The Research On Fault Diagnosis Of Analog Circuits Based On Ridgelet Neural Network

Fault diagnosis of analog circuits is faced with new and challenged tasks with the rapid development of modern electronic technology and increasing need of industrial production. Currently, the representative neural networks as a way of modern intelligent information processing technologies present a new and effective road to fault diagnosis of analog circuits, and attract the attention of many of researchers and practitioners.In this paper, in terms of kernel principal analysis (KPCA) and ridgelet theory, the neural-network fault diagnostic system based on novel feature extraction method and a new fault diagnostic system of ridgelet neural network are studied in depth. The main contributions of this paper include as follows:(1) A neural-network analog fault diagnosis method based on KPCA with maximal class separability is studied in detail. The novel KPCA combines the conventional KPCA with maximal class separability method in order to form an improved KPCA with maximal class separability which leads to a small within-class scatter and a large between-class scatter of fault data set. Compared to PCA and KPCA, this method can not only extract the most effective and a minimal number of features, but also result in a neural network classifier with minimal structure and higher rate of correct classification.(2) The application of ridgelet neural network to fault diagnosis of analog circuit is studied and the detailed training algorithm based on steepest gradient descent method is given. The ridgelet neural network is constructed by substituting ridglelet function into the hidden activation function of BP network. The fault diagnosis system of ridge let neural network has smaller structure, more rapid convergence rate and higher accuracy of fault diagnosis than those of BP network and wavelet neural network.(3) The fault feature extraction method based on wavelet fractal analysis and the selection method of neural network structure based on PCA are presented. Wavelet fractal analysis can extract the most discriminant features of fault signal, which allows the subsequent classifier to implement efficient fault diagnosis. The selection method of neural network structure based on PCA requires only algebra operations and has the advantages of a small number of computations and simple implementation. The methods reduce the number of hidden ridgelet neurons and the operation time and computational amount of network, and improve the diagnostic performance.(4) A linear ridgelet neural network and the associated training algorithm are studied. The linear ridgelet neural network is an extension of ridgelet network by adding linear terms. Compared with ridgelet network and wavelet network, the linear ridgelet network has more rapid convergence rate and less amount of computations and higher rate of correct classification.The simulation results show the correctness and effectiveness of the methods presented in this paper.