Adaptive Multi-domain Feature Fusion Remaining Useful Life Prediction of Rolling Bearings Using Temporal Convolution Long Short-Term Memory Network

Accurate bearings’ remaining useful life (RUL) prediction can effectively preserve the reliability of rotating machinery. However, some key problems remain in RUL prediction, including effective screening signals’ multi-domain features and mining the nonlinear relationships between features and RUL. To solve the above problems, an adaptive multi-domain feature fusion RUL prediction method based on temporal convolutional long Short-term memory (TCN-LSTM) network is proposed. Firstly, multi-domain features extraction of bearings’ monitoring sensor signals are carried out in the time, frequency, and time-frequency domains. Then, an adaptive feature selection method is utilized to reduce redundant features. Finally, the multi-domain feature matrices are utilized as the model’s input, and a TCN-LSTM model is proposed to effectively learn the direct complex relationships between the multi-domain feature matrices and the bearings’ RUL. Experiments on PHM2012 bearing datasets demonstrate the designed model’s prediction accuracy and better generalization ability.