Remaining Useful Life Prediction of Rolling Bearings Based on Multi-Scale Time-Frequency Features and Attention-BiLSTM

Abstract

Remaining useful life (RUL) prediction of rolling bearings is essential for condition-based maintenance and reliability management of rotating machinery. To improve degradation representation and temporal modeling ability, this paper proposes a bearing RUL prediction method based on multiscale time-frequency features and an Attention-BiLSTM model. Time-domain features, including root mean square, peak value, kurtosis, skewness, and standard deviation, are extracted together with frequency-domain features, including band energy and spectral entropy, from horizontal and vertical vibration signals. The extracted features are standardized and organized into degradation sequences through a sliding-window strategy. BiLSTM is then used to encode bidirectional temporal dependencies, and an attention module adaptively weights the hidden states within each input window for RUL regression. Experiments on the PHM2012/PRONOSTIA bearing degradation dataset show that the proposed method achieves competitive overall performance on Bearing1_3 compared with SVR, LSTM, and BiLSTM in terms of RMSE, MAE, stable MAPE, and PHM Score. Feature ablation results further indicate that the fusion of time- and frequency-domain features provides more effective degradation representation than either feature group alone. Multi-bearing and per-condition tests also reveal that bearing-to-bearing variation and operating conditions still have a clear influence on prediction difficulty. This study provides a practical feature construction and attention-based sequence modeling scheme for bearing prognostics, offering a useful reference for data-driven RUL prediction in condition-based maintenance of rotating machinery.