Research on Industry-LLM Fault Diagnosis and Decision-Making with Knowledge Augmentation and Multimodal Alignment

With the increasing complexity of intelligent manufacturing systems, fault diagnosis and maintenance decision-making for key production equipment face severe challenges such as heterogeneous data silos, scarce fault samples, and a lack of interpretability in the decision-making process. This paper proposes an Industry-LLM architecture integrating domain knowledge enhancement, multimodal perception, and temporal reasoning. The architecture incrementally pre-trains through the construction of an industrial equipment knowledge graph, endowing the model with domain commonsense; designs a specialized multimodal temporal-text alignment encoder to map sensor signals into the semantic space of large language models; and introduces a Chain-of-Thought (CoT) mechanism to guide the model in generating interpretable maintenance decision reports containing logical reasoning processes. Experiments demonstrate that this method achieves a 12.5% improvement in F1 score for few-shot fault diagnosis scenarios compared to traditional convolutional neural networks, maintains high robustness in strong noise environments, and produces maintenance solutions evaluated by experts as highly feasible and safe, significantly reducing the mean time to repair (MTTR).