Generative Adversarial Network Based Methods For Rolling Bearing Fault Diagnosis

As electromechanical equipment evolves toward diversification,complication and intellectualization,the fault diagnosis and early warning of equipment have been paid more attention and studied.The progress of computer and big data technology has greatly promoted the development of fault diagnosis algorithm based on deep learning.However,due to the difficulty of obtaining massive labeled fault samples and the different distribution of training data and testing data caused by different working conditions and data acquisition positions,there are two kinds of problems in the application of deep learning in the field of fault diagnosis:(1)due to insufficient sample size and unbalanced data,the model is undertrained;(2)due to different working conditions and collection positions,the generalization ability of model is insufficient and the diagnosis accuracy is reduced.Taking the rolling bearing as the research object,this thesis explores the feasibility of generative adversarial network in solving the above fault diagnosis problems,and puts forward the bearing fault diagnosis method based on generative adversarial network.The completed work is as follows:(1)This thesis introduces the research background and status quo of generative adversarial network,convolutional neural network and transfer learning theory.Emphatically,the structure and principle of network layer used in this thesis is analyzed,as well as the existing problems of generative adversarial network.Moreover,the unified representation form of transfer learning is established.(2)An improved auxiliary classifier generative adversarial network with Wasserstein distance(I-WACGAN)is proposed.By means of sample generation and data enhancement,the fault diagnosis under the condition of insufficient sample size and unbalanced dataset is solved.By combining the advantages of ACGAN and WGAN-GP algorithms,optimizing the discriminator and adjusting the range of gradient penalty,IWACGAN can produce high-quality generated samples.In addition,the evaluation system of generative model is further proposed,and the perfect evaluation index of generated samples is established.Finally,the effectiveness of the proposed algorithm is verified under the condition of insufficient sample size and unbalanced dataset.(3)A Domain-Adversarial neural network with threshold-control transfer regularization parameters(TC-DANN)is proposed to solve the problem of fault diagnosis in the case of cross working conditions and cross collection positions.TC-DANN forms the training mode of adversarial training between feature extractor and domain classifier,and label predictor is used for auxiliary classification.By introducing the strategy of adaptive transfer regularization parameters and dynamically adjusting the proportion of classification loss and loss of domain discriminant in parameter update,the ability of feature extractor to find the relationship between source domain and target domain in feature space is strengthened,and TC-DANN shows better domain adaptation performance.Finally,the effectiveness of the proposed algorithm is verified in two different transfer scenarios: cross working conditions and cross collection positions.The algorithms proposed in this thesis are verified by the bearing fault dataset of Case Western Reserve University.Under different data settings,the fault diagnosis accuracy of the algorithm in different scenarios is verified.Compared with other algorithms.The proposed algorithm I-WACGAN and TC-DANN in this thesis show better performance.Figures 37,Tables 25,References 67.