Motor Bearing Fault Diagnosis Of Variable Working Condition Based On Adversarial Transfer Learning

In the actual industrial production,bearing and other rotating machinery which can work in health condition is especially important.The working condition of the bearings could not invariable.With the alteration of the actual situation on site,the bearings usually work in different working conditions.Because the data distribution of bearing data in different working conditions is different,it means a critical challenge for the fault diagnosis method based on deep learning.As a method which can achieve classify and predict tasks under different working conditions,Transfer Learning has widely used in fault diagnosis with variable working conditions.The existing Transfer Learning models have the following problems:(1)The accuracy of the source domain is extremely high during the training phase,while the accuracy of the target domain is very low,and as the process of training deepens,the model will gradually collapse and thus shift to an under-fitting state.(2)Some Transfer learning models suffer from slow training speed and even difficulty in convergence in some tasks.(3)In the target domain,there exists the problem of categories confusion.From the visualization of the diagnosis results,the decision boundaries between different categories are not clear enough,and there occurs even the phenomenon of the confusion of a large number of data points.For the first problem,a classification module consisting of an Up-sampling Layer,a Convolution Layer,a Fully Connected Layer and Self-Attention Mechanism is proposed.The processing of the data has an impact on the data information,and the Up-sampling layer is used to supple information to the data after feature extraction.This method is used to solve the problem caused by the large difference in accuracy between the target and source domains during the training phase.To address the second problem,the Self-Attention Mechanism introduced in the classification module is used to solve the problem of slow and difficult convergence of the Transfer Learning models in certain tasks by taking advantage of its stronger ability to capture temporal signals.For the third problem,the Adversarial Mechanism is introduced into the model and the domain discrimination module is constructed.Among other things,the module is designed to enhance the metric effect of data alignment,thus enhancing the supervision of the learning process and the alignment effect of different data during the iterations.Firstly,the vibration signals from different working conditions are extracted with features.The source domain data after feature extraction are fed into the classification module to obtain the classification Loss and the Loss based on the attention score;then the features from source and target domain data are fed into the domain discrimination module to obtain the difference of data distribution;these Losses are used to the iteration of the parameters during the training process.Finally,based on the trained parameters,the test set is fed into the saved model to obtain fault diagnosis results.The experiments demonstrate that the MSATL model proposed in this paper can outperform the Adversarial Transfer Learning model with single Adversarial Mechanism,and the ATLSAM framework based on MSATL outperforms other methods for variable condition fault diagnosis,with better accuracy and fault diagnosis results.