A Single-Domain Generalization Method for RUL Prediction via Gradient-Guided Feature Disentanglement

In practical industrial settings, training data are often available from only a single source (single domain), whereas deployment must contend with changes in operating conditions and domain shift, which degrade performance. To improve the out-of-domain generalization of remaining useful life (RUL) prediction from single-domain training, we propose an attribution-guided framework for learning domain-invariant features. Using gradient attribution, we jointly quantify task significance and domain significance, and accordingly decompose features into domain-invariant and domain-specific components. We further design a timeseries representation module that combines parallel convolutions with multi-head attention, and adopt a momentum teacher-student dual-branch architecture to provide stable paired views. Building on this design, we introduce a composite loss comprising an RUL supervision term, a domain-discrimination term, maximum mean discrepancy (MMD)-based alignment of the invariant subspace, and cross-covariance regularization to disentangle invariant and specific factors. Experiments on the N CMAPSS dataset demonstrate that, under single-domain training with cross-domain evaluation, the proposed method delivers strong performance.