Application Of Deep Auto-encoder In Bearing Fault Diagnosis

Rotating machinery is developing with the trend of precision,high speed,high efficiency and integration.Bearings are an important part of rotating machinery.The running state of bearings directly affects the operation of the entire machine.Therefore,it is necessary to study the fault diagnosis of bearings.With the increase of the number of bearing monitoring and the increase of sampling frequency,the health condition monitoring of bearings has entered the "big data" era.Traditional expert diagnostic models are difficult and inefficient to process big data,and cannot adapt to this era.Deep learning models can automatically learn the internal characteristics of the data to achieve accurate classification results.The current deep network model can obtain high diagnostic accuracy for the condition of balanced samples,but the classification effect of unbalanced samples is worse.Based on Deep auto-encoder(DAE),this paper studied the data preprocessing of bearing vibration signal,feature extraction and fault classification algorithm under the condition of unbalanced data.The main research contents are as follows:(1)Aiming at the noise in bearing data,an improved wavelet threshold function combined with artificial fish swarm algorithm(AFSA)is proposed.The superiority and practicability of this method are verified by the non-stationary test signal and the bearing data set of Case Western Reserve University(CWRU).From the final results of noise reduction,this method is superior to other existing methods.The SNR of this method is 13%?16% higher than other methods.The root mean square error is 10%?41% lower than other methods.(2)The improved wavelet threshold function combined with DAE is used to diagnose the bearing data in the strong noise environment.Denoise the noisy data by improving the threshold function,and then extract the wavelet packet energy of the denoised data by wavelet packet transform.Finally,the fault classification result is obtained by DAE.The experiments on the CWRU bearing data set show that the model can get more accurate classification results under the background of strong noise.(3)Aiming at the problem of unbalanced samples in bearing data,a controlled weight coefficient is introduced into the loss function of DAE.The weight coefficient is adjusted according to the number of samples for each category.The smaller the number of samples,the larger the coefficient,which makes the model focus more on small sample during training.In addition,there is an internal covariate shifting problem in deep model which raises the complexity and training time of saturation nonlinear model.At present,prevailing deep learning models use Batch Normalization(BN)layer to accelerate model training and improve model performance.But it needs to calculate intermediate parameters(mean and variance)according to the batch in training process,which makes the BN layer heavily dependent on the batch size and causes inconsistency between training and testing results.Therefore,a deep filter response normalized auto-encoder(Dfrn AE)is proposed to overcomes the dependency problem of batch size.Two experiments on balanced and imbalanced datasets show that the proposed model is superior to existing deep learning models.The diagnostic accuracy is higher than BN no matter the batch size is too large or too small.