Research On Fault Diagnosis Of Power Electronic Circuit Based On Deep Learning

With the development of the technology of power electronics,the requirements of stable and reliable operation have been increasingly promoted for the new power electronics devices.The power electronic converters act an important role in the power converting and controlling,thus the entire system will be seriously damaged or even shut down in case of failures occurring.Therefore,the study of rapid intelligent diagnosis and precise fault localization of power electronics converter circuits is a critical task in theoretical and practical fields.The traditional fault diagnosis methods contain two main stages,i.e.,feature extraction and fault identification.The conventional feature extraction methods include several approaches,e.g.the fast Fourier transform(FFT)and principal component analysis(PCA),where FFT is based on the signal transform from time-domain to frequency domain,and PCA is an optimal linear mapping based on minimizing mean squared error ignoring the class attributes.However,there are not only the influences of manually feature extracting,but the ignored components may contain the imperative information to distinguish the faults.In addition,the traditional neural network used for fault identification is seriously unsuitable in the case of confusion or absence of fault labels,noise interference,etc.To overcome the problems mentioned above,an stacked auto-encoder based deep neural network has been proposed in this thesis for the fault diagnosis of power electronics circuits.The auto-encoder can automatically learn and extract the intrinsic features of the original data from the unlabeled data set.The SAE-DNN can fuse the automatic feature extraction and fault identification,and meanwhile realize the layer-by-layer greedy extraction and fault identification from the deep hidden features of high-dimensional data.In this thesis,the traditional shallow neural network is briefly introduced firstly,the concept,theory and basic training process as well as the common models of the deep neural network are described,and then the proposed SAE-DNN approach is addressed in detail.To study the fault of power rectifier,the fault models of the three-phase fully-controlled and uncontrolled rectifier circuit are established.The failure analysis and classification are presented to confirm the position of information acquisition and the coding strategy of the faults,and the fault generation modules are designed for the specific faults.Finally,in order to verify the performances of accuracy and anti-disturbing of the fault diagnosis method,the proposed SAE-DNN based method is compared with the traditional FFT-BPNN and PCA-BPNN methods in the various anti-noise performance experiments.It is obvious that the SAE-DNN based method shows better characteristic in the similar fault aggregation and dissimilar fault alienation,and the fault diagnosis results have higher and more stable evaluation accuracy.In summary,the method proposed in this thesis has better performances in the feature extraction,identification and anti-disturbing,which effectively improves the traditional fault diagnosis of power electronics.