LLM-enhanced multi-agent reinforcement learning for human-centric prescriptive maintenance in manufacturing

摘要

原始英文摘要

Industry 5.0 emphasizes human-centric manufacturing and creates demand for maintenance systems that move beyond passive failure prediction toward economically grounded prescriptive decisions. Prescriptive maintenance strategies aim not only to anticipate failures but also to optimize intervention policies under practical economic and operational constraints. While deep reinforcement learning has shown promise in automating such decisions, existing approaches rely primarily on numerical data and cannot process the qualitative diagnostic information stored in maintenance logs. By missing these crucial historical fault patterns, they operate as opaque systems that are difficult for practitioners to interpret and trust. This paper addresses these gaps by proposing a large language model (LLM) enhanced hierarchical multi-agent reinforcement learning framework for prescriptive maintenance with three main contributions. First, it adopts a heterogeneous architecture where a policy-gradient Manager handles station-level maintenance prioritization and a value-based Worker selects from a predefined taxonomy of prescriptive maintenance actions, reflecting how real maintenance teams are organized. Second, locally deployed LLMs translate textual maintenance context, including short-form maintenance reports and technician notes, into diagnostic semantic state variables with no dependence on cloud services. Third, the framework introduces a six-component cost-model reward based on production economics, covering tool, material, equipment, downtime, maintenance, and handling costs, so that automated decisions can be understood in straightforward economic terms and the effect of reward design on maintenance behavior can be examined explicitly. The framework is evaluated on a calibrated digital shadow of a lab-scale drone assembly line. Semantic state augmentation is associated with a 25.1% higher converged reward and 42% lower learning variance compared to the numerical-only baseline for the strongest LLM configuration. Under joint noise at σ = 0.25, the LLM-augmented agent still retains a performance advantage over the baseline, indicating that text-derived semantic variables provide a relatively stable auxiliary signal when numerical observations become unreliable. The cost-model formulation also reduces total operational cost compared to the overall equipment effectiveness based alternative while making the resulting policies easier to interpret in economic terms. These results show that integrating human diagnostic knowledge, reflecting real maintenance workflows, and improving economic transparency can build the necessary trust to support effective human-AI teaming in industrial settings.

原始来源: SSRN ↗