Wavelet Neural Network And The Application On Analog Circuit Diagnosis

Wavelet neural network (WNN) can be divided into two kinds: loose integrated Wavelet Neural Network and inlay model Wavelet Neural Network. WNN has the time-frequency local characteristic and focus-changing characteristic specially owned by wavelet transform, and also has the characteristics of self-learning, self-adapting, robust, fault tolerance and generalization ability which are specially owned by neural networks. As a result, it has wide application foreground in the domain of analog circuit fault diagnosis. Traditional inlay model WNN is generally used to diagnose analog circuit fault using test points'voltage. As a result of adopting local search strategy to adjust parameter, it has disadvantages of tending to slump into local extremums and having comparatively great training error. Traditional loose integrated WNN is generally applied in analog circuit fault diagnosis using signal curve. Influenced by continuity of analog circuit signal, component tolerance and circuit structure, in the process of extracting feature of circuit nodes'voltage curve, the scopes of each fault mode are overlapped. So that problems of nonequilibrium and redundancy exist in the feature samples. Aimed at these problems, this dissertation introduces chaos particle swarm arithmetic and association rule to improve the diagnosis efficiency of WNN in analog circuit fault diagnosis. Author's main work is listed as follows:Firstly, subject background, basic theory and methods involved in analog circuit fault diagnosis are introduced. In this part the dissertation focuses on the analysis of application and problems existed of wavelet neural networks in analog circuit fault diagnosis.Secondly, two integrating manners between wavelet transform and neural network are discussed. And then the theory background, basic principle and arithmetic of these two kinds of WNN are introduced. Thirdly, aimed at the problem of local extremum introduced by gradient descent method in inlay model WNN, chaos particle swarm optimization is proposed to optimize network weights and other parameters, making network training converge faster and jump out of local extremum easier. As to the problem of weakness and infirmness of loose integrated WNN in processing sample, association rule is introduced to pretreat sample, optimize sample's nonequilibrium and redundancy, and improve diagnosis precision. Experiments based on benchmark data indicate that those two kinds of improved WNN get better performance than traditional WNN.Finally, the dissertation discusses the application of improved WNN in analog circuit fault diagnosis software system. The diagnosis result of benchmark circuit indicates that, after optimization by chaos particle swarm optimization, inlay model WNN performs better in astringency and training error. Meanwhile, after introducing association rule into loose integrated WNN, input sample gets optimized. As a result of this, loose integrated WNN gets higher precision and efficiency in fault diagnosis based on voltage curve.