Fault-diagnosis using neural networks with ellipsoidal basis funcions

In this paper a fault detection scheme for applications in the automotive industry is presented. The detection scheme has to process up to several hundreds of different measurements at a time and check them for consistency. Our fault detection scheme works in three steps: First, principal component analysis of training data is used to determine nonsparse areas of the measurement space. Fault detection is accomplished by checking whether a new data record lies in a cluster of training data or not. Therefore, in a second step the distribution function of the available data is estimated using kernel regression techniques. In order to reduce the degrees of freedom and to determine clusters of data efficiently in a third step the distribution function is approximated by a neural network. In order to use as few basis functions as possible a new training algorithm for ellipsoidal basis function networks is presented: New neurons are placed such that they approximate data points in the vicinity of their centers up to the second order. This is accomplished by adapting the spread parameters using Taylor's theorem. Application to measured data from a real automotive process show that the proposed algorithm yields good results.