Research On Deep Learning Method For Analog Circuit Fault Diagnosis

The rapid development of artificial intelligence,big data,and other technologies has increased the complexity of electronic systems.Circuit health status detection has become the fundamental guarantee for the system’s safe operation and reasonable maintenance.Due to the tolerance characteristics of components,analog circuits are prone to soft faults,which brings significant challenges to system health detection.Therefore,developing efficient analog circuit fault diagnosis technology has become an inevitable choice for electronic system health detection.Traditional fault diagnosis methods are mainly based on machine learning algorithms,which require manual feature extraction,and obtaining in-depth feature information of circuit signals is not easy.The deep learning algorithm has the unique advantage of automatically completing feature extraction and classification.This paper introduces the deep learning algorithm into the analog circuit fault diagnosis technology,and the fault diagnosis method based on deep learning is explored.The diagnosis of double fault states is studied based on single fault diagnosis.It can further improve the efficiency of fault diagnosis.This paper takes the analog circuit fault as the research object and combines the deep learning algorithm to study the analog fault diagnosis.The main research contents are as follows:1.Aiming at the problem that the fault diagnosis method based on machine learning algorithm relies heavily on feature extraction technology and low accuracy of double fault diagnosis,a one-dimensional convolutional neural network(1D-CNN)fault diagnosis method,the advantages of this method in the fault diagnosis of analog circuits are:(1)the convolution layer automatically extracts data features,(2)the batch normalization algorithm solves the problem of covariate shift within the network,(3)96 A convolution kernel of size1 replaces the fully connected layer for global feature processing.(4)The Adam gradient optimization algorithm improves the convergence speed of the loss function.The above four advantages improve the diagnosis method’s learning efficiency and fault identification ability.Based on the above advantages,this paper selects the Sallen-Key band-pass filter circuit for Monte Carlo simulation,establishes nine kinds of single-fault data sets and 24 double-fault data sets,and tests and analyzes the proposed diagnosis method.The single-fault diagnosis accuracy rate reaches 99.44%,and the double-fault diagnosis accuracy rate can reach97.22%.The test results show that the 1D-CNN diagnostic method has a short training time and high diagnostic efficiency.However,1D-CNN cannot obtain the time information of fault signals,and the double fault diagnosis rate is low.2.To solve the problem that the one-dimensional convolutional network method cannot obtain the time information of fault signals and the double fault diagnosis rate is low,a onedimensional convolutional long short-term memory(1D-CLSTM)neural network fault diagnosis method is studied and proposed.The method uses 1D-CNN to extract features to improve learning efficiency,uses LSTM to obtain temporal information,and further extracts hid feature information.Finally,the Sallen-Key circuit data set is used for verification and analysis.The new model has a diagnostic accuracy of 99.63% on the single-fault data set and 98.56% on the double-fault data set.The verification results show that 1D-CLSTM can better obtain the hidden feature information of fault signals,and the double fault diagnosis rate is higher than that of 1D-CNN.