Abstract:

As universal deep learning object detection technology is difficult to achieve real-time and efficient fault detction during high-voltage transmission and transmission line inspection, a real-time detection method is proposed for inspecting the fault of high-voltage transmission line based on CenterNet improved algorithm. The three kinds of common inspection faults such as insulator self-explosion, shock hammer falling off and bird’s nest are detected. An efficient DLA-SE deep feature extraction network is designed based on Deep Layer Aggregation（DLAnet）, Squeeze-and-Excitation modules, and deformable convolution. In CenterNet architecture, as heat map of center point, width height and offset information of the object are obtained through the feature extraction network of DLA-SE, which gets the object bounding box. Experimental results show that under the nvidia GTX 1080 test conditions, the proposed method has a mAP of 0.917 and an inference speed of 27.03 frame/s in comparison with CenterNet, SSD and YOLOv3 method, and detection accuracy has been greatly improved, which proves the method effectiveness.

Key words: CenterNet, Deep Layer Aggregation（DLAnet）, electric power inspection, fault detection, real-time detection

摘要：

针对通用深度学习目标检测技术难以在高压输变电线路巡检时实现实时高效的故障检测，提出一种改进CenterNet的高压输变电线路巡检故障实时检测方法，对绝缘子自爆、防震锤脱落、鸟巢三类常见巡检故障进行检测。该方法基于深层特征融合网络（DLAnet，Deep Layer Aggregation）、挤压-激励SE（Squeeze-and-Excitation）模块、可形变卷积进行高效深层特征提取网络DLA-SE的设计。在CenterNet架构下通过DLA-SE特征提取网络获取对象的中心点热力图，回归对象的宽高、偏移信息，得到对象边界框。实验结果表明，在Nvidia GTX 1080测试条件下该方法的mAP达到0.917，推理速度达到27.03?frame/s，与CenterNet、SSD与YOLOv3方法相比在检测精度上取得大幅度提升，证明了该方法的有效性。

关键词: CenterNet, 深层特征融合, 电力巡检, 故障检测, 实时检测