PowerGPT‐R1: Vision‐Language Reinforcement Fine‐Tuning With Verifiable Reward for Power Inspection

Vision‐driven intelligent power inspection systems have long faced the challenge of scarce high‐quality datasets in specialised domains, leading to limited performance of traditional deep learning methods (e.g., Faster‐RCNN) in few‐shot learning scenarios. Recent breakthroughs in language models, particularly the open‐source DeepSeek‐R1 model, have provided new technical pathways for vision‐language models. This study proposes the PowerGPT‐R1 framework—the first innovative architecture to successfully transfer the reinforcement fine‐tuning paradigm to multimodal learning in power inspection. The core contributions include: (1) a cold‐start strategy based on supervised fine‐tuning, integrating expert‐annotated electrical image descriptions and GPT‐4o‐generated reasoning chains to establish a foundation for multimodal semantic understanding; (2) reinforcement learning fine‐tuning, employing a verifiable reward mechanism to score reasoning responses and utilising the Group Relative Policy Optimisation algorithm for model parameter updates; (3) establishment of a multi‐dimensional evaluation system, with experimental results showing a significant improvement in detection accuracy under few‐shot conditions compared to traditional visual models while demonstrating multimodal reasoning capabilities absent in visual perception models. The proposed technical framework offers an innovative solution for developing domain‐specific vision‐language models with high interpretability and low data dependency, holding significant application values in the autonomous inspection of smart grid infrastructure.