Health assessment is crucial for the reliable and safe operation of rolling bearings, yet existing methods struggle to unify complementary health stage information and health indices, adapt to varying contexts, and maintain interpretability. This paper proposes HA-LLM, a context-augmented online multifunctional large language model (LLM) framework that unifies adaptive health stage division and health index prediction, providing interpretable, multi-granular health assessment under diverse computational constraints. Specifically, a context-augmented autoregressive LLM prediction paradigm is developed to introduce temporal and health priors into degradation trend sequence (DTS) prediction, enabling lightweight adaptation of a frozen LLM backbone to support variable lookback lengths and prediction horizons without lookback-specific retraining. In addition, a cascaded prediction-informed functional chain integrates intra-window anti-aliasing decimation, adaptive health stage division, and autoregressive DTS prediction to establish a future-trend-informed dual health-index scheme for maintenance decision support. Extensive experiments on the IEEE PHM 2012 and XJTU-SY bearing datasets demonstrate the effectiveness of HA-LLM in degradation trend prediction and multifunctional health assessment, together with its generalization to unseen operating conditions and adaptability across different LLM backbones.
Health assessment is crucial for the reliable and safe operation of rolling bearings, yet existing methods struggle to unify complementary health stage information and health indices, adapt to varying contexts, and maintain interpretability. This paper proposes HA-LLM, a context-augmented online multifunctional large language model (LLM) framework that unifies adaptive health stage division and health index prediction, providing interpretable, multi-granular health assessment under diverse computational constraints. Specifically, a context-augmented autoregressive LLM prediction paradigm is developed to introduce temporal and health priors into degradation trend sequence (DTS) prediction, enabling lightweight adaptation of a frozen LLM backbone to support variable lookback lengths and prediction horizons without lookback-specific retraining. In addition, a cascaded prediction-informed functional chain integrates intra-window anti-aliasing decimation, adaptive health stage division, and autoregressive DTS prediction to establish a future-trend-informed dual health-index scheme for maintenance decision support. Extensive experiments on the IEEE PHM 2012 and XJTU-SY bearing datasets demonstrate the effectiveness of HA-LLM in degradation trend prediction and multifunctional health assessment, together with its generalization to unseen operating conditions and adaptability across different LLM backbones.