Fault Diagnosis Of Analog Circuits Based On Neural Networks

Comparing with Digital Circuits, the fault diagnosis of analog circuits is difficult due to the tolerance of elements, the non-liner of circuits and the difficulty of modeling. It is a significant work to locate the fault elements and find what happens to it. This work makes us do the sustaining works easily not only during the running time of circuits but also when they are just being produced.Through almost 40 years, many researchers bend themselves to fault diagnosis of analog circuit and produce a large amount of production. Along with the theories, this area is in progress now into applications to improve productivity, and a practicable flow of diagnosis system is need at the level of industry. Besides, the widely used Artificial Neural Network with single network in this area can not provide the better generalization performance for the more and more complex circuits.In this thesis, Artificial Neural Networks are used to fault diagnosis because they can solve the problems well due to the learning ability. A neural network based fault diagnosis system is proposed in this thesis from the practice in projects, and then the diagnosis performance is improved using ensemble neural networks. The following problems are solved in this thesis:1. A practicable diagnosis system with flow is proposed based on neural networks. Each of the components are discussed in detail: the selection of testing circuit, the selection of faults set, solving tolerance using Monte Carlo simulation method, feature exaction based on Primary Component Analysis and the design of neural network. Then a systemic diagnosis flow is proposed based on the components listed above, and neural networks can be used in analog circuit fault diagnosis with this flow in industrialized automatic diagnosis.2. Ensemble neural networks are used to enhance diagnosis performance. Theories are discussed firstly that the error of ensemble classifier can be decreased by comparing to single neural network, and then an experiment is processed to prove it. The widely used Bagging and AdaBoost ensemble methods are used to improve the classification performance in the diagnosis flow proposed above, and the experiment results indicate the feasibility of this idea. At meantime, the ensemble neural networks can make the design easier because we do not need to adjust the parameters and network structure ceaselessly during the training phase of neural networks.