A comprehensive examination of neural network architectures for analog fault diagnosis

We have performed a detailed comparison of single-output and 1-of-k neural-network architectures used for analog fault diagnosis. Training and testing data are obtained from linear and nonlinear circuits simulated by SPICE. We have used wavelet and Fourier transforms, data normalization and principal component analysis to preprocess the node voltages and extract optimal features for training the neural networks. Comparison of the two architectures using the sample circuits reveal that 1) for circuits with overlapping features across fault classes, the 1-of-k is the superior architecture for analog fault diagnosis, 2) the 1-of-k architecture requires a significantly smaller training set to accurately diagnose faulty behavior, 3) because of its bigger size, the 1-of-k architecture imposes a high demand on computer resources and may become prohibitive for problems with large number of faults and 4) the training time per epoch is longer for the 1-of-k architecture but it requires a smaller number of epochs to reach the prespecified error goal. These results indicate that the 1-of-k architecture provides a more robust and stable analog fault diagnostic system unless it is prohibited by its increased demand on computer resources.