架空输电线路是电网的主要构成部分，因长期暴露于自然环境下工作，容易发生金 属结构件的断裂、变形、锈蚀并存在异物附着的现象，对电力系统的安全稳定构成重大 威胁。为保证电网安全可靠的运行，必须定期巡检，以便及时更换金具、清除异物。因 此研究基于深度学习的多目标缺陷与异物识别算法，可以为输电线路故障诊断提供决策 依据。本文主要研究内容如下： （1）对无人机巡检图像进行数据增强与数据优化。使用改进的 SRCNN 卷积神经 网络对数据集进行图像超分辨率重建，得到更高分辨率的图像数据集，重建后数据集图 片的峰值信噪比相对原始低分辨率图片高出了 4.29 dB。 （2）针对输电线路缺陷与异物检测模型精度不够高的问题，构建改进 YOLOv7 目 标检测模型。首先将 GhostNet 作为替换的主干网络，保持模型检测精度高的同时，参 数量和计算量尽可能少；其次在特征融合阶段使用改进 BiFPN 多尺度特征融合结构代 替原始 PANet，提升小目标检测能力；然后将随机池化的方法加入 CBAM 注意力模 块，得到更完整的注意力图。实验结果表明：改进后的 YOLOv7 算法平均精度达到 98.4%，相比原始 YOLOv7 算法提高了 5.4%，检测速率达到 63.7 f/s。 （3）针对改进 YOLOv7 目标检测模型体量大，构建 YOLOv7-tiny 轻量化检测模 型。使用模型压缩中知识蒸馏的方法，通过学习蒸馏损失使 YOLOv7-tiny 学生网络的 预测结果接近改进 YOLOv7 教师网络。得到的轻量化 YOLOv7-tiny 蒸馏模型大小为 14.0 MB，仅为改进 YOLOv7 模型的 1/5 左右，平均精度达到 95.0%，相比蒸馏前原始 YOLOv7-tiny 高出 4%，检测速率达到 93.4 f/s，较改进 YOLOv7 教师网络提高了 29.7 f/s。 本文所构建的两种模型具备良好的实时性、精确性和鲁棒性。改进的 YOLOv7 算 法准确率高，可以提高智能巡检精度。而 YOLOv7-tiny 蒸馏模型速度快、体积小，适 用于实际电力巡检部署，可为推动智能电网的发展，保证电力系统的安全运行提供应用 基础。

Aerial transmission lines are essential to the power grid, yet they are vulnerable to metal structural failures and the accumulation of foreign objects due to prolonged exposure to natural elements, posing a substantial risk to the grid's safety and stability. To guarantee the grid's secure and dependable operation, routine inspections are essential to promptly replace components and remove foreign objects. Consequently, research into multi-target defect and foreign object identification algorithms grounded in deep learning can offer crucial decisionmaking support for diagnosing transmission line faults. The main research of this thesis is as follows: (1) Enhancing and optimizing UAV inspection images. An enhanced SRCNN convolutional neural network is utilized for image super-resolution reconstruction of the dataset, resulting in a dataset with higher-resolution images. The peak signal-to-noise ratio of the reconstructed images is 4.29 dB higher than that of the original low-resolution images. (2) Developing an improved YOLOv7 target detection model for transmission line defects and foreign objects. GhostNet serves as the primary backbone network to maintain high detection accuracy while minimizing parameters and computational requirements. An enhanced BiFPN multi-scale feature fusion structure is implemented in the feature fusion stage to boost the detection of small targets. Random pooling is integrated into the CBAM attention module to produce a more comprehensive attention map. Experimental outcomes demonstrate that the enhanced YOLOv7 algorithm achieves an average precision of 98.4%, a 5.4% improvement over the original YOLOv7 algorithm, and a detection rate of 63.7 frames per second. (3) Constructing a lightweight YOLOv7-tiny detection model for transmission line defects and foreign objects. Knowledge distillation techniques in model compression are utilized to train the YOLOv7-tiny student network to predict outcomes similar to those of the enhanced YOLOv7 teacher network. The distilled YOLOv7-tiny model is 14.0 MB in size, roughly one-fifth the size of the enhanced YOLOv7 model, with an average precision of 95.0%, a 4% improvement over the undistilled YOLOv7-tiny, and a detection rate of 93.4 frames per second, a 29.7 frames per second increase over the enhanced YOLOv7 teacher network. The two models constructed in this thesis have good real-time, accuracy and robustness. The enhanced YOLOv7 algorithm has high accuracy and can improve the smart inspection accuracy. And the YOLOv7-tiny distillation model is fast and small, which is suitable for actual power inspection deployment, and can provide an application basis for promoting the development of smart grid and ensuring the safe operation of the power system.