基于轻量级卷积神经网络的地铁轨道线路状态检测系统

Track line status detection system for subway based on lightweight convolutional network

轨道作为承载车辆运行的重要部件,其工作状态对地铁运营的安全性有重要影响。传统人工巡检或者采用轨检车的检测模式只能在正线停运后进行作业,工作效率低。针对该问题,提出一种基于地铁运营列车的轨道线路状态检测系统,该系统采用高速线扫相机对轨道线路进行实时图像采集,并将采集数据送入轻量级RegNet骨干神经网络提取图像深层特征。在此基础上,加入双向特征金字塔网络进行多层次特征融合。最后将融合特征输入目标检测头实现轨道线路状态的实时检测。结合基于云边协同的困难样本挖掘以及模型部署加速技术,算法可实现高准确率、高实时性的检测性能。试验表明,该系统针对11类钢轨伤损及扣件状态的检测平均准确率(mAP)达到0.951,推理速度大于20 f/s,满足地铁在载客运营同时对轨道线路状态进行实时检测的需求。

As an important component to carry running train,the working state of the track has an important impact on the safety of the subway operation.Traditional manual detection or rail inspection vehicles can not operate until subway operation ends,leading to low work efficiency.Aiming at this problem,an track line status detection system based on subway passenger trains was proposed.A high frequency line-scanning camera was used to collect real-time images and a lightweight backbone RegNet was used to extract the deep features of track line images.Then,a BiFPN was added to fuse multi-layer features.Finally,the features were send to a object detection head for track line status real-time detection.Combined with the hard sample mining based on cloud-edge collaboration and model acceleration technologies,the algorithm can meet the detection performance of high accuracy and real-time.The experiments show that the mean average precision of detection for 11 types of rail failure and fastener status reaches 0.951,and the inference speed is greater than 20 f/s,which proves that the system can realize real-time track line status detection during the subway passenger operation.