In order to enhance the probability of correct quality diagnosis, it is useful to be able to identify the statistical manner in which the quality signal has changed, i.e. identify change structure. Specifically we wish to distinguish between changes in mean, variance and lag one autocorrelation. Because these change structures yield significant similarities in their corresponding output, a multistage decision tree is necessary. A multistage classification system with a neural network and quadratic discriminant functions is used, where neural network output is ana prioridistribution for the Bayesian quadratic discriminant function. Experimental results show that this multistage decision strategy performs significantly better than its single stage counterpart, with an overall success rate of 84%.