LSTM-based Residual Life Prediction Of Rolling Bearings

Remaining useful l ife(RUL)prediction of rolling bearings as part of rotating machinery failure pr ediction and health management has received widespread attention and research from domestic and foreign scholars.The prediction of bearing life can timely reflect the current working condition and future degradation trend of rolling bearings,and can prov ide a good understanding of the degree of damage and failure during the operation of the bearings.Through the prediction of their service life,it can provide a basis for preventive maintenance decisions,and to a certain extent avoid failures and accidents,and ensure the safety of operators and mechanical equipment.The data-driven approach based on condition monitoring is one of the most important methods for predicting the remaining life of rolling bearings in recent years,and consists of the followin g three main parts: data acquisition,feature extraction and feature selection,and the construction of a prediction model.The paper does not describe the data acquisition aspect too much,but focuses on the latter two parts and carries out experimental work.In order to make the extracted features reflect the degradation trend of the bearings objectively and across the board,and to match the actual degradation state of the bearings better,the thesis picks up the characteristics of the time,frequency a nd time-frequency domain,and the evaluation index is used for characteristics selection.In the process of characteristics selection,time and frequency domain features are selected respectively,while complementary empirical mode decomposition(CEEMD)is used to disintegrate the primeval vibration signal into a certain number or amount intrinsic mode function(IMF)components,and the energy characteristics,energy ratio and energy entropy of each mode are extracted as time-frequency domain features,from which the three together form a high-dimensional feature vector set.In characteristics selection,by calculating the comprehensive evaluation index,the selected feature vector should contain enough information to accurately reflect the degradation trend of bearing,and construct the feature vector set as the input of subsequent life prediction model.In order to better deal with the original signals of bearings in different fault types and extract the deeper features of vibration signals,the extracted high-dimensional feature vectors are used as the input of convolutional self encoder(CAE)to reduce the dimension of feature set and extract the deeper features,It is used as the input of the prediction model.Using CAE instead of the traditional dimensio n reduction method principal component analysis(PCA)for feature extraction,the feature data dimension reduction at the same time,has a better performance in the extraction of deep features.In the process of constructing the prediction model,as the be aring vibration data is a kind of series about time,in order to achieve the prediction of time series more efficiently and with higher accuracy,the thesis proposes a prediction model based on particle swarm optimised LSTM,which improves the ability of t he prediction model to capture the existing degradation trend features of the bearing by optimising the key parameters of the LSTM through PSO.Experiments are conducted by using some of the well-performing features as input.The paper shows that the PSO-LSTM has higher prediction accuracy compared to the LSTM,avoiding the prediction errors caused by manual selection of parameters and making it more adept at handling complex non-linear problems.Experiments show that the real-life prediction model based on CAE and PSO-LSTM can well reflect the monotonicity of bearing degradation trend,and the prediction error of the model is small,which is approaching the true real-life value.