Research on the application framework of multi-agent technology-driven failure analysis for industrial products

Abstract Accurate failure analysis in intelligent supercharging stations is critical for operational safety and efficiency. However, traditional methods suffer from poor cross-condition adaptability, limited few-shot learning capability, and insufficient cross-model generalization. This paper proposes a Multi-Agent Failure Analysis Execution System (FAES), which synergistically integrates a Multimodal Workbench, an Intelligence Hub, and a modular agent cluster. The Multimodal Workbench fuses heterogeneous sensor data (voltage, current, temperature, acoustic) through a progressive three-level fusion strategy with cross-modal attention and Dempster-Shafer evidence theory. The Intelligence Hub formulates task allocation as a constrained optimization problem, which is relaxed to a differentiable softmax routing mechanism. This enables real-time, load-aware scheduling. A fine-tuned large language model (Qwen3-Omni) provides domain-adaptive diagnostic reasoning through closed-loop reinforcement learning from human feedback. Extensive experiments on real-world charging station data demonstrate that FAES achieves 83.2% diagnostic accuracy under few-shot cross-dataset transfer-a 15.8 percentage point improvement over training from scratch and reduces average response time by 42.9% (from 4.2 s to 2.4 s) via dynamic multi-agent scheduling. Ablation studies quantify the contribution of each component, while hallucination analysis confirms a low rate of just 0.4%. These results establish FAES as an explainable, generalizable solution for industrial fault diagnosis, filling a critical gap in the application of LLM-based multi-agent collaboration to charging station maintenance.