Trend contrast features-based bearing remaining useful life prediction method

The accuracy of data-driven bearing remaining useful life (RUL) prediction is highly dependent on input degradation features. These degradation features, extracted from original signals, should effectively represent the degradation state of bearings. However, these degradation features tend to exhibit low monotonicity and correlation, which reduces prediction accuracy. To address this issue, a new RUL prediction approach is proposed, called temporal associated contrastive learning-long short-term memory (TACL-LSTM). The TACL-LSTM approach mainly comprises four steps: (1) original signals are converted to the frequency domain to reduce noise interference; (2) the proposed TACL approach is used to extract the features; (3) a comprehensive evaluation metric is used to select key features called trend contrast features; (4) an LSTM neural network model is established based on trend contrast features for bearing RUL prediction. The PHM2012 dataset experimental results reveal that the TACL-LSTM method achieves higher RUL prediction accuracy compared with other traditional methods.