基于深度学习的高铁接触网顶紧螺栓的缺陷检测

针对高铁接触网顶紧螺栓目标小、缺陷检测困难的问题,提出一种基于深度学习的顶紧螺栓缺陷检测方法。根据4C检测车拍摄的高铁接触网图片(大小为6600*4400 pixels),首先对特征信息更多的斜撑套筒进行定位,采用TDM模块与SSD相结合的算法提升算法对小目标的检测精度,并通过改变默认框的尺寸以得到更好的检测精度和速度;然后利用DeepLab v3 plus算法对顶紧螺栓部分进行语义分割;最后提出一种阈值法对顶紧螺栓的缺陷情况进行判别。为满足实际工程的速度需求,对训练好的模型进行优化。实验结果表明:相较于经典的SSD,本文改进的SSD方法对斜撑套筒的定位精度和速度都有提升;对于6600*4400 pixels的原始样本,本文提出的顶紧螺栓缺陷检测方法精度上达到95.9%,速度达到17.9 fps。

Defect detection of the puller bolt in high-speed railway catenary based on deep learning

To deal with the defect detection of the puller bolt in the high-speed railway catenary,a new method based on Deep Learning was proposed to detect the detects of puller bolts.The high-speed railway catenary images(size is 6600*4400 pixels)were taken by the 4C inspection vehicle.Firstly,the slanting sleeve with more characteristic information was located,and the detection precision of the small target was improved by using the TDM module and the SSD Algorithm,got a better detection accuracy and speed by changing the size of the default box.Then,using the DeepLab v 3plus algorithm to complete the semantic segmentation of the Bolt Part.Finally,a threshold value method was proposed to judge the defects of the Bolt.The trained model was optimized to meet the actual engineering requirements.Compared with the classical SSD,the improved SSD can improve the precision and speed of the positioning of the diagonal sleeve.For the original sample of 6600*4400 Pixels,in this paper,a method for detecting the defects of Jacking Bolts was presented,which has the precision of 95.9%and the speed of 17.9 fps.