Classifying Process Behavior with Neural Networks: Strategies for Improved Training and Generalization

Artificial neural networks trained by backpropagation have recently been applied to fault diagnosis problems. Backpropagation produces-decision surface that effectively separate training examples of different cases. However, separation of the training examples does not necessarily result in the most plausible or robust classifier. The decision surfaces may deviate from optimality and take on non-intuitive shapes because the decision surfaces are based on hyperplanes. This also results in backpropagation networks having no mechanism to detect when a case to be classified falls in a region with no training data. To remedy these difficulties a network is proposed which uses Radial Basis Functions instead of sigmoid threshold units. This network is compared to the standard backpropagation network on an prototypical example problem.