Decision Support System for Equipment Fault Diagnosis Based on Large Language Model: Dynamic Construction of Unstructured Industrial Knowledge Graph for Industrial Equipment

With the deepening of industrial digital transformation, equipment fault diagnosis faces challenges including low utilization of unstructured data, weak cross‐modal semantic association, and lagging knowledge updates. Traditional methods relying on artificial rules and static knowledge bases struggle to effectively integrate multimodal information such as text, images, and sensors. This paper proposes a decision support system for equipment fault diagnosis based on large language models (LLMs), with an unstructured industrial knowledge graph (UIKG) as the core knowledge carrier. We construct a dynamic UIKG driven by multimodal data, abandoning static manual construction modes. The system employs a domain‐adaptive LLM combined with contrastive learning and low‐rank fine‐tuning to enhance professional terminology understanding and contextual reasoning. A cloud‐edge collaborative architecture balances real‐time response and computational constraints through lightweight deployment and dynamic knowledge subgraph transmission. Temporal weight update and distributed graph incremental mechanisms ensure dynamic knowledge evolution. Experiments on port cranes and wind turbines show that the system reduces the missed detection rate to 1.5% and maintenance costs by over 20%, while supporting rapid cross‐device migration. The core novelties of this work are threefold: (1) proposing a domain‐adaptive UIKG construction method driven by multimodal data, breaking the static manual construction mode of traditional industrial knowledge graphs; (2) designing a cloud‐edge collaborative dynamic UIKG update and reasoning framework, balancing the real‐time performance of edge fault diagnosis and the depth of cloud knowledge analysis; and (3) integrating LoRA fine‐tuning and contrastive learning to realize cross‐modal semantic alignment of industrial text, image, and sensor data, solving the problem of semantic discretization of unstructured industrial data. Future research will focus on federated learning frameworks and multimodal generation technologies to further optimize knowledge sharing mechanisms and adaptive reasoning capabilities.