RUL Prediction of Rolling Bearings With MVGA-DSAL: A Multiscale Variant Gaussian Attention Model With Depth-Wise Sparse Attention LSTM

The performance of mechanical systems relies heavily on the condition of rolling bearings, with any damage potentially leading to catastrophic failures. Accurate bearing health monitoring and remaining useful life (RUL) prediction are crucial. This article proposes a data-driven framework, the multiscale variant Gaussian self-attention (MVGA), integrated with a depth-wise sparse attention long short-term memory (DSAL) mechanism. Specifically, the MVGA module employs a segmented multiscale Gaussian distance mask self-attention mechanism, which prioritizes critical positions and reveals hidden cyclic frequencies, addressing the positional limitations of conventional self-attention methods. The DSAL module, in turn, extracts temporal features across multiple scales, from local to global, facilitating effective multiscale temporal feature processing. Depth-wise propagation minimizes fluctuations in health curves and mitigates information loss and feature confusion from deep stacking, thereby enhancing model stability and robustness. Experimental results demonstrate that our approach effectively addresses limitations within Transformer networks, yielding more accurate predictions and outperforming existing methods in terms of prediction accuracy and model robustness.