Research On The Inverter Fault Diagnosis Method Based On Compressed Sensing And Convolutional Neural Network

At present,the inverter fault diagnosis technology based on traditional method to extract feature deficiencies and redundant information,which lead to diagnostic accuracy is not high,and big data in the information age brings the problem of signal processing technology is increasing,especially for complex fault inverter,based on the traditional Nyquist sampling theorem of signal sampling method gradually no longer applies.So the compressed sensing(CS)sampling method in combination with convolution neural network(CNN)applied in the large data complex inverter in the study of fault diagnosis,which can effectively solve the traditional method is difficult to deal with large data information storage and transmission pressure,and realize adaptive feature extraction and integrated fault diagnosis technology,to improve high accuracy and intelligent of the inverter complex fault diagnosis technology has important research significance.Firstly,the simulation model of the inverter was built to simulate all the output voltage waveforms of the open-circuit IGBT faults of the inverter with three tubes and below,the numerous normal and fault voltage waveforms of inverters with different voltage levels and load powers were obtained,and select the three-phase output voltage with more characteristic signal quantity as the research signal parameter.Then,the sparse matrix of CS theory is designed.Delay and extension factors are added to the orthogonal Fourier sparse basis to make it become the non-orthogonal extended Fourier sparse basis.The sparse performance and reconstruction performance are verified by Matlab simulation software.Then,the fault and normal output voltage sample set of large-capacity inverter is established by using simulation software,and the CNN model suitable for the voltage signal sample characteristics of the inverter is designed with Tensorflow software,and the fault diagnosis effect of the inverter based on CNN network is obtained through experiments.Experiments show that the CNN network has realized adaptivity and intelligence of feature extraction and fault type recognition in the research and application of inverter fault diagnosis,and the fault diagnosis rate is improved obviously by using the method of DWT combined with BP neural network,but the model training time and sample cost are multiplied.In order to further improve the fault diagnosis,designing an adaptive regularization coefficient of CNN model,the adaptive regularization coefficient convolution neural network(Arc-CNN),will be regularized operation role in each round of iteration process,the stochastic gradient descent values as reference,to do the regularization coefficient of Adaptive adjustment.The feasibility and effectiveness of the method are verified by experiments.Finally,using include expanding Fourier CS signal processing method of sparse matrix of fault and no fault of the three-phase inverter voltage signal compression dimension reduction,some samples with new characteristics and data dimension to input the network,with compression sensing convolution neural network(CS-CNN)model of inverter fault diagnosis experiment.Experiments show that,compared with the original signal CNN and Arc-CNN model,the CS-CNN model appropriately simplifies CNN model structure and parameter requirements,improves the comprehensive deep learning and extraction ability of fault features,accelerates model convergence speed,reduces model training time,and finally improves diagnostic accuracy to 99.82%.