Health state assessment and remaining useful life prediction are two key tasks for predictive maintenance of rotating machinery. To address limitations of existing studies-such as the separation of health assessment and lifetime prediction, insufficient modeling of temporal degradation features, and limited interpretability-this study proposes a deep temporal learning framework that jointly implements both tasks. The method takes raw monitoring signals as input, constructs temporal samples using a sliding window, and employs a shared temporal feature encoder to extract degradation representations. Based on these shared features, a health state assessment branch generates a continuous health indicator, while an RUL prediction branch estimates the remaining useful life, enabling collaborative modeling of the degradation process within a unified feature space. A multi-task joint loss is introduced to jointly optimize health state modeling and lifetime regression, enhancing the representation of local degradation patterns, long-term trends, and stage-wise characteristics. Experimental results on public run-to-failure datasets show that the proposed method produces health indicators with strong monotonicity, trendability, and robustness, and achieves superior prediction accuracy compared to several baselines. These findings verify the effectiveness of the proposed framework and highlight its potential for intelligent and predictive maintenance applications.

rotating machinery, deep temporal learning, remaining useful life prediction, health state assessment, health indicator