前端化目标检测技术在电力巡检中的应用研究

Research on the Application of Front⁃end Target Detection Technology in Electric Power Inspection

电力巡检机器人存在巡检图像分析和缺陷检测识别必须依赖后台服务器、电力巡检工作实时性较低等问题。针对此问题,研究前端化目标检测技术,提出层间合并、参数量化的深度学习模型轻量化方法,模型体积压缩为原模型的60%,模型推理效率提升10~20倍,解决了深度学习模型网络层数深、结构复杂、存储占用多、推理效率低的问题。将轻量化模型部署在边缘侧设备中,设计了电力巡检图像前端化目标检测系统,实现输变电站机器人的全自主巡检与缺陷自动采集,同时与后端服务器进行协同交互,建立了“云-边-端”协同运检体系,能够提升电力巡检效率、降低巡检成本,为电力巡检机器人推广应用提供参考。

In the way of electric robot inspection,the inspection image analysis and defect detection and identification must rely on the background server,and the real⁃time performance of electric power inspection is low.To solve this problem,the front⁃end target detection technology was studied,and a lightweight method of deep learning model based on layer merging and parameter quantization was proposed.The model volume is compressed to 60%of the original model,and the model reasoning efficiency is improved by 10~20 times,which solves the problems of deep learning model network layers,complex structure,large storage occupation and low reasoning efficiency.The lightweight model was deployed in the edge side equipment,and the power inspection image front⁃end target detection system was designed to realize the full autonomous inspection and defect automatic acquisition of the robot in the power station.Moreover,it cooperated with the back⁃end server and established a“cloud edge end”collaborative operation inspection system,which can improve the efficiency of power inspection and reduce the cost of inspection,and provide a reference for the promotion and application of power inspection robot.