Research On Analog Circuit Fault Detection Technology Based On Deep Learning Algorithm

In the field of modern electronic circuit manufacturing,circuit failures caused by electronic component defects may cause safety problems for electronic products.Most of the faults in the circuit system occur in the analog circuit part,therefore,scholars from all over the world have studied on the analog circuit malfunction test technology in the more depth.In the field of analog circuit malfunction test,deep learning technology is one of the emerging research directions.However,due to the tolerance and non-linearity of the component parameters in the analog circuit,it is very difficult to extract the eigenvalues of the deep learning model,and need to choose a suitable diagnostic model to fault test.For the goal of fault test,this paper designs a deep learning feature parameter extraction method,and selects two classification models to distinguish faults,designed the experiment and verified the analog circuit fault detection scheme based on circuit signal detection technology and deep learning technology.The main content of this article is as follows:1.The selection of the eigenvalues of deep learning models.Aiming at the selection of the eigenvalues of the deep learning model,according to the principle of low-frequency noise and noise superposition theorem in the analog circuit,this paper selected the low-frequency noise power spectrum density as the fault characteristic information.Designed 10 single failure modes of active low-pass filter,used Pspice software to simulation circuit,calculated the equivalent output noise of the circuit.The experiment show that 10 kinds of the equivalent output noise of single fault is different,proved that circuit faults can be detected by low frequency noise power spectral density.2.The proposition of the improved average period cross-spectrum method.In order to collect the noise power spectrum density,this paper analyzes the characteristic of the average periodogram method and the Welch modified periodogram method.According to the principle of data overlap in the Welch modified periodogram method and the resources of the FPGA platform,proposed an improved average cycle cross-spectrum method.This method uses low-pass filter,low-noise amplifier,ADC sampling and FPGA hardware platform to collect the noise spectral density under different circuit status as characteristic value.Set up a comparative experiment.Compared with the average period cross-spectrum method,the variance of the collected eigenvalues is reduced by 3.8686 times,improved the variance,improved the quality of noise spectrum estimation.3.Analog circuit fault diagnosis based on multiple Logistic regression algorithm and K neighboring algorithm.Aiming at the problem of fault diagnosis this paper chose the multivariate Logistic regression algorithm and the K neighbor algorithm to construct a classification model,and used the power spectrum density of the fault noise as the characteristic value,achieved the purpose of fault detection by training the classification model.According to the features of the noise spectrum obeys the normal distribution,proposed a sample preprocessing method based on the normal distribution theorem and sample expectation filling method,improved the quality of the sample set by cleaning abnormal samples.Setting up the contrast experiment,proved that after data preprocessing,the variance of the sample set is reduced by 2.3648 times,improved the discreteness of the sample set effectively.After experimental testing,The experimental results show that the correct rate of fault diagnosis is 92.8%,the diagnosis time was 2S,and the accuracy rate increased as the number of sample sets increased.