Research On Fault Diagnosis Of Analog Circuits Based On Deep Learning

With the development of deep learning technology,fault diagnosis based on deep learning has been widely used.However,the fault knowledge of such diagnosis methods is generally obtained by training fault samples,so the effectiveness of fault features in fault samples has a direct impact on the accuracy and effect of fault diagnosis.At present,fault feature extraction still mainly relies on professionals' self-extraction according to the characteristics of UUT(Unit Under Test).However,as the fault data in the analog circuit is generally continuous response signal,and due to the existence of component parameter tolerance and noise,the feature engineering needs to build a complex mathematical model,and at the same time requires testers with strong professional knowledge to extract more effective fault features.In order to improve the generality of feature engineering and reduce the labor cost of fault diagnosis feature extraction,an automatic fault feature extraction Method is proposed by directly inputting sampling data into convolutional neural network.In this paper,lenet-5,a classical structure of high-efficiency pattern recognition convolutional neural network designed for handwriting recognition,is selected,and we modified the implicit layers,convolution kernel and pooling layers of Lenet-5.The quad-op high-pass filter circuit was selected as the experimental circuit,the voltage response simulation data of each fault mode in the circuit was obtained by using THE AC analysis in Pspice combined with monte Carlo analysis.And then input the voltage response simulation data into the network.after that,we improve the network parameters.Two different improvement schemes were adopted.Scheme 1 reduces the size of the convolution kernel and increases the number of the convolution kernel.Scheme 2 keeps the original size and number of the convolution kernel of Lenet-5.Scheme 1 and scheme 2 both use RELU function to replace the original tanh activation function of Lenet-5.The classification results show that the accuracy of the network structure which improved by scheme 1 is up to 98.3333%.The research shows that the use of convolutional neural network to directly read fault response data can effectively classify faults,not only automatically extract the features in fault data,but also improve the universality of feature engineering,thus greatly reducing the labor cost and complexity of fault feature extraction.