Residual Life Prediction Of Rolling Bearings Based On Deep Learning And Big Data

In recent years,rolling bearings have become more and more widely used,and have gradually become an indispensable standard component of various complex machinery.The reliability of rolling bearings directly affects the safety of rotating parts and even the entire mechanical equipment.Therefore,it is very meaningful to predict the remaining useful life(RUL)of rolling bearings.In view of the problems of complex feature extraction process and low prediction accuracy in the current bearing residual life prediction method,it is necessary to solve the problems of complex feature extraction process and low prediction accuracy.Convolutional Neural Networks(CNN)and Bi-Directional Gated Recurrent Unit(Bi GRU)in deep learning are used in this paper.The CNN-Bi GRU model is directly used to predict bearing RUL based on original one-dimensional vibration signals.CNN-Bi GRU model extracts spatial features from input signals through CNN,while Bi GRU extracts temporal features.Finally,the validity of the model is verified through experiments,and the comparison experiment proves that the model has better prediction effect than CNN,CNN-GRU and other models.In order to further improve the RUL prediction performance of the CNN-Bi GRU model,the attention mechanism is added to the CNN-Bi GRU model.The final model consists of a CNN layer superimposed by five convolutional layers,a bidirectional GRU layer,and an attention mechanism layer.Attention mechanisms enable models to reduce the loss of historical information and enhance the impact of important information.In this paper,the CNN-Bi GRUAttention model is verified using the IEEE PHM 2012 bearing dataset,and the experimental results show that under the RMSE and Score indicators,CNN-Bi GRU-Attention has higher prediction accuracy than CNN-Bi GRU and other deep learning models under the same conditions.Experimental results show that the attention mechanism can improve the accuracy of the model to predict bearing RUL.In order to improve the prediction ability of the CNN-Bi GRU-Attention model,a large data set is currently used for training,which leads to the problem that the model training takes more time.This paper implements the parallel training of the CNN-Bi GRU-Attention model with the help of the Tensor Flow On Spark framework on the big data platform built by four Alibaba Cloud servers.In order to verify the effect of parallel training,this paper uses the IEEE PHM 2012 dataset to verify the parallel training effect of the CNN-Bi GRU-Attention model.Then the experimental results are analyzed in detail from the perspectives of prediction accuracy,running time and speedup ratio.Finally,it is proved that the CNN-Bi GRU-Attention model will not have much impact on the prediction accuracy after parallelized training,but it can effectively shorten the model training time.In order to verify the generalization of the proposed prediction method,this paper uses the XJTU-SY dataset to conduct generalization experiments.Comparing the experimental results of IEEE PHM 2012 and XJTU-SY datasets,it is found that in the XJTU-SY experiment with a larger training set,the speedup ratio is greater.It shows that the larger the data set during model training,the better the acceleration effect of parallel training.In order to put the research content into practical application more conveniently and effectively,at the end of this paper,a practical software system for predicting the life of rolling bearing is designed and developed using the Py Qt5 module of Python,and the availability of the system is proved by a running example.