Bearing Remaining Useful Life Prediction and Maintenance Decision Support Using a Large Language Model

Bearing remaining useful life (RUL) estimates require causal inputs and temporal consistency before supporting maintenance. This study develops a large language model (LLM)-assisted pipeline connecting online degradation evidence, prediction applicability, structured state output, and maintenance rules. Historical vibration evidence is used to construct a health indicator and determine the domain in which RUL prediction is applicable. The LLM then generates health stages and RUL predictions, with direct token and regression head methods compared for numerical RUL prediction. A causal median filter and monotonic temporal projection are subsequently applied, and deterministic risk rules are used to generate maintenance recommendations. Direct token and regression head readouts are evaluated separately; only the regression head path supplies nonincreasing temporal processing and the frozen risk mapping. Five bearing disjoint folds cover 15 XJTU-SY bearings and 9216 observations. Seven numerical baselines are evaluated on validation and test data. In the seed-42 test, long short-term memory achieves the lowest sample-pooled mean absolute error (MAE), 0.134, while direct token achieves the lowest bearing-equal MAE, 0.174. The regression head’s validation advantage over direct token is not retained. Across three LLM seeds, temporal processing reduces mean overestimation proportions by 35.762% and 55.666%. The primary rule reaches high on 12 bearings. Complete LLM inference takes a median 10.744 s per observation at batch one on an RTX 3090. The framework makes prediction and rule trade-offs traceable through a common structured interface. Its cross-bearing evidence supports conservative temporal processing and explicit inspection rules, while compact models retain important accuracy and computational advantages.