ABSTRACT Deep learning approaches have been widely used for predicting the remaining useful life (RUL) of rotating machinery. However, commonly used data‐driven targets and generic time‐frequency features may reduce sensitivity to degradation and lead to physically inconsistent RUL trajectories. To address these issues, this study proposes a physics‐guided Swin‐KAN Transformer with a three‐channel log‐linear filterbank energy (LLFE) representation for bearing RUL prediction. To highlight monotonic degradation patterns, the feature map comprises log‐energy, its temporal difference, and a per‐bearing trend channel, all relative to a healthy baseline. A physics‐guided loss promotes monotonic consistency of RUL predictions throughout deterioration, whereas KAN modules substitute the final multilayer perceptron regression head and the feed‐forward network sublayers of Swin blocks. Experiments on the XJTU‐SY bearing dataset show that the proposed PGSwin‐KANT2 achieves an average MAE of 0.069 and RMSE of 0.083, reducing both errors by approximately 30.3% compared with the best baseline TT‐ConvLSTM. Ablation and sensitivity analyses further validate the contributions of the three‐channel LLFE representation, physics‐guided regularization, and KAN‐based components.
Karimjonov et al. (Wed,) studied this question.