In response to the operational challenges posed by the rapid growth of the photovoltaic industry and to address issues of low efficiency and difficulties in knowledge retrieval in traditional models, this study has designed an intelligent fault diagnosis system that combines anomaly detection with a knowledge graph. The system utilizes large language model (LLM) to automatically extract knowledge from operational documents, thereby creating a photovoltaic operational knowledge graph. Additionally, it uses a lightweight (MobileViT) visual model to perform fault identification from fault images. The system features a closed-loop diagnostic model where “image serves as the query”: users upload fault images, and the model’s recognition results directly update the knowledge graph to retrieve the fault cause and suggest recommendations. This research provides the photovoltaic operations and maintenance field with an automated, intelligent diagnostic tool that can significantly enhance operational efficiency and reduce reliance on expert experience.