Research On Fault Detection And Diagnosis Methods For Analog Circuits Based On Deep Learning

As the critical functional units of electronic devices,analog circuits are extensively used in many fields such as aerospace,military,communication and industrial control.The reliability of analog circuits directly determines the safety level of electronic devices.Due to the harsh operating environment and increasing use intensity of electronic devices,the parameter values of components in analog circuits gradually deviate from the allowable tolerance range under environmental stress resulting in soft faults.The degradation of component performance is low in the soft fault stage.If not handled in a timely manner,analog circuits will suffer from the occurrence of hard faults including short circuits or broken circuits,which can cause serious failures or complete damage to electronic devices.To address the soft faults of analog circuits caused by component performance degradation,this dissertation conducts effective detection and accurate identification throughout the life cycle for the problems of nonlinear structure,faint fault features,environmental noise,and weak dataset scenarios that exist in the actual operating environment,so as to establish the foundation for the preventive maintenance and discretionary repair of electronic devices.The main research work of this dissertation is as follows.1.A fault detection method based on the vector quantization sparse autoencoder(VQSAE)and the comprehensive statistic is proposed to address the problem that nonlinear structures of analog circuits makes it difficult to achieve accurate status representation and precise identification by traditional fault detection methods.A vector quantization mechanism with self-normalized properties is constructed to extract the nonlinear structural features of analog circuits stably and efficiently.A construction rule based on local Mahalanobis distance and K nearest neighbors is designed to calculate the K Mahalanobis nearest neighbor(KMN)metric with the features extracted by the VQSAE model.A comprehensive statistic is constructed by fusing the VQSAE loss(VL)metric,which represents the residual information,and the KMN metric,which represents the system change information,to comprehensively and accurately reflect the operating status of analog circuits.The generalized extreme value(GEV)distribution models of the comprehensive statistic are established to determine the fault detection thresholds,thereby achieving accurate status identification for analog circuits.2.An end-to-end incipient fault diagnosis method based on the complex convolutional self-attention autoencoder(CCSAE)is proposed to address the problem of poor discrimination for incipient faults in analog circuits due to the inherent tolerance of components and the faint nature of incipient fault features.A backbone network based on complex convolutional autoencoder(CCAE)is designed to explore the effective features reflecting the amplitude information and phase information of the analog circuit responses.A feature enhancement strategy is constructed to to cope with the faint nature of incipient faults by strengthening relevance structure information in features.A step-by-step training mechanism covering feature training and classification training is designed for CCSAE model,where the critical operation is the construction of supervised contrast loss(SCL)to disperse dissimilar features while aggregating similar features,thus improving the distinction of incipient faults in analog circuits.3.A fault diagnosis method based on denoise sparse deep autoencoder and generalized discriminant analysis(DSDAE-GDA)is proposed to address the problem that the signal data reflecting the operating status of analog circuits in the actual operating environment contains a large number of noise components leading to the annihilation of fault features.A denoise mechanism is designed by incorporating corrupted noise in the input layer to enhance the ability to extract effective features from signals contaminated by noise.A sparse strategy is constructed to extract the consistency information of signal data under different noise intensities and reduce the redundant information interference.GDA is employed for supervised compression and dimensionality reduction of the features,which saves computational cost while improving the feature clustering properties.The sine cosine algorithm(SCA)with global optimization capability is introduced to find the optimal hyperparameters of the generalized multiple kernel learning support vector machine(GMKL-SVM)classifier to achieve accurate fault identification for analog circuits under environmental noise.4.To address the problem of low accuracy of fault identification for analog circuits due to insufficient sample supervision information in weak dataset scenarios,cross-domain adversarial transfer learning is introduced to reduce the feature distribution discrepancies among signal data with various fault degrees,and a cross-domain fault diagnosis method based on energy adversarial transfer network(EBATN)is proposed.A spatio-temporal feature extraction mechanism is constructed by fusing spatial convolution and temporal recurrent units to achieve spatio-temporal feature exploration in the source and target domains.A target unsupervised loss is constructed in accordance with the entropy minimization principle to improve the recognition of target domain sample data by the classifier.The maximum mean discrepancy(MMD)metric is adopted to achieve feature distribution alignment between the source and target domains.An energy discriminator is developed to efficiently drive the adversarial transfer learning of the whole model.By introducing a flexible energy function instead of explicit probability calculation,the training efficiency and stability of the model are enhanced while strengthening the domain discriminatory ability,thus improving the fault diagnosis performance of analog circuits in weak dataset scenarios.The standard analog circuits in the benchmark circuit set and the real analog circuit in the water jet propulsion device are selected as the simulation and physical experimental objects of this dissertation.The experimental results indicate that the fault detection and diagnosis methods proposed in this dissertation can effectively monitor the operating status of analog circuits,and have the feature extraction capability to decipher samples with strong noise interference,the fault classification capability to identify samples with faint features,and the cross-domain recognition capability to deal with weak dataset scenarios,so as to achieve prompt and accurate fault identification for analog circuits in the whole life cycle of practical application scenarios.The research on this dissertation can effectively reduce the hazards caused by the analog circuit faults and reduce the maintenance cost of electronic devices,which has significant application values and potential economic benefits.