Research On The Prediction Of Remaing Useful Life Of Rolling Bearings Based On Deep Learning

Prognostics and Health Management(PHM)uses monitoring data,models,and software to perform initial fault detection,status assessment,and remaining useful life(Remaining Useful Life,RUL)prediction,which ensures the reliability and safety of the system It has shown strong advantages in terms of sex and economy,and has become a hot spot for many industries.On the one hand,traditional data-driven machine learning methods require sufficient prior knowledge to build degradation models and feature extraction is complex.On the other hand,deep learning provides a strong impetus to the development of RUL prediction with its powerful ability and applicability of automatic feature extraction.In view of this,this paper takes the rolling bearing as the object,and uses the deep learning method to extract the degradation features,construct the degradation model and predict the remaining service life,which has good engineering application value.In view of the insufficient ability of traditional features to express the bearing degradation trend,this paper uses stacked sparse auto-encoder(Deep Sparse Auto-encoder,DSAE)for deep feature extraction,and preliminarily verifies the effectiveness of deep features in bearing degradation representation by means of bearing fault diagnosis.First,the original vibration signal of the bearing is preprocessed and fast Fourier transformed to obtain the frequency domain signal;secondly,the frequency domain signal of different fault types is input into DSAE for deep feature extraction,and the fault diagnosis model is constructed.Fault diagnosis is carried out on the bearing data set of Case Western Reserve University(CWRU),and the expressiveness of deep features on bearing degradation information is preliminarily verified,which provides the basis for the feature extraction method of remaining service life.Aiming at solving the problems of network saturation degradation in deep learning and noise interference in signal,this paper proposes a bearing remaining service life prediction method based on attention residual noise reduction network and improved two-way long and short-term memory network.The original vibration signal of the bearing is preprocessed and Hilbert-Huang Transform(Hilbert-huang Tranfer,HHT)is used to obtain the marginal spectral time-frequency domain signal;in order to further enhance the network’s ability to resist noise and alleviate degradation,a CBAM(Convolution Block Attention Module)Attention Residual Network(ARN)extracts deep degradation features and constructs the Health Indicator(HI)of the bearing degradation trend;input the health indicator into the two-way long-short time optimized by the Sparrow Search Algorithm(SSA)Memory network(Bi LSTM)for remaining useful life prediction.The life prediction experiment was carried out on the PHM 2012 bearing data set.The results show that the attention residual denoising network can effectively avoid noise interference,SSA can effectively and quickly optimize the network parameters,and the overall proposed method can effectively improve the accuracy of the remaining service life.sex.Aiming at solving the problem of inconsistency in the distribution of bearing data in different working conditions,this paper proposes a migration learning method based on Wasserstein Domain Adverdrial Training Neural Networks(WDANN)based on Wasserstein distance improvement.Firstly,the original vibration signal of the bearing is preprocessed and HilbertHuang Transform(HHT)is used to obtain the marginal spectral time-frequency signal;secondly,the domain discriminator in DANN is improved by using the Wasserstein distance,and the marginal spectral signal is used as input for anti-migration training;finally,lifespan prediction is performed using the transferred deep features.The life prediction experiment was carried out on the PHM 2012 bearing data set.The results show that,compared with the original DANN and several other deep transfer learning methods,the proposed method can effectively reduce the impact of distribution differences on life prediction,and improve the model’s performance in different industries.The applicability of the situation and the accuracy of the prediction.