A Zero-Shot Multimodal Agent Framework for Power Line Inspection via Large Language Models

With the widespread deployment of unmanned aerial vehicles (UAVs) in smart grids, visual inspection based on deep learning has significantly advanced. However, conventional object detection paradigms heavily rely on massive, high-quality annotated datasets, struggling with poor generalization and a fundamental “data-hungry” dilemma caused by the extreme scarcity of long-tail defect samples. To overcome these bottlenecks, this paper proposes a novel zero-shot multimodal agent framework driven by large language models (LLMs). Adopting a centralized star topology, the architecture strictly decouples physical visual perception from high-level business cognition, utilizing an LLM as the cognitive hub and a vision-language model (VLM) as the perception tool. Crucially, the system introduces a retrieval-augmented generation (RAG) mechanism to dynamically align visual findings with authoritative grid regulations, supporting rule-referenced report generation. Experimental results demonstrate that, without any domain-specific fine-tuning, the proposed framework reduces false alarms under the tested configuration, with a detection F1 score of 82.52%. In a common-report comparison repeated three times, the framework achieves 65.30% correct reference-clause coverage, compared with 43.84% for a direct Qwen-VL-Max call supplied with the same visual priors and rules. The framework extends visual detection to structured, human-readable reports linked to supplied industrial rules; its grading accuracy and repeatability remain dependent on the visual observations and subsequent tool execution.