Temporal-physics aware dynamic network for remaining useful life prediction of rolling bearings

Abstract Accurate Remaining Useful Life (RUL) prediction is vital for industrial reliability. While deep learning methods like Temporal Convolutional Networks (TCN) have shown potential, they suffer from static convolutional kernels that fail to adaptively capture complex degradation dynamics and often ignore frequency-domain physical priors. To address these limitations, we propose a Temporal-Physics Aware Dynamic Network (TPADN). TPADN adopts a Deep-stage Decomposition Strategy that splits deep stages into an Overview-Net for global degradation context generation and a Focus-Net with one-dimensional Context-Mixing Dynamic Convolution, enabling input-dependent causal kernel generation. A physics-aware input strategy fuses raw vibration signals with their Envelope Spectrum to inject fault characteristic frequency information, while a learnable Soft Thresholding mechanism suppresses noise adaptively. Experiments on FEMTO and XJTU-SY datasets demonstrate that TPADN achieves superior accuracy and robustness compared to state-of-the-art methods.