PJDAN: progressive joint domain adaptation with multi-scale temporal modeling and dual MMD alignment for remaining useful life prediction

作者：Jiarui Cui, Lingyi Liu, Minggang Wang, Qun Yan, Jian Huang, Xu Yang, Baowei Zhang, Q S Li · 发表于：Measurement Science and Technology · 年份：2026 · DOI：10.1088/1361-6501/ae8e97 · 研究领域：Domain Adaptation and Few-Shot Learning、Machine Fault Diagnosis Techniques、Face recognition and analysis

Abstract Remaining useful life (RUL) prediction is often undermined by pronounced distribution discrepancies in monitoring data acquired under diverse working conditions, which substantially degrade the generalization capability of prognostic models. To overcome this challenge, a novel progressive joint domain adaptation network (PJDAN) that addresses severe distribution discrepancies is proposed for RUL prediction. This framework incorporates a multi-scale feature generation module that integrates a convolutional block attention module (CBAM) with a parallel temporal convolutional network (TCN) for capturing both discriminative local features and long-term temporal patterns from raw data. To ensure global domain invariance, a progressive dual-layer maximum mean discrepancy mechanism is embedded within the backbone, enforcing statistical alignment across both the TCN outputs and CBAM-enhanced representations. Beyond merely utilizing the progressive metric mechanism for global alignment, a conditional adversarial module is incorporated for fine-grained subdomain adaptation. To stabilize the optimization process, a stage-wise training scheme is introduced to provide deterministic gradient guidance for adversarial training. Experimental evaluations on two public bearing datasets demonstrate that the proposed PJDAN outperforms existing prior methodologies in cross-domain RUL estimation.