A Defect Detection and Early Warning System for Transmission Line Insulators Based on the Integration of Lightweight YOLOv11 and Large Language Model

To address the limitations of existing object detection algorithms such as low accuracy, slow processing speed, and the inability to generate timely alerts for maintenance personnel in complex environments, this paper proposes an insulator defect detection and early warning system that integrates an enhanced lightweight YOLOv11 with the GPT-4o large language model, tailored to practical engineering scenarios. An improved PM image denoising algorithm is first applied to preprocess the dataset, effectively mitigating issues such as small defect targets and complex backgrounds, thereby enhancing robustness in field applications. The YOLOv11 backbone is augmented with a Haar Wavelet Downsampling (HWD) module, while a channel attention (squeeze-and-excitation networks, SENet) module is embedded in the Neck to boost feature extraction efficiency. These enhancements reduce the model’s complexity while simultaneously improving accuracy and inference speed. Furthermore, based on the detection outputs, structured natural language reports are generated via interaction with GPT-4o through OpenAI’s API, enabling automated fault description and intelligent early warning. Experimental results demonstrate that the proposed method achieves a detection accuracy of 96.52% and a realtime processing speed of 102 FPS, outperforming existing approaches in accuracy, responsiveness, and ease of deployment, and thus holds strong potential for engineering applications.