, real, time- domain Yk values into n discrete, real-valued ‘frequencies’, Hi.I I I I I I I I I I I I Deconvolution and manipulation I I I of deconvoluted signal I I ----------- I I deconvoluting signal I I to multipeak response Rotate and translate I the sianal I I I Derive the FHT of the I - - - - - - - - - - - - - 1 I i I Derive the IFHT I I 1 i I deconvoluting signal Filter the Lconvoluted Display I I signal I + ‘ I---------- I I Display r x l ! I Inverse rotate-translate the deconvoluted signal I ! Fig.1 The flow diagram of the program for deconvolution using the FHT1016 Analyst, August 1996, Vol. 121 1.2- 1 .o- 0.8- 0.6- 0.4- 0.2- 0 -0.2- -0.4- -0.6- -0.8- -1 .o - -1.2- The inverse FHT IFHT is given by the relationship: Y , = ('::J Hi [cosi?) +sin [T)] (2) fork=O, 1, 2, ..., n-1. /-\\ ( 4 I \ I \ I \ I --, I \-'-\ / , I I I I \ \ \ \ I / \\ '..- , , , , , , , , , The IFHT reconverts the frequency-domain signal into the initial y1 time-domain, real, Y k values. It is interesting to note that, unlike the FFT, the FHT and its inverse are identical in form (except for the multiplication factor l / n ) and, in addition, the FHT does not involve complex notation.The computational advantage of the FHT, as defined in eqns. (1) and (2), is that it uses half the memory to produce the same information as the FFT and at the same time is faster than the FFT by a factor of two.",'2 Some minor benefits can be further gained by using the FHT such as fewer roundoff errors (due to fewer calculation^)^^ and limited hard disk swapping (as result of the limited amount of data). Thus, the FHT presents an advance over the traditional FFT for representing real-valued data in the frequency domain. The utility of the FHT can be extended to the faster calculation of the power spectrum, P , , the magnitude spectrum, M,, and the phase spectrum, +L, of a signal according to the formulae: 1,12 L fori=O, 1, ..., n-1. The FFT can also be computed directly from the FHT by means of the formula: F,r = H , + H,* -, F,'m = H , - H n - , (4) for i = 0, 1, ..., n - I , where F i r and Flim are the real and imaginary coefficients of the FFT, respectively.Finally, the FHT can be used in the frequency domain smoothing of real data in a way similar to FFT filtering, i.e., by multiplying the FHT of the noisy signal with the FHT of a suitable filter function and then calculating the IFHT. In the context of the present work, it can be shown that the deconvolution of two real, sampled-data signals corresponds to a simple division of the discrete coefficients in their FHT spectra, provided that the deconvoluting signal is even;I2 this property can be exploited in the present work, since all the deconvoluting peak functions used are even functions. Software-programming and Execution of Programs The software selected for the program was LabVIEW for Windows, Version 3.0 (National Instruments, Austin, TX, USA).A flow diagram of the program developed for deconvolution is shown in Fig. 1. Four main operation modules can be (4 600.0 2 500.0 L 3 300.0 *OO.O1;JJJ 100.0 1 k 0.0 I I I I -0.4 -0.5 -0.6 Potential/V( versus Ag/AgCI) 800.0 700.0 0 300.0- 200.0 - -50.0 1 I I I -0.4 -0.5 -0.6 Potent i alN( versus Ag/Ag C I) Fig. 3 Deconvolution of voltammetric peaks obtained for the determina- tion of Pb" and U"' by adsorptive stripping voltammetry on a HMDE. ( a ) Original data; and (h) deconvoluted data. Traces 1, 2 and 3 represent increasing concentrations of 10, 20 and 30 mmol of dm-3 of both U"' and Pb".Deconvolution was performed with a voltammetric peak with WHMH of 40 mV. The deconvoluted data were low pass filtered with a Hamming FIRF with 120 coefficients and cutoff at the 82nd harmonic. Voltammetric conditions: electrolyte, 0.1 mol dm-l PIPES buffer, pH 6.8; ligand, 8-hydroxyquinoline, 0.1 mmol dm-3 preconcentration time, 30 s; square wave scan at 60 Hz; scan increment, 4 mV; pulse height, 30 mV. The potentiostat was a PARC 273 and the electrode was a PARC 303 SMDE. -1.24 I , , , , , , , , 150 175 200 225 250 275 300 325 350 375 4 -1.24 , , , , , , , , , ) ' 150 175 200 225 250 275 300 325 350 375 400 Variable, x Fig. 2 Deconvolution using the FHT of three overlapping peaks with WHMH 25 points located at 250,275 and 300 points and height ratio of 2 : 1 : 1, using a deconvolution peak of the same type.Traces are before (---) and after (-) deconvolution (the latter multiplied by -1). ( a ) Voltammetric peaks, deconvoluting peak with WHMH 20 points; (h) Gaussian peaks, deconvoluting peak with WHMH 22 points; and (c) LorentLian peaks, deconvoluting peak with WHMH 21 points.Analyst, August 1996, Vol. 121 1017 distinguished: (i) the generation of the simulated multipeak response (for simulated peaks) or the import of data (for real data) and the subsequent manipulation of the analytical signal; (ii) the generation and manipulation of the deconvoluting function; (iii) the deconvolution operation and manipulation of the deconvoluted signal; and (iv) the display of the results.The number of points, peak types, peak widths and peak locations for the simulated multipeak analytical signal were user-defined, as were the peak type and peak width for the deconvoluting function. White noise of varying amplitudes and linear baselines could be added to the simulated signal. Some initial signal manipulation (wrapping around of the deconvoluting function,l3 rotation-translation of the signal to be deconvo- luted14 and filtering of the final deconvoluted signals) were necessary as illustrated in Fig. 1. Results and Discussion Four types of deconvolution function have been assessed in this work: (i) in chromatography? and some forms of spectroscopy, the peaks are usually assumed to be Gaussian curves;9 (ii) in spectroscopy (notably NMR and IR), Lorentzian peaks are usually encountered;5,15 (iii) in differential pulse voltammetry, as well as in stripping voltammetric techniques, the response can also be simulated by a peak-shaped function;1,4 and (iv) a simple triangular peak, The most critical parameter in all these functions, representing peaks of different forms, is the width at half the maximum height of the peak, designated as WHMH.As in the case of FFT deconvolution,4~9 the use of a deconvoluting function of the same mathematical type as the peak to be deconvoluted preserves the peak types in the deconvoluted signal as shown in Fig. 2; note the different + In chromatography. peaks are usually tailing due to first-order extra-column effects. However, the Gaussian shape is a useful approximation in most cases.0 ‘ I 1 -0.4 -0.5 -0.6 Potential/V( versus Ag/AgCI) “ W -0.4 -0.5 -0.6 Potential/V( versus Ag/AgCI) shapes of the peaks. In the absence of information regarding the exact type of overlapping peaks, any peak-shaped deconvolut- ing functions can be used but the exact initial peak shapes will not be preserved. The wider the deconvoluting peak, the better the resolution enhancement achieved after deconvolution, provided that the mathematical and physical restrictions derived earlier are con~idered.~39 However, better resolution will be accompanied by a poorer S/N ratio in the final deconvoluted signal which will require a more stringent filtering regime with the associated risk of generation of artefacts due to ‘leakage’.s It is interesting to note that digital filtering introduced a group delay into the filtered data.l6 This delay causes a shift of the peaks in the deconvoluted signal that must be compensated for.For multiple peaks with different WHMHs, the ratios of the different peaks heights before and after deconvolution were not preserved. Nevertheless, when subjecting any one (or more) peak(s) to a linear height increase (corresponding to a standard additions in practical situations), the linearity of the peak height(s) was maintained after deconvolution. This property allows quantitative analysis to be carried out. It was found that, in practice, deconvolution based on the FHT was around 35% faster than the same operation using the FFT to deconvolute a 8192-point signal on a 486 66 MHz PC with 16 MB of memory.The speed advantage was less than the theoretically expected value of 2 due to the various overheads (such as plotting and baseline correction) but can be important in cases of larger data sets. Filtering of the deconvoluted data makes for a significant proportion of the total computation time; in the case of the FHT, the amount of data to be filtered is half than the amount in the FFT and no overhead arises from this operation. More importantly, the greater storage requirements of the FFT, as opposed to the FHT, consistently caused memory overflow unless the memory reserved for LabVTEW was set to more than 8 MB. This advantage of the FHT becomes more 250.0 a 2 3 150.0 0 100.0 50.0 s 2oo-o 0 1 -0.4 -0.5 -0.6 Potential/V( versus Ag/AgCI) 300.0 250.0 a 2 5 150.0 0 100.0 50.0 + *OO.O -0.4 -0.5 -0.6 Potential/V( versus Ag/AgCI) Fig.4 Deconvolution of voltammetric peaks generated by the reduction of 10 pmol dm-3 Pb” and 10 pmol dm- SnI1 on a HMDE. Deconvolution by: (a) voltammetric peak (40 mV WHMH); (h) Gaussian peak (40 mV WHMH); (c) triangular peak (40 mV WHMH); (d) Lorentizian peak (32 mV WHMH). Traces before (---) and after (-) deconvolution. The deconvoluted data were lowpass filtered with a Hamming FIRF with 120 coefficients and cutoff at the 82nd harmonic. Voltammetric conditions: electrolyte, 0.1 mol dm-3 acetate buffer, 4.5; differential-pulse scan at 10 mV s-1; pulse height, 100 mV; pulse width, 50 ms; step time, 0.4 s; scan increment, 4 mV. Potentiostat and electrode as in Fig. 3 .1018 Analyst, August 1996, Vol.121 serious in cases of several programs sharing computer resources and when large data sets are manipulated. Disregarding the computational advantages of the FHT over the FFT, the results produced by the two methods in terms of resolution improve- ment, S/N and baseline characteristics were identical. The proposed deconvolution method has been applied to voltammetric data for the determination of trace metals. The first application is concerned with the simultaneous determi- nation of Pb" and Uv' by adsorptive stripping voltammetry. For this purpose Pb" and Uvl were complexed with S-hydrox- yquinolone, the complex was preconcentrated by adsorption on a hanging mercury drop electrode (HMDE) and, finally, the adsorbed complex was reduced by a negative-going scan, in the square wave mode, of the working electrode potential.Ur- anium(v1) and Pb" produce reduction peaks with the peak potential characteristic of the species and the peak height proportional to the species concentration in the sample. Since these peaks partially overlap, the FHT deconvolution method was applied to improve the resolution of the data for three voltammetric scans of increasing metal concentrations. The results are shown in Fig. 3. An increase in resolution was achieved with the linearity of the response preserved. It is interesting to point out that there was a shift of the peak potentials to the positive direction upon increase in concentra- tion of the metal ions, a phenomenon common in electroche- mistry due to systematic changes of the electrode surface.However, the deconvoluted peak potentials are identical to the original peak potentials. Another example was the simultaneous determination of Sn" and Pb" by differential pulse, solution- phase voltammetry on a HMDE. Again, the reduction peaks of Pb" and Sn" occur within a narrow potential window and deconvolution can be used to separate the two peaks. A voltammetric data array was subjected to FHT deconvolution using voltammetric, Gaussian, triangular and Lorentzian decon- voluting peaks as illustrated in Fig. 4. All of the peaks produced similar results both in terms of resolution enhancement and background characteristics. This last point demonstrates that, in practical situations, the criterion of preservation of the exact peak shapes might be of limited importance as opposed to the benefits gained by increasing the resolution.References 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Pizeta, I., Anal. Chim. Acta, 1994, 285, 95. Engblom, S. O., and Ivaska, A. U., in Electrochemistry, Sensors and Analysis, ed. Smyth, M. R., and Vos, J. G., Elsevier, Amsterdam, Pizeta, I., Lovric, M., and Branica, M., J . Electroanal. Chem., 1990, 296, 395. Engblom, S. O., J . Electroanal. Chem., 1990, 296, 371. Kauppinen, J. K., Moffat, D. J., Mantsch, H. H., and Cameron, D. G., Appl. Spectrosc., 1981, 35, 271. Kauppinen, J. K., Moffat, D. J., Cameron, D. G., and Mantsch, H. H., Appl. Opt., 1981, 20, 1866. Mann, C. K., and Vickers, T. J., Appl. Spectrosc., 1989, 43(1), 33. Chapados, C., Trudel, M., and Miletic, J., Chemom. Intell. Lab. Syst., 1994, 22, 209. Kirmse, D. W., and Westerberg, A. W., Anal. Chem., 1971, 43, 1035. Lynn, P., and Fuerst, W., Introductory Digital Signal Processing with Computer Applications, Wiley, New York, 1994, pp. 270-279. Bracewell, R. N., Science (Washington, D.C., I883-), 1990, 248, 697. O'Neil, M. A., Byte, 1988, 13(1), 293. Press, W. H., Flannery, B. P., Teukolsky, S. A., and Vetterling, W. T., The Art of Scientific Computing, Cambridge University Press, Cambridge, 1989, pp. 452454. Hayes, J. W., Glover, D. E., Smith, D. E., and Overton, M. W. S., Anal. Chem., 1973, 45(2), 277. Baker, C., Johnson, P. S., and Maddam, W. F., Spectrochim. Acta, 1978,34A, 683. LahVIEW Analysis Reference Manual, Austin, TX, 1992, pp. 4.1- 4.24. 1989, pp. 49-54. Paper 6103670H Accepted May 28,1996

ISSN:0003-2654    

DOI:10.1039/AN9962101015

出版商:RSC    

年代:1996

数据来源: RSC

摘要:

Analyst, August 1996, Vol. 121 (1019-1024) 1019 Spline Wavelet Multi-resolution Analysis for High-noise Digital Signal Processing in Ult raviolet-Visi ble Spect rophotomet ry Xiao-Quan Lu and Jin-Yuan Mo Department of Chemistry, Zhongshan University, Guangzhou 51 0275, China The application of B-spline wavelet multi-resolution analysis (MRA) in UV spectrophotometry is presented, with a modified approach to signal processing. The results indicate that if the optimum B-spline wavelet and truncation frequency are selected, the absolute values of the peak relative errors are less than 2% when the signal-to-noise ratio is not less than 0.5, which means that useful signals can be separated from high-frequency noise. Most often third-order B-spline wavelet MRA is the most effective. In this respect, significant differences are observed for the different B-spline wavelets.The processed results of experiment data are also satisfactory. Keywords: Spectrophotometry; wavelet multi-resolution analysis; noise reduction; chemometrics Introduction The use of chemometric methods is playing an increasing role in modern analytical chemistry because of the rapid progress in computer technology, the wide availability of powerful micro- computers and the introduction of the elements of computer science at various levels of education. The rapid development of the application of computer science and chemometrics in analytical chemistry is well documented in reviews on chemo- metrics. 1 Computer control of complex measuring instrumenta- tion, computer databases and computer signal processing are commonly used in research and routine analytical laboratories.Among the various possible applications, signal processing is often utilized. Signal processing in computerized analytical instruments can be used for the following: data acquisition and storage, especially for signal correction and graphics; trans- formation of signals (e.g., Fourier transformation) and calcula- tions; improvement of signals by background reduction and digital filtering; and analysis of experimental data. In common analytical practice, each measured signal is deformed by the presence of noise mostly from the measuring instrument, the environment and analogue-digital (AD) con- version; such information noise usually has a very complex origin and sometimes submerges useful information.Hence it is important to acquire useful information from noise-deformed signals.2 One of the possibilities for signal processing is digital filtering. The most often applied processing usefully includes Fourier transformation (FT) and Kalman filtering (KF), exam- ples of such data processing can be found in the literature.3-6 One of the keys of FT is to determine the frequency separation value of the useful signal from noise in both the frequency and time domains. However, at the same time it is not possible to obtain precise separation in frequency and time domains. Hence FT can only be employed to process signals containing certain types of noise.7 On the other hand, in order to study the spectral behaviour of a signal from its FT, full knowledge of the signal in the time domain must be acquired.This even includes future information. In addition, if a signal is altered in a small neighbourhood of some time instant, then the entire spectrum is affected. Hence, in many applications such as the analysis of non-stationary signals and real-time signal processing, FT alone is completely inadequate.8 Recently, the subject of 'wavelet analysis' has attracted much attention from mathematicians and engineers. Wavelets have become a very popular topic these days. Some view wavelets as a technique for time-frequency, some consider it as a new basis for representing functions, and others think of it as a new mathematical subject. Of course, all of these are right, since wavelets are a versatile tool with a very rich mathematical content and great potential for application.This paper shows the advantages of the use of a signal processing technique called wavelet transformation for tackling these problems. The main characteristics of the technique are (i) it can decompose the signal directly according to the frequency and give the signal representation of the frequency domain distribution state in the time domain; in this transform, time and frequency information of the signal is retained, and with proper identification of 'scales' with 'frequency', wavelets provide an adaptive window by virtue of convolution with a scaled 'wavelet,' in the sense that it has zoom-in and zoom-out capability at any frequency; (ii) the development of signals in the frequency domain can be constituted with a flexible choice of waveforms rather than with trigonometric functions as a basis, and therefore it is a more powerful method to depict the relationships among the variables.The aim of this study was to separate the UV spectral signal from noise by proper choice of the form of the B-spline wavelet. Theory The purpose of the wavelet transformation procedure is to decompose a signal into localized contributions that are characterized by a so-called scale parameter. The mathematical techniques for performing the decomposition have been described in a number of papers.9-14 Here only a short review of the main equations is given. Supposing the introduced disturbance is stochastic noise: S,(t) = M(0.5 - RND) (1) where M is the amplitude of noise and RND is any random data between 0 and I , then the equation for UV spectrophotometry with noise is described by where S,(t) is useful information.function of the unit interval (O,l), namely, The first-order cardinal B-spline N1(t) is a characteristic 1 o a t < 1 0 otherwise (3) and for m 3 2, the mth-order B-spline N , is defined recursively by convolution:1020 Analyst, August 1996, Vol. 121 N,(t) = N , - I(t)Nl(t) = ji N , - I(t - x)dt (4) The wavelet function Vm(t) may be expressed as follows (Fig. 1): . 3rn-2 where ZV$z(t) is the mth-order derivative of the function N2rn(t): m ~ ? ( t > = c (-lY(y)Nm(t - j ) (6) j = O From eqns. (4)-(6), we have v j , m ( t ) = v,(2h - k) k E ZZ (integer) (7) $j,m(t) = Nm(2/t - k ) k E 22 (integer) (8) where @i,m(t) is called the scaling function.For the energy-limited signal S ( t ) represented by eqn. (2), wavelet MRA can separate it into many orthogonal components so that (9) S(t) =: g-&) + g-z(t) +.--+ g-,(t) + S-,(t) is a spline series and gj(t) = c d',W(2it - k ) (11) k e 22 is a wavelet series. Explicitly, cjk and dj are obtained from cp' by (12) ' C j+l c J ~ = a,-2k~'~ djk = C bn-2,gp1 (13) Schematically, the decomposition can be represented by where c.j+l is the original sampling data representation of the signal S(t); {a,, } and { bn } are called weight sequences which can be calculated from the literature.9 In this manner, S(t) is decomposed into many orthogonal components so that each component represents a portion of the original signal at a given resolution level.Another interpretation is that each component represents the portion of signal energy in that particular frequency band. We remark here that the sequences { a , } and { b, }are infinite sequences; however, the amplitudes of the coefficients decay exponentially fast so only 0.~-0.4 4 I 0 1 2 3 4 5 6 7 t Fourth-order B-spline wavelet v4 (t). Fig. 1 a few coefficients are needed for the computations. By truncating the sequences { a, } and { b, } , the decomposition algorithms operate on signals as a finite impulse response filter (FIR). Eqn. (1 1) represents the higher frequency component of a signal (bandpass) and eqn. (10) shows the lower frequency component of the signal (blurred version). In most analytical measurements, a noise-deformed signal has a much higher frequency than an analytical signal, and therefore at the optimum truncation frequency 1 (cut-off frequency) the signal separated from noise can be represented approximately by eqn. (1 0).10 Experimental Chemicals A standard solution of 1.00 X mol I-' Gd3+ rare earth ion was prepared by dissolving the oxide (99.99% Gd203; Hulan, analytical reagent grade) in 5 ml of perchloric acid, evaporating the solution nearly to dryness and then diluting to 100 ml with de-ionized water.The solution was standardized titrimetrically. Solutions of 1.0 X 10-3 moll-1 chloprophosphonazo I11 (CPA 111) and 0.1 mol I-' sodium nitrite were prepared in the normal way with distilled water. All other reagents were of analytical- reagent grade and doubly distilled water was used throughout.Apparatus All UV spectra were recorded with a UV-3400 spectrophoto- meter (Shimadzu, Kyoto, Japan). In measurements with com- puter data acquisition, the strip-chart recorder was replaced with a 386 DX-40 MHz microcomputer compatible with a 14-bit AD-D/A converter (laboratory-assembled). A pHS-3C acidi- meter was used to adjust the pH of the sample solutions. Results and Discussion UV Spectral Character of CPA III and Its Rare Earth Complex CPA 111 is a well known water-soluble developer, used extensively in analytical chemistry. The structure of CPA 111 has two azo chromophores, -N = N-, linking two aromatic systems (Scheme 1). The absorption spectrum of Gd3+-Sr2+-CPA 111 was recorded (Fig. 2). The main instrumental parameters affecting the spectrum are the wavelength, scanning rate, the wavelength increment and the response time.All of them need to optimized to give a good peak in the determination; preliminary observations revealed that the best result was obtained with a wavelength interval (Ah) of 2 nm arid a scan rate of 120 nm min-1. The response time is automatically selected by the spectrophotometer. Fig. 2 indicates that CPA I11 has one absorption peak at 570 nm, but after Gd3+ was added two peaks which have an equivalent area were observed (curve b); this indicates that a homodinuclear complex was formed between Gd3+ ion and CPA III.15 When Sr2+ ion was added to the above system, the peak height increased and shifted to a higher wavelength. One of peaks increased considerably at 663 nm.These results show that a heterodinuclear complex was formed (curve c). A more detailed description can be found in a previous paper,16 including electrochemical and UV spectral studies. "N B .CI Scheme 1 Structure of CPA 111.Analyst, August 1996, Vol. 121 1021 The aim of this study was to compare several B-spline wavelet procedures using the experimental data corresponding to the signals observed in Fig. 2 (curve c) with added artificially high-frequency noise according to eqn. (1); noise was repre- sented by S,(t). In this study, curve c was recorded with a microcomputer compatible with a 14-bit AD-D/A converter and used as a simulated signal after padding the spectra with zero to the nearest larger power of 2. The whole simulated signal consisted of 512 points, Therefore, S(t) is a sequence of discrete data by eqn.(2). Prior to adding artificial high- frequency noise, values of the absorbance of 0,053 16 at 612 nm and 0.06388 at 663 nm (curve c in Fig. 2) were employed as theoretical peak values for detection in the procedures. The simulated noisy signal obtained in Fig. 3a (scatter points) was filtered using the above-mentioned procedures after preliminary optimization of all parameters for each mth-order B-spline wavelet. The effects of mth-order B-spline wavelets, the parameter I, which represents the cut-off (or truncation) frequency value between the useful signal and noise, and the change in S/N have been discussed. Our laboratory-written software package of MRA algorithms17 in True BASIC and Turbo C was used to select the optimum values of 1, rn and S/N and to process the UV signal with high-frequency noise.The results of this test produced the solid line in Fig. 3a. A general strategy assumed for the optimization of MRA parameters was to remove as much as possible of the high-frequency noise. In order to compare the results obtained by B-spline wavelet procedures, the errors of the peaks were calculated according to the expressions (AP - Ao) X 100 A,, % = A0 (Ap - h,) x 100 b he, % = where A, is the peak absorbance value obtained after processing, Ao is the theoretical value of the peak absorbance, which equals to 0.05316 at 612 nm and 0.06388 at 663 nm, and h, and ho are wavelengths, corresponding to A, and Ao, respectively. The second, third and fourth B-spline [N,(t), m = 2, 3, 41 wavelets were applied to process the UV signal.The influence of m, I and S/N on the UV signal processing is discussed below. --c ....., b 2 8 500 650 800 WAVELENGTHhm Fig. 2 UV spectrum of Gd'+-Sr2+-CPA 111 complex in 0.1 mol I-' NaN02 solution: a, 1.70 X 10-5 moll-' CPA 111; b, a + 6.0 X 10-7 moll-' Gd'+ ion; c, b + 1.70 X 10-5 mol 1-1 Sr2+ ion. pH, 5.0; scan rate, 120 nm min-l; and Ah, 2 nm. Effect of diferent mth-order B-spline wavelets If S/N = 0.5, the truncation frequency l = 4 and the number of sampling points k = 29 are fixed. The value of m changed from 4 to 2. The processed results are shown in Fig. 3. In each instance the solid line shows the results of the application of a given MRA procedure to the simulated signal (scatter points).If rn = 3 is chosen, the processed curve is similar to the pure signal and the errors are also very small, as can be found from Table 2. For no other reason than that, we chose rn = 3 as the optimum value to process the UV spectra. Eaect of different truncation frequency (1 = 3) If S/N = 0.5, the order of B-spline m = 3 and number of sampling points k = 29 are fixed, the truncation frequency l ranges from 6 to 3. The processed results are shown in Fig. 4. One can overprocess data, as illustrated in Fig. 4b, by selecting a value 1 that is too small. This is equivalent to rejection of a portion of the desired signal along with the high-frequency noise, producing a peak rather than two peaks that is generally broader and lower in height than warranted.Conversely, by choosing too large a value of I , much of the mid-frequency noise is retained. Fig. 4a serves as an example of such an V 500 650 800 0.06 0.03 " 500 650 800 WAVELENGTHhm Fig. 3 Unprocessed (points) and processed (solid lines) signals at 1 = 4 and k = 29 with different B-spline wavelets: a, m = 2; b, m = 3; and c, m = 4.1022 Analyst, August 1996, Vol. 121 underprocessing situation. In our experiments it was found that the best result was obtained when 1 = 4 (shown in Fig. 3b). Eflect of diflerent SIN The practical significance of the noise in analytical chemistry is expressed by S/N which is commonly used to estimate the detectability of a given method. The lower the S/N is, the less useful information the signals include. If the order of B-spline m = 3, truncation frequency 1 = 4 and number of sampling points k = 29 are fixed, S/N varies from 100 to 0.5.The processed results are listed in Table 2, From the tables, even if S/N decreases to 0.5, the results exhibit a better reduction of the high-frequency noise. The errors of the processed signal are very small. Results of A,,.,.% and h,,-,.% The influence of different m, 1 and S/N on A,,% and he,% are illustrated in Tables 1-3. From the tables, the satisfactory results of A,,% G2.426 (if S/N 30.5) and he,% G 2 can be obtained if the 3rd B-spline wavelet is selected while truncation frequency 1 =4. According to the above-mentioned discussion, we came to the conclusion that the third-order B-spline wavelet and truncation frequency 1 = 4 are optimum parameters.Their use results in a reduction of high-frequency noise from the UV signal. In other words, the third-order B-spline wavelet MRA offers potentially large benefits in the extraction of the useful signals from noise. Experimental Results In order to establish the effectiveness of the wavelet procedure, determinations were carried out for the above heterodinuclear complex system. When the concentrations decreased and the scanning rate increased, a noise-deformed signal was ob- tained. Fig. 5(h) shows the processed UV spectrum obtained with the third-order B-spline wavelet. The reagents used in this determi- nation were 5.0 X 10-8 moll-' Gd3+ ion, 4.0 X 10-7 mol 1-1 Sr2+ ion, 3.0 X 10-7 mol 1-1 CPA I11 and 0.1 mol 1-1 NaN02. The raw spectrum for this system was recorded from 500 to 850 nm and digitized at 2 nm intervals and at a scanning rate of 400 nm min-1.In spite of the poor resolution in Fig. 4(a), wavelet processing produced very satisfactory results. The most efficient process- ing in this instance was obtained using the third-order B-spline wavelet at a frequency 1 = 4. 0.m 0.00 0 WAVELENGTHhm Fig. 4 1 = 3. Effect of different truncation frequencies 1: points, without processing; solid lines, after processing. S/N = 0.2, m = 3 and k = 29. a, 1 = 6 and b, Table 1 Error data with the second-order B-spline wavelet basis at different S/N 1 = 4 1 = 5 S/N 100 80 60 50 30 20 10 5 3 1 0.8 0.5 Peak 1 he,% A,,,% 1.1729 -2.6497 1.1728 -2.6083 1.1728 -2.6684 1.1728 -2.5047 1.1728 -2.6892 1.1728 5.6381 1.1728 -2.6274 1.1728 -4.86 16 1.1728 -3.2754 -2.0908 -14.129 1.1729 - 1.3028 1.1729 5.4366 Peak 2 - her,% A,,% -0.4712 -0.2097 -0.4712 -0.2106 -0.47 17 -0.2284 -0.4712 -0.2024 -0.4712 -0.2363 -0.4712 0.3960 -0.4712 -0.2752 -0.4712 -0.5110 -0.4715 -2.9778 -0.4712 -2.5399 -0.4712 -1.3209 2.5447 -6.4847 Peak 1 Peak 2 he,% -2.1021 -2.0144 - 1.9985 -0.8549 2.6874 -1.1025 - 1.1058 - 1.2478 - 3.2042 2.6588 3.1244 -5.5488 Aem% -1.8216 -0.8581 - 1.3563 -0.3591 - 1.6320 -0.5759 - 1.1355 --2.6808 -3.2737 4.8926 5.8320 -8.1 173 hem% -0.3698 -0.4436 -0.5814 -0.2147 -0.2199 -0.4597 -0.9874 - 1.2488 2.1544 3.6655 8.2014 6.8216 A,,% 0.2567 0.9790 0.3680 1.2610 1.2559 2.4558 3.4452 3.3641 3.9388 4.48 18 -7.4217 -6.9945Analyst, August 1996, Vol.121 002 a 1023 0.02 b Conclusions removal of high-frequency noise, and we expect that this W 0.01 1 I 0 0 .strategy will be useful fo; processing other signals such as voltammetric signals. This study of the application of wavelet MRA in UV spectrophotometry with three different B-splines demonstrates I that wavelet MRA, with a judicious choke of 1 and rn, to separate signals from high-frequency noise appears to be a promising technique. Among the procedures examined, the third-order B-spline wavelet MRA is the most effective in the This work was supported by the National Natural Science Foundation of China and the Ma Can-an Educational Founda- tion of Zhongshan University. Table 2 Error data with the third-order B-spline wavelet basis at different S/N 1 = 4 1 = 5 S/N 100 80 60 50 30 20 10 5 3 1 0.8 0.5 he,% -0.2030 0.0000 - 0.203 9 -0.2039 -0.2014 -0.0326 -0.2 12 1 0.0000 -0.2 120 -0.1893 -0.1883 -0.2039 Aerr% -3.2030 -3.1044 -3.1220 -3.2133 -3.1254 -2.9592 -2.9870 -4.9906 - 1.1240 0.7430 3.0474 1.0066 0.7070 0.7070 0.7070 0.7070 0.7070 0.7070 0.7070 0.7070 0.7070 0.7070 0.7070 0.7070 Aerr% - 3 S450 -3.7429 -3.620 - 3.6097 -3.1014 -3.6234 -3.1046 -3.8715 -3.1242 -4.7940 - 1.3096 -2.4260 Peak 1 Peak 2 kern% 0.0000 0.2038 0.2038 0.2038 0.2038 0.2038 0.2038 0.2038 0.2038 0.7079 1.8358 1.0197 Peak 1 Peak 2 Aerr% -0.8216 -0.5581 -0.3563 -0.5591 -0.6320 -0.2759 -0.1 355 -0.6808 -0.2737 1.8926 0.8320 -4.1073 kern% 0.1415 0.1414 0.1414 0.14 14 0.1414 0.1414 0.1414 0.1414 0.1414 0.7 147 0.7070 -0.2357 Aem% 0.3561 0.2796 0.3650 0.2614 0.2558 0.2558 0.4422 0.3541 1.9355 6.0794 6.0793 -2.478 1 Table 3 Error data with the fourth-order B-spline wavelet basis at different S/N 1 = 4 1 = 5 Peak I Peak 2 Peak 1 Peak 2 S/N hem% 100 0.8157 80 0.8157 60 0.8157 50 0.8157 30 0.8157 20 0.6119 10 0.8157 5 0.8174 3 0.8157 1 0.8157 0.8 0.2038 0.5 0.8157 A,,% -3.1096 -3.1012 -3.1262 -3.1550 -3.1016 -3.4362 -0.1734 6.3866 -3.1016 6.3849 4.777 1 1.0372 he,% 0.1414 0.1414 0.1414 0.1414 0.3290 0.1530 0.1414 0.1414 0.2144 0.5 184 0.1414 -0.0470 A,,% -2.2560 -2.3060 -0.28 15 -2.1867 -2.3146 -2.1631 -2.2804 -2.3691 1.6316 2.2591 1.3613 0.2910 hem% 0.0000 0.2038 0.0000 0.0000 0.0000 0.0000 -0.05 1 1 -0.2550 0.1121 0.3569 0.5608 -0.2525 A,,% - 1.0600 -0.13 19 -0.9 146 -0.8584 -0.8585 - 1.4368 - 1.4483 - 2.3907 -2.7596 6.5484 2.2863 10.8029 kern% 0.1414 0.1414 0.1414 0.1414 0.1414 0.1414 0.092 0.0942 0.0942 0.8483 0.2828 -0.2828 A,,% -0.2655 -0.4450 -0.1783 -0.2024 -0.2024 -0.3918 -0.4540 -0.4094 2.5246 6.763 I - I .2609 -4.9399 WAVELENGTHhm Fig.5 and b, after processing. UV spectrum of Gd3+-Sr2+-CPA 111 complex in 0.1 moll-’ NaN02 solution. pH, 5.0; scan rate, 400 nm min-l; Ah, 2 nm. a, Without processing;1024 Analyst, August 1996, Vol. 121 References 12 1 2 3 4 5 6 7 8 9 10 11 Brown, S. D., Bear, R. S., and Blank, T. B., Anal. Chem., 1992, 64, 22R. Miller, R. M., Anal. Proc., 1988, 25, 350. Heyes, J. W., Glover, D. E., and Smith, D. E.,Anal. Chem., 1973,45, 277. O’Halloram, R. J., and Smith, D. E., Anal. Chem., 1978, 56, 1391. Bialkowski, S. G., Anal. Chem., 1988, 60, 355A. Brereton, R. G., Anal. Proc.. 1989, 26, 3 11. He, X., and Shi, H., Fenxi Nuaxue, 1994, 22, 94. Chui, C. K., An Introduction to Wavelets, Academic Press, San Diego, 1992. Liu, G. Z., and Di, S. L., Wavelet and Its Application, Xi’an University Press, Xi’an, 1992 (in Chinese). Chui, C. K., and Wang, J. Z., Proc. Am. Math. Soc., 1991, 113, 758. Grossman, A., Morlet, J., and Paul, T., J . Math. Phys., 1985, 27, 2473. 13 14 15 16 17 Meyer, Y., Wavelets: Algorithms and Application, Society of Industrial and Applied Mathematics, Philadelphia, PA, 1993. Daubechies, I., Ten Lectures on Wavelets, Society of Industrial and Applied Mathematics, Philadelphia, PA, 1992. Grossman, A., and Morlet, J., SIAM (Soc. Ind. Appl. Math.) J . Math. Anal., 1984, 15, 723. Zhang, S. W., Orgunic Reagents in Analytical Chemistry, Chinese Academic Press, Beijing, 1983, p. 156. Lu, X. Q., Kang, J. W., Gao, J. Z., and Mo, J. Y., .I. Elecmxmal. Chem.. in the press. Mallat, S. G., Trans. Am. Math. So(.., 1989, 315, 69. Paper 6/00798H Received February 2 , 1996 Accepted April 25, 1996

ISSN:0003-2654    

DOI:10.1039/AN9962101019

出版商:RSC    

年代:1996

数据来源: RSC

摘要:

Analyst, August 1996, Vol. 121 (1025-1029) 1025 Robust Multivariate Calibration Algorithm Based on Least Median of Squares and Sequential Number Theory Optimization Method ~~~ ~~ Yi-zeng Liang" and Kai-Tai Fangh a Department of Chemistry and Chemical Engineering, Hunan University, Changsha, China Department of Mathematics, Hong Kong Baptist UniversiQ, With' the help of constraints on the concentrations to be estimated in direct multivariate calibration, an algorithm for least median of squares (LMS) was developed. The sequential number theory optimization (SNTO) method developed in statistics was used for robust multivariate calibration in order to reach the global optimization. The computational complexity of LMS is dramatically reduced by means of the constraints on the concentrations to be estimated and the SNTO method.The algorithm was applied to a simulated data set and to two sets of real data from two- and/or three-component analytical systems. Comparisons with the least squares technique show that the proposed method is efficient in terms of computational complexity and is robust to a large number of outliers in the data. Keywords: Global optimization; least median of squares; multivariate calibration; number theory method; robust estimator; sequential number theory optimization Introduction Most of the methods commonly used in chemometrics are based on the least squares (LS) technique, e g . , multiple linear regression (MLS), principal component regression (PCR), partial least squares (PLS) and principal component analysis (PCA).It is well known that the LS technique is not robust,1-2 since its objective function is the sum of residual squares, i.e., Minimize %el2 (1) Here rl ( i = 1, ..., n) are residuals from LS for direct multivariate calibration. The first step towards a robust regression estimator was taken by Edgeworth. I He argued that outliers have a very large influence on LS because the residuals rl are squared. 1 Therefore, he proposed a least absolute values regression estimator, which is determined by (2) This technique is often referred to as L, regression, whereas LS is L?. Unfortunately, L1 regression is still very sensitive to the other type of outlier, the so-called bad leverage point. An example of this can be found in the paper of Rousseeuw.3 In order to evaluate the robustness of methods used in statistics, a definition of the breakdown point was first introduced by Hodges4 and a much more general equation was given by Hampel.5 It is defined as the smallest fraction of contamination (outliers) that can totally offset the estimator from the real estimator.The breakdown point for LS (L2) is l/n, which means that if there is a single outlier in a set of n observations, the estimates from LS will be totally destroyed. The breakdown point of L1 is still lln because of the existence Minimize C I Y, I Hong Kong of a leverage point, which is defined as the point which lies far away from the bulk of the observed points in the sample. Thus, Rousseeuw6 developed the so-called least median of squares (LMS) estimator using the following objective in order to obtain a robust method with a higher breakdown point Minimize med ( ~ ~ 2 ) c i = 1, ..., n (3) Here med(.) denotes the median of the residual squares.Its breakdown point is 50%, producing the highest possible value achieved by robust methods. On the other hand, the LMS estimator is robust with respect to outliers in both the y and X direction, an example of which can be found in the paper of Rousseeuw.3 Unfortunately, the LMS technique has an abnor- mally slow convergence rate. An algorithm for LMS was originally presented in ref. 2. A study of the problem of calculating the LMS regression line can be found in ref. 7. The results showed that the time complexity for LMS is in general of order (n3), and the algorithm proposed in ref.7 appears to be expected to have a time complexity of order [nlog(n)12. Robust multivariate calibration has attracted increasing attention in chemistry, since there are always some outliers found in chemical measurement data.8.9 Philips and Eyring10 first applied an M-estimator, which is a type of robust regression technique developed in statistics, to the regression analysis of 38 data sets. They showed that in most instances the efficiency of robust regression was about the same as, or better than, LS regression.10 Wolters and Katemanl 1 applied another M- estimator to study the influence of different distributions on parameter estimation with Monte Carlo simulations. Wei et al. 12 applied a sine function M-estimator to multicomponent analysis of UV data.The results showed that when the noise distribution is not normal, a robust method can provide better estimated concentrations than LS.12 Xie et al. l3 applied several M-estimators to multicomponent analysis for treating a partial non-linearity problem. The wavelengths with a non-linear deviation were regarded as outliers. The results showed that a robust method can provide better estimated concentrations than LS. l3 A robust principal component regression (RPCR) proce- dure was also recently developed by Walczak and Massart.14 The approach was based on ellipsoidal multivariate trimming (MVT)6,15 and LMS methods. The MVT method was utilized to obtain robust principal component vectors in order to reveal all of the outliers in the X data set in the first step of PCR, and the LMS regression technique was then employed in the second step of PCR to identify the outliers in the y data set using standard residuals from the robust model.The aim of this paper was to provide an algorithm to allow analytical chemists to use the LMS technique with less computation time to process the chemical data. With the help of constraints on the concentrations to be estimated in direct1026 Analyst, August 1996, Vol. 121 multivariate calibration, the computation of the LMS becomes a constrained optimization problem. The sequential number theory optimization (SNTO) method developed in statistics was used for robust multivariate calibration in order to find the global optimum in the region investigated. The computational complexity of LMS is dramatically reduced by means of the constraints on the concentrations to be estimated and the SNTO method.The algorithm was applied to a simulated data set and to two sets of real data for two- and/or three-component analytical systems. Satisfactory results were obtained. Compar- isons with LS show that the proposed method is efficient in terms of computational complexity and is robust to a large number of outliers in the data. Theory and Algorithm The computation of the LMS regression coefficients is not obvious because the median of residual squares is introduced into the objective function. It is probably impossible to derive a straightforward equation for the LMS estimator. In general, the algorithm2 proceeds by repeatedly drawing all the possible sub- samples of p (p < n and is a sub-set of n observations) different observations, and then calculating their LMS objective func- tions as defined in eqn.(3). This is essentially a combinatorial optimum problem which makes the time complexity of the algorithm of LMS very high, e.g., of order (173). The improved variant of the algorithm developed by Steele and Steiger7 still has a time complexity of order [nlog(n)]2. Fortunately, the coefficients to be estimated in direct multivariate calibration are always the concentrations of the chemical components in the analytical systems studied. This enables the computation of LMS to be simplified. Constraints of Concentration Estimates in Multivariate Calibration As the coefficients to be estimated in multivariate calibration are always the concentrations of the chemical components existing in the mixture, they cannot be less than zero.On the other hand, the maximum possible values of the concentrations can also be confined. For direct multivariate calibration, the mathematical model is as follows: YI = c'lxll + * . . Ct?lk'I,7, + el (44 (4b) or in matrix form: yn x 1 = x,1 x 171 c,, x 1 + en x 1 here y, (i = 1, . . ., n) are the response observations, x,, (i = 1, . . ., n ; j = 1, . . ., m) the so-called calibration variables, cI (j = 1, . . ., nz) the relative concentrations to be estimated and e, the measurement noise. It should be emphasiLed that there are already some assumptions behind eqn. (4), viz., it is assumed that there are no interactions between the chemical components (linear additivity system) and that all the pure component spectra are known.The problem to be solved in this case is how to deal with the outliers existing in the data mentioned above. In general, the number of observations, n (wavelengths for spectrometry), is always much greater than the number of chemical components, m. Note that x , ~ (i = 1, . . ., n; j = 1, . . ., m) are always positive values except for some small noisy measurements in spectrometry. Thus, we have Y1 > L j X - l J + e/ ( 5 ) (6) Here a is a threshold value expressing the noise influence. A positive value much greater than the noise level is taken in this This leads to L', d ( y l + a ) / x l j d max [CyI+a)/-~,1 ic{ 1,. . .,n) work. Hence, the relative concentrations of the chemical components to be estimated will lie in the following ranges: (7) With these constraints for concentration estimates, the computa- tion of the LMS estimator becomes an optimum problem.One can investigate all the ranges for the possible concentrations of the chemical components simply by using the objective function defined in eqn. (3). As n is always much greater than m, the computational complexity for LMS will be reduced dramat- ically. 0 d c; d max [(yi + a)/x,,] Sequential Number Theory Optimization Method The objective function defined by eqn. (3) is not obvious in a mathematical sense. It would probably be impossible to derive a straightforward differential equation for the LMS estimator, because the median of the square residuals is taken as its objective function.Hence, one can never be sure whether there are local minima existing jn the range investigated or not. This calls for a global optimization method to solve the problem. The sequential algorithm for optimization based on number theory and/or quasi Monte Carlo methods developed independently by Fang and Wangl6 and Nieden-eiter and Peart17 seem to be good candidates for this purpose. The idea behind the SNTO method is first to find a set of representative points which most uniformly scatter in a closed and bounded domain D ( = [a, b]) in m-dimensional real number space Rm. Note that the vectors a and b are the minimum and maximum possible concentration vectors for each component [see eqn. (6)]. The set of representative points is the so-called good lattice point (glp) set, which can be generated with the help of the number theory method (see ref.18 for further details). The domain Dk ( = [ak, bk]), based on the obtained location of the best approximation of the objective, is then reduced sequentially until the required precision is reached. The algorithm for SNTO is as follows: Step 1: Generate a gZp set. Use the number theory method to generate h points p(f) (see Appendix B) uniformly scattered on D(t) = [a('), b(f)]. Step 2: Compute a new approximation. Find dt) E pif)uc(r - and optimum approximation f(c(f)) such that f(c(')) < f ( c ) c(') E p(')Uc(t - 1) (8) Here is the empty set, f(c(1)) and c(t) are the best approximation to the global optimum f(c*) and its correspond- ing location c*. In this work, the objective function f ( c ) is defined by eqn.(3). Step 3: Termination criterion. Let d(') = (b(t) - a@))/2. If all the elements in d('), say dJ(t) (i = 1, . . ., m), are less than a pre- assigned small number, then D(t) is sufficiently small, and f(c(')) and c* are acceptable and the algorithm is terminated. Otherwise, proceed to the next step. andAnalyst, August 1996, Vol. 121 1027 where y is a pre-defined contraction ratio (in this work, we use 0.5). Set t = t + 1. Go to step 1. Experimental Several simulated data sets with different numbers of outliers in the observations were first created in order to check the proposed method. The outliers were artificially added to the data by contaminating a normal distribution according to the following equation: noise = (1 - E)N(O,O.O~)UEU([O,S]) (10) where N(0,0.01) denotes the normal distribution with 0 and 0.01 as the mean value and the variance; respectively, and U([0,6]) is the uniform distribution within the range [0,6].E is a percentage value denoting the contaminated part in the uniform distribution. The larger E, the greater the number of outliers present in the data. Standard solutions of rn-nitroaniline, phenol and pyro- catechol were prepared with 40% ethanol solution. Two sets of mixture samples were prepared under the same experimental conditions. Set 1 were mixture samples of two chemical components, e.g., rn-nitroaniline and phenol, with different concentration ratios. Set 2 were mixture samples of three chemical components. The concentrations of these two mixture sets are listed in Table 1.All the solutions were prepared with distilled water. The chemicals were of analytical-reagent grade and were used as received. The measurements were carried out on a Beckman (Fullerton, CA, USA) Du-7 UV/VIS spectro- photometer. The data were recorded in the wavelength range 190-289 nm at 1 nm intervals. The pure spectra for the three chemical components are shown in Fig. 1. Table 1 Concentrations of each component for two sets of mixtures Set I : Two component analytical system- M-Nitroanilinel Phenol/ No. 10-4 moll-' 10-4 moll-' 1 1.2650 0.8563 2 0.6325 0.8563 Set 2: Three-component analytical system- m-Nitroanilinel Phenol/ Pyrocatechol/ No. mol 1-I 10-4 moll-' moll-' 1 0.1236 0.40 16 0.2189 2 0.3090 0.1907 0.2 189 2 1 .a 1.6 1.4 8 1.2 m $ 1 CJY n a 0.8 0.6 0.4 0.2 0 A 190 200 210 220 230 240 250 260 270 280 Wavelengthhm Fig.1 phenol; and C, pyrocatechol. Pure spectra of three chemical components. A, m-Nitroaniline; B, The programs implementing the proposed algorithm were written in Matlab for windows (version 4.0). All the calcula- tions were carried out on an IBM 486 PC. Results and Discussion The results obtained for the simulated data set are collected in Table 2, from which it can be seen that as the values of E and 6 increase, the estimates from LS become worse, whereas the results from LMS are not affected even when the value of E is as high as 45%. The response surface for LMS in the domain investigated for a simulated two-component system is shown in Fig. 2, from which it is obvious that there are indeed several local minima in the domain.Hence it is necessary to apply a global optimization technique in order to use the LMS estimator in this way. Of course, other stochastic global optimization methods, such as the genetic algorithm (GA) 19,20 and simulated annealing (SA)213 can also be used for this purpose. The most important step in the SNTO method is to generate a set of representative points scattered uniformly in a closed and bounded domain. There are several techniques to create such a point set.18 The most commonly used glp set based on number theory was adopted in this work. The method for generating a glp set of h points is shown in Appendix B. It should be noted Table 2 Results of simulation with different degrees of contamination Relative error (%) No.1 2 3 4 5 6 7 8 E 10 5 10 20 30 40 45 3 6 0.1 0.2 0.2 0.2 0.2 0.2 0.2 0.5 Estimated by LS -6.34 -2.13 -1.7 -1.4 - 10.1 -4.2 -20.1 -6.6 -25.3 -6.32 -30.3 -11.2 -34.2 - 12.4 -9.02 -3.41 Estimated by LMS 0.1 -0.1 0.1 -0.1 0.02 0.8 -0.6 0.1 -0.2 0.5 0.1 -0.1 -0.12 0.1 0.08 -0.02 0 -5 in E 5 8-10 E Y -15 1 Fig. 2 Response surface of LMS for a simulated data set of a two- component system with several outliers. The logarithm of the response values of LMS is used for ease of viewing.1028 Analyst, August 1996, Vol. 121 that the number of points included in the glp set can be chosen by the user for different purposes (see Appendix B). It seemed necessary to confirm further whether the final results obtained from SNTO were in fdct the global optima or not.Hence two optimum computations were performed with the adjacent number of points in the glp set table. The global optimum is reached if the results for the two computational procedures are almost identical. The two real data sets described under Experimental were analysed by both the classical LS and the proposed LMS algorithm. The results obtained are shown in Table 3, from which it can be seen that the results from LS are not satisfactory. The relative estimated errors are high, which indicates that there must be some outliers (non-linear observations) in the data. The results from the LMS estimator confirmed this conclusion. This illustrates that the LMS estimator is superior to LS, if there are some outliers present in the data.On the other hand, LMS can be used for the detection of outliers. 14 This may be important information in multivariate spectral analysis, because it can be used for the selection of good wavelengths of linear additivity. A typical result for the detection of outliers in a mixture spectrum is shown in Fig. 3, Table 3 Results for two sets of real data Set I : Two-component analytical system- Actual concentration/ Estimated Estimated Relative error (%) No. moll- by LS by LMS 1 1.2650* -6.49 0.4 0.8563+ -6.38 2.23 2 0.6325* -7.44 -2.12 0.8563* -15.99 0.93 Set 2: Three-component analytical system- Actual concentration Estimated Estimated Relative error No. 10-4mo11-' by LS by LMS 1 0.1236* 25.2 -0.80 0.40161 16.2 1.21 0.2 1 89t 38.6 0.00 2 0.3090* 14.2 2.67 0.1907 70.1 - 1.60 0.2 189; 38.4 4.65 * m-Nitroaniline.+ Phenol. * Pyrocatechol. * + t + ' 190 200 210 220 230 240 250 260 270 280 290 Wavelengthhm Fig. 3 Use of LMS as a tool for outlier detection. The outlier observation points are denoted by -I-. Solid line, estimated mixture spectrum by LMS; broken line and +, measured mixture spectrum. from which it can be seen that LMS is very robust and is able to detect the outliers in a mixture spectrum. The main advantages of the sequential algorithm for optimization based on number theory are its simplicity and efficiency. It can be easily understood and programmed. The calculations for the LMS estimates of two- and three- component analytical systems can be performed in just a few minutes on an IBM 486 PC. The Crouch Foundation of Hong Kong Baptist University, the Fok Ying Tong Foundation of the Educational Commission of China and the National Natural Science Foundation of China are gratefully thanked by Y.-z.L.Appendix A List of Notations Used in This Work r-i (i= 1, ..., n) n Number of observations m Number of chemical yi (i = 1, ..., n) Response observations xlj (i = 1, ..., n; j = 1, ..., m) Calibration variables cj o'= 1, ..., m) Relative concentrations to be ei (i= 1, ..., n) Measurement noise Y Vector of response X Matrix of calibration variables C Vector of component e Vector of measurement noise a Threshold value expressing the noise influence D ( = [a, bl) to be investigated a Minimum concentration vector for each component b Maximum concentration vector for each component P Point set scattered uniformly in the domain to be investigated Number of points in p Residuals from least squares components estimated observations concentrations Closed and bounded domain h Y Pre-defined contraction & ratio for SNTO observations Percentage of outliers in the Appendix B Generation of the glp Set With the help of tables of the glp set, the generation of the glp set can be achieved by the following steps: (i) Find the generating vector (h; h l , .. ., h,) with hi-1 and the other integral components satisfying 1 d hi G h, h, # hJ (i # j ) and the greatest common divisors (h, h,) = 1. Here h denotes the number of points to be generated and the subscript rn the number of dimensions of the space investigated (ii) The corresponding glp set can be created using the following equation: Here {x} represents the fractional part of x and k is an integral between 1 and h.The glp points produced are scattered in the domain of [0, 11 of rn-dimensions. Here 0 = [0, . . ., 0It and 1 = [ 1, . . ., 1 1 7 . Tables of the glp set are available in Appendix A of ref. 18. The tables for two- and three-dimensional space are given here for reference. [(hlk - 0.5)/h, { h& - 0.5)/h}, . . ., { (h,k - 0.5)/h ) ] (1 6 Gh)Analyst, August 1996, Vol. 121 1029 Table A1 (m = 2) h 34 55 89 144 233 377 610 987 1597 h , 1 1 1 1 1 1 1 1 1 h2 21 34 55 89 144 233 377 610 987 Table A2 (m = 3) h 35 101 135 185 266 418 597 828 1010 h2 1 1 40 29 26 27 90 63 285 140 h3 16 85 42 64 69 130 169 358 237 hl 1 1 1 1 1 1 1 1 1 References 1 2 3 4 5 6 7 8 Edgeworth, F.Y., Hermathena, 1887, 6, 279. Rousseeuw, P. J., and Leroy, A. M., Robust Regression and Outlier Detection, Wiley, New York, 1987. Rousseeuw, P. J., J . Chemornetr., 1991, 5, 1. Hodges, J. L., Proc. Fifth Berkeley Symp. Math. Stat. Prohah., 1967, 1, 163. Hampel, H. R., Ann. Math. Stat., 1971, 42, 1887. Rousseeuw, P. J., J. Am. Stat. Assoc., 1987, 79, 871. Steele, J. M., and Steiger, W. L., Discrete Appl. Math., 1986, 14, 93 * Clancey, V. J., Nuture (London), 1974, 159, 339. 9 10 11 12 13 14 1s 16 17 18 19 20 21 22 Harris, E. K., and DeMets, D. L., Cfin. Chem., 1972, 18, 605. Philips, G. R., and Eyring, E. R., Anal. Chem., 1983, 55, 1 134. Wolters, R., and Kateman, G., J . Chemometr., 1989, 3, 329. Wei, W. Z., Zhu, W. H., and Yao, S., Chemometr. Intell. Luh. Syst. 1993, 18, 17. Xie, Y. L., Liang, Y. Z., Jiang, J. H., and Yu, R. Q., Anal. Chim. Acta, 1995, 311, 185. Walczak, B., and Massart, D. L., Chemometr. Intell. Lab. SyJt., 1995, 27, 41. Devlin, J. S., Gnanadesikan, R., and Kettenring, J. R., .I. Am. Stut. Assoc., 1981, 76, 354. Fang, K. T., and Wang, Y., in Lecture Notes in Contemporary Mathematics, eds. Yang, L., and Wang, Y., Science Press, Beijing, Niederreiter, H., and Peart, P., SIAM J. Sci. Stat. Comput., 1986, 7, 660. Fang, K. T., and Wang, Y., Numher-Theory Method5 in Statktics, Chapman and Hall, London, 1993. Kirkpatrick, S., Gelatt, C. D., and Vecchi, M. P., Science, 1983,220, 671. Bohachevsky, I. O., Johnson, M. E., and Sein, M. L., Technometrzc,s, 1986, 28, 209. Holland, J. H., Adaptation in Naturul and Artifcial Systems, University of Michigan Press, Ann Arbor, MI, 1975. Handbook of Genetic Algorithms, ed. Davis, L., Van Nostrand Reinhold, New York, 1991. Paper 6l00635C Received January 26,1996 Accepted April 25, 1996 1990, pp. 17-28.

ISSN:0003-2654    

DOI:10.1039/AN9962101025

出版商:RSC    

年代:1996

数据来源: RSC

摘要:

Analyst, August 1996, Vol. 121 (1031-1036) 103 1 Flow Injection-Fourier Transform Infrared Spectrometric Determination of Oil and Greases : Preliminary M icrowave-assisted Extraction Studies Yasmina Daghbouche, Salvador Garrigues and Miguel de la Guardia* Department of Analytic-a1 Chemistry, University of Valencia, SO Dr. Moliner St., 461 00 Burjussot (Valencia), Spain A flow injection-Fourier transform infrared method has been developed for the determination of oil and greases previously dissolved or extracted in CC14. The use of the area under the FTIR absorbance spectra, from 3058 to 2780 cm-1, corrected for a baseline established between 3200 and 2700 cm-l, provides a sensitive and precise means for the determination of oils of different types, the only requirement being the use of standards of the same type as the samples.Using a 300 p1 injection volume and a carrier flow rate of 1.5 ml min-I, a sampling frequency of 60 h-1 can be obtained, with a limit of detection of 1 pg m1-l and an RSD of the measurements between 0.5 and 4 %. Preliminary results for the microwave-assisted extraction of aqueous samples with CC14 in a PTFE reactor provided quantitative recoveries for mineral oil, edible oil, gasoline and petroleum when an irradiation time of 6 min and an exit power of 420 W were employed. Keywords: Fourier transform infi-ared specti-ometry; flow injection; oil; greases; mic.I-owa~,e-ussisted extraction Introduction Hydrocarbon oils and greases are widely used in agriculture, industrial and domestic activities. Hence, the presence of oils in waste waters has different causes and origins and the determina- tion of these products is difficult owing to the large number of compounds involved.The typical absorbance of hydrocarbons in the IR region provides an excellent means for both the identification of oil wastes and their quantitative analysis, being a good alternative to the standard protocol for the determination of oil and grease in waste water, which currently utilizes liquid-liquid extraction with 1,1,2-trichloro- 1,2,2-trifluoroethane (Freon 1 13) and gravimetric determination.2 Following the pioneering study of Simard et u Z . , ~ a mixture of isooctane, hexadecane and benzene (37.5 + 37.5 + 25) was employed for the determination of oils in waste water^.^ However, from 1985 the ASTM method prompted the elimina- tion of benzene for calibrations and mixtures of hexadecane, isooctane and chloroben~ene,~37 hexadecane, pristane and toluenes or hexadecane and isooctane"." were employed for calibration in the determination of hydrocarbons by IR spectrometry.It seems clear that these calibration standards are not representative of the different types of oils and greases that can be found in waste waters, and for this reason only relative values of hydrocarbon concentrations in water can be obtained following the aforementioned procedures. In consequence, the ASTM5 suggested that the preferred material for calibration is a * To whom correspondence should be addressed. sample of the same oil and grease that is known to be present in the sample of waste water. Another problem with the FTIR determination of hydro- carbons and oils is the use of Freons6.7.9 or CC143,4,8 as solvents for the initial dissolution or extraction of these compounds from waters and the urgent necessity to reduce the use of ozone- depleting solvents, specially chlorofluorocarbons,lo in order to develop environmentally friendly analytical procedures.I 1 The aim of this work was to carry out a revision of the experimental conditions for the determination of oils of different types by FTIR spectrometry, with particular con- sideration of the interesting qualitative information provided by the IR spectra. Additionally, we have tried to improve this determination, based on our previous experience with the automation of FTIR procedures, by means of flow injection, which makes possible a substantial decrease in reagent consumption and an improvement in the productivity of laboratories, 2,11 simplifying the steps of filling and cleaning the measurement cells and also providing a good way for the on- line decontamination of wastes14 and the on-line recycling of solvents. To solve the problem of the extraction of oils from water samples, which is not easy by liquid-liquid extraction or solid- phase extraction, we employed a microwave-assisted extraction procedure based on our previous experience in this field with the solvent extraction of PAHs, linear hydrocarbons, PCBs and pesticides from sediments for their subsequent determination by gas chromatography. I 6 Experimental Apparatus and Reagents A Magna IR 550 system from Nicolet (Madison, WI, USA) equipped with a DTGS temperature-stabilized detector with a KBr beamsplitter and precise digital signal processing (DSP) was used to carry out the 1R measurements using a Ultra-Micro quartz flow-through cell of 1 cm pathlength and 30 p1 internal volume from Hellma (Mullheim, Germany). The time required to obtain and store the whole interferogram of samples is 1.03 s, using a spectral resolution of 8 cm-l and a speed of the moving mirror of the interferometer of 0.6329 cm S - I .Version 1.2 of the Omnic software developed by Nicolet was employed for processing the FTIR absorbance data in both the stopped and flow modes. The manifold shown in Fig. 1 was employed for the analysis of real and synthetic samples.An HPLC pump from Hewlett- Packard (Waldbronn, Germany) was employed to prevent the formation of air bubbles inside the system. A Rheodyne (Cotati, CA, USA) injection valve including injection loops of PTFE with volumes from 100 to 600 p1 was employed for theI032 Analyst, August 1996, Vol. 121 0.00 -- introduction of samples and standards. All the connecting tubes in the manifold were made of PTFE and had an id of 0.8 mm. A Sharp (Seoul, Korea) R-2V16 domestic microwave oven of 700 W maximum exit power and laboratory-made PTFE reactors of 115 ml internal volume were employed for the microwave-assisted extraction of water samples. Carbon tetrachloride (99.7%) from Scharlau (Barcelona, Spain) was employed as the carrier and for the dilution of samples and standards.Analytical-reagent grade isooctane and benzene were ob- tained from Panreac (Barcelona, Spain) and hexadecane from Aldrich (Steinhein, Germany). A reference oil was prepared from a mixture of hexadecane, isooctane and benzene (37.5 + 37.5 + 25) and 50 mg of this mixture were diluted to 25 ml with CC14 in order to obtain a 2 mg ml-1 stock standard solution. Working standard solutions in the concentration range 0.059-0.227 mg ml-' were prepared from the stock standard solution. General FI-FTIR Procedure Portions of 300 pl samples, previously dissolved or extracted in CC14 as indicated, are injected into the FI manifold (Fig. 1) using a carrier (CC14) flow rate of 1 .5 ml min-l. Absorbance measurements are made in the area mode, from transient signals between 3058 and 2780 cm-l using a baseline established between 3200 and 2700 cm- 1 .The peak-height values of the FI recording obtained for samples are interpolated in a regression line established using CC14 solutions of standards of the same type as the samples. Alternatively, absorbance measurements at a fixed wavenumber of 2928 cm-J can be employed to obtain the FI recordings. Provisional Microwave-assisted Extraction Procedure Water samples (25 ml) are introduced into a laboratory-made PTFE reactor with an internal volume of 1 15 ml, 10 ml of CC14 are added and the reactor is closed and irradiated at 420 W for 6 min. After irradiation, the reactor is cooled and the organic phase is removed and dried using Na2S04. Results and Discussion FTIR Spectra of Standard Mixtures and Real Samples Fig.2 shows the spectra of standard solutions of isooctane, hexadecane and benzene dissolved in CC14 and that of a solution containing these three compounds in the proportions 37.5 + 37.5 + 25. As can be seen, hexadecane provides an intense band at 2928 cm-I due to the CH2 groups, the CH3 bands corresponding to ETIR instrument HPLC pump Fig. 1 Manifold employed for the FI-FTIR determination of oils. isooctane being located at 2959 cm-1 and the CH aromatic bands at 3040 cm-l in both pure and mixed solutions. Comparing the above spectra with those found for real samples (Fig. 3), it can be concluded that the IR spectrum in the wavenumber range 3250-2600 cm-1 provides good infor- mation about the nature and amount of oils present in CC14 solution.It is clear that the relative proportions of the different absorbance bands in this region vary from sample to sample, thus preventing the possibility of using peak-height absorbance measurements of synthetic standards, with a specific composi- tion, for the quantification of oils of different origin, but open up possibilities for the characterization of the type of oils and greases present in a sample. Analytical Characteristics of the FTIR Determination of Oil Using Isooctane-Hexadecane-Benzene Standards Using, for calibration, ternary mixtures of isooctane, hexade- cane and benzene, excellent analytical features of the calibra- tion lines obtained in CC14 were found using single absorbance measurements at each of the most representative bands of the three compounds employed, and also for the regression between the sum of the absorbance values obtained at the characteristic bands at 2928,2959 and 3040 cm-1 and the total hydrocarbon concentration in mg ml-l, and for area values under the absorbance spectra in the wavenumber range 3058-2780 cm-1.For single absorbance measurements, the calibration equa- tions obtained at 2928, 2959 and 3040 cm-1 are A = 0.002 + 2.114C ( r = 0.9999), A = 0.000 + 1.873C ( r = 0.99999) and A = 0.0002 + 0.202C ( r = 0.9996), respectively, and provide 0 0 0 0 Waven u m ber/cm-' Fig. 3 Absorbance spectra of a gasoline sample (.-) 0.1321 mg ml-l, a mineral oil (15W/40) (---) 0.1320 mg ml-l, an olive oil (-) 0.1357 mg ml-' and a petroleum sample (-) 0.1299 mg ml-I, diluted in CC14.Analyst, August 1996, Vol.I 2 I 1033 limits of detection of 0.08, 0.08 and 1.2 pg ml- 1 for a k = 3 probability level of 99.6%, the RSD of 10 independent measurements of a 0.115 mg m - l solution of the three compounds being 0.07, 0.09 and 0.7%, respectively. On the other hand, using the sum of the peak-height absorbance at the three wavenumbers considered, the calibra- tion equation obtained is A = 0.002 + 4.188C ( r = 0.9999) and provides a limit of detection of 0.16 pg ml-1, the RSD of absorbance measurements being 0.08%. The sensitivity of the FTIR determination of oils, using isooctane, hexadecane and benzene as standards and CC14 as the solvent, can be improved by using area values of the absorbance spectra in the wavenumber range 3058-2780 cm-1, the calibration equation found under these conditions being A = 0.014 + 174.9C ( r = 0.9999), but the limit of detection remains of the order of 0.17 pg ml-' and the RSD is 0.08%.In all the aforementioned cabes the absorbance data were corrected for a baseline established between 3200 and 2700 cm-1. Analysis of Real Samples Using fsooctane-Hexadecane-Benzene Standards The analysis of real samples of different types of compounds, such as edible oils, mineral oils, petroleum and gasoline dissolved in CCld, using isooctane-hexadecane-benzene stan- dard solutions, provided a poor recovery of the oil concentration in all cases. Table 1 shows, as an example, the values found for area measurements in the wavenumber range 3058-2780 cm-1, which indicate that only semiquantitative data could be obtained using a general calibration line for the analysis of samples of different types.Table 1 FTIR determination of different types of oil samples dissolved in CCL. Area values between 3058 and 2780 cm-l were used for calibration with a baseline correction established between 3200 and 2780 cm-1 Analysis A ' Analysis B" Sample Edible oilx- 01 i ve Refined olive Sunflower Miner111 oils- 1 5 W/40 20W/50 (used) 20w/50 Petrolem- QUOA MAYA Un lecrdeu' extru gmolitx- Gas.- I Gas.-' Added Found Error /mg ml-1 /mg m - l (%) 0.0467 0.040 1 - 14.1 0.1393 0.1 150 -17.4 0.0705 0.0556 -21.1 0.1036 0.0824 -20.5 0.0446 0.0360 - 19.3 0.0704 0.0649 -7.8 0.0404 0.0538 33.2 0.0871 0.1177 35.1 0.0255 0.0329 19.6 0.0762 0.0964 26.5 0.041 I 0.0526 28.0 0.0766 0.1020 33.2 0.0376 0.0422 12.2 0.0682 0.0779 14.2 0.0303 0.0337 11.2 0.0669 0.0744 11.2 0.0557 0.0378 -32.1 0.1066 0.0782 -26.6 0.0585 0.0384 -34.4 0.1 118 0.0802 -32.0 Found Error /rng ml-' (96) 0.0500 7.1 0.1424 2.2 0.0686 -2.7 0.1020 -1.5 0.0442 -0.9 0.0802 -0.3 0.0400 -1 0.0869 -0.2 0.0246 -3.6 0.0713 -6.5 0.0391 -4.9 0.0754 -1.6 0.038 I 1.3 0.0693 1.6 0.0307 1.3 0.0662 --I 0.0537 -3.6 0.1059 -0.7 0.0545 -6.8 0.1085 -3 "A: Using calibration lines obtained with standards of ternary mixtures of isooctane.hexadecane and benzene prepared in CCI4. B: Using standards of the same type as the samples. Surprisingly, using peak-height absorbance measurements at 2928 cm-' and a general calibration line (results not shown), errors <S% were found for the analysis of edible oil samples.However, for mineral oils the average errors were of the order of 47% and for petroleum samples an average error of 17% was found using absorbance measurements at the typical wave- number of the CH2 groups. As a consequence of the data obtained, it is clear that a general discussion about the adequacy of the standard solutions required for accurate IR determination of oils is required. Adequacy of Standards for the Determination of Oil by FTIR As can be seen in Table I , accurate results for the FTIR analysis of different types of oil samples can be obtained when standards of the same type as the samples are employed. Thus, using standards prepared for olive oil, 15W/40 mineral oil, petroleum and gasoline, average relative errors of the order of 2.6% were obtained for the determination of oil dissolved in CC14 at concentrations from 0.025 to 0.1 1 ing ml-1.The aforementioned results indicate that IR spectrometry is an appropriate technique for the determination of oils, it being necessary, in order to obtain accurate results, to perform a preliminary extraction of solutions of the samples in CC14 to characterize the type of oils present in the samples, and, after this preliminary characterization step, to determine the concen- tration by using appropriate standards of the same type as the samples. When appropriate standards are employed, accurate results can be obtained also using peak-height absorbance measure- ments at 2928 cm-1 and for the analysis of the aforementioned samples an average relative error of the order of 1.4% is found.The use of different reference standards also changes slightly the analytical features of the FTTR determination. Table 2 summarizes, as an example, the main figures of merit obtained using olive oil, mineral oil, petroleum and gasoline as standards and compares them with those obtained using synthetic mixtures of isooctane, hexadecane and benzene for both peak- height absorbance measurement at 2928 cm- I and area measurements. As can be seen, the limit of detection is, in all cases, between 0.1 and 0.3 pg ml- and the RSD varies from 0.1 to 0.3%, better sensitivity being obtained using area measure- ments than peak-height values. FI-FTIR Determination of Oils The use of a fast FTIR instrument allows us the possibility of determining oils, previously dissolved or extracted in CC14, automatically, with injection of the samples into a carrier stream of CC14 and analytical treatment of the transient signal obtained.Using the manifold shown in Fig. 1, different volumes of standards and samples were injected using different carrier flow rates. Experiments carried out using peak-height absorbance measurements at 2928 cm-l, corrected with a baseline established between 3200 and 2700 cm-I, indicated that the highest sensitivity and reproducibility of the FIA recordings can be found at carrier flow rates of 1 and 1 .5 ml min- I , the latter providing the best sampling frequency with high sensitivity (Fig. 4). At a flow rate of 1.5 ml min-1, the injection of increasing volumes of samples provides increasing sensitivity values, as can be seen in Fig, 5.However, the use of injection volumes 2500 pl leads to poor precision of the experimental measure- ments, so an injection volume of 300 pl was selected.1034 Analyst, August 1996, Vol. I21 Analytical Features of the FI-FTIR Determination of Oils As can be seen in Fig. 6, highly reproducible FI recordings can be obtained by injecting 300 1-11 sample volumes and using a 1.5 ml min-I carrier flow rate. The different points in this figure correspond to the area values below the IR absorption spectrum, obtained as a function of time. The wavenumber range employed for measurement was 3058-2780 cm-1 corrected using a baseline established between 3200 and 2700 cm-I, as was also used in batch analysis. Table 3 summarizes the main figures of merit of the FI-FTIR method obtained using different types of standards and working in the peak height and area modes. Comparing the character- istics of the FI-FTIR method with those found in batch measurements, it can be concluded that the use of transient signals decreases the sensitivity and provides a lower precision than the use of conventional measurements, owing to the use of only four cumulated spectra to obtain the IR data in FI as compared with the 25 spectra employed for batch measure- ments. However, the FI method provides a sampling frequency of 60 h-I and strongly improves the steps of filling and cleaning the measurement cell, and also lowers the consumption of reagents and the amount of waste.Additionally, as we have demonstrated recently, 1s the incorporation of a small distillation unit at the exit of the measurement cell permits on-line distillation of the carrier solvent, thus decreasing the consump- tion of CC14 and providing an environmentally friendly method that avoids the accumulation of toxic residues in the labora- tory.The analysis of samples of different types provides accurate results using either peak height or area values of the FI recording, as can be seen in Table 4, taking into account that the calibration standards employed for the analysis of edible oil samples were prepared from olive oil, those for mineral oil with unused 15W/40 oil, those for petroleum MAYA crude oil and those for gasoline samples unleaded extra petrol standards. On the other hand, when analyses were carried out using ternary mixtures of isooctane, hexadecane and benzene, inaccurate results were found and the average errors for edible oil analysis were of the order of 3.3%, mineral oil 42%, petroleum 45% and gasoline 26%, indicating the necessity for using standards of the same type as the samples in order to obtain accurate results.Determination of Oils in Water Samples The previous extraction of oils from water samples is, together with the appropriate selection of adequate standards and the reduction of the amounts of solvents and wastes, one of the main problems of the determination of oils and greases. The traditional liquid-liquid extraction and the modern solid- phase extraction, at room temperature, are unable to provide quantitative recoveries of all types of oils and greases from water samples.However, as demonstrated previously for solvent extraction of organic pollutants in sediment sam- ples,l&l* the use of microwave-assisted procedures can im- prove the extraction steps, providing a fast means for the quantitative removal of organic compounds. The heating inside a microwave oven is due to the absorption of the microwave energy by molecules with a permanent or induced dipole moment and, in the case of oil extraction from water, the presence of the aqueous phase provides the Table 2 Analytical features of the FTIR determination of oils using different standards Isooctane-hexadecane- Petroleum Gasoline Parameters benzene Olive oil Mineral oil Mode of measurement 2928 cm-I Area 2928 cm-' Area 2928 cm-1 Area 2928 cm-1 Area 2958 cm-1 Area Regression 0.005 + 0.33 + 0.008+ 0.40+ 0.005+ 0.22+ 0.001 + 0.06+ -0.002 + -0.34 + line 2.11c 174.3C 2.181C 140.2C 3.121C 237.2C 2.587C 199.8C 1.592C 135.0C coefficient 0.9995 0.9998 0.9990 0.9991 0.9998 0.9999 0.9998 0.9999 0.9997 0.9994 Correlation LOD (n= 10) 0.17 0.2 0.2 0.3 0.11 0.2 0.14 0.2 0.2 0.3 RSD (%) (n = 10) 0.2 0.2 0.1 0.3 0.1 0.2 0.1 0.1 0.1 0.2 Dynamic range/ mg ml-' 0.01684.3359 0.01434.30 16 0.0 1 134.2247 0.01 11-0.2362 0.0201-0.3326 Table 3 Analytical features of the FI-FTIR determination of oils using different types of standards.Area values of the FTIR spectra recorded for transient signals were established between 3058 and 2780 cm- l Isooc tane-hexadecane- Parameters benzene Olive oil Mineral oil Petroleum Gasoline Mode of measurement Height Area Height Area Height Area Height Area Height Area Regression 0.3 + 0.06 + 0.5 + 0.2 + 0.3 + 0.2 + 0.4 + 0.2 + 0.007+ 0.02+ Correlation line 126.2C 43.91C 119.4C 40.4C 180.8C 60.6C 168.8C SS.9C 100.54C 33.86C coefficient 0.9996 0.9994 0.9986 0.9972 0.9993 0.9994 0.9992 0.9974 0.9992 0.9989 LOD (n= lo) 0.8 0.9 0.8 1 0.6 0.6 0.6 0.7 1 1.1 (n = 10) 0.5 0.6 1.4 1.4 1.1 1.6 2.9 4.1 0.2 1.7 concentration/ RSD (%) Range of mg ml-' 0.0 1684.3359 0.0 1434.30 16 0.01 134.2247 0.01 11-0.2362 0.020 1-0.3 326Analyst, August 1996, Vol.121 1035 G E .c -%0.04 ~ I absorbent, thus increasing the temperature of the mixture between the sample and the organic phase and improving the efficiency of extraction by both the increase in the extraction constant and the speed of this procedure.Preliminary studies carried out using closed PTFE reactors, a sample volume of 25 ml and 10 ml of CC14 as extractant 25 9 g 20- 3 t 4- 15- 10- 0.12 1 10 c5 6 I 4 c4 4 9 0 2 4 6 8 1 0 1 2 1 4 c3 c2 cn 0.08 0.04 g o 2 c t 2 0 4 8 1 2 1 6 Tim e/m in 0.00 ' I I 0.0 1 .o 2.0 3.0 Flow rate/ml min-' Fig. 4 Effect of the carrier flow rate on the peak height of the FI recording. Measurements were carried out at 2928 cm-1 corrected with a baseline established between 3200 and 2700 cm-1. The results were obtained with an injection volume of 300 pl of a 0.068 mg ml-1 standard solution of isooctane-hexadecane-benzene in CCI4. 0.06 1 'F ,100 150 200 300 400 500 600 0 100 200 300 400 500 600 Injected vol u me/pI Fig.5 Effect of the sample volume injected on the peak area of the FI recording. Measurements were carried out at 2928 cm-l corrected with a baseline established between 3200 and 2700 cm-l. The results were obtained with a flow rate of 1.5 ml min-1. The concentration of the synthetic standards was 0.075 mg m1-I. 14, I , I 0 2 4 6 8 10 Time/min Fig. 6 FI recordings obtained for different standard solutions of a ternary mixture of isooctane, hexadecane and benzene in CC4 containing (C 1) 0.0270, (C2) 0.0622, (C3) 0.0962, (C4) 0.1388 and (C5) 0.1833 mg ml-I. Data were obtained from area measurements between 3058 and 2780 cm-' with a baseline established between 3200 and 2700 cm-1. The inset corresponds to 10 independent injections of a standard solution of 0.068 mg ml-' .All the data were obtained with a carrier flow rate of 1.5 ml min- I and an injection volume of 300 1-11. indicated (Table 5 ) that mineral oil, edible oil, gasoline and petroleum added to water samples can be quantitatively recovered by using an irradiation time of 6 min and an exit power of 420 W. The quantitative recovery values found in water analysis in this study are due to the combination of heating and the use of an inert beaker for extraction, which increase the extraction efficiency and also reduce the analyte retention on the walls of the extraction vessel. The extraction takes place in a closed system and requires only 10 ml of CC14 (which probably could be lowered; additional studies of this aspect are in progress), thus providing a reduction of the amounts of solvent and wastes.Improvement of the Sensitivity of the FTIR Determination of Oils Based on the strong transparency of CC14 to IR radiation, it could be possible to improve the sensitivity of FTIR measure- ments by increasing the optical pathlength. Therefore, 5 cm micro glass flow-through cells were tried in the FTIR determination of oils using absorbance measurement in the wavenumber range 3000-2700 cm-1. Under these conditions the sensitivity measurements were improved by a factor of 5 compared with those obtained using 1 cm ultra-micro quartz flow-through cells. However, the transparency of glass is lower than that of quartz, so the blank measurements obtained with the 5 cm glass cells were higher and it was therefore impossible to obtain better limits of detection than those indicated in Tables 2 and 3.Conclusion These studies have demonstrated that quantitative results can be found for the determination of oils by FTIR spectrometry using appropriate standard solutions and that the systematic use Table 4 FI-FTIR determination of oils in different types of samples. Standards of the same type as the samples were employed in all cases. FT traces obtained from area between 3058 and 2780 cm-I Height Area Added/ Found/ Error Found/ Error Sample mgml-1 mgml-1 (%) mgml-1 (%) Edible oils- Olive Refined olive Sunflower Mineral oils- 1 5 W/40 20W/50 (used) 20W/50 Crude petroleum- QUOA MAYA Unleaded extra gasoline- Gas.-1 Gas.-2 0.0547 0.1006 0.0725 0.1393 0.0357 0.0766 0.0474, 0.0882 0.0408 0.0766 0.0498 0.0802 0.0356 0.0644 0.0349 0.0637 0.0587 0.1002 0.0628 0.1119 0.0534 0.0978 0.0736 0.1367 0.037 0.079 1 0.0477 0.0884 0.0399 0.0745 0.0489 0.0799 0.0388 0.0702 0.0379 0.0683 0.0600 0.1032 0.0593 0.1067 -2.4 -2.8 1.5 -1.9 3.6 3.3 0.6 0.2 -2.2 -2.7 -1.8 -0.6 9.0 9.0 8.6 7.2 2.2 3.0 -5.6 -4.6 0.0527 0.0980 0.0736 0.1365 0.0386 0.0792 0.0486 0.091 7 0.0401 0.0752 0.0488 0.0799 0.0398 0.0683 0.0373 0.0685 0.0604 0.1018 0.0586 0.1043 -3.7 -2.6 1.5 -2.0 8.1 3.4 2.5 4.0 -1.7 -1.8 -2.0 -0.4 11.8 6.1 6.9 7.5 2.9 1.6 -6.7 -6.81036 Analyst, August 1996, Vol.121 Table 5 Recovery of oil and greases from spiked water samples by microwave-assisted extraction with CC14 and FTIR determination. The extraction of 25 ml of spiked water samples was carried out with 10 ml of CC4 in closed PTFE reactors of 115 ml internal volume using 6 min irradiation at 420 W and after drying the organic layer with Na2S04.For calibration, the equations in Table 2 were employed Area between Pollutant added Added/mg ml-l 3058 and 2780 cm-' Found/mg ml-I Recovery (5%) Mineral oil 0.20 47.5 f 0.1 0.195 101 f 4 0.19 47.5 f 0.3 0.199 0.22 52.7 k 0.2 0.222 0.23 34.10 k 0.04 0.241 0.27 36.07 f 0.07 0.255 0.18 22.58 k 0.08 0.170 0.15 18.9 f 0.1 0.143 Olive oil 0.25 36.5 -t 0.1 0.264 102 f 6 Gasoline 0.14 19.3 f 0.5 0.146 9 8 f 6 Petroleum 0.18 0.19 0.20 33.2 f 0.4 0.168 41 .5 k 0.4 0.207 37.40 f 0.02 0.187 9 9 f 9 of isooctane-hexadecane-benzene does not provide accurate results. The FI-FTIR method proposed in this paper is simple and rapid and can be applied to the analysis of samples previously dissolved in or extracted with CC14. The method provides a limit of detection of the order of 1 pg ml-', an RSD 0.54% and a sampling frequency of 60 h-l.Additionally, the FI method- ology reduces the consumption of reagents and the incorpora- tion of a distillation unit provides on-line recycling of CC4, thus avoiding environmental contamination associated with the use of chlorinated hydrocarbons. The use of a microwave-assisted extraction procedure for the determination of oils in water samples provides a rapid analysis and a decrease in the volume of solvent required from 100 to 10 ml, using a closed system which also avoids solvent losses. The authors acknowledge the financial support of the Spanish DGICYY (Project PB920870) and of the Generalitat Valenciana (Project GV 1021/93).Y.B. acknowledges a grant from the Patronato Sud-Nord of the University of Valencia. References Rotter, S., Boertzler, H., Dubernil, J. P., Evers, J.. Ilsbroux, J., Remstedt, H. G., Somerville, H. J., and Van Strien, W., Conc.awJe Rep., 1984, 1, 1. American Public Health Association, American Water Works Association and Water Environment Federation, Standard Methods for the Examination qf Water and Wastewater, American Public Health Association, Washington, DC, 18th edn., 1992. Simard, R. G., Hasegawa, I., Bandaruk, W., and Headington, C. E., Anal. Chmi., 1951, 23, 1384. 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 Chen, T., Lin, C., and Chen. C., Lihna-Jianyun Huuzue-Fence, 1987, 23, 143. American Society for Testing and Materials, Standard Test Method for Oil and Grease und Petroleum Hydrocarbons in Water, Method 03921, ASTM, Philadelphia, 1985. US Environmental Protection Agency, Oil and Grease, Total Recoverable, EPA 413.2, Environmental Protection Agency, Wash- ington, DC, 1978. Gaches, A. J., and Wilson, P. S., Spectrosc. World, 1991, 3, 22. Department of the Environment and National Water Council, MethodJ Exum. WuterJ Assoc. Muter., 1984, 21. McGrattan, B. .J., Perkin-Elmer IR Bull., 1989, No. 114. Noble, D., Anal. Chenz., 1993, 65, 695A. de la Guardia, M., and Ruzicka, J., Analyst, 1995, 120, 17N. de la Guardia, M., Garrigues, S., Gallignani, M., Burguera, J . L., and Burguera, M., Anal. Chim. Acta, 1992, 261, 53. Garrigues, S., Gallignani, M., and de la Guardia, M., Tulanfa, 1993, 40, 1799. de la Guardia, M., Khalaf, K., Hasan, B. A., Morales-Rubio, A., and Carbonell, V., Anulyst, 1995, 120, 231. Sinchez-Dasi, J., Cervera, M. I., Garrigues, S., and de la Guardia, M., paper presented at the First Mediterranean Basin Conference on Analytical Chemistry, Cordoba, November 5-10, 1995. Pastor, A., Vazquez, E., Ciscar, R., and de la Guardia, M., paper presented at the First Mediterranean Basin Conference on Analytical Chemistry, Cordoba, November 5-10, 1995. Ganzler, K., Salgo, A., and Valko, K., J . Chromatogr., 1986, 371, 299. Lopez-Avila, V., Young, R., and Beckert, W. F., Anal. Chem., 1994, 66. 1097. Paper 6J03454C Accepted May 28, 1996

ISSN:0003-2654    

DOI:10.1039/AN9962101031

出版商:RSC    

年代:1996

数据来源: RSC

摘要:

Analyst, August 1556, Vol. 121 (1037-1041) 1037 Shipboard Determination of Dissolved Cobalt in Sea-water Using Flow Injection with Catalytic Spectrophotometric Detection Alexander Malahoff", Irina Ya. Kolotyrkinab and Lilly K. Shpigunb a University of Hawaii, I000 Pope Road, Honolulu, HI 96822, USA b Kurnakov Institute of General and Inorganic Chemistry, Russian Academy of Sciences, Leninsky Prospect 31, Moscow 11 7507, Russia A flow injection spectrophotometric technique for the determination of cobalt in sea-water and hydrothermal sea-water solutions is described. It is based on the catalytic effect of cobalt(I1) on the oxidation of N,N'-diethyl-p-phenylenediamine by hydrogen peroxide in the presence of Tiron as an activator. The catalytic activity of cobalt(rr) was found to be significantly enhanced by the presence of sea-water matrix components, especially calcium ions, improving the sensitivity of cobalt determination in sea-water.The comparatively weak basic medium (pH 8.7-9.0) of the reaction and its relative freedom from co-existing ions allowed the direct analysis of sea-water to be conducted effectively without any preliminary steps. The sampling rate was 50 h-1. The limit of detection for cobalt was 1 ng 1-1 and the relative standard deviation at the level of 10-100 ng 1-I was 2 4 % ( n = 5). The technique was confirmed to be accurate on the basis of the analysis of the standard sea-water solutions CASS-2 and NASS-2. Artificial hydrothermal solution samples were analysed by the proposed technique and the results were in good agreement with data obtained by ICP-MS.Keywords: Cobalt determination; flow injection; catalytic reaction; sea-water; spectrophotometry Introduction Cobalt exists in sea-water mainly as the CO" ion and in its chloro and carbonate complexes. The oceanic cobalt concentration is extremely low; it has a maximum of about 10-20 ng 1-l in surface water and is depleted to less than 1 ng 1-1 at depths below 1000 m.132 Cobalt enrichment in hydrothermal fluids on the sea-floor up to 227 ng 1-1 has been found.3 Therefore, this metal is interesting from the geochemical point of view because hydrothermal activity can be recognized chemically by indicat- ing anomalous concentrations of cobalt in deep sea-water around hydrothermal fields." Obviously, it is important to find quick, sensitive and selective procedures for determining cobalt in deep sea-water. Unfortunately, shipboard methods preclude well known techniques such as ICP-MS and ETAAS because the instruments are bulky and sensitive to the constant vibrations, and in addition the determination with ETAAS requires at least 250 ml of a sample.There are only a few methods which are available for shipboard determinations of trace amounts of cobalt in sea-water.5-10 Sensitive techniques using adsorptive cathodic stripping voltammetry based on the accumulation of its complexes with dimethylglyoxime at a hanging mercury drop electrode have been reported.5 Recently, a rapid method for the determination of cobalt in sea-water with a detection limit of about 0.4 ng 1-1 was described that combines high scan-rate staircase voltammetry and the nitrite catalytic effect to enhance Co"-dimethylglyoxime reduction currents.6 The reported method was used for the determination of the oceanic profile of cobalt in the Central East Atlantic Ocean.However, the necessity for preliminary irradiation of sea-water samples with a UV lamp for 3 h, oxygen interfer- ences, and the use of the mercury drop electrode mode are obvious disadvantages of this approach, especially for monitor- ing applications. Significant progress has been made in applying continuous flow and flow injection (FI) catalytic methods for trace metal determinations in sea-water. This approach has proved to be extremely suitable for marine chemical research on-board a ship.11-14 Several such systems have been developed for the detection of cobalt in sea-water samples.7-'0 Most of them were based on the spectrophotometric monitoring of the known catalytic reactions of the oxidation of o-dihydroxybenzene derivatives, such as Tiron7 and protocatechuic acids in a basic medium.An FI system with chemiluminescence detection based on the catalytic effect of cobalt on the oxidation of gallic acid by hydrogen peroxide was also proposed by Sakamoto- Arnold and Johnson.9 It was possible to reach a limit of detection of 0.5 ng 1-l with a sampling rate of approximately 8 h-1. Unfortunately, the strongly alkaline medium of the indicator reactions (pH > 10) and, sometimes, a lack of selectivity and an insufficient sensitivity do not allow the direct determination of cobalt in sea-water and a preliminary separa- tion/preconcentration step was required.The use of chem- iluminescence detection has some additional disadvantages such as the relatively high price of the detector and the need for a photomultiplier cooler. l 5 The catalytic oxidation of N-phenyl-p-phenylenediamine by hydrogen peroxide activated by Tiron was used by Kawashima et ~ 1 . 1 0 for the FI determination of cobalt. It was shown to be applicable to the determination of more than 250 ng 1-I cobalt in coastal sea-water owing to interference by alkaline earth metal ions in the strongly alkaline media of the reaction. In order to provide the direct determination of dissolved cobalt in sea-water without any preliminary steps, a search for sensitive and selective catalytic reactions that proceed in the pH range close to that of sea-water is required.This paper describes a catalytic FI technique which was specifically developed for determining background trace levels and the anomalous hydrothermal plume concentration of cobalt in sea-water. A detailed study was carried out of the oxidation reaction of N,N'- diethyl-p-phenylenediamine (DePD) by hydrogen peroxide catalysed by cobalt(I1) and activated by Tiron (Tr) in sea-water matrix solutions. Experimental Reagents Cobalt(I1) stock standard solutions (1, 10 and 100 pg 1-1) were prepared by suitable dilution of a Sigma (St. Louis, MO, USA),1038 Analyst, August 1996, Vol. 121 1000 mg 1-1 atomic absorption standard solution with 0.1 mol 1-I HCI, stored in polyethylene bottles and kept in a refrigerator. Working standard solutions of cobalt(I1) were prepared daily by diluting the stock standard solutions with water, sodium chloride and cobalt-depleted sea-water. A mixed organic reagent solution (R1) was prepared daily by dissolving appropriate amounts of N,N'-diethyl-p-phenylenediamine sul- fate (Sigma) and Tiron (Sigma) in 200 ml of water.A working hydrogen peroxide solution was prepared daily from 30% reagent by dilution with a borate buffer. All the inorganic reagents used were of analytical-reagent grade. Doubly distilled water was used in all experiments. NASS-2 Sea-water (Open Ocean) and CASS-2 Sea-water (Coastal) certified reference materials were obtained from the National Research Council of Canada (Toronto, Canada).Sea- water Sampling Oceanic sea-water samples were collected in PTFE-coated Niskin bottles, filtered immediately, placed in polyethylene bottles, acidified to pH = 2 and then stored at a low temperature (4 "C) until required for analysis. Apparatus and FI Manifold Configuration An FIAstar 5010 flow injection analyser (Tecator, Hoganas, Sweden) was used. The absorbance was monitored by an FIAstar 5023 spectrophotometer equipped with an 18 pl flow- through cell (light path 10 mm) and a Model 5032 controller or IBM computer. All tubing was made from polytetrafluoro- ethylene (0.7 mm id). A Chemifold I1 (Tecator) was used in order to provide the proposed FI manifold configuration (Fig. 1). The reagent streams pumped through the FI manifold were R1 = DePD + Tr and R2 = NaOH or buffer solution + H202.The following carriers (C) were tried: water, 0.5 moll-' NaC1, sea-water matrix solution and cobalt-depleted sea-water acid- ified to pH 2. Results and Discussion FI Study of the Catalytic Reaction Chemical system The catalytic effect of cobalt on the oxidation reactions of a mixture of DePD and Tr (1 + 1) by hydrogen peroxide was studied by the injection of 20 pg 1-l cobalt(I1) standard solution into the FI system (Fig. 1). The absorption spectra and absorbance values were found to be strongly dependent on the matrix composition and pH in the reaction zone. Fig. 2(a) shows the absorption spectra for the baseline solution (without cobalt) and for the injected solution zone at pH 9. The spectra were n I I T°C Fig. 1 Schematic diagram of FI manifold used for the study of the catalytic reaction: C, carrier; R1 and R2, reagents; S, sample; v l , v2, v3, flow rates; PI and P2, pumps; I, injection valve; L I , mixing coil; L2, reaction coil; D, detector; and W, waste.obtained by scanning in the stopped-flow mode (Tstop = 30 s). In the case of baseline solution scanning during this time, no development of coloration was noticed in the wavelength range 400-700 nm, but the spectrophotometric reaction was ac- celerated in the presence of cobalt(I1) [Fig. 2(a), curve 21. It was also noticed that the presence of sea-water matrix compounds in the injected cobalt solution significantly enhanced the catalytic oxidation [Fig. 2(a), curve 31. By comparing the spectra obtained with the absorption spectra of the individual products of the oxidation of DePD and Tr under the same flow conditions [Fig.2(b)], the primary oxidation of DePD in the reaction mixture at pH 9 was identified. It was noticed that maximum absorbance values obtained at 5 18 and 554 nm for the oxidation of the mixture of DeDP and Tr in the presence of 1 pg 1-1 of cobalt(I1) were much higher than the values obtained with DePD alone in the presence of 20 pg 1-1 of cobalt(u) [see Fig. 2(a), curve 3, and Fig. 2(b), curve l)]. This effect could be attributed to the activating behaviour of Tr in the oxidation process of DePD by hydrogen peroxide. Thus, establishing that the cobalt- catalysed oxidation of DePD took place at pH = 9 and that it could be significantly accelerated by the sea-water matrix compounds gave us the opportunity to continue further investigations of the reaction using sea-water matrix solution as a carrier, without any risk of precipitation of alkaline earth metal hydroxides in the FI manifold.All of the following spec- trophotometric measurements were made at 554 nm. In order for the reaction to be applicable to the determination of cobalt in sea-water, the influences of chemical variables on the analytical signal for cobalt(I1) were studied. For this purpose, the FI system (Fig. 1) was used under the following fixed conditions: v1 = 1.2; v2 = 1.5; v3 = 0.6 ml min-1; length of L1 = 0.6 m; length of L2 = 2.0 m; injection volume V, = 200 I 1 3 n I I",'" /" 1 Wavelength f nm Fig. 2 Absorption spectra of solutions (pH 9) containing (a) DePD-Tr and H20z : I, without cobalt; 2, with injection of 20 pg 1-1 cobalt in water; and 3, with injection of 1 pg 1-1 cobalt in sea-water matrix solution; and (6) 1, Tr and H202 with injection of 20 pg 1-1 cobalt in sea-water matrix solution. and 2, DePD and H202 with injection of 20 pg 1- I cobalt in sea-water matrix solution.Analyst, August 1996, Vol.121 1039 pl; and temperature in reaction coil L2 = 30°C. With the chosen hydrodynamic parameters, the reaction time was 14 s. Effect of pH A detailed study of the effect of pH on the cobalt-catalysed oxidation of DePD in the presence of Tr was carried out for different concentrations of cobalt(rr) in the sea-water matrix solutions. The effect of pH on the catalysed reaction was examined over the range 8-10.The pH of the system was changed by varying the concentration of sodium hydroxide in R2. The maximum signal for more than a 0.5 pg 1-1 concentration of cobalt was obtained in the pH range 8.6-9.0, but the optimum pH range shifted slightly to 8.9-9.2 as the concentration of cobalt decreased to 0.1 pg 1-1 (Fig. 3). It was known7 that a small amount of ammonium ion and large amounts of carbonate and borate ions were not acceptable for use as buffers in cobalt-catalysed reactions. However, we demonstrated that a concentration of borate of up to 0.02 mol 1-1 in solution R2 did not interfere significantly with the catalytic action of cobalt(II), giving a recovery of at least 90%, and could be used to maintain pH stability in the reaction coil. Effect of reagent concentrations The dependence of the catalytic activity with varying DePD and Tr concentration in the R I stream from 0 to 0.02 mol 1-1 was examined for 1 pg 1-1 cobalt solution. It was found that the peak height increased rapidly with increasing concentration of DePD up to 0.005 mol 1-1 at the fixed concentration of Tr (0.005 and 0.01 mol 1-1).Higher concentrations of DePD resulted in only a slight increase in absorbance but at the same time showed a poorer reproducibility because of increased baseline ab- sorbance. A concentration of 0.005 moll-' DePD was selected as optimum. With a fixed concentration of DePD (0.005 and 0.01 moll- l), the peak height was observed to increase with an increase in concentration of Tr up to 0.005 rnol 1-1 and remained almost constant over the range 0.005-0.02 mol 1-1.Hence a 0.005 moll-' Tr concentration was adopted for the procedure. When the hydrogen peroxide concentration in the R2 stream was varied over the range 0.0-1 .O%, the optimum concentration of this reagent was found to be 0.1-0.3%. At higher concentra- tions of hydrogen peroxide, especially more than 1%, poor reproducibility of the signals was observed, because of the decomposition of peroxide to oxygen gas. O.' T 0 4 I 8 8.2 8.4 8.6 8.8 9 9.2 9.4 9.6 9.8 PH Fig. 3 Effect of pH of the reaction mixture on the FI signal for sea-water matrix solutions with different concentrations of cobalt: l , O . l ; 2,0.2; 3,0.5; 4, 1.0, pg 1-1. Effects of temperature, reaction time and FI variables The effect of temperature on the catalytic signal was examined in the range 20-60°C.It was found that the peak height increased with increase in temperature from 20 to 40 "C, but the baseline absorbance also increased, as expected. A temperature of 30°C was chosen to optimize high sensitivity and baseline stability. The development of coloration in the oxidation of DePD by hydrogen peroxide in the presence of Tr with time was studied by using the stopped-flow mode. The kinetic curves for a cobalt- catalysed reaction and an uncatalysed reaction were compared (Fig. 4). The strongest catalytic effect of cobalt was observed during the first 15-30 s of the reaction, where curve 1 had a maximum typical for an intermediate reaction; after that time, the uncatalysed reaction proceeded very quickly and after a reaction time of more than 100 s the difference between the two reactions was either unmeasurable or insignificant.The de- scribed kinetic curves again demonstrate the advantages of the FI technique over batch techniques, as it is obvious that analytical measurements of cobalt concentrations by the catalytic indicator reaction described in this paper would be impossible, at least with the same sensitivity and reproducibil- ity, without the FI technique. The FI signal increased on increasing the length of reaction coil L2 from 0.3 m to 3.4 m. In our experiments, the most significant rise in the signal level was observed when L2 was increased to 2.0-2.2 m, after which a peak change was also noticeable, but the baseline increased significantly. An L2 coil length of 2.2 m, corresponding to a 15 s reaction time, was chosen.The relationship between cobalt peak and injection volume was also studied. On increasing the volume of the 1 pg 1-I cobalt solution from 40 to 200 p1, the absorbance of the reaction changed from 0.225 to 0.500. Further increases in the injection volume up to 400 pl made the peak only 0.080 absorbance units higher. An injection volume of 200 pl was chosen as optimum. Different variations of the velocities of R1 and R2 reagent streams and their sequence of pumping into the FI system were also examined. The best results were achieved with v1 set at 1.5, v2 at 0.8 and v3 at 0.6 ml min-I. All the chosen optimum variables were used in subsequent experiments. Effect of macrocomponents of sea-water As was noticed before (Fig.2), the catalytic activity of cobalt(I1) increased on addition of sea-water matrix components in the reaction zone. In order to clarify this observation, the effect of the individual macrocomponents of sea-water was investigated. Standard solutions with different concentrations of sodium chloride and a fixed concentration of cobalt (4 pg 1-l) were 0.7 0.6 0.5 0.4 5: 0.3 0.2 0 +? 0 Y 100 200 300 400 500 600 700 0 Time/s Fig. 4 Kinetic curves for the oxidation of DePD by hydrogen peroxide in the presence of Tr: 1, uncatalysed reaction; 2, cobalt-catalysed reaction obtained with injection of 1 yg I-' of cobalt.1040 Andyst, August 1996, Vol. 121 0 1 . - prepared and injected into the carrier stream with the corre- sponding salt concentration. All measurements were made at pH 9.0 f 0.1.Fig. 5(u) shows that increasing the ionic strength up to 0.5 increased the cobalt peak height four-fold. A variation in ionic strength from 0.4 to 0.6 made the peak height deviate by +15%. A negligible salt effect on the investigated oxidation reaction in the absence of cobalt was observed on injecting the pure sodium chloride solutions into water [Fig. 5(a), curve 11. Another macrocomponent of sea-water is magnesium, the average concentration of which in the ocean is 0.053 moll-1.16 The effect of this component on the catalytic reaction was studied by injecting solutions containing a fixed amount of cobalt (4 pg l-I), 0.5 rnol 1-1 NaCl and different concentrations of MgClz into the same salt mixture without cobalt as a carrier [Fig.5(b)]. The pH in the reaction mixture was maintained at 8.9 k 0.1. All 0.5 rnol I-' NaCl + MgC12 solutions without cobalt were also measured relative to 0.5 in01 1-1 NaCl solution as a carrier. As shown in Fig. 5(h), an increase in magnesium concentration from 0 to 0.05 mol 1- I almost doubled the peak height,which was more than would be expected from the effect of ionic strength according to Fig. 5(a). Previously,IS a similar enhancement of manganese catalytic activity was noticed on addition of magnesium to a reaction mixture of leucomalachite green and potassium periodate. The next important macrocomponent of sea-water is calcium, with an average concentration of 0.01 mol I-l.16 The effect of calcium on the catalytic reaction was studied by injecting solutions with a fixed concentration of cobalt (0.5 pg 1-l) containing 0.5 mol 1-1 NaCl and different concentrations of 1 , 2 A 0 I 0 0 2 0 4 06 0 8 NaCl concentration/mol I-' 0 4 -r 2 :: I 0 C' 0 0 111 0 02 D 0 ; Ca" concentration/mol I-' Fig.5 Effect ofrnacrocomponents in sea-water on the FI signal. ( a ) NaCl: I , without cobalt; and 2, with 4 pg 1- of cobalt. ( h ) Mg": 1, without cobalt; 2, with 4 pg I-' of cobalt. ( c ) Ca": 1, without cobalt; 2, with 0.5 pg I-' of cobalt. CaC12 into the corresponding salt mixture without cobalt as a carrier [Fig. 5(c)]. At the same time, each matrix solution without cobalt was also measured relative to 0.5 rnol 1-1 NaCl solution. It was found that calcium ions have a dramatic effect on the rate of the catalytic reaction because the addition of small concentrations of calcium increases the cobalt peak by several- fold.A 15-fold increase in the peak height for 0.5 pg 1-1 cobalt was achieved by the addition of 0.01 mol 1-1 CaC12 solution to the matrix solution. A very small increasing effect of pure calcium chloride solutions (without cobalt) on the rate of the oxidation reaction was also observed and measured [Fig. 5(c) curve 11. Similar effects of calcium ions on the catalytic activity of cobalt in the oxidation of Tiron have been observed previously.7 Considering the conservative character of the macrocomponents in the ocean, it can be emphasized that there is no risk of interferences from the sea-water matrix on the results of the analysis of real sea-water samples. Interferences of sea-water microcomponents The effect of various &transition metal ions on the determina- tion of cobalt(r1) at the 100 ng 1-1 level was examined, with respect to possible metals enrichment in hydrothermal plumes.As can be seen from Table 1, Cr, Cu, Al, Pb and Zn at levels up to 500 pg 1- I did not interfere with the determination of cobalt. Manganese, nickel and iron showed interferences with max- imum tolerable concentrations of 3, 10 and 150 pg 1-I, respectively, with tolerance limits for diverse ions estimated as a 5% relative error. Calibration Graph A calibration graph for the analysis of acidified sea-water samples was obtained under the chosen hydrodynamic condi- tions and the following optimum concentrations of the reagents: R I = 0.01 mol 1-1 DePD + 0.01 moll-l Tr; R2 = 0.02 mol I-' Na2B4O7.l0H20 +0.04 mol 1-l NaOH + 0.2% H202.In this case, acidified sea-water depleted of cobalt was used as a carrier and as a matrix for preparing cobalt standard solutions. A typical calibration graph of cobalt obtained under the chosen otpimum conditions was linear over the range 0-300 ng 1-1 of cobalt(1r). The regression equation of the graph was AA = 6.026 + 3.370C, where A is absorbance X 10-3 and C is the concentration of cobalt (ng I-l), and the correlation coefficient Table 1 Tolerance to diverse ions in the determination of 100 ng 1-' of cobalt Ion lMeJ/lCol ratio Recovery (70) 1 :30 I :so 1 : 150 1 : 100 1 : 200 1 :400 1 : so0 1 : 2000 1 : 1000 1 : 2000 1 :so00 1 : 2000 I : 5 000 1 : 10000 1 : 2000 1 :5000 1 : 10000 1 : 20 000 1 : 20 000 I : 20000 104 108 118 10s 109 I16 120 130 103 106 123 102 108 119 105 114 12s 1 04 100 98Analyst, August 1996, Vol.I21 1041 was 0.99939. The detection limit was 1 ng I - l , calculated as three times the blank signal. The sampling rate was 50 h-1. Reproducibility, Accuracy and Analytical Applications The data for the recovery experiments, obtained from calibra- tion graphs for cobalt standards prepared with acidified sea- water depleted with cobalt, are given in Table 2. The relative standard deviation was 2 4 % for the concentration range 10-100 ng 1-1 of cobalt. The accuracy of the method was confirmed by analysing Coastal Atlantic Standard Sea-water-2 (CASS-2) and North Atlantic Sea-water-2 (NASS-2) reference materials.The results obtained were in good agreement with the certified cobalt concentrations (Table 3). The proposed F1 method was also tested by analysing artificial hydrothermal solutions obtained during geochemical experiments with basalt-sea-water systems. The concentrations obtained by our method were in good agreement with the values obtained by ICP-MS using sample splits (Table 3). The applicability of the proposed method was also demonstrated by the analysis of real sea-water samples. More than 100 samples of deep sea-water collected not far from the island of Hawaii were analysed, in order to detect cobalt anomalies in the area of Pele hydrothermal vents. It was found that almost all of the Table 2 Recovery of cobalt added to sea-water matrix solutions 0, = 0.95; n = 5 ) Co" added/ Co" found/ RSD Recovery ng I-' ng I-' ("/.I (%) 4 4.8 + 0.5 10.6 113 10 10.1 f 1.1 8.5 101 30 31.3 It 1.7 4.4 104 100 103.0 k 2.0 1.6 103 Table 3 Results for the determination of cobalt in sea-water samples (p = 0.95) Cobalt(ri) concentration/ng 1- 1 Sample n NASS-2 4 CASS-2 14 Sea-water (Pacific Ocean) 4 Artificial 5 Solutions 5 5 Hydrothermal 9 FI method 3.1 + 0.5 27.1 f 0.4 10.5 + I .6 160 k 7 200 f 5 35 f 3 175 If: 5 RSD Certified (YO) ICP-MS value 10.4 - 4 + 1 2.8 - 25 & 6 9.4 - 3.4 151 + 5 - 2.7 2 1 0 f 6 - 5.8 33 * 1 - 2.5 165 + 5 - - collected samples had cobalt concentrations less than 4 ng 1-1 and only two contained about 10 ng 1-1 of cobalt.Conclusion An FI spectrophotometric method based on the catalytic activity of Co" on the oxidation of DePD by hydrogen peroxide in the presence of Tr was developed for rapidly tracing anomalous concentrations of dissolved cobalt in sea-water.The method is relatively free from interferences from co-existing micro- components in sea-water within the open ocean sea-water salinity range and provides results, without any preliminary separation, that have satisfactory reproducibility and accuracy. It was successfully applied to the determination of cobalt concentrations in both sea-water samples and hydrothermal sea- water solutions. The authors express their gratitude to J. A. Resing (USA) for providing the reference materials. References 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Bruland, K. W., in Chemical Oceunography, ed. Riley. J. P., and Chester, R., Academic Press, London, 1983, vol. 8, ch. 45. Knauer, G. A., Martin, J. H., and Gordon, R. M., Nature (London), 1982, 297, 49. Von Damm, K. L., Edmont, J. M., Grant B., Measures, C. I., Walden, B., and Weiss, R. F., Geochim. Cosmochim. Acta, 1985, 49, 2197. Sakai, H., Tsubota, H., Nakai, T., Ishibashi, J., Akagi, T., Gamo, T., Tilbrook, B., Igarashi, G., Koreda, M., Shitoashiima, K., Nakamura, S., Fujiioka, K., Watanabe, M., McMurtry, G., Malahoff, A., and Ozima, M., Geochem. J . , 1987, 21, 1134. Zhang, H., Wollast, R., Vire, J.-C., and Patriarche, C;. J., Analyst, 1989, 114, 1597. Herrera-Melian, J. A., Hemandez-Brito, J., Gelado-Cabballero, M. D., and Perez-Pena, J., Anul. Chim. Actu, 1994, 299, 59. Isshiki, K., and Nakayama, E., Tulanta, 1987. 34, 277. Yamane, T., and Watanabe, K., Anul. Chim. Acta, 1988, 207, 331. Sakamoto-Arnold, C.M., and Johnson, S., Anul. Chenz., 1987, 59, 1789. Kawashima, T., Minami, T., Ata. M., Kamada, M., and Nakano, S., Flow Inject. Anal., 1985, 2, 40. Shpigun, L. K., Kolotyrkina, I. Ya.. and Zolotov, Yu. A., Anal. Chim. Acta, 1992, 261, 307. Kolotyrkina, I. Ya., Shpigun, L. K., Zolotov, Yu. A., and Tsysin, G. I., Analyst, 1991, 116, 707. Johnson, K. S . , Coale, K. H., and Jannasch, H. W., Anal. Chem., 1992, 64, 1065A. Kolotyrkina, I. Ya., Shpigun, L. K., Zololov, Yu. A., and Malahoff, A., Anulyst, 199.5, 120, 201. Resing, J. A., and Mottl, M. J . , Anal. Chem., 1992, 64, 2682. Popov. N. I., Fedorov, K. N., and Orlov, V. M., in Seu-Water, ed. Monin, A. S., Nauka, Moscow, 1979, pp. 27-29. Paper 6102593K Accepted April 15, 1996

ISSN:0003-2654    

DOI:10.1039/AN9962101037

出版商:RSC    

年代:1996

数据来源: RSC

摘要:

Analyst, August 1996, Vol. 121 ( I 043-1 046) 1043 Flow Injection-Fluorimetric Method for the Determination of Ranitidine in Pharmaceutical Preparations Using 0- P h t halalde h yde Carmen Lopez-Erroz, Pilar Viniis, Natalia Campillo and Manuel Hernandez-Cordoba* Department of Analytical Chemistry, Faculty of Chemistry, University of Murcia, E-30071 Murcia, Spain A flow injection-fluorimetric procedure for the determination of ranitidine is proposed. The assay is based on the reaction of the drug with sodium hypochlorite, to produce a primary amine, which reacts with o-phthalaldehyde and 2-mercaptoethanol to form highly fluorescent derivatives. The calibration graph, based on peak area, is linear in the range 20-500 ng ml-l of ranitidine with an accuracy of 3.4%. The corresponding detection limit is 13 ng ml-l (3.7 pmol).The method was applied to the determination of ranitidine in pharmaceuticals. The recovery was quantitative and no interferences from excipients were observed. Keywords: Flow injection; fluorimetry; ranitidine; pharmaceuticals Introduction Ranitidine is a histamine H2-blocker commonly used in clinical practice for the treatment of duodenal ulcers. It provides a new and effective therapeutic approach to gastric hypersecretory diseases.' The drug is a hydrophilic molecule containing a substituted furano ring. The determination of ranitidine is usually carried out by means of liquid chromatography2-2g with spectrophotometric detection. However, to our knowledge, fluorimetric detection has not been applied in this system. Flow injection (FI) fluorimetric methods could be useful for the determination of ranitidine because they can be easily auto- mated and are very sensitive. The fluorescent detection of primary amines and amino acids is commonly carried out by means of o-phthalaldehyde (OPA) in combination with a thiol compound, usually 2-mercaptoethanol (ME).29-31 For the determination of secondary amines, a prior oxidation step using hypochlorite to oxidize the secondary amine to a primary amine"2.33 is currently included in the method.However, the sensitivity achieved for secondary amines is usually lower than that for primary amines. In this work, a highly sensitive procedure for the determina- tion of ranitidine is proposed after its oxidation with hypo- chlorite. The primary amine obtained is then allowed to react with OPA and ME and the fluorescence is measured. The system was automated using FI and provided greater selectivity and sensitivity than the usual UV detection method.The method permits the determination of ranitidine even at low picomolar levels and was applied to the determination of the drug in pharmaceutical preparations. * To whom correspondence should be addressed. Experimental Apparatus A Kontron (Zurich, Switzerland) SFM 25 fluorescence detector set at wavelengths of 350/450 nm (excitation/emission) and a PC integration pack (Kontron) to record and integrate the peaks were used. The FI system consisted of a Gilson (Worthington, OH, USA) Minipuls HP4 peristaltic pump, an Omnifit (Cam- bridge, UK) injection valve, a Hellma (Jamaica, NY, USA) 176.052-QS fluorimetric flow cell, 0.8 mm id PTFE tubing and various end-fittings and connectors (Omnifit).Thermostating of the reactor coil was carried out using a laboratory-made electronic device. Reagents High-quality water obtained using a Milli-Q system (Millipore; Milford, MA, USA) was used throughout. A 0.02 moll-* OPA (Fluka, Buchs, Switzerland) solution was prepared in 2% ethanol and 0.5 moll-' borate buffer of pH 10.5. A 0.1 moll-' ME (Sigma, St. Louis, MO, USA) solution was prepared in 0.5 mol 1-1 borate buffer of pH 10.5. Solutions were kept in dark bottles at 4 "C. The 5 mmol 1- sodium hypochlorite was diluted from the commercial product (Probus; Barcelona, Spain; 12%) in 0.05 moll-' sodium acetate-acetic acid buffer of pH 4.5.A 500 pg ml-1 ranitidine hydrochloride (Sigma) stock solution was prepared in water and also kept at 4 "C; working standard solutions were prepared by dilution with water immediately before use. Analytical Procedures The flow manifold is shown in Fig. 1. Ranitidine (200 pl) was injected into a water channel and mixed with the sodium hypochlorite; this stream was then passed through a thermo- stated reaction coil (0.8 m X 0.5 mm id) at 25 "C. The oxidized r - - - - - - - - - 7 carrier c10- OPA W L _ _ _ _ - _ - _ _ J T U Fig. 1 FI manifold for the determination of ranitidine. P, Peristaltic pump (total flow rate 2 ml min-1); V, injection valve (sample loop, 200 pi); R, reactor coil (0.8 m X 0.8 mm id); R2, reactor coil (2.5 m X 0.8 mm id); T, thermostat at 25 "C; D, fluorimeter 350/450 nm (excitation/emission); PC, personal computer; and W, waste.I044 Analyst, August 1996, Vol.I21 ~ ~ _ _ _ _ _ effluent was mixed with the fluorogenic reagent in a T-piece. This reagent was obtained by merging the 0.02 moll-1 solution of OPA in 0.5 moll-l borate buffer of pH 10.5 with a stream of 0.1 mol 1-1 ME in 0.5 mol I--] borate buffer. The resulting solution flowed through a second reactor coil (2.5 m X 0.5 mm id) thermostated at 25 "C and then passed into the flow cell for the fluorescence to be recorded. All carrier streams were pumped at the same flow rate by means of a peristaltic pump with a total flow rate of 2.0 ml min-I. Calibration graphs were obtained by plotting peak area against the ranitidine concen- tration.For the determination of the drug in pharmaceutical prepara- tions, the total powdered tablet was dissolved with 250 ml of water or the injectable solution was directly diluted with water as necessary. An aliquot was filtered through a 0.2 pm nylon Millipore chromatographic filter and analysed by the FI method. Results and Discussion The molecule of ranitidine contains no primary amino groups and, therefore, this drug does not react directly with OPA to form fluorescent products. However, the molecule has two secondary amino groups. In order to detect ranitidine by fluorescence, the secondary amine was converted into a primary amine by treatment with sodium hypochlorite. This subse- quently reacted with OPA and ME to give a fluorescent species. The mechanism of the reaction has not been elucidated; however, it appears reasonable to assume that the mechanism of the conversion of ranitidine into a primary amine and the subsequent reaction are similar to those for other secondary amines which have been de~cribed.3333~ As this is a time- dependent process, to improve reproducibility and handling the reaction was carried out using FI methodology.Optimization of the Oxidizing Reagent The experimental parameters were optimized to obtain max- imum fluorescence, i.e., maximum sensitivity in the chemical analytical procedure. The influence of sodium hypochlorite concentration on the reaction which converts ranitidine into a primary amine was studied first. Concentrations ranging from 1.6 x 10-4 to 4 X 1 0-2 moll- 1 diluted in 0.05 moll- sodium acetate-acetic acid buffer of pH 4.5 were tried (Fig.2A). The signal increased up to 1 X 10-2 moll- I and then decreased for higher concentrations. This was probably caused by destruction of the aniine formed from ranitidine by excess of the reagent. A 5 x 10-3 rnol 1-1 sodium hypochlorite concentration, which gave a constant signal, was chosen. The decrease in fluores- 60 I 1 40 I I I I I 0 1 2 3 4 5 2 4 6 8 1 0 1 2 [hypochlorite]/l 0-2 rnol I-' PH Fig. 2 lnfluence of the oxidizing agent on the fluorescence. A, Variation of the sodium hypochlorite concentration (5 X moll-' OPA and 0.1 mol 1-1 ME, both in 0.5 in01 1-I phosphate buffer of pH 10.5; ranitidine injected, 1 pg ml-I); B, effect of the pH of the buffer (5 X mol 1-1 sodium hypochlorite, 5 X lo-? mol I-' OPA and 0.1 mol 1-1 ME, both in 0.5 rnol 1-1 phosphate buffer of pH 10.5: ranitidine injected, 1 pg ml-1).cence caused by the excess of hypochlorite has been described in previous studies,33 and this undesirable effect has been overcome by adding 2,2'-thiodiethanol (TDE).3*,33 It was therefore decided to investigate the effect of adding TDE after the ranitidine and hypochlorite had reacted. The results indicated that the sensitivity was similar both in the presence and absence of TDE (using 5 X 10-3 moll-' hypochlorite), and hence the addition of TDE was unnecessary. The pH of the conversion reaction markedly affected the fluorescence. Solutions of sodium hypochlorite were prepared with different pH buffers. Fig. 2B shows that the maximum signal was achieved at pH 4-5.Thus, a pH of 4.5 was fixed and hypochlorite was diluted using a sodium acetate-acetic acid buffer. The concentration of the buffer was varied from 0.02 to 0.4 mol 1-l and no significant variations were obtained; hence, a value of 0.05 mol 1-' was selected. Optimization of the Fluorogenic Reagents The influence of the OPA concentration was studied between 1.2 X and 0.04 mol 1-l, while maintaining a 0.1 rnol 1-1 ME concentration. Both OPA and ME solutions were prepared in 0.5 mol I-' phosphate buffer of pH 10.5. The fluorescence rapidly increased up to an OPA concentration of 0.02 mol 1-1 (Fig. 3A) and this was chosen as the optimum. Fig. 3B shows the variation of the signal with the ME concentration in the range from 5 X 10-3 to 0.4 mol 1-l, when the OPA concentration was fixed at 0.02 mol 1-1.Concentrations above 0.1 moll- led to a decrease in the fluorescence and hence this concentration was selected. The influence of the pH of the reagents was then studied using borate buffers in the pH range 8-11 in both the OPA and ME streams. Fig. 3C demonstrates that the fluorescence continuously increased up to pH 10. Consequently, 0.5 mol 1-' borate buffer of pH 10.5 was added to the reagents. An increase in temperature had a favourable effect on the signal up to 30 "C; however, the fluorescence decreased when higher temperatures were tried (Fig. 3D). Therefore, the reactor coils were thermostated at 25 k 0.5 "C. B 30 - $ 0 1 2 3 4 0 1 2 3 4 5 [o-phthalaldehyde]/l OV2 rnol I-' [2-mercaptoethanol]/I 0-' rnol I-' 30 t 120 90 so I 30t4 0- 0- 8 9 10 11 0 20 40 60 PH Temperature/"C Fig.3 Influence of the fluorogenic reagents. A, Variation of the OPA concentration; B, variation of the ME concentration; C, variation of the pH; D, influence of the temperature. Experimental conditions: 5 x 10-3 moll-' sodium hypochlorite in 0.05 mol 1-' acetate-acetic acid buffer of pH 4.5; ranitidine injected, 1 pg ml-1.1045 Analyst, August 1996, Vol. 121 Influence of the FI Variables The influence of the length of the first reactor coil in which the oxidation reaction occurred was studied in the range 0.03-2.5 m. Fig. 4A shows that the fluorescence increased up to 1 m; longer lengths of the reactor generated a decrease in the fluorescence, probably caused by sample dispersion.An optimum value of 0.8 m was selected. The second reactor coil was introduced for mixing both the OPA and ME fluorogenic reagents with the oxidized amine. As Fig. 4A demonstrates, variation of the reactor length between 0.5 and 5 m produced an increase in the signal up to 2.5 m, this being the length selected. Longer lengths caused a slight decrease in the fluorescence owing to band spreading resulting from an increase in the solute residence time in the reactor coil, which inevitably led to dispersion. The effect of the sample loop size was examined between 60 and 660 pl, and linearity was observed up to approximately 200 pi, this being the loop size chosen (Fig. 4B). Variation of the total flow rate between 0.6 and 2.7 ml min-I produced a decrease in the fluorescence as the flow rate increased because it considerably affected the time for reaction. A low flow rate increased the fluorescence yield because the reaction time 120 a $ 90 v) 0 2 2 60 120 90 60 30 0 0 2 4 6 0 200 400 600 Reactor length/m Sample loop/pI Fig.4 Effect of FI variables. A, R I and R2 reactor coil lengths; B, sample loop size. Experimental conditions: 5 X mol I-' sodium hypochlorite in 0.05 moll-' acetate-acetic acid buffer of pH 4.5,s X 10-3 moll - OPA and 0.1 mol I-' ME, both in 0.5 mol 1-1 phosphate buffer of pH 10.5; ranitidine injected, 1 pg ml-1. 5 rnin 1 - 0.1 I 0.075 a, a, Time Fig. 5 correspond to pg m - ' of the drug. FI peaks for calibration of ranitidine. Numbers on the peaks Table 1 Determination of ranitidine in pharmaceutical preparations Ranitidine Product (laboratory) Reported Found" Ranuber (ICN-Hubber) Zantac (Glaxo) 150 mg per tablet 50 mg per 5 ml 15 1.7 k 3.0 mg per tablet 50.4 k 2.6 mg per 5 ml Mean f standard deviation ( n = 6).between the analyte and the fluorogenic reagents was longer. Conversely, at high flow rates, the residence time of the sample in the mixing coil was too short and the extent of the reaction was too low, which severely reduced the detector response. A 2 ml min-l value (0.5 ml min-1 for each channel) was selected as a compromise between a good signal and an adequate sampling frequency. Calibration, Interferences and Applications Calibration of ranitidine was performed by plotting concentra- tion against peak area and was linear between 20 and 500 ng ml- I .Fig. 5 shows the FI peaks obtained. The detection limit calculated on the basis of 30 was 13 ng ml-I. The precision and accuracy of the method were demonstrated by repetitive analyses and the average relative standard deviation calculated for ten replicate determinations of 0.15 pg ml-1 of ranitidine was 3.4%. Interferences caused by common tablet excipients were studied by injecting solutions containing ranitidine (0.3 pg ml-I) and different amounts of the other compounds. No interferences were found for lactose, glucose, saccharose, saccharin, maltose, ascorbic acid, sorbitol or starch at [inter- ferent] : [ranitidine] ratios up to 100 : 1. Higher concentrations were not tested. The tolerance limit was taken as the concentration causing an error of not more than +3% in ranitidine recovery. The reactivi ties and possible interference of other therapeutic agents usually employed in the treatment of peptic ulcers, such as cimetidine and famotidine, were also investigated.It was found that these drugs also yielded fluorescent products; however, differences in the reaction rate and the fluorescent quantum yield due to these species resulted in different detection limits for each drug and interference appeared only when cimetidine or fdmotidine was present in a 50-fold excess. The clinical use of ranitidine is based on the inhibition of gastric acid secretion. 1 For oral use, ranitidine is administered as tablets containing 150 mg of the drug. There is also an injectable dosage form containing 25 mg ml-I of ranitidine.The reliability of the proposed method was tested by analysing two pharmaceutical preparations in these different physical forms. The results obtained are shown in Table 1. The recovery of ranitidine was quantitative because all values were in good statistical agreement with the values supplied by the manu- facturers, and there was no interference from the excipients. Conclusion The proposed FI-fluorimetric method is a significant improve- ment over other previously reported methods in terms of sensitivity, selectivity and simplified procedure. The FI system provides a less expensive and more versatile system with considerably reduced analysis times compared with manual methods, making it useful for the routine determination of ranitidine in pharmaceuticals.The sensitivity is higher than that obtained using spectrophotometric procedures, and the method permits the determination of ranitidine even at low picomolar levels. Linearity, precision and recovery are also satisfactory. The authors are grateful to the Spanish DGICYT (Project PB93- 1138) for financial support. N. C. holds a fellowship from Consejeria de Cultura, Comunidad Autonoma de la Region de Murcia. References 1 The Phuimuc.ologic~crl Basis of Tlwapeutic,s, eds. Goodman, 6 . A., Goodman, L. S., Rall, T. W., and Murad, F., MacMillan, New York, 7th edn., 1985.1046 Analyst, August 1996, Vol. 121 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 Carey, P. F., and Martin, L. E., J . Liq. Chromatogr., 1979, 2, 1291. Mihaly, G. W., Drummer, 0. H., Marshall, A., Smallwood, R.A., and Louis, W. J., J . Pharm. Sci., 1980, 69, 1155. Vandenberghe, H. M., MacLeod, S. M., Mahon, W. A., Lebert, P. A., and Soldin, S. J., Ther. Drug Monit., 1980, 2, 379. Carey, P. F., Martin, L. E., and Owen, P. E., J . Chromatogr. Biomed. Appl., 1981, 14, 161. Boutagy, J., More, D. G., Munro, I. A., and Shenfield, G. M., J . Liq. Chromatogr., 1984, 7, 165 1. Ficarra, P., Ficarra, R., and Tommasini, A., J . Pharm. Biomed. Anal., 1984, 2, 119. Carey, P. F., Martin, L. E., and Evans, M. B., Chromatographia, 1984, 19, 200. Sticht, G., Kaeferstein, H., Oehmichen, M., and Staak, M., Beitr. Gerichtl. Med., 1986, 44, 263. Guiso, G., Fracasso, C., Caccia, S., and Abbiati, A., J . Chromatogr. Biomed. Appl., 1987, 57, 363. Mullersman, G., and Derendorf, H., J .Chromatogr. Biomed. Appl., 1986, 54, 385. Rustum, A. M., Rahman, A., and Hoffman, N. E., J . Chromatogr. Biomed. Appl., 1987, 65, 418. Kames, H. T., Opong-Mensah, K., Farthing, D., and Beightol, L. A., J . Chromatogr. Biomed. Appl., 1987, 66, 165. Rustum, A. M., J . Liq. Chromatogr., 1988, 11, 2315. Evans, M. B., Haywood, P. A., Johnson, D., Martin-Smith, M., Munro, G., and Wahlich, J. C., J . Pharm. Biomed. Anal., 1989, 7, 1. Beaulieu, N., Lacroix, P. M., Sears, R. W., and Lovering, E. G., J. Pharm. Sci., 1988, 77, 889. Das Gupta, V., Drug Dev. Ind. Pharm., 1988, 14, 1647. Salem, M. S., Gharaibeh, A. M., Alkaysi, H. N., and Badwan, A., J . Clin. Pharm. Ther., 1988, 13, 351. Kaka, J. S., J . Liq. Chromatogr., 1988, 11, 3447. 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 Prueksaritanont, T., Sittichai, N., Prueksaritanont, S., and Vongsaroj, R., J . Chromatogr. Biomed. Appl., 1989, 82, 175. Rahman, A., Hoffman, N. E., and Rustum, A. M., J . Pharm. Biomed. Anal., 1989, 7, 747. Arafat, T., Al-Saket, M., Awad, R., Saleh, M., Gharaibeh, M., and Sallam, S., Alexandria J. Pharm. Sci., 1990, 4, 11. Segelman, A. B., Adusumalli, V. E., and Segelman, F. H., J . Chromatogr., 1990, 535, 287. Lloyd, T. L., Perschy, T. B., Gooding, A. E., and Tomlinson, J. J., Biomed. Chromatogr., 1992, 6, 3 11. Lau-Cam, C. A., Rahman, M., and Roos, R. W., J . Liy. Chromatogr., 1994,17, 1089. Smith, M. S., Oxford, J., and Evans, M. B., J . Chromatogr., 1994, 683, 402. A1 Khamis, K. I., El Sayed, Y. M., A1 Rashood, K. A,, and Bawazir, S. A., J . Liq. Chromatogr., 1995, 18, 277. Hoyer, G. L.. LeDoux, J., and Nolan, P. E., Jr., J . Liq. Chromatogr., 1995,18, 1239. Ichinose, N., Schwedt, G., Schnepel, F. M., and Adachi, K., Fluorometric Analysis in Biomedical Chemistry, Wiley, New York, 1991. Roth, M., Anal. Chem., 1971, 43, 880. Allenmark, S., Bergstrom, S., and Enerback, L., Anal. Biochem., 1985, 144, 98. Himuro, A., Nakamura, H., and Tamurd, Z., J. Chromatogr., 1983, 264, 423. Himuro, A., Nakamura, H., and Tamura, Z., Anal. Chim. Acta., 1983, 147, 317. Myers, H. N., and Rindler, J. V., J . Chromutogr., 1979, 176, 103. Paper 6102308H Received April 2, 1996 Accepted May 28,1996

ISSN:0003-2654    

DOI:10.1039/AN9962101043

出版商:RSC    

年代:1996

数据来源: RSC

摘要:

Analyst, August 1996, Vol. 121 (1047-1050) 1047 Spectrophotometic Flow Injection Determination of Lead in Port Wine Using In-line Ion-exchange Concentration Teresa I. M. S. Lopesa, Antonio 0. S. S. Rangela," Raquel P. Sartinib and Elias A. G. Zagattob a Escola Superior de Biotecnologia, Universidade Catblica Portuguesa, RLLU Dr. Ant6nio Bernardino de Almeida, 4200 Porto, Portugal h Centro de Energia Nuclear nu Agricultura, Universidade de Sfio Paulo, P. 0. Box 96, 13400-970, Piracicaba SP, Brazil A flow injection system with in-line ion exchange is proposed for the spectrophotometric determination of lead in Port wine. A Chelex-100 resin mini-column (200-400 mesh) is used for lead concentration, and the chromogenic reaction is based on the formation of a ternary complex between lead, Malachite Green and iodide.Forty-five digested wine samples can be run per hour (25-500 pg 1-1 Pb) and results are comparable to those obtained by ETAAS. Other features are the measurement precision (RSDs lower than 2.6 %), a detection limit of 12 pg 1-1 and a sampling rate of 45 h-1. Keywords: Flow injection analysis; lead determination; Port wine; spectrophotometry; ion exclzange Introduction There has been growing interest in the control of the heavy metal content in wines owing to the public health aspects involved. Among the heavy metals usually present in wines, lead is the most important toxicologically. Winery equipment, lead-containing pesticides and environmental pollution, such as car exhausts, are considered to be the major sources of contamination. These concerns led to the establishment of maximum permissible lead content levels; in Europe, the Office International de la Vigne et du Vin (ON)* set a maximum limit of 0.3 mg 1- and the USA and Canada set maximum values of 0.1 and 0.2 mg 1- l, re~pectively.~ The 01V reference procedure2 for the determination of lead in wines is ETAAS.Some modifications of this protocol have been suggested to meet the specificity of some wines such as Port.3 Other procedures involving flame AAS or spectro- photometric methods with previous sample digestion have also been reported.4 The main purpose of this work was to develop an alternative procedure for lead determination in Port wine, using a flow injection system with spectrophotometric detection. The method was based on a previously described reaction for the determination of cadmium,' in which lead was an interferent agent.The reaction involved the formation of a ternary complex between Malachite Green (MG), iodide and the metal cation. The flow injection manifold included a Chelex-100 mini- column to concentrate lead from the digested wine samples. Experimental Reagents and Solutions All solutions were prepared with de-ionized water and analyt- ical reagent-grade chemicals. * To whom correspondence should be addressed. BufSer solution (RI). Ammonium acetate (93 g) and glacial acetic acid (14 ml) were dissolved in water to 1 1. Malachite Green (RZ). Prepared weekly by dissolving 52.4 mg of MG in 250 ml of water; the solution was stored in an amber-glass bottle. Potassium iodide (Rs). Potassium iodide (83 g) and ascorbic acid (1 .O g) were dissolved in 250 ml of water; the solution was protected from light.Conditioning bufSer (R4). Glacial acetic acid (97 ml) was added to 54 g of sodium acetate trihydrate previously dissolved in water and diluted to 1 1 (pH == 4). BufSerlmasking solution (Rs). Ammonium acetate (54 g), sodium hydroxide (20 g) and picolinic acid (1.1 g) were dissolved in water to 1 1. Working standard solutions. Solutions within the range 25-500 mg 1-1 were prepared daily from a 1000 mg 1-1 lead(rr) stock standard solution (Merck Ltd., Poole, Dorset, UK) in 0.05 moll-' HNO?. Sample Preparation Wine samples were digested according to the method described for lead determination in Port wine.6 To 100 ml of sample, 1 ml of sulphuric acid (1 + 1) was added and the sample was allowed to evaporate in a water-bath.The residues were then miner- alized at 500-550 "C until white ash was obtained. The ash was dissolved in 1 ml of nitric acid (1 + 4) plus 5 ml of water. The solution was heated to boiling, filtered through cotton and diluted with water to 100 ml. The final concentration of nitric acid (0.05 mol 1-1) was about the same as in the standard solutions. Apparatus The propulsion system consisted of an Ismatec (Zurich, Switzerland) Mini S 830 and two Gilson (Villiers-le-Bel, France) Minipuls 3 peristaltic pumps. The injection system was a laboratory-made commutator7 and the detector was a Philips (Eindhoven, The Netherlands) PU 8625 UV/VIS spcctropho- tometer with a Hellma (MullheimDaden, Germany) 178.7 1 1 -QS flow cell ( 1 0 mm path length) coupled to a Kipp & Zonen (Delft, The Netherlands) BD 111 recorder.The wave- length was set as 690 nm. Omnifit PTFE tubing (0.8 mm id) was used for the manifold conduits. Reactors B3 (Fig. 1) and Bs (Fig. 2) were immersed in an ice-bath in order to maintain the temperature at about 4 "C. All the reactors were helically coiled to promote radial mixing. The resin mini-column consisted of a piece of a Gilson PVC pump tubing (2 cm X 2 mm id) filled with Chelex-100 ion- exchange resin (Bio-Rad Labs., Richmond, CA, USA, 2004001048 Analyst, August 1996, Vol. 121 mesh, sodium form). The resin was dispersed in the condition- ing buffer and further introduced into the column by means of a syringe.Ordinary dishwashing foam was placed at both ends of the mini-column to entrap the resin, which was thereafter conditioned in the flow injection system by consecutive injections of 0.05 rnol 1-1 HN03 solution. The determinations related to the reference method were carried out in a Perkin-Elmer (Norwalk, CT, USA) Model 4100 ZL atomic absorption spectrometer with electrothermal atomi- zation, using the operating conditions described elsewhere.' Procedure The manifold for studying the chromogenic reaction is presented in Fig. I . The standards (S) were introduced into a carrier stream (C) with the same acidity (0.05 rnol 1-1 HNO') as the digested wines. The injected plug was mixed with a buffer solution ( R , ) at confluence a in order to achieve a pH value in the'range 3.5-4.Subsequently, the dispersed zone merged with the reagents MG and KI at confluence b, and the chromogenic reaction started inside reactor B3, which was immersed in an ice-bath. Thereafter, the processed plug passed through the flow cell and the transient absorbance was recorded as a peak with height proportional to the lead content in the injectate. The influence of the water-bath temperature (4-40 "C), pH (2.0-4.5) for reaction development, concentrations of KI (1.0-4.0 mol 1-1) and MG (1.0 X 10-4-1.0 X 10-3 mol 1 - 1 ) and reactor lengths (3-200 cm) on the analytical signal were assessed by using a 500 pl sampling loop (L,) and fixed flow rates (Fig. 1). Variations in pH were attained by using acetic acid-acetate buffer solutions with different concentration ratios.The flow injection system incorporating the chelating ion exchange resin is shown in Fig. 2. In the concentration position, the sample plug was carried by 0.05 rnol 1-1 HN03 solution (C), merged at confluence d with the conditioning buffer (R4) and flowed through the column, where lead was retained and concentrated. The in-line concen- tration step was time controlled. When the commutator was switched, the eluent (0.5 rnol 1-l HN03) passed through the column and washed out the lead ions. The resulting plug was mixed at confluence a with R5 solution to adjust the pH to 3.5-4 and to suppress copper interference (picolinic acid). Potassium iodide and MG were then added at confluence b and the plug was processed similarly as mentioned above.This system (Fig. 2) was optimized in relation to the preconcentration pH and concentration time and as the eluent concentration. A detailed re-evaluation of the parameters related to the chromo- Fig. 1 Flow injection aystem for the study of the chromogenic reaction. S = standards; C = carrier (HN03 0.05 rnol I-I); R I = buffer solution (1.2 moll-' ammonium acetate-0.2 moll-' acetic acid); R2 = Malachite Green (5.0 X mol I-,); R? = potassium iodide (2.0 mol 1-I); B, = reactors (B, = 30 em, B2 = 50 cm, B1 = 150 cm); L1 = 0.5 ml; a, b, c = confluences; W, = waste; numbers in parentheses are flow rates (ml min-I); h = spectrophotometer set at 690 nm. The manifold components within dashed lines were kept immersed in an ice-bath (T = 4 "C). The shaded area is an alternative permissible position of the commutator.genic reaction (length of Bs and flow rates of the reagents MG and KI) was performed, to adjust them to the changes introduced by the preconcentration step. After system optimization, the main analytical characteristics were evaluated: detection limit, precision and sampling rate. Thereafter, standards and samples (wine digests), already analysed by ETAAS, were injected in duplicate into the flow system and the concentrations were calculated from the cali bration graph. Results and Discussion A flow injection system was first developed and optimized in order to maximize the sensitivity (slope of the calibration curve) and obtain a low detection limit. The concept of this assembly was based on a previous one developed for the determination of cadmium using the same reaction.8 However, the detection limit of this system was not low enough for lead determination in Port wines (expected concentrations below 300 pg 1-I), and so a preconcentration step was introduced into the system to allow the determination of lower concentration levels.Optimization of the Flow Injection System Without Ion Exchange The system is very dependent on temperature variations. Within the tested temperature range (4-40 "C) the signal increased as the temperature decreased (Fig. 3). The temperature was set to about 4 "C as this corresponds to the temperature obtained with an ice-bath. Regarding the acidity inside reactor B3 (Fig. I ) , a pH in the range 3.5-4 was used. For a 500 pg 1-1 lead standard, the absorbance increased 3.3-fold with increase from pH 2.0 to 3.0, remained approximately constant up to pH 4.2 and then decreased 1.2-fold with an increase to pH 4.5.On increasing the KI concentration from 1 .O to 4.0 rnol 1- I , the signal increased up to 2.0 mol 1 - 1 , with a further slight increase being observed for higher KI concentrations. More- over, at iodide concentrations 22.0 rnol l- ' precipitation effects were observed when MG concentrations 27.5 X 10F moll-' were used. The concentration of KI was then selected as 2.0 Fig. 2 Flow injection system with in-line ion-exchange concentration for the determination of lead in Port wine. S = standards/samples; C = carrier (0.05 mol 1--' HNO,); E = eluent (0.5 mol 1-i HNO,); R2 = Malachite Green (5.0 X lop4 mol 1-I); R3 = potassium iodide (2.0 mol I-'); R4 = conditioning buffer solution (1.4 mol 1-1 acetic acid-0.4 mol 1-1 sodium acetate); R5 = buffedmasking solution (0.7 moll- ammonium acetate-0.4 mol I-' sodium hydroxide-0.02 mol I-' picolinic acid); Bj = reactors (B, = 30 cm, B2 = 50 cm, B4 = 50 cm, B5 = 200 cm); L2 = 1 ml; a, b, c, d = confluences; Wj = waste; numbers in parentheses are flow rates (ml min-I); h = spectrophotometer set at 590 nm.The manifold components within broken lines were kept immersed in an ice-bath (T = 4 "C); column filled with Chelex 100 resin (Bio-Rad), 200-400 mesh (2.2 cm X 2.2 mm id). The shaded area is an alternative permissible position of the commutator.Analyst, August 1996, Vol. 121 1049 moi 1-I. The MG concentration was set at 5.0 X 10-4 mol 1-1, since above this value an increase in the baseline noise was observed, probably owing to the precipitation of MG on the flow cell and tubing walls. The length of reactor B, should be as small as possible to minimize dispersion of the injected plug, yet sufficient to allow the mixing of the plug with the buffer stream; it was set as 30 cm.The length of B2 should not be selected too short, since its main function was to provide a given time interval for KI-MG interaction prior to their addition to the sample.y It was varied from 3 t o 100 cm and a value of 50 cm was chosen as this corresponded to the maximum sensitivity. The decrease of sensitivity observed with longer tubes was probably due to precipitation effects as a consequence of micellar growth inside the reactor.The length of B3 was selected as 150 cm as the best compromise between sampling rate, sensitivity, possibility of precipitation and measurement precision. System With the Ion-exchange Mini-column In the flow system incorporating the chelating ion exchange resin (Fig. 2), for a fixed flow rate and loop (L2) volume, the amount of lead flowing through the resin was controlled by the time interval between commutations required for sample injection and elution (also loading position). The signal underwent a 200% increase when this interval was changed from 15 to 30 s, and only 7% from 30 to 90 s. Therefore, a concentration time of 30 s was set. Regarding the three pH values tested for concentration (4.3, 8.5, 9.3, maximum sensitivity was observed at pH 4.3, which confirmed details supplied by the manufacturer.The eluent concentration was investigated in the range 0.25-1.0 mol 1-1 HNOI. A concentration of 0.5 mol I-' was adopted as since there was no significant improvement in signal intensity and return to baseline (elution efficiency) at higher concentrations. Moreover, the use of higher acid concentrations would make it more difficult to buffer the resulting solution. In the system in Fig. 2, the length of reactor B5 was set as 200 cm. Maximum sensitivity was attained with this length, which provided a narrow baseline (< 0.002 A). When the flow rates of the MG and KI streams were increased from 0.4 to 0.8 ml min-I, a systematic increase in the signal was observed, but beyond 0.7 rnl min-l the baseline became unstable.The flow rates of the two reagents were therefore set to this value. Interference from copper was observed as it also forms a ternary complex with MG and KI.5 In order to overcome this interference, picolinate, which forms a very stable, colourless complex with copper,x was included in solution Rs. A picolinic 0.20 A 0.15 0.10 0.05 0.00 bath temperature I0C tnlluence of tempcrature on the analytical signal ( A ) for a lead Fig. 3 concentration of 500 pg I-', obtained with the system in Fig. 1 . acid concentration of 1 X mol 1 - i was used and copper concentrations as high as 3 mg 1-1 were suppressed. This is sufficient as the maximum copper content in wines2 should not exceed 1 mg 1-1. The interference of cadmium was not studied as its content in wines4 is usually substantially lower than that of lead.In addition, at the preconcentration pH used, lead retention in the ion-exchange resin is much higher than that of cadmium. Application to Wine Analysis In order to assess the quality of the results obtained with the developed flow injection system for the determination of lead in Port wines, 20 digested samples were analysed by the proposed procedure and by ETAAS.3 The values obtained and their differences are presented in Table 1. From a linear regression between the results obtained with the two methods, the 95% confidence limits obtained' I for 18 degrees of freedom (t-value = 2.10) were 0.8 1 k 10.29 pg 1-I for the intercept and 1.021 f 0.080 for the slope; these results indicate good agreement between the two methods.The results obtained by the proposed procedure were precise, with RSDs (n = 5 ) of 2.6,2.5 and 1.2% for wines with concentrations of 35, SO and 80 pg I-', respectively. The detection limit was determined as 12 pg 1- 1 according to IUPAC recommendations. 12 The sampling rate achieved was about 45 determinations per hour. Conclusions The flow injection system with spectrophotometric detection proposed for lead determination in Port wine is an alternative to more expensive methodologies such as AAS. It provides good- quality results, in terms of accuracy and precision (RSDs lower than 2.6%), and allows 45 determinations per hour for the digested samples. With some adaptations, the described methodology can be applied to the determination of low levels of lead in other matrices.The authors acknowledge support from the European Union through project AIR PL 94 2468. T. I. M. S. Lopcs acknowledges grant PRAXTS XXI/BD/53 18/95. Agostinho A. Almeida, Jos6 Anchieta G. Neto and C. C. Oliveira are thanked Table 1 Comparison between the results obtained by FI and by ETAAS Lead concentration/yg I-' Difference Sample No. FI ETAAS w> 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 66 64 189 67 63 61 52 75 57 44 100 85 so 72 68 56 57 70 54 64 63 67 78 68 56 58 49 79 51 46 04 85 so 68 71 60 50 63 47 55 4.8 -4.5 6.2 -1.5 12.5 5.2 6.1 -5.1 11.8 -4.3 -3.8 0 0 5.9 -4.2 -6.7 14.0 11.1 14.9 16.41050 Analyst, August 1996, Vol. 121 for critical comments. The collaboration of Cockbum Smithes (Gaia, Portugal) is also acknowledged. 6 7 References 8 9 Ough, C. S., Am. J . End. Vitic., 1993, 44, 464. Office International de la Vigne et du Vin (OIV), Recueil des 10 Mkthodes Internationales d’Analyse des Vins et des MoQts, OIV, Bordeaux. 1990. 11 Almeida. A. A., Bastos, M. L., Cardoso, M. I., Ferreira, M. A., Lima, J. L. F. C., and Soares, M. E., J . Anal. At. Spectrom., 1992, 7, 12 1281. Ough, A., Methods .for Analysis of Musts and Wines, Wiley, New York, 1980. Lopez Garcia, I., Navarro. P., and Hernandez Cordoba, M., Talanta, 1988, 35, 885. Peres, M. R. L. A., and Pereira, J., Separata dos Anais do Instituto do Vinho do Porto, 1975, p. 6. Bergamin F@, H., Medeiros, J. X., Reis, B. F., and Zagatto, E. A,, Anal. Chim. Acta, 1978, 101, 9. Gomes Neto, J. A., Bergamin F@, H., Zagatto, E. A. G., and JSrug, F. J., Anal. Chim. Acta, 1995, 308, 439. Gomes Neto, J. A., Bergamin F@, H., Sartini, R. P., and Zagatto, E. A. G., Anal. Chim. Acta, 1995,306, 343. Chelex 100 and Chelex 20 Chelating Ion Exchange Resin, Instruction Manual, Bio-Rad Laboratories, Richmond, CA. Miller, J. C., and Miller, J. N., Statistics for Analytical Chemistry, Ellis Horwood, Chichester, 3rd edn., 1993. International Union of Pure and Applied Chemistry (IUPAC), Anal. Chem., 1976,48, 2294. Paper 6102225A Received March 29, 1996 Accepted May 13,1996

ISSN:0003-2654    

DOI:10.1039/AN9962101047

出版商:RSC    

年代:1996

数据来源: RSC