Research On Ensemble Fault Diagnosis Method Based On Generative Adversarial Network

With the continuous expansion of the information intelligence scale,the complexity of the instruments and equipment we are exposed to is increasing day by day,so the requirement of fault diagnosis technology is also getting higher and higher.As one of the important parts of these instruments,the health of motor bearing determines whether the equipment can run normally.At present,fault diagnosis methods based on deep learning are widely used in engineering practice,but most of these methods are directly supervised learning of onedimensional time-domain vibration signals with insignificant features,which requires a lot of energy to mark the data set and the diagnosis results are not ideal.To obtain higher accuracy of fault diagnosis in the case of scarcity of marker data,this thesis proposes a new method of integrated fault diagnosis based on 2D-BN-SGAN.Firstly,the onedimensional vibration signal of the bearing is converted into a 2D grayscale image,which is used as the input for the semi-supervised generative adversarial network SGAN.Secondly,to prevent the over-fitting and improve network performance,batch normalization layer and dropout layer are introduced into SGAN network.Then,softmax is used as the output layer of discriminator network to output fault categories and optimize and update network parameters under different label proportions.Finally,the trained discriminator network is used for bearing fault diagnosis.In the actual industrial process,when the motor operates in a complex environment,it will be accompanied by a large amount of noise interference,and these noise signals will be converted into the gray scale and brightness of the image,affecting the feature extraction.In addition,when solving different practical application problems,different normalization operations should be designed in principle.However,if each normalization layer is designed manually,a lot of experiments are needed,which consumes energy and time.To solve the above problems,this thesis proposes a set fault diagnosis method based on de-noising 2D image and adaptive normalized semi-supervised generative adversarial network: VMD-2D-SN-SGAN.This method firstly converts one-dimensional vibration signal into 2D image after noise reduction by VMD as the input of SGAN network.Secondly,to improve the normalization ability of networks,the batch normalization is replaced to switchable normalization(SN)in SGAN networks.Then,optimize and update network parameters under different label proportions.Finally,the trained discriminator network is used for bearing fault diagnosis.To verify the effectiveness and generalization ability of the proposed method,this thesis selects three sets of motor bearing data sets from Western Reserve University(CWRU),Southeast University(ML)and Xi ’an Jiaotong University(XJTU-SY)for experimental research.In the training of SGAN network,the parameters of generator G are fixed first,and the random gradient is promoted by Adam algorithm to optimize the weight parameters of discriminator D.Then,the parameters of discriminator D are fixed,the output of the full connection layer is selected as the feature of the middle layer,and Adam algorithm is adopted to reduce the random gradient and update the weight parameters of generator G until the number of iterations is reached.Experimental results show that the proposed 2D-BN-SGAN fault diagnosis method achieves semi-supervised learning,is suitable for running in large numbers of times,and has high fault identification accuracy.The VMD-2D-SN-SGAN fault diagnosis method has strong normalization ability and anti-interference ability,and can obtain high fault diagnosis accuracy under any batch size,which significantly improves the generalization ability and robustness of the model.