Analog Circuit Fault Diagnosis Method Based On Parameter Transfer Convolutional Neural Network

In recent years,with the increasingly close relationship between electronic power equipment and social industrial production and human daily life,the demand for fault diagnosis of circuit system is higher and higher.Timely and accurate detection and repair of circuit faults can avoid greater economic losses and reduce potential accidents.At present,there are relatively mature automatic diagnosis technologies for digital circuit fault diagnosis.For analog circuits,due to the problems of complex nonlinear feedback,large fault diversity and limited measurable nodes,its fault diagnosis is difficult,and there is still no perfect solution.The rapid development of computer technology and the emergence of deep learning provide new ideas for the research of analog circuit fault diagnosis methods.In the traditional machine learning fault diagnosis methods,complex signal processing and analysis methods are usually used to extract fault features,and shallow classifier is used to complete fault classification.The overall fault diagnosis process is complex.With the increase of circuit scale and fault types,the accuracy of the fault diagnosis mode composed of artificial feature extraction and classifier will decline significantly.In order to improve the above problems,an analog circuit fault diagnosis method based on two-dimensional convolutional neural network is proposed.The fault diagnosis is realized by analyzing the gray images corresponding to the output voltage signal of circuit under test.The specific work contents are as follows:(1)An analog circuit fault diagnosis method based on 2D-CNN.The output voltage sequence of the circuit is transformed into a two-dimensional fault gray image(FGI)through matrix reconstruction and numerical conversion.Build a 2D-CNN model,in which the batch normalization method(BN)is used to regularize the data to alleviate the impact of data distribution offset.The gradient descent process is optimized by adaptive moment estimation(Adam)to improve the gradient dispersion problem and local optimal solution problem when the network layer becomes deeper.The convergence performance of the network is improved by adaptively adjusting the change of learning rate.Taking the FGI data set as the input of 2D-CNN,the convolution layer is used to automatically extract the deep features of the fault,and the fully connection layer is used to complete the fault classification.This method integrates the fault diagnosis steps into the same network and simplifies the fault diagnosis process.Sallen-key bandpass filter circuit and four-opamp biquad high-pass filter circuit are used as the circuits under test.The experimental results show that the proposed method is effective and has strong generalization ability.(2)An analog circuit fault diagnosis method based on parameter transfer.Aiming at the problem that the increase of the scale of circuit data set increases the amount of calculation and modeling cost of 2D-CNN,the idea of transfer learning is introduced,that is,some parameters of the pre-training model are transferred to the fault diagnosis model of circuit to be transferred(T-circuit).The data set of T-circuit is used to train the transfer learning model(TLCNN),and the new network layer parameters are obtained by selecting the appropriate parameter transfer strategy and initial learning rate,so as to minimize the loss function of the transfer learning model.The diagnosis model of Sallen-key bandpass circuit is used as the pre-training model,two-stage four-opamp biquad high-pass filter circuit and a nonlinear rectifier circuit are used as T-circuits.Experimental results show that the proposed method can achieve good fault diagnosis rate and reduce modeling cost in a shorter time under a small data set,and can also be effectively applied in nonlinear circuits.