基于深度学习的电力缺陷检测研究综述

A Review of Deep Learning-Based Power Defect Detection

摘要: 电力设备缺陷检测是输电线路、变电站和新能源设备智能运维中的重要任务。随着无人机、巡检机器人、可见光相机、红外热像仪等设备的应用，电力巡检图像数据快速增长，推动了基于深度学习的自动检测方法发展。本文围绕电力缺陷检测对象与任务特点，梳理两阶段目标检测、一阶段目标检测、Transformer视觉模型和多模态融合技术在电力缺陷检测中的研究进展。现有研究表明，两阶段方法在定位精度和区域分析方面具有优势，但推理速度和部署复杂度较高；YOLO等一阶段方法具有实时性和工程部署优势，但在小目标、复杂背景和全局语义建模方面仍存在不足；Transformer能够增强长距离依赖和高层语义表达，多模态融合则为图像、文本、红外和点云信息协同利用提供了新的方向。最后总结当前研究存在的问题及未来的发展趋势。

Abstract: Power equipment defect detection is an important task in intelligent operation and maintenance of transmission lines, substations and photovoltaic facilities. With the increasing use of UAVs, inspection robots, visible cameras and infrared sensors, massive inspection images have promoted the development of deep learning-based automatic detection methods. This review summarizes the detection objects and task characteristics of power defects, and discusses two-stage detectors, one-stage detectors represented by the YOLO series, visual Transformer models and multimodal fusion techniques. Existing studies show that two-stage methods are advantageous in localization and region-level analysis but are relatively complex for real-time deployment. One-stage YOLO-based methods are efficient and deployment-friendly, while small defects, cluttered backgrounds and limited global semantic modeling remain challenging. Transformer models provide long-range dependency modeling, and multimodal fusion offers a possible way to combine visual, textual, infrared and point-cloud information. Finally, summarize the existing problems of current research and future development trends.