AgentPDM: Context-Driven Agentic Reasoning for Predictive Maintenance

Abstract

Predictive Maintenance (PDM) uses sensor signals to assess machine health and guide actions such as continued operation, monitoring, or repair. Existing data-driven classifiers and knowledge-driven condition-monitoring tools each capture only part of the expert diagnostic workflow, which also requires reasoning over sensing modality, operating context, bearing geometry, spectral signatures, model uncertainty, and maintenance implications. LLM/VLM agents offer a way to coordinate these heterogeneous inputs, but free-form fault prediction remains unreliable. We present AgentPDM, an interactive multi-agent system for context-aware bearing diagnosis. It combines a four-backbone signal ensemble, bearing-physics tools, structured visual evidence, a conservative Decision Agent, and fixed-label report generation. AgentPDM proposes fault hypotheses, verifies them against operating context, spectral evidence, and model disagreement, and then accepts, overrides, or escalates the diagnosis before reporting. On cross-domain scenarios from CWRU, HUST, and Paderborn, AgentPDM improves accuracy from 84.3% to 87.7% with zero correct-to-wrong regressions. These results show that LLM/VLM agents are most reliable in industrial PDM when embedded in a bounded diagnostic workflow rather than used as unconstrained fault predictors.