An intelligent maintenance decision-making method for power equipment based on multimodal large language model

Abstract

In intelligent maintenance of power equipment, engineers analyze collected infrared images of the equipment and draft maintenance reports, which suffers from low efficiency and high subjectivity, and inconsistent standards. To address these issues, an intelligent maintenance decision-making method based on the large language model is proposed. The Qwen2- VL-7B-Instruct is used, which is characterized by robust visual capabilities, while its compact parameter scale enables efficient deployment to be achieved. Low-rank adaptation (LoRA) is used to reduce training costs while preserving generalization, and retrieval-augmented generation (RAG) with knowledge graphs (KG) is employed to construct power domain knowledge, whereby factual hallucinations are effectively mitigated and reliable, traceable diagnostic reasoning is ensured. A multimodal dataset of 3636 reports is constructed, and experimental results show that our method outperform base model in report consistency and standardization, which improve the efficiency of intelligent maintenance.