A Research For Rolling Bearing Fault Diagnosis Based On Deep Learning With Two--dimensional Data

As a key component of rotating machinery,rolling bearing plays a crucial role.Due to the consecutive and high-load operation of the rotating machinery,the rolling bearing will inevitably fail,which may affect the performance of the equipment and lead to accidents.Therefore,it is of great practical significance to study the fault diagnosis technology of the rolling bearing.The traditional intelligent diagnosis method requires extensive expert knowledge and complex feature extraction process,which limits its development in the field of fault diag-nosis,while deep learning has opened up a new way of thinking in the field of fault diag-nosis due to its end-to-end recognition capability.As deep learning has powerful ability of feature extraction in processing two-dimensional data,the better diagnostic results can be obtained when the original vibration signal is processed into two-dimensional data.How-ever,the related research about taking two-dimensional data as input is not thorough,and the report for two-dimensional data in complicated cases analysis is also rare.Therefore,this paper mainly studies the method of deep learning diagnosis with two-dimensional data as input.Taking rolling bearing as the research object,three different deep learning methods based on two-dimensional data training are proposed,which respectively solve the problem of fault diagnosis of rolling bearing when the sample size of training data set is small,the data is unbalanced,and the speed is inconsistent.The main work of this paper is as follows:(1)For the fault diagnosis of rolling bearing under small sample size,transfer learning is introduced into the training process of deep learning model.A diagnosis method based on time-frequency representation and transfer learning is proposed.The proposed method can utilize less training data set while get high diagnosis accuracy,which greatly reduce the sample size for the training of deep network model.(2)Data imbalance is an important problem to be solved in the fault diagnosis.Fo-cused on this problem,a conditional generative adversarial network model based on Wasser-stein distance is designed.The model can generate the specified two-dimensional time-frequency images when the label is given,thus to extend the unbalanced data sets.The proposed method can improve the effect of unbalanced data for diagnosis.(3)Focused on the problem of fault diagnosis under different rotation speed,the sig-nals from two sensors data are fused into two-dimensional matrix,and a capsule network is designed and training.An integrated fault diagnostic scheme is proposed.The proposed scheme can achieve an end-to-end fault diagnosis under different rotational speed,which can eliminate the effect of speed to a certain extent.