A Combined Clustering and Neural Network Approach for Analog Multiple Hard Fault Classification

A new neural network-based fault classification strategy for hard multiple faults in analog circuits is proposed. The magnitude of the harmonics of the Fourier components of the circuit response at different test nodes due to a sinusoidal input signal are first measured or simulated. A selection criterion for determining the best components that describe the circuit behaviour under fault-free (nominal) and fault situations is presented. An algorithm that estimates the overlap between different faults in the measurement space is also introduced. The learning vector quantization neural network is then effectively trained to classify circuit faults. Performance measures reveal very high classification accuracy in both training and testing stages. Two different examples, which demonstrate the proposed strategy, are described.