RP-LLM: A semantically conditioned cross-modal large language model framework for remaining useful life prediction

Remaining useful life (RUL) prediction supports maintenance decision-making and risk control in industrial systems, but its performance is often limited by few-shot and cross-domain scenarios. Recent progress in large language models (LLMs) offers a flexible interface for instruction generation and contextual modeling. However, current LLM-based studies for RUL prediction either use large models as powerful regressors or rely on textualized signal descriptions, while the intermediate organization of degradation states before value generation remains insufficiently explored. To address this issue, this paper develops RP-LLM, a cross-modal large language model framework with health-state semantic conditioning for multivariate time series RUL prediction. RP-LLM first learns degradation representations through contrastive pretraining between time series windows and structured textual descriptions. It then maps these representations to an ordered health prototype space encoded by text, where the resulting membership distribution serves as a semantic bridge between temporal patterns and value generation. Finally, a Prompt Learner converts the prototype distribution into soft prompts, and an LLM adapted with LoRA generates structured RUL responses under a numeric-aware training objective. The framework is evaluated under few-shot and cross-domain transfer protocols on public battery and aero-engine datasets, together with a strict zero-shot case study on a private on-orbit cryocooler trajectory. Across the evaluated protocols, it achieves up to 27.1% lower prediction error while supporting natural-language RUL queries and structured textual prediction beyond conventional scalar regression.