FaultCollaborator: A Collaborative Intelligent Agent Integrating Large Language Models and Domain-Specific Models for Intelligent Fault Diagnosis

In recent years, deep learning has demonstrated strong potential in measurement-driven intelligent fault diagnosis (IFD) for industrial systems. However, conventional diagnostic models are constrained by fixed network architectures and task-specific designs, which limits adaptability to diverse operating conditions. Moreover, such models often operate as isolated data-driven components and lack effective human-machine interaction and reasoning capabilities, reducing transparency and practical reliability. Although large language models (LLMs) exhibit strong abilities in semantic understanding, reasoning, and task planning, existing studies frequently employ them merely as passive feature processors, without fully exploiting their cognitive decision-making potential. To address these limitations, this paper proposes FaultCollaborator, a human-machine collaborative IFD agent that elevates the LLM to an active decision-making role. In the proposed agent, the LLM serves as the cognitive core, responsible for diagnostic intent understanding, high-level task planning, semantic reasoning, and human-machine interaction. Based on user-described operating conditions, the LLM dynamically selects and invokes appropriate domain-specific models to perform low-level fault diagnosis on measurement data. This collaborative design establishes a closed-loop workflow that tightly integrates measurement-driven signal analysis, model inference, and human-in-the-loop interaction. Experimental results demonstrate that FaultCollaborator can accurately identify diagnostic scenarios and select suitable domain-specific models, achieving high diagnostic accuracy and robustness under diverse operating conditions.