Cognitive-State-Driven LLM Decision Agent for Fault Diagnosis of Rotating Machinery Under Incomplete Evidence

Abstract

Reliable fault diagnosis of rotating machinery under incomplete evidence involves identifying the fault and determining whether the available evidence is sufficient for a decision. However, existing methods mainly map fixed inputs to fault labels and do not systematically evaluate evidence sufficiency. To address this limitation, this paper proposes a cognitive-state-driven large language model (LLM) decision agent for fault diagnosis under incomplete evidence. First, an evidence-state space and a masked representation organize the available evidence, while a unified diagnostic judge provides calibrated class probabilities and a candidate label. Second, uncertainty, stability, out-of-distribution (OOD), and case-memory signals form an augmented cognitive state that describes the current diagnostic belief and its reliability. Finally, a rule-based expert policy provides reference decisions for supervised fine-tuning (SFT), and Group Relative Policy Optimization (GRPO) adjusts the LLM policy to balance submission, further evidence acquisition, abstention, and expert review. Experiments on 4,018 decisionstate records from the Paderborn University (PU) bearing dataset show that the 14B GRPO policy reduces potentially premature automatic submissions. These cases occur when the model submits a diagnosis even though the reference policy recommends acquiring more evidence, and their number decreases from 181 to 119, corresponding to a relative reduction of 34.3%. The wrong automatic diagnosis rate decreases from 4.16% to 3.56%. These results suggest that the proposed agent achieves a more favorable balance between automatic submission and further evidence acquisition at the decision-state level, thereby supporting more reliable diagnostic decisions.