Bearing Remaining Life Prediction Based On LSTM And Transformer Improved Model

Aiming at the specific problems in Remaining Useful Life(RUL)prediction of bearings,such as difficulty in feature extraction and memory decline of prediction algorithm,this paper carries out the research on the method of remaining life prediction of bearings based on LSTM improvement model and Transformer improvement model,which is mainly as follows:1.A model using LSTM-SA for predicting the RUL of rolling bearings is suggested.Most traditional feature extraction methods cannot effectively extract the bearing degradation characteristics that can precisely indicate the process of failure.For this reason,Degradation features are extracted by enveloping and demodulating the vibration signal,segmenting the envelope spectrum,and calculating the Pearson correlation coefficient with the standard sample.In order to map the degradation features into bearing health indicators(HI),LSTM time series networks are mostly used,but when the data time span is too long,LSTM will experience memory decay.Therefore,combining the Self-Attention mechanism into LSTM can solve the phenomenon of memory decay.During the training process,an improved algorithm is proposed to guide the learning process to away from the saddle point and tend to the global optima.Accurately predict the trend of bearing health indicators through the double exponential life model and particle filter algorithm,and finally realize the RUL prediction.The LSTM-SA model is verified on public and On-site rolling bearing datasets.According to the results,the suggested method can efficiently extract degradation features and precisely forecast the RUL of rolling bearings in complex operating conditions.2.A RUL prediction model based on CNN and Transformer is constructed.Accurate life prediction can provide a theoretical basis for decision-making to the greatest extent and improve economic benefits.There are still problems such as "serial computation" and "gradient explosion" for life prediction using recurrent neural networks and their variants.At the same time,the short-term and long-term trends of equipment degradation may not be consistent.Aiming at the above problems,a remaining life prediction model based on CNN and Transformer was established.CNN is used to extract local features,while Transformer is responsible for extracting global features,considering both short-term and long-term trends.Meanwhile,multiple parallel CNN-Blocks are used to extract local short-term trends to achieve "parallel computing" and improve the computing speed.The experimental outcomes demonstrate that the constructed model can adaptively extract degradation features and precisely forecast the RUL of rolling bearings.3.Apply CNN-Transformer model to the remaining life prediction of sliding thrust bearings.Nowadays,most of the monitoring of rolling bearings is to collect its vibration acceleration or speed data,while sliding thrust bearings are more monitored by collecting physical quantities such as shaft displacement and temperature.Apply CNN-Transformer to the On-site sliding thrust bearing dataset,and replace the input data from vibration signals with shaft displacement,temperature and AC.The experimental outcomes indicate that the CNN-Transformer model exhibits strong generalization capability and can precisely forecast the RUL of sliding thrust bearings.