Research On Intelligent Method Of Analog Circuit Fault Diagnosis Based On Machine Learning

In recent years,electronic circuits have been widely used in communication,industrial control,medical equipment,household electronic products,aerospace equipment and military industry.With the increasing integration and complexity of electronic circuits,it requires higher and higher reliability of electronic equipment.At present,the widely used circuit system is based on the digital-analog hybrid circuit.Digital circuit diagnosis methods have been widely used in digital and analog hybrid circuits,and the diagnostic methods have been developed.Because the analog circuit fault diagnosis model is relatively complex and the number of measurable nodes is also very limited,the analog circuit fault diagnosis method often appears errors in the actual diagnosis,which affects the diagnosis rate.At the same time,the modern analog circuit fault diagnosis presents the intelligent diagnosis system,the integration of the diagnosis system and the diagnosis method.If through the traditional testing methods and theories,it is unable to meet the needs of modern.This is because there is no general expression model in the traditional analog circuit fault diagnosis.Therefore,this paper in the process of the study of analog circuit fault diagnosis methods were analyzed,can adapt to the current analog circuit development,its research has important significance and relatively large practical value.The theoretical basis of this paper is deep Learning,Extreme Learning Machine(ELM)and Auto Encoder(AE).Nonlinear circuit is the test object,and a fault diagnosis model of analog circuit based on representation Learning is proposed.By extracting the output signal of the circuit,The fault diagnosis of analog circuit is analyzed and studied.The main research results are as follows:1.For fault diagnosis of analog circuit fault information characteristic,high noise and fault diagnosis of a slower time problem,this paper proposes a model of analog circuit fault diagnosis based on characterization of learning,the representation is the basic unit of study ELM-AE,its main work is designed by the transformation of the data to obtain meaningful data characteristics.SELM-AE and DELM-AE network models are built through in-depth study of relevant theories of ELM and AE,which combines the advantages of both ELM and AE.ELM has high generalization ability and fast learning ability,while AE can reconstruct input signals.Therefore,SELM-AE and DELM-AE can achieve dimensional compression and sparse representation of data.Based on the characteristics of SELM-AE model,H-SELM model was formed by cascading SELM-AE.H-SELM can make the output infinitely close to the original input by reducing the reconstruction error as much as possible,and the output of SELM-AE is the same as the original input information to prevent the loss of important features of the original information.After each characterization learning of SELM-AE,more compact features of the input can be learned.It can realize the feature representation of different dimensions of data to meet the requirements of processing high-dimensional data.2.Since SELM-AE has only one hidden layer,it is possible that the data features learned are not compact enough.Therefore,a hidden layer is added on this basis to build a DELM-AE model.The architectural unit of this model is a deep extreme learning machine with double random hidden layers.Two random hidden layers are used for encoding features,and one output layer is used for decoding features.First,H-DELM randomly maps the lower dimensional data to the higher dimensional representation space to obtain the extended dimensional data,and then randomly transforms the features from the higher dimensional space to the lower dimensional space to obtain the compressed features,namely the learned feature code.The H-DELM model was constructed by stacking DELM-AE in a hierarchical structure,because DELM-AE could represent features and the output was the same as the original input information.As a result,H-DELM can replicate as much of the original input data as possible,thereby learning more expressive and compact features.However,in the process of DELM-AE representation learning,the first hidden layer has a large number of nodes,which ensures a strong capability of feature mapping,while the second hidden layer reduces the number of nodes and realizes the compact feature.Therefore,as the basic unit of H-DELM,it can realize more complete feature representation to save more information,and then achieve the rapidity of training.The ultimate goal of H-DELM is to seek faster training speed,more compact representation learning and higher classification performance.3.For the characterization of learning analog circuit fault diagnosis model is suitable for analog circuit suitability analysis,the characterization of the emergence of learning has broken the traditional analog circuit fault diagnosis method,the feature extraction and classifier,simplifies the traditional methods,the characterization of learning can deal with high dimensional failure data,so the failure characteristics of high-dimensional data as input,the characterization of learning framework Through data transformation to obtain more meaningful representational data for classification.However,the fault data extracted from the circuit of the four-op amp double-quadratic high-pass filter and the two-op amp double-quadratic low-pass filter have higher dimensions,which is conducive to the realization of the nonlinear fitting of the representation of learning fault diagnosis and learning network.4.Since representation learning can meet the requirements of processing high-dimensional data and learning more compact fault features,the applicability of Quadruple operational amplifier double-order high-pass circuit and more complex Two-stage four-op-amp biquad lowpass filter circuit to the learning model proposed in this paper is verified.The general applicability of the nonlinear rectifier circuit is analyzed and verified.Experimental results show the feasibility of the proposed algorithm in analog circuit fault diagnosis,and compared with other algorithms,it is proved that the proposed model has higher robustness,the classification speed can reach about 1s,the fault classification accuracy can reach more than 99%,and even the single fault diagnosis rate of some circuits can reach 100%.