关键词： 管道巡检机器人, 深度学习, 目标检测, 卷积网络, 管道

Abstract: To solve the problem of low accuracy of pipeline inspection robot in identifying target sludge inside pipeline, an improved YOLO model based on a new learning rate updating strategy was is proposed. The model adopteds deep learning convolutional network to conduct target image learning and training. For the problem of inaccurate location, a distance location model is proposed to achieve accurate location of target distance. Comparative training tests were conducted on YOLO model (Model A) under the new learning rate updating strategy, YOLO model (Model B) under the constant attenuated learning rate updating strategy, YOLO model (Model C) under the exponential attenuated learning rate updating strategy, and YOLO model (Model D) using the traditional gradient descent method. The results showed that the improved YOLO model (Model A) achieved 96.1% accuracy of target detection with location error less than 2 cm.

Key words: pipeline inspection robot, deep learning, target detection, convolutional network, pipeline operation