Research On Fault Diagnosis Method For Analog Circuits With Multi-scale Convolutional Neural Network

The operating state of analog circuits has a significant impact on the reliability and stability of electronic circuit systems and equipment.The characteristics of analog circuit components,including tolerance effects,high degree of nonlinearity,and limited measurable points,pose many limitations on the research of fault diagnosis methods.The differences in signal characteristics between the early fault state and normal operating state of an analog circuit are small,and even signal mixing occurs,resulting in long diagnosis times and high costs.In order to overcome the research difficulties caused by the characteristics of analog circuit components and improve the accuracy of fault diagnosis,this paper introduces the convolutional neural network algorithm from the field of deep learning into the fault diagnosis technology of analog circuits,and studies the soft fault diagnosis method of analog circuits from the perspective of multi-scale feature information of fault signals and multiscale network structure.The main research contents are as follows:(1)To enhance the fault diagnosis accuracy of analog circuits by extracting richer and more effective fault features from limited sample data,a fault diagnosis method based on MS-CNN(Multi-scale Convolutional Neural Network,MS-CNN)is proposed.The method transforms the one-dimensional timing signal of the circuit into a two-dimensional circuit spectrum map using short-time Fourier transform.The MS-CNN model is built using Tensorflow,and the 1D timing signals and 2D circuit spectrum images are fed into the 1DCNN and 2D-CNN of the MS-CNN to extract rich fault features of the circuit.The timedomain and frequency-domain feature information is then stitched together and delivered to the Softmax layer of the network for fault classification.The model is a two-branch convolutional neural network structure that takes into account the different feature extraction capabilities of different scales of convolutional neural networks for one-dimensional signals and two-dimensional images,effectively extracting the feature information of fault data at both scales,ensuring the richness and effectiveness of the extracted fault features,and facilitating further completion of analog circuit fault diagnosis.The simulation results show that MS-CNN has good feature extraction and fault diagnosis ability for Sallen-Key bandpass filter circuit and four-operator second-order high-pass filter circuit,obtaining 100.00%and 99.62.% diagnosis accuracy,which proves that MS-CNN can obtain richer and more effective feature information of the circuit,thus making the circuit more distinguishable among various types of faults The MS-CNN can achieve accurate diagnosis of circuit faults.(2)To tackle the challenge of diagnosing faults in more complex analog circuits,this paper proposes a fault diagnosis method based on residual neural networks.The approach leverages multi-scale feature information from overlapping fault signals while avoiding the issue of gradient disappearance associated with deep networks.Specifically,the method employs short-time Fourier transform to obtain two-dimensional spectral maps of various fault signals,and utilizes the concept of residual networks to extract feature maps from different layers in a "shortcut" manner and splice them together.This approach ensures the extraction of multi-scale signal features and mitigates the problem of gradient disappearance that arises in deeper networks.The experimental results show that the analog circuit fault diagnosis method based on Res Net12 has good feature extraction and fault identification capability for fault signals of more complex circuits,and shows more than 99% fault diagnosis accuracy for the second-stage quad op-amp second-order low-pass filter circuit and Leap-frog filter circuit.