Research On Prediction Method Of Remaining Useful Life Of Rolling Bearing Based On Long-time Sequential Analysis

Rolling bearing,as one of the main components in rotating machinery system,runs in a harsh environment and is prone to failure during equipment operation.Once the rolling bearing is damaged,the machine will stop working slightly and serious accidents will occur.If bearing faults can be found early,a more reasonable maintenance plan can be made to ensure safe,smooth and efficient operation of the equipment.Therefore,the prediction and analysis of bearing remaining useful life(RUL)is one of the important contents of health monitoring research of rotating machinery system.The emergence of in-depth learning technology allows bearing RUL to be predicted without building complex mathematical model.Based on this,this thesis constructs a deep neural network model based on long temporal sequence analysis for rolling bearing life temporal sequence signal,and conducts in-depth research on rolling bearing RUL prediction method.The main research contents are as follows:(1)Multi-scale Temporal Convolutional Network(MSTCN)was established for rolling bearing RUL prediction.First,Temporal Convolutional Network(TCN)is embedded into multi-scale module.MSTCN blocks with two different structures,Mstcn Pool-Block and Mstcn-Block,are selected to construct.By changing the size of expansion factor in expansion convolution,the multi-scale module is embedded to learn feature information on different scales and fuse it to realize feature learning and prediction on different scales of bearing life data.Secondly,residual connection is introduced to prevent over-fitting and under-fitting.Finally,the bearing RUL prediction experimental analysis is completed,and the validity of this method in bearing RUL prediction is verified by comparing the prediction results with those of several life prediction models.(2)MSTCN model still has some room for improvement in bearing RUL prediction accuracy and stability.Therefore,a combined MSTCN-Transformer network model is proposed for bearing RUL prediction.MSTCN is used to extract the feature of bearing long sequential temporal sequence as the feature input of Transformer prediction model,and t-SNE visual analysis and robustness and trend calculation are carried out to further verify the effectiveness of MSTCN temporal sequence feature extraction.Based on self-attention mechanism and combining the structure of coding layer and decoding layer,the Transformer model is constructed,and the extracted temporal features are input into the Transformer model to further mine the potential interdependence of features.By comparing three groups of experiments,the proposed combined model achieves more accurate prediction results based on MSTCN.(3)Transformer’s self-attention mechanism is characterized by large calculation amount,low work efficiency and long running time of the model when performing matrix operation.Based on this,Informer network is used instead of Transformer network to build a combined MSTCN-Informer network model to predict bearing RUL.Informer network is based on sparse self-attention mechanism.When calculating attention value,instead of choosing all Q matrices to participate in the matrix operation,it uses KL divergence to select important Q matrices to participate in the calculation,which greatly reduces the calculation amount.In addition,Informer network uses distillation mechanism to downsample features,simplifies the model structure,overcomes the problems of large computing memory and increased network parameters caused by the combination of Transformer multi-layer encoder layer and decoder layer.The experimental analysis shows that the training time of MSTCN-Informer network model is 48% shorter and the RMSE is 8% lower than that of MSTCN-Transformer network model when used for bearing RUL prediction.