Research On The Fault Diagnosis Method Of Rolling Bearing Unbalanced Data Based On Time-frequency Sample Generation

As the core component of rotating machinery,rolling bearings play a pivotal role.This will affect the operation of the device and may cause safety accidents in serious cases.Therefore,the fault diagnosis of rolling bearings is a very practical work.Conventional intelligent diagnosis methods require high-level expert knowledge,and their development has been greatly restricted.Deep learning has the characteristics of "end-to-end",which provides a new method for fault diagnosis.However,under real operating conditions,most rolling bearings are in normal working condition,and the probability of failure is very small,resulting in imbalance in the collected data,which will greatly affect the accuracy and stability of the deep learning model.Secondly,deep learning has strong feature extraction capabilities in image processing.The rolling bearing fault diagnosis learning method using twodimensional data input is more suitable for practical use.In view of the above problems,based on convolutional neural networks,combined with time-frequency transformation,generative adversarial networks and other methods,this paper carries out the fault diagnosis of rolling bearings under non-equilibrium data scenarios from the perspectives of optimization model and amplified samples,and the main research contents are summarized as follows:In order to solve the problem of low fault diagnosis due to unbalanced bearing data,a new fault diagnosis method is proposed,which uses time-frequency images as input and adds a modified version of two-dimensional convolutional neural network(WKFL-Improved2DCNN)with higher size convolutional kernels to enhance the accuracy of the model.The original vibration signal is transformed into a time-frequency picture using continuous wavelet transform,and the modified 2-D convolutional neural network uses a large scale first layer convolutional kernel to extract the deep features of the picture,and adds BN layers behind the linear and convolutional layers to adjust the data distribution,and then uses a focal loss function instead of the traditional cross-entropy loss function to offset the effect of unbalanced data distribution on the network.The proposed method can effectively improve the diagnosis of bearing faults with low imbalance rate samples,and has higher Recall and F1-score values.A rolling bearing fault diagnosis method based on Wasserstein Generative Adversarial Net(WGAN)and convolutional neural network is designed from the perspective of augmented samples to address the problem that the model cannot be fully trained due to the high nonequilibrium rate.In order to eliminate the influence of initial phase on the generative adversarial network,the continuous wavelet transform is used to convert the time-frequency spectrum samples into WGAN which can effectively divide the data set into three parts: training set,validation set and test set,and these parts can be implemented by the WGAN method,and more new samples can be generated by comparison training,thus effectively enhancing the diversity of the training set.By analyzing the enhanced training set,a more efficient convolutional neural network can be constructed,and the performance of the system can be tested and evaluated by using these trained models.Experimentally,the model is shown to have powerful generalization ability in dealing with sample imbalance.It is shown that the traditional adversarial network training has some problems,such as unstable gradient disappearance,low sample quality and only one class of samples can be generated at a time.In order to solve these problems,a new rolling bearing fault diagnosis method is proposed,which employs an adversarial network based on Wasserstein distance conditional gradient penalty(CWGAN-GP)and a residual network.By introducing the gradient penalty term,the objective function of the adversarial network is significantly changed,thus replacing the traditional gradient clipping method and making the network more efficient and reliable and integrating the Res Net50 network to identify the fault types.The model is proven to generate high quality samples quickly,which greatly reduces the generation time of fault samples,and the model has excellent robustness and can show good diagnostic results even in the case of extreme data set imbalance.