Research On Fault Diagnostic Methods For Analog Circuits Based On Time-frequency Features

Analog circuits are an important part of electronic equipment,and testing and fault diagnosis is a hot research topic at present.However,due to the fact that analog circuits are much susceptible to external noise,there are many problems in components,such as parameter continuity,tolerance and high non-linearity,which make the development of fault diagnosis and testing technology relatively slow,far from meeting the actual demand of electronic industry for high reliability of equipment.Aiming at analog circuits fault diagnosis,based on cross-wavelet time-frequency analysis,the theory of matrix decomposition,image feature analysis,statistical classification methods and generative adversarial networks(GANs),this dissertation focuses on the feature extraction and classification strategies,then dicusses the application of deep learning in analog circuits fault diagnosis.The research contents of the work are described as follows:(1)Feature extraction for analog circuits based on cross wavelet transform(XWT)and matrix decomposition is proposed.By analyzing the time-frequency characteristics of cross-wavelet spectrums,singular value decomposition(SVD)and variational Bayesian matrix decomposition(VBMF)are introduced into fault feature extraction,then two methods for obtaining fault features are proposed.First,the circuit output response signals of normal states and fault states are sampled,and the response signals are transformed into two-dimensional spectrums by using XWT.Then the spectrum matrices are decomposed by utilizing SVD and VBMF techniques.Finally,the information entropies of SVD singular value sequences and the statistical parameters' of VBMF singular value sequences in all fault states are calculated to construct the fault feature vectors.The simulation results show that the fault features based on XWT and matrix decomposition have high discrimination.(2)Feature extraction for analog circuits based on XWT and image information analysis is proposed.By analyzing the geometric moment characteristics and texture structure information of cross-wavelet spectrum images,Krawtchouk moment and local binary mode(LBP)are introduced into fault feature extraction of analog circuits to construct fault features.First,the cross-wavelet spectrums of the output signals in all fault states are obtained;subsequently,the weighted Krawtchouk moment and local optimal direction pattern(LOOP)are applied to process the spectrum images,and finally the fault eigenvectors are generated.Among them,the weighted Krawtchouk moment method is used to solve the numerical instability in the traditional Krawtchouk moment,while LOOP is designed to solve the shortcoming of the traditional LBP method which is too over-dependence on direction.The simulation results demonstrate that the method based on XWT and image information analysis achieves better performances in fault feature extraction.(3)Fault location methods for analog circuits based on statistical classification models are proposed.In traditional fault classification,neural network is used as a fault classifier.The method based on empirical risk minimization is easy to fall into local optimum.Considering the excellent performance of statistical classification models in pattern recognition,fault diagnosis methods for analog circuits based on feature weighted kernel linear discriminant analysis(FWKFDA),support vector machine(SVM)and vector-valued regularized kernel function approximation(VVRKFA)are proposed.In light of the good parameter optimization ability of the flower pollination algorithm(FPA)and the quantum particle swarm optimization algorithm(QPSO),the FPA is used to optimize the SVM classifier and QPSO is utilized to optimize the VVRKFA classifier.The simulation results prove that the three methods can achieve high fault diagnosis accuracies.(4)The fault diagnosis method for analog circuits based on generative adversarial networks(GANs)is proposed.In traditional fault diagnosis methods,when dealing with different diagnostic tasks,different fault feature methods and classification strategies need to be selected manually.On the contrary,deep learning method can effectively solve the problems above.With its powerful non-linear mapping ability,feature extraction and classification can be realized in different network layers.In this thesis,the basic principle of GAN is described.Deep convolution neural network,sigmoid and softmax classifier are utilized to further improve its ability in feature extraction and classification.Meanwhile,wavelet cross-spectrum and coherence spectrum data are used as input vectors to reduce the burden of GAN.The simulation results indicate that this method can effectively fulfill the task of fault diagnosis.