Analyst, June 1996, Vol. 121 (743-748) 743 Interpreting Signals From an Array of Non-specif ic Piezoelectric Chemical Sensors* Patricia McAlernon, Jonathan M. Slatert, Philip Lowthian and Mark Appleton Centre for Analytical Science, Birkbeck College, University of London, Gordon House, 29 Gordon Square, London, UK WCIH OPP Pattern recognition methods were tested using a gas sensor array consisting of eight interchangeable quartz crystals coated with different sorbent layers. The system is designed in such a way as to allow headspace sampling from a jar or vial. Using a two-sample t-test the two sampling methods were found to be significantly different at the 95% confidence level. It was found that the application of principal component analysis, multivariate analysis of variance and discriminant function analysis to the gradient of the initial responses and over-all response magnitudes allowed hexane, o-xylene, toluene, dodecane and tetradecane to be distinguished.Keywords: Array sensor; piezoelectric sensor; olfaction; pattern recognition Introduction The array sensor concept was first proposed around 10 years ago as an approach for overcoming some of the limitations of existing discrete sensors.' The principle of using a number of sensors simultaneously and analysing the combined responses has some analogies with human olfaction and hence the term 'electronic nose' has become popular in certain quarters. Often the 'nose' analogy can be misleading because most systems generate and match physico-chemical fingerprints and do not attempt to map the human olfactory percept.Nevertheless, the promise of such systems is enormous because they offer the potential for generating unambiguous standards for measure- ments previously described only in subjective terms. There are many methods of implementing the array concept, the most significant variations being the choice of transduction platform and the pattern recognition regime. However, the physical embodiment of the method, e.g., the sampling system, is often critical in determining the quality of information available and hence the outcome of any measurement. The objective of this paper is to demonstrate how the sampling regime and initial signal processing methods effect the success of the technique for a particular transducer platform.A range of sensor technologies have been employed, including bulk wave and surface acoustic wave quartz crystal resonators,2-5 metal oxide gas sensors,6 electrochemical cells7 and conducting polymer chemiresistors.8 Recent research on multilayer conducting polymer gas sensors arrays has also proved promi~ing.~ All of these transducers have different operating characteristics and the physical mode of interaction with the gas phase is understood to a greater or lesser extent. In this study, we chose bulk wave quartz crystal microbalance sensors because the amount of information available about their mode of interaction is c ~ n s i d e r a b l e ' ~ ~ ~ and this may be utilized in optimizing the array method. Experimental Reagents and Materials The sample test panel comprised two aromatics (toluene and o- xylene) and three aliphatics (hexane, dodecane and tetrade- cane).All were AnalaR-grade reagents from Merck (Poole, Dorset, UK). The sample containers were either 1.5 ml polypropylene vials (Sarstedt, Leicester, UK) or modified 60 ml glass jars (Fisons, Loughborough, UK). Instrumentation The sensor response data were collected on a ScanMaster I1 sensor array system (Array Tec Chemical Sensors, Chesham, Buckinghamshire, UK). A block diagram of the system is shown in Fig. 1. The sensor chamber contained eight piezo- electric quartz crystals with different sorbent layers ACS2 to ACS9 (Array Tec Chemical Sensors) designed for organic vapours. The system generated comma-separated variable (CSV) data files containing the response of each individual sensor over the time course of an experiment.These data files were exported to Excel (Microsoft UK) and MINITAB (Minitab, PA, USA) for data analysis. Sampling Methods Two sampling regimes were examined: (A) 10 ml of sample were placed in a 60 ml glass jar and left to equilibrate for 10 min at 25 "C; and (B) 0.5 ml of sample was placed in a 1.5 ml polypropylene vial at 25 "C. In both cases the instrument was configured to sample the headspace for 60 s in a recirculating mode and then switched to a purge cycle for 120 s. Results and Discussion Array sensor devices contain a number of individual sensors which respond to stimuli. The extent to which each sensor responds will depend on its affinity for a given analyte. Thus the pattern generated is essentially a chemical 'fingerprint'. The fingerprint may equally well be in response to a complex mixture as to a single compound.In the case of a mixture, the 'fingerprint' will change as the composition changes. * Presented at Sensors and Signals 111, Malahide, Co. Dublin, Ireland, October 26-27, 1995. + To whom correspondence should be addressed. Fig. 1 Block diagram of the ScanMaster 11 sensor array system.744 Analyst, June 1996, Vol. 121 The quartz crystal microbalance array is essentially a set of sorption sensors where the response is determined by the diffusion of gases and vapours into absorbent ‘solvent’ be employed. polymers. A particular advantage of the sorption sensor platform is that for a wide range of samples, the sensor response is linearly additive,16 so linear pattern recognition routines can Statistical Analysis of Data Prior to application of pattern recognition methods, the quality of the data was assessed.Individual sensor responses taken at Fig. 2 Response (at 60 s) of the quartz crystal gas sensor array to aliphatic and aromatic compounds contained in 1.5 ml polypropylene vials: (a) relative frequency (%); and (b) fractional frequency change (96). Fig. 3 Response (at 60 s) of the quartz crystal gas sensor array to aliphatic and aromatic compounds contained in 60 ml glass jars: (a) relative frequency (%); and (b) fractional frequency change (%). Table 1 Correlation matrices of standardized % fractional frequency responses for jar sampling regime at different times Method Sensor 3 s 1 2 3 4 5 6 7 8 60 s 1 2 3 4 5 6 7 8 1 2 3 4 5 1.00000 0.87868 0.85956 0.87414 0.87495 1.00000 0.98743 0.98396 0.94641 1 .00000 0.98975 0.93556 1 .00000 0.93682 1 .ooooo 1 .OOOOO 0.97477 0.99094 0.42201 0.97871 1 .OO000 0.99414 0.5 1302 0.99829 1 .OOOOO 0.44827 0.99493 1 .OOOOO 0.49884 1 .ooooo 6 0.88 150 0.9 1482 0.93 149 0.942 17 0.957 19 1 .00000 0.99 195 0.97342 0.98723 0.45020 0.97728 1 .ooooo 7 0.86724 0.87438 0 39901 0 30810 9.9 1779 ~1.98874 1 .00000 0.99175 0.97334 0.98657 0.46647 0.97626 0.99917 1 .00000 8 0.5 1649 0.39737 0.31591 0.40425 0.49 160 0.42438 0.37480 1 .ooooo 0.53 1 18 0.6092 8 0.56355 0.7 103 1 0.60587 0.58125 0.58435 1 .oooooAnalyst, June 1996, Vol.121 -2 - 745 0 A 60 s were found to have relative standard deviations (RSD) of between 2 and 5% (n = 5).The RSD was correlated with the magnitude of the response, which suggests that there may be useful residual information available in the noise generated by different sensor coatings. Using a two-sample t-test, the two sampling methods were found to be significantly different at the 95% confidence level. Sensor responses for jar headspace measurements were found to be up to 50% higher than those for corresponding vial headspace measurements. The headspace of the jar is 50 cm3 whereas the headspace of the vial is 1 cm3. The total value of the recirculating flow system is around 15 cm3, corresponding to a 30% dilution of the jar headspace sample and a 1500% dilution of the vial headspace sample. A consequence of the measuring regime is that the concentration of the vapour- phase sample will increase throughout the measurement cycle, although at a greater rate for the vial than the jar. On exposure to a vapour, the frequency of the sensor (termed the gradient of the initial response) decreases rapidly, then reaches an equilib- rium frequency (termed the over-all magnitude).The RSD for sensor responses taken at 3 s (essentially the gradient of the initial response) varied from 30 to 50% for both jar and vial measuring regimes. Signal Pre-processing Previous studies suggest that the pre-processing algorithm employed is important in determining the performance of the pattern recognition method. Various pre-processing parameters have been used in the field of gas sensing, for example, difference models, relative models, fractional difference models and normalization procedures. *7-l9 However, for the bulk acoustic wave quartz crystal microbalance, data extraction from sensor response curves has not yet been fully exploited.The over-all response magnitude is the most commonly used descriptor of sensor response, but the gradient of the initial response or the sensor recovery may also be a useful variable in pattern recognition regimes. Nanto et a1.20 used parameters which characterize the transient responses of quartz crystals to discriminate aromas. Edmonds et a1.21 also used initial rate measurements of quartz crystals and compared them with equilibrium shift values. Saunders et ~ 1 . 2 2 examined the time- dependent frequency responses (termed kinetic signatures) of sensors.Using the initial sensor response has two significant advantages: (1) it will shorten the analysis time and (2) it may increase the lifetime of the sensor. Table 2 Eigenanalysis of the correlation matrix for the various sampling methods Vial method Jar method Eigenvalues 3 s 60 s 3 s 60 s 6.0678 (75.8%) 1.1520 (1 4.4%) 0.4986 (6.2%) 0.225 1 (2.8%) 0.0301 (0.4%) 0.0 134 (0.2%) 0.0094 (0.1 %) 0.0037 (0% ) 7.2559 (90.7%) 0.5802 (7.3%) 0.1361 (1.7%) 0.0 192 (0.2%) 0.0044 (0.1 %) 0.0030 (0%) 0.001 1 (0%) 0.0000 (0%) 6.7433 (84.3 9%) 0.8337 (10.4%) 0.2019 (2.5%) 0.1512 ( I .9%) 0.0562 (0.7%) 0.008 1 (0.1 %) 0.0036 (0%) 0.0022 (0%) 6.6059 (82.6%) 1.0543 (1 3.2%) 0.2794 (3.5%) 0.0509 (0.6%) 0.0068 (0.1%) 0.0020 (0%) 0.0005 (0%) 0.0003 (0%) Since the efficiency of pattern recognition methods relies on samples having different response patterns, it is crucial to maximize these differences using the most appropriate pre- processing algorithm.In this work three pre-processing methods were investigated using responses obtained at 60 s only: (i) % relative frequency = [AF(t)/AFo(t)] X 100 (ii) % fractional frequency change = [AF(t) - AFo(t))/ (iii) normalized fractional frequency change = [AF(t) - AFo(t)J X 100 AFo(t)/AFo(t)] X loo/(, 2 [AF(t) - A F O ( ~ > ] ~ / A F O ~ ( ~ ) } ‘” I = 1-8 where AF(t) is the frequency after exposure ( H z ) at time t s (t = 60) and AFo(t) is the initial frequency (Hz) at time t s ( t = 0). The % relative frequency and the % fractional frequency response patterns obtained from the sensor array system at 60 s using the vial and jar regimes are shown in Figs.2 and 3, respectively. The latter pre-processing method maximized the differences in response patterns taken at 60 s and was employed for subsequent pre-processing of responses at t = 3 s. Choosing this pre-processing method for sensor responses at 3 s will essentially give the gradients of initial sensor responses. Pattern normalization and subsequent autoscaling to a mean of zero and standard deviation of unity were not found to extract any extra information or enhance the response of any of the pattern recognition methods employed. Pattern recognition is some- times employed to remove the effects of concentration and the sensitivity of one vapour relative to another. Autoscaling has the disadvantage that sensors which have weak responses to an individual compound are often subject to greater relative experimental error but have an equal influence on the analysis, leading to noise propagation within the response pattern.23 ‘1 0 0.0 0.5 :: “15, I I c X c O.0 - 1.0 1 I I 1 I 1 1 - 4 -3 - 2 - 1 0 1 2 3 FC1 Fig. 4 Principal component analysis of data for (a) 3 s and (b) 60 s response of the quartz crystal gas sensor array to (0) hexane, ( X ) o-xylene, ( A ) toluene, (0) dodecane and (0) tetradecane contained in 1.5 ml vials.746 Analyst, June 1996, Vol. 121 4 - 3 - 3--r ( b ) A 2 - 4a 2 - (v 1 - 0 - 1 - -1 QD -1 - A i? 0 - -2 - 3 - -1 - r - 2 - 1 Pattern Recognition Methods Pattern recognition can be divided into unsupervised (classifi- cation) and supervised (discrimination) approaches.24.25 These methods differ in that a training set is required for supervised learning in order to establish a response model.Unsupervised learning is commonly used for preliminary investigations to determine the natural groupings of data in two-dimensional space. The unsupervised method employed for this work was principal component analysis (PCA). Principal component analysis has been successfully applied to analyse the response of piezoelectric devices2C28 and more recently to classify odours. The object of principal component analysis is to take p variables XI, X2, ..., X p and find combinations of these to produce indices Z,,Z,, ..., 2, that are uncorrelated and measure different ‘dimensions’ in the data. The indices are also ordered so that Z1 displays the greatest variation, 2 2 displays the second largest and Zp the least. Since the best results are obtained when the original variables are very highly correlated, this multivariate technique is ideally suited to gas sensor arrays which have partially overlapping sensitivities to different sample components. Our implementation of PCA involved four steps.First, the variables were coded to have zero means and unit variance. Second, the correlation matrix was calculated for the different sampling regimes. The correlation matrices for the jar sampling regime are shown in Table 1. Some sensors are more highly correlated at the 3 s response than the 60 s response, which may suggest that different interaction mechanisms predominate at different times in the sensor response profile.The third step involved calculating the eigenvalues and corresponding eigenvectors. The eigenvalues and corresponding percent variance for the different sampling regimes are shown in Table 2. The first principal component is by far the most important of the eight components for X ( b ) X x x x A * a d .4 1 1 1 1 1 1 I I 1 AA I I I I I I 1 1 Ic I - 6 - 5 - 4 - 3 - 2 - 1 0 1 2 3 4 A r .C, I I I I I I I I I -5 - 4 -3 - 2 - 1 0 1 2 3 FC1 representing the variation in the measurements of the five samples. The eigenvectors, more commonly called the loadings, provide the coefficients of the principal components. The loadings for the sampling methods were found to be similar for all variables, suggesting that all have an impact on classifica- tion.The corresponding scores for each sample were plotted in two-dimensional space to complete the PCA and these are shown in Figs. 4 and 5. The PCA plot obtained when gradient and steady-state signals were mixed, shown in Fig. 6, reveals less within-group scatter for the jar measurements. Comparison with Figs. 4 and 5 suggests that measurements at 3 s have an important contribution when gradient and steady-stage signals were mixed. This may in turn suggest that an over-all ‘kinetic signature’ may yield valuable information overlooked when gradient and steady-state measurements are considered sepa- rately. In Figs. 4-6 hexane, dodecane and tetradecane cluster in the same area of the plots, which suggests similarities in response patterns.The location of these clusters may indicate similarities in vapour-coating interaction mechanisms for the three ali- phatic compounds. Toluene and o-xylene were found to take up a different location, which may suggest different vapour- coating interaction mechanisms. Comparison of PCA plots for different sampling methods show greater scatter in the 3 vial responses, which is probably related to the increased im- portance of exact timing for initial data collection. Multivariate analysis of variance (MANOVA) has been used to determine several linear combinations (canonical dis- criminant functions) for separating groups.27 The first dis- criminant function gives the maximum possible F ratio (mean 3-l 1 2 -I” 0 a 0 -7 -2 A , I A 1 I I I -6 -1 4 PC1between groups divided by the mean within groups) on a one- way analysis of variance for the variation within and between groups.The second function gives the maximum possible F ratio for a one-way analysis of variance provided that there is no correlation between the first function and the second function within groups. The relationship between groups can be visualized by plotting these two functions for each individual in a similar manner to PCA when the principal components are plotted. Application of MANOVA to the % fractional frequency responses for the eight sensors revealed distinct groupings, as shown in Figs. 7 and 8. The jar headspace responses at 3 s revealed distinct groupings. Results of eigenanalysis for class and corresponding percentage variance are shown in Table 3.The first two canonical variates account for a high percentage of variation in the data. The coefficients of the canonical variates are shown in Table 4. The values suggest that the variables are not highly correlated. The five groups (hexane, o-xylene, 1.0 0.5 0.0 2 Q - 3 -8 Analyst, June 1996, Vol. 121 747 0 1 2 21 f I I I - 20 - 10 0 zl Fig. 7 Application of MANOVA to responses obtained at (a) 3 s and (b) 60 s for (e) hexane, (X) o-xylene, (A) toluene, (0) dodecane and (0) tetradecane contained in 1.5 ml vials. toluene, dodecane and tetradecane) were found to be sig- nificantly different using Wilks' lambda and Bartlett's chi- squared approximation: Wilks' lambda = I Wo 1/1 Bo + WO I where Wo is the sum-of-squares and products within groups and Bo is the sum-of-squares and products between groups.One approach to supervised (discrimination) learning is hard modelling through linear discriminant analysis (LDFA). The disadvantage of this method is that there is no allowance for overlapping of class models. The data do not need to be standardized prior to analysis as is the case with PCA. The first step involves setting up a model for all classes using a training set. The efficiency of the model is then determined using a test set to identify unknowns. Unknowns are identified by calculat- ing the Mahalanobis distance to group centroids and allocating it to the group to which it is closest. LDFA was carried out on the four sets of measurements. Since the data sets were small, 4 3 Q 2 1 0 r- ~ - 1 I I I I -5 -4 -3 -2 - 1 0 '1 f 1 I 0 10 zl 20 Fig.8 Application of MANOVA to responses obtained at (a) 3 s and (b) 60 s for (*) hexane, (X) o-xylene, (A) toluene, (0) dodecane and (0) tetradecane contained in 60 ml glass jars. Table 3 Results of eigenanalysis for class (MANOVA) of the various sampling methods Vial method Jar method Factor 3 s 60 s 3 s 60 s 18.5 136 3.6128 2.4549 0.1984 0.00 0.00 0.00 0.00 (74.71%) (14.58%) (9.91%) (0.8%) (0%) (0%) (0%) (0%) 2303.46 235.77 27.64 12.31 0.00 0.00 0.00 0.00 53.9269 3 1.3439 1.9825 0.1625 0.00 0.00 0.00 0.00 (61.69%) 1532.88 (35.86%) 133.02 (2.27%) 27.37 (0.19%) 1.73 (0%) 0.00 (0%) 0.00 (0%) 0.00 (0%) 0.00748 Analyst, June 1996, Vol. 121 cross-validation was employed to determine how well groups could be separated using the available variables (i.e., sensors).The results of this procedure (also known as leave-one-out) are shown in Table 5. The success rate for the vial and jar methods when 60 s responses were used was 100%. For the 3 s responses a 96% success rate was found. In a further comparison, the data set was halved, which resulted in training with only ten samples and testing with the remaining ten. This caused problems with LDFA owing to highly correlated variables. It was necessary to remove at least five variables which were considered to be too highly correlated. This high correlation was confirmed using multiple linear regression (MLR), which is useful to isolate redundancy of sensors in the array. When highly correlated variables were removed and LDFA was implemented for the 3 s vial and jar responses, six and eight out of ten samples were correctly predicted, respectively.For 60 s vial and jar responses, ten and nine out of ten samples were correct. The results are shown in Table 5. Conclusions The over-all response magnitude results suggest that 1.5 ml polypropylene vials provide the optimum sampling regime on the basis of higher success rates obtained from linear dis- criminant function and less within-group scatter observed in MANOVA plots. These findings may be due to the correlation between the RSD and the magnitude of the response. Sensor responses for jar headspace measurements were much greater Table 4 Coefficients of the canonical variates for the various sampling methods Vial method Jar method Canonical variate Variable 1 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 2 3 s 6 0 s -0.191 -0.912 0.016 0.411 0.003 0.740 0.162 -0.566 -0.108 -1.738 2.314 -6.359 -2.962 -0.172 1.646 0.303 0.045 1.025 0.030 1.301 -0.141 -1.571 0.425 -4.539 -2.993 1.329 2.824 1.774 -0.599 3.004 - 1.923 - 1.998 3 s 60 s 0.413 -0.919 -0.356 0.743 0.270 -0.154 -0.041 0.048 0.301 -1.113 -4.957 8.547 3.139 -2.798 0.450 -0.122 0.420 -1.495 0.309 0.620 0.270 0.034 -0.416 -0.433 -0.052 1.050 -0.899 -0.804 0.550 -1.528 0.529 0.123 Table 5 Results of linear discriminant analysis for the various sampling methods Success rate (%) Using whole data set Method with cross-validation Using part of data set 60 100 80 90 than those for vial headspace measurements, resulting in higher RSDs and consequently more scatter in two-dimensional plots.This sampling regime gave 100% success rates for discriminat- ing between hexane, o-xylene, toluene, dodecane and tetra- decane when correlated variables were removed and LDFA was implemented. For the ‘gradient of response’ approach, the results suggest that sampling from the headspace of a 60 ml glass jar was the best. This sampling regime gave 96% success rates for discrimination using cross-validation. PCA plots obtained by mixing gradient and steady-state measurements suggested that the former has an important contribution ( i e . , the gradient of the response is simply not ignored but is providing additional information). This may in turn suggest that over-all ‘kinetic signatures’ may yield valuable information overlooked when gradient and steady-state measurements are considered separately.The authors thank Array Tec Chemical Sensors for supplying the instrumentation and for their valuable support and advice. References 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 Stetter, J. R., Jurs, P. C., and Rose, S . L., Anal. Chem., 1986, 58, 860. Zaromb, S., and Stetter, J. R., Sens. Actuators, 1984, 6, 225. Gardner, J . W., Sens. Actuators B , 1991, 4, 109. Hidehito, N., Kawai, T., Sokooshi, H., and Usuda, T., Sens. Actuators B , 1993, 13-14,718. Nakamoto, T., Fukuda, A., and Moriizumi, T., Sens. Actuators B , 1993, 10, 85. Shurmer, H. V., Gardner, J. W., and Chan, H. T. Sens. Actuators, 1989, 18, 361. Stetter, J. R., Jurs, P. C., and Rose, S . L., Anal. Chem., 1986, 58, 860. Pelosi, P., and Persaud, K., in Sensors and Sensory Systems for Advanced Robots, ed. Dario, P., NATO AS1 Series, vol. F42, Springer, Berlin, 1988, pp. 361-381. Slater, J. M., Paynter, J., and Watt, E . J., Analyst, 1993, 118, 379. King, W., Anal. Chem., 1964,36, 1735. Alder, J. F., and McCallum, J. J., Analyst, 1983, 108, 1169. Guilbault, G., Zon-Sel. Electrode Rev., 1980, 2, 3. McCallum, J. J., Analyst, 1989, 114, 1173. Lai, C. S. I., Moody, G. J., and Thomas, J. D. R., Analyst, 1986, 111, 511. Slater, J. M., and Paynter, J . , Analyst, 1994, 119, 191. Janghorbani, M., and Freund, H., Anal. Chem., 1973,45, 325. Heiland, G., Sens. Actuators, 1982, 2, 343. Yannopoulos, L. N., Sens. Actuators, 1987, 12,287. Homer, G., and Hierold, C., Sens. Actuators B , 1990, 2, 173. Nanto, H., Kawai, T., Sokooshi, H., and Usuda, T., Sens. Actuators B , 1993,13-14,718. Edmonds, T. E., Hepher, M. J., and West, T. S . , Anal. Chim. Acta, 1986,187, 293. Saunders, B. W., Thiel, D. V., and Mackay-Sim, A., Analyst, 1995, 120, 1013. Gardner, J. W., and Bartlett, P. N., in Techniques and Mechanisms in Gas Sensing, ed. Moseley, P. T., Norris, J., and Williams, D. E., Adam Hilger, Bristol, 1991, p. 335. Manly, B. F. J., Multivariate Statistical Methods: a Primer, Chapman and Hall, London, 2nd edn., 1994, p. 76. Brereton, R. G., Analyst, 1987, 112, 1635. Carey, W. P., Beebe, K. R., Kowalski, B. R., Illman, D. L., and Hirschfeld, T., Anal. Chem., 1986, 58, 149. Ide, J., Nakamoto, T., and Moriizumi, T . , Sens. Actuators B , 1993, 13-14, 35 1 . Gardner, J. W., Shurmer, H. V., and Tan, T. T., Sens. Actuators B , 1992, 6 , 71. Paper 5108341 I Received December 22, I995 Accepted March I I , 1996