GAUSSIAN MIXTURE MODELING VERSUS AUTO ASSOCIATIVE NEURAL NETWORK FOR ANALOG CIRCUITS FAULT DIAGNOSIS

Abstract

A new fault Diagnosis procedure for analog circuits is presented. The remarkable abilities of the Gaussian Mixture Model (GMM) and the Auto Associative Neural Network (AANN) to model arbitrary densities are exploited in isolating analog circuits' faults. The innovation aspect of the proposed approaches is the use of new training techniques for the GMM and the introduction of the AANN model in the analog fault diagnosis problem.

Three different learning algorithms for the GMM are investigated: the regular Expectation Maximization (EM) algorithm; and two unsupervised learning algorithms: Figueiredo and Jain (FJ) Algorithm, and the Greedy EM (GEM) algorithm. These two algorithms are capable of selecting the number of components of the mixture, they don't require careful initialization, and they avoid the possibility of convergence toward the boundary of the parameter space. AANN can be viewed as an alternative to the current approaches based on GMMs especially when the distribution is not Gaussian. It is a feedforward neural network performing identity mapping of the input space. The distribution capturing capability of the AANN models is studied using a probability surface derived from the training error surface. A new error normalization procedure for the AANN model is presented.