Circuit multi-fault diagnosis and prediction error estimation using a committee of Bayesian neural networks

There is currently interest in developing the computerised diagnosis of analogue circuits. The presence of both soft faults and feedback effects make the diagnosis difficult. One approach is to measure the response of the faulty circuit to a test input signal. Faults in the circuit may cause the response to be modified. It may then be possible to diagnose the faults by digitally processing their associated responses. One technique for diagnosing the faults is to use the pattern recognition and nonlinear regression properties of multilayer perceptron (MLP) artificial neural networks. While the MLP approach yields the values of the faulty components, both accuracy of prediction and some estimation of the confidence in the predictions are needed. More accurate predictions are obtainable by training the MLPs using Bayesian methods than by using the conventional back-propagation algorithm. Confidence estimates in the predictions may be formed by computing the error bars associated with each prediction, based upon a theory of errors applicable to MLPs derived from Bayesian statistical theory. In practical terms the prediction errors may then be calculated by using a number of committees of MLPs, where each committee contains several MLPs. This technique is illustrated for the case of diagnosing multi-faults in an integrated differential amplifier