基于卷积神经网络的钢轨测量廓形畸变动态识别

针对车体多自由度振动对基于激光图像技术的钢轨廓形动态测量所造成的影响,提出一种新颖的钢轨测量廓形畸变识别方法.首先根据钢轨廓形特征和畸变前后的几何差异,设计了一种三通道且参数独立的卷积神经网络结构用于畸变识别,其输入分别为原始廓形图像的降采样、轨鄂点周边裁剪图像和轨底点周边裁剪图像.为了有效训练该网络,通过采集大量正常廓形图像和畸变廓形图像来构建带标签训练样本库.利用训练后的卷积神经网络,在室内钢轨廓形动态测量平台上进行大量的测量廓形畸变识别实验.实验结果表明本文识别方法的精度和查全率均能达到92%以上,验证了该方法的有效性和可靠性.

Dynamic Recognition of Deformation on Measured Rail ProfileBased on Convolutional Neural Network

This paper presents a novel method of recognizing deformation on the measured rail profile based on laser image system,which is caused by multiple degrees of freedom vibration on the vehicle.According to the characteristic and distinction of the normal profile and distorted one,a convolutional neural network with triple channels,whose parameters are independent,is designed to recognizing deformation.The down samples of original profile image,cropped images around the rail jaw and rail foot are taken as input for each channel,respectively.To train the network effectively,a dataset consisting of labeled samples is established via capturing a large amount of normal and distorted profiles.Utilizing the trained convolutional neural network,a comprehensive recognizing experiment was implemented on the indoor simulated experiment platform for dynamically measuring rail profile.The results show that the proposed method achieves high precisions and recalls more than 92%,which verifies its effectiveness and reliability.