A Two-Stage Neural Network DC Fault Dictionary

A new concept of neural network DC fault dictionary is presented. The neuraI network is combined with two networks: Learning Vector Quantization network and Mu~tiIayerPerceptron. Simulation results are presented and they show that the proposed approach is an improvement as compared with classicaI nearest neighbour role approach. INTRODUCTION AND BACKGROUND Two different approaches to fauIt location in analog circuits can be distinguished: Simulation After Test and Simulation Before Test, so-caUed fault dictionary approach. However SAT approach bas received far more attention in recent years than SBT approach, at present the DC fauIt dictionary is the onIy one that is used in practice [2]. This approach consists of two distinct stages: 1. Pretest analysis to compute a fauIt dictionary, 2. Posttest analysis to identify fault(s). The basic problems that are faced in Stage 1 are fault definition and measurements selection, and these problems have boon discussed in Refs [14]. In this paper we confine ourselves to Stage 2, i.e. to fault isolation. The most widely used criterion for fault isolation is the Nearest Neighbour Rule criterion. The NNR is given by minimizing, over aII simulated faults, the value of Euclidean distance betwoon measured voltages y and stored voltages yf, where superscript f denotes the f-th fault case (f= l,..,n). !J.Vf = 0f. yf )1 0f. y~ The NNR approach has many deficiences: It assumes that onIy voltages are measured and that they aU are of the same order. In practice, also source and output currents can be measured, and it may happen that some voltages are much smaUer than the others. In such case, a contribution of smaUer voltages to the Euclidean distance can be disregarded, however the information carried by them may be crucial for fault detection. It focuses onIy on searching the minimat value of !J.Vfover aU f= l,..,n and does not take into accountthe values of aU other deviations !J.y<l (q=l,..,n; q;ct). The information carried by such deviations may be helpful in fauIt isolation if two or more fauIts give approximately the same minimaI Euclidean distance. A nature of NNR excIudes the possibility of multiple fauIts isolation. The proposed neural network approach overcomes aU of these deficiences. (1) NEURAL NETWORK FAULT DICTIONARY FauIt detection via dictionary approach is essentially the task . of pattem recognition.The applicationof a new tool, a neuraI network, to solve this task has boon extensively studied in recent years. In Ref. [3] we have proposed a neuraI network fault dictionary. This dictionary bas utilized a MuItiLayer Perceptron with nodal voltages . and source current as inputs, and fauIts as outputs. However the first results were promising, the tumed upside down pyramide structure (number of inputs much less than outputs) of a network caused problems at the learning pbase. As it has boonpointed out by many authors, a narrowing-up or at least uniform structure is very advantageous at both learning and recall phases. Now, we propose a two stages neural network fault dictionary. The first stage is the NNR cIassifier, while the second stage network recognizes a pattern given by aU EucJidean distances !J.Vf, f= l,..,n. Structure of the whole network is shown in Figure 1.