基于多尺度深度学习的接触网吊弦异常检测及应用

Detection of Abnormality of OHL Droppers based on Multi-scale Deep Learning and its Application

提出一种基于多尺度深度学习的接触网吊弦异常检测算法,并研究其在受电弓打弓预判中的应用。该算法由深度神经网络提取图像特征,根据特征图确定吊弦图像感兴趣区域并得到候选框,再对候选框进行分类和回归,确定吊弦状态并得到吊弦位置。通过吊弦的松、脱、断等异常状态,进一步预判是否会出现受电弓打弓,从而及时给出预警。通过实际应用验证,该算法可有效对吊弦异常状态进行检测,可提前预判打弓隐患。

The paper proposes a detection algorithm based on multi-scale deep learning for OHL droppers and studies its application in prediction of pantograph striking. The process of the algorithm is: extract the image feature from the deep neural network, determine region of interest of the dropper image and obtain the candidate frame according to the feature mad, classify and reset the candidate frames, determine the status of droppers and obtain the dropper positions. With reference of status of abnormality of looseness, falling off and breakage of the droppers, predict further whether the pantograph striking may occur or not, and alarm may be presented timely if there is any. Through practical application, the algorithm is verified to be able to inspect the abnormality of droppers and is able to predict the hidden risks of pantograph striking.