Degradation stage-aware domain adaptation and individualized error compensation for remaining useful life prediction across operating conditions

Accurate prediction of equipment’s remaining useful life (RUL) is crucial for enhancing its operational reliability. However, complex and variable operating conditions pose significant challenges to traditional deep learning prediction models in practical applications. Currently, multi-source domain adaptation (MSDA) approaches demonstrate substantial potential for RUL prediction across operating conditions. However, existing MSDA-based approaches suffer from inconsistent source domain quality, neglect of the degradation stage (DS) information, and private information in the target domain. To address these, an RUL prediction framework that combines degradation stage-aware domain adaptation and individualized error compensation is proposed. First, a multi-scale migration degree measurement method is developed to identify optimal source domains and subdomains, constructing a high-quality dataset matching the target domain. Second, a Multi-Scale Mamba (MS-Mamba) model is established. A multi-kernel parallel dynamic convolution architecture is adopted to enhance the MS-Mamba model’s ability to extract multi-scale temporal degradation features from source domains. Combined with DS information and a simplified self-attention, adaptive weighting for degradation features is realized. The multi-kernel Maximum Mean Discrepancy incorporating temporal information is leveraged to reduce distribution discrepancies among samples within the identical degradation stage. Finally, an individualized error compensation module is designed based on self-supervised learning and physical laws of equipment degradation, utilizing target domain’s private features to compensate for the loss of degradation information during domain adaptation. The proposed method’s effectiveness was validated across ten transfer tasks using two bearing datasets and one servo turret power head system dataset, providing a novel technical route for across-domain RUL prediction of industrial machinery.