面向变电站机器人的仪表检测与场景理解技术

Instrument detection and scene understanding technology for substation robot

为提升变电站巡检机器人巡检效率和环境适应能力,将深度学习算法应用于变电站巡检机器人仪表检测和道路场景理解中,提出一种多视觉任务交替实现的轻量级卷积神经网络。该网络由骨干结构、控制模块、任务分支3部分串联组成,通过改进的Inception结构结合注意力模型提取图像特征信息,引入基于分类思想的控制模块实现仪表检测和场景理解支路交替运行,使网络充分利用平台计算资源,避免对无效信息的处理。实验结果表明,所提网络与传统网络相比,其精度与效率都有较大的提升,同时,在实际变电站场景中,该网络也体现出更高的适应性,可以更好辅助机器人完成巡检任务。

To improve the inspection efficiency and environment adaptability of the inspection robot in substation,a deep learning method was applied to instrument detection and road scene understanding of substation inspection robots,and a lightweight convolution neural network with multiple vision tasks alternately was proposed.The network was composed of backbone structure,control module and task branch in series.The improved inception structure was combined with the attention model to extract the image feature information,and the control module based on the classification idea was introduced to realize the alternate operation of the instrument detection and scene understanding branches,so that the network could make full use of the platform computing resources and avoid the processing of invalid information.Experimental results show that the network proposed has greater improvement in accuracy and efficiency compared to traditional networks.At the same time,in the actual substation scene,the network also shows higher adaptability,which can better assist the robot to complete the inspection task.