The monorail crane is important in mining operations, and its operation affects both safety and efficiency. Currently, fault diagnosis for monorail cranes has several challenges, such as heterogeneous mixing of multimodal data, poor use of knowledge, low real-time requirements, and high deployment costs for large-scale models. To solve these problems, we present an agent framework using a multimodal knowledge graph and a lightweight large model. In particular, we construct a fault knowledge graph for monorail cranes, organizing professional knowledge about components, failure modes, symptoms, and maintenance. By employing retrieval-augmented generation (RAG) technology, the knowledge graph is merged with the Qwen lightweight large model (low-rank adaptation) for fine-tuning to develop a diagnostic agent with task planning, tool invocation and memory. The experimental results show that the agent framework reduces “machine hallucination” and outperforms conventional diagnostic accuracy, response speed and resource efficiency, thus offering a safe and efficient solution for intelligent operation and maintenance of mining equipment.