DRIFT CORRECTION FOR PATTERN-CLASSIFICATION WITH NEURAL NETWORKS

A neural network may be used as a pattern classifier to assign objects, which are characterized by different variables, to certain classes (J.R.M. Smits et al., Chemom. Intell. Lab. Syst., in press). One of the problems that might be encountered with pattern classification is instrumental drift. Drift may cause the neural network to misclassify the objects, if the clusters of the different classes lie relatively close to each other. Therefore it is necessary to calibrate the analytical instrument that is used to measure the variables. However, calibration of an instrument is not always straightforward. Correction of the measurements for drift would make it permissible to calibrate less often. In this paper the effects of different drift correction strategies are studied. Simulated data sets are used, with different amounts of drift of a non-linear nature. A multi-layer feed-forward neural network is used to classify the simulated objects. The performance of the neural network is determined for three situations: no drift correction, correction for the drift by subtracting it, and correction for the drift by using the amount of drift as an extra input variable for the neural network. For the type of non-linear drift used in this study, the last drift correction strategy appears to give the best results.