Prediction of bearing remaining life based on STFT-SWT and Transformer-BiLSTM

Abstract

Rolling bearings are critical components in rotating machinery; therefore, the accurate prediction of their remaining useful life (RUL) is essential for efficient predictive maintenance. While deep learning enables the extraction of valuable insights from large datasets to facilitate equipment condition forecasting, challenges remain in early fault detection and long-term sequence modeling. This study proposes an enhanced bearing RUL prediction framework integrating advanced data preprocessing and a novel deep learning architecture. First, an improved wavelet threshold denoising method is developed to process raw vibration signals, effectively suppressing background noise. Subsequently, a novel short-time Fourier transform and synchrosqueezed wavelet transform time-frequency fusion strategy is introduced for high-resolution feature extraction. The resulting time-frequency diagrams are scaled via bilinear interpolation to ensure network compatibility while reducing computational redundancy. Finally, a Transformer-BiLSTM hybrid model is proposed, which replaces the traditional Transformer decoder with a bidirectional long short-term memory (BiLSTM) layer to capture both global degradation trends and local temporal dynamics. Comprehensive experiments conducted on the XJTU-SY, IEEE PHM 2012, and UC Cincinnati IMS datasets demonstrate that the proposed framework significantly outperforms baseline models (GRU, LSTM, and Transformer-LSTM) in RUL prediction accuracy, exhibiting robust generalization capabilities across various operating conditions.