面向巡检机器人的电力设备状态检测算法研究

为更高效利用变电站巡检机器人开展电力巡检工作,满足电力行业发展对智能化巡检的需求,研究了面向电力巡检机器人的电力设备状态检测算法。首先,根据深度网络部署硬件芯片应用情况与性能对比,选择海思Hi3559A芯片作为算法移植的嵌入式平台。然后综合考虑各种检测算法的精度与速度,选用YOLOv3算法作为设备状态检测的基本判别模型。为了提升检测算法速度并减少模型体积,开展模型压缩算法及轻量型YOLOv3模型设计研究,分别提出了改进的小型化YOLOv3模型和基于通道剪枝与层剪枝结合的模型压缩方法,提高模型上下层的语义信息及剪枝后模型的精度保持。根据测试结果选择最优的模型在机器人前端部署,提出的轻量化YOLOv3模型很好地保持了设备目标与异物检测的精度,检测速度提升了4倍。

Research on Algorithm for Equipment Condition Monitoring Based on Inspection Robot

In order to implement power equipment inspection work using substation inspection robot efficiently and meet the demand of the development of electric power industry for intelligent inspection,the status detection algorithm of electric equipment for electric power inspection robot is researched.Firstly,Hi3559A chip is chosen as an embedded development platform to implement arithmetic transplant according to the comparison of properties and application between different hardware chips.Taking the accuracy and speed of various detection algorithms,YOLOv3 algorithm is selected as the basic discriminant model of device state detection.In order to improve the speed of detection algorithm and reduce model size,the model compression algorithm and the lightweight YOLOv3 model are designed and studied.Both improved tiny YOLOv3 and model compression algorithm based on channel pruning and layer pruning are proposed to improve semantic information extraction and holding accuracy of original model.The optimal model is chosen to deploy on the hardware platform according to the test results.The experiment demonstrates that the proposed lightweight YOLOv3 model maintains the detection accuracy of equipment targets and foreign bodies well,and the detection speed is increased by 4 times.