计算机科学
降级(电信)
领域(数学分析)
域适应
可靠性(半导体)
可靠性理论
适应(眼睛)
人工智能
可靠性工程
数据建模
长期预测
数据挖掘
期限(时间)
机器学习
选择(遗传算法)
稳健性(进化)
作者
Jiawei Xiong,Jing Lin,Jian Zhou,Haoyuan Shen,Yizhong Ma
标识
DOI:10.1109/tr.2026.3744341
摘要
Accurate remaining useful life (RUL) prediction under cross-domain conditions is essential for reliability-oriented maintenance of industrial assets. In practice, target domain degradation trajectories are typically incomplete and unlabeled due to operational constraints that prevent assets from running to failure, while distribution discrepancies across domains further degrade prediction performance. Existing domain adaptation (DA) methods mainly rely on global feature alignment and therefore often suffer from semantic misalignment and negative transfer when only partial degradation trajectories are available. To address this challenge, this paper proposes a semantic-level segment-wise domain adaptation (SSD-Adapt) framework for cross-domain RUL prediction under incomplete target degradation trajectories. An adaptive degradation trajectory segmentation module is first developed to automatically identify multi stage degradation structures without requiring manual stage annotations. Based on the segmented trajectories, a multi-level semantic alignment strategy jointly performs global feature alignment together with degradation stage-level and healthy stage-level alignment, while a contrastive learning-based feature preservation module mitigates excessive alignment and preserves target-specific degradation characteristics. Extensive experiments on two turbofan degradation datasets demonstrate that the proposed framework consistently outperforms state-of-the-art DA methods under multiple cross-domain prediction scenarios. The framework is further validated on an electric bus battery dataset collected from two real metropolitan transit systems, demonstrating its effectiveness for fleet-level cross-city RUL prediction under realistic operating conditions with incomplete degradation observations. These results indicate that the pro posed framework provides an effective solution for cross-domain RUL prediction with incomplete degradation observations and offers strong potential for reliability-oriented prognostics in real world industrial applications.
科研通智能强力驱动
Strongly Powered by AbleSci AI