Prognostics and health management (PHM), built upon the seven-layer OSA-CBM standard architecture, underpins the lifecycle operation and maintenance of industrial equipment. Conventional PHM systems construct functional units with fixed procedures and dedicated machine-learning models, and typically implement layered task scheduling through rule-based multi-agent systems; they suffer from inherent limitations such as weak generalization, insufficient autonomous decision-making, and the absence of natural-language interaction. Large language models (LLMs) endow agents with general-purpose semantic understanding, dynamic task planning, external tool invocation, and self-reflective iteration, giving rise to LLM-driven autonomous agents and offering a new paradigm for fully automated PHM workflows. This survey first reviews the OSA-CBM seven-layer architecture and the theoretical foundations of PHM multi-agent systems, classifies PHM agents into two generations, namely conventional multi-agent systems and LLM-driven autonomous agents, and compares their architectural logic, operational mechanisms, and application limitations. It then systematically summarizes the research progress of LLM autonomous agents in data processing, fault diagnosis, remaining useful life prediction, and maintenance decision-making, and further identifies key challenges spanning hallucination, multi-modal fusion, edge deployment constraints, and multi-agent coordination conflicts. Finally, future research directions, including lightweight domain foundation models, digital-twin integration, causality-enhanced reasoning, and standardized evaluation, are outlined, with the aim of providing a reference for research and engineering practice of LLM agents in intelligent industrial equipment maintenance.

prognostics and health management,autonomous agents,OSA-CBM,Intelligent operation and maintenance