MINIMISATION OF DECISION ERRORS IN A PROBABILISTIC NEURAL NETWORK FOR CHANGE POINT DETECTION IN MECHANICAL SYSTEMS

Probabilistic neural nets have been applied in the detection of structural damage. These networks, which rely upon approximating the multivariate density of the training data, have been shown to be effective in some applications. However, quantification of decision errors, which must ultimately be used to rank their effectiveness, has received little attention. In this paper, a two-dimensional probabilistic pattern classifier (PPC) based on an L2detector is studied. Each dimension is modelled as a Gaussian random variable in order to directly study the effects introduced by a commonly used density estimator. The region of acceptance is examined for different parameters of the kernel density estimator. The detector behaviour of the original PPC is compared to theory. It is demonstrated that performance is affected by errors introduced by a kernel constructed using a finite set of data, and also from theoretical limitations inherent in the test statistic. A modification of the test statistic is suggested to improve the sensitivity for structural damage detection.