Diagnosis using backpropagation neural networks—analysis and criticism

Abstract Artificial neural networks based on a feedforward architecture and trained by the backpropagation technique have recently been applied to static fault diagnosis problems. The networks are used to classify measurement vectors into a set of predefined categories that represent the various functional and malfunctional states of the process. While the networks can usually produce decision surfaces that correctly classify the training examples, regions of the input space not occupied by training data are classified arbitrarily. As a result, the networks may not accurately extrapolate from the training data. Although extrapolation is not required under ideal circumstances, in practice the network may be required to extrapolate when undersized training sets are used, when parent distributions of fault classes undergo shifts subsequent to training, and when the input data is corrupted by missing or biased sensors. These situations cause relatively high error rates for the neural classifier. A related probem is that the networks cannot detect when they lack the data for a reliable classification, a serious deficiency in many practical applications. Classifiers based on distance metrics assign regions of the input space according to their proximity to the training data, and thus extrapolation is not arbitrary but based on the most relevant data. Distance-based classifiers perform better under nonideal conditions and are to be preferred to neural network classifiers in diagnostic applications.