Applying mutual information theory to behavioural analogue fault modelling

To assess the effectiveness of a testing strategy for an integrated circuit, the potential structural faults in a circuit must be modelled. Analogue fault simulation is conventionally done at the transistor level. Behavioural fault models are desirable to speed up the simulations. Behavioural fault modelling needs faults to be grouped. However, it is not easy to group faults using a Euclidean measurement of the distance between faults, if the populations of the circuit faults have distributions with differing variances. Mutual information theory is suggested here as a robust method for clustering circuit faults. The bootstrap technique is proposed to speed up the process of generating statistical data. Statistical data on the performance of circuits under fault conditions is generated using HSPICE. A software program has been written to implement clustering of responses using mutual information theory and to generate statistical data using bootstrap. The technique is shown to generate a suitable set of parameters for a regression function. The simulation results for the behavioural models are close to those of the full circuit model. Mutual information theory is a useful technique for clustering responses of circuits under fault conditions.