Integrating Large Language Model and Lightweight Network for Robust Multimodal Fault Diagnosis in Wheeled Robots

As robots are increasingly integrated into industrial applications, robust fault diagnosis is becoming increasingly crucial to ensure operational safety and minimize unplanned downtime. Graph neural networks have emerged as a promising solution for handling multisensor fault diagnosis tasks, but conventional approaches for constructing spatial–temporal graphs from robotic sensor data tend to rely on either static physical knowledge or purely statistical relationships, each with inherent limitations in changing operational contexts. To address these challenges, this article proposes a novel hybrid architecture driven by large language model (LLM), which integrates global physical knowledge with localized statistical patterns to guide the dynamic construction of spatial–temporal graphs. A concrete operational method is further proposed to structure textual inputs for the LLM, allowing it to integrate domain knowledge, task objectives, and fusion logic into graph representations that are physically grounded and responsive to real-time conditions. Building on the outputs of this LLM-driven process, we design a lightweight spatial–temporal graph network that incorporates the dynamically generated graphs into fault inference. Finally, we validate our approach on two real-world robot platforms, and demonstrating notable improvements in diagnostic accuracy, adaptability, and robustness compared with existing state-of-the-art methods.