基于深度学习的钢轨伤损智能识别方法

Intelligent Detection Method for Rail Flaw Based on Deep Learning

基于钢轨探伤车检测数据通道设计、B显数据生成原理和钢轨伤损分类,对比钢轨探伤车检测数据伤损识别与普通图像识别特点的不同,将检测数据视为由16个通道二进制矩阵叠加成的图像;设计包含1个输入层、3个卷积层、3个池化层、2个全连接层、1个输出层的深度学习架构,并通过噪声和通道预处理,将钢轨伤损的"物体检测"问题转换为"分类"问题。以某地人造钢轨伤损检测数据扩充后作为训练集,得到基于深度学习的钢轨伤损智能识别模型,以另一地的人造钢轨伤损检测数据作为测试数据分析该模型的识别效果,并与钢轨探伤车既有系统识别结果和人工分析结果进行对比。结果表明:基于深度学习的钢轨伤损智能识别模型在准确率、误报率指标上均优于钢轨探伤车既有系统,达到人工分析的指标要求,提高了准确率。

Based on the design of the test data channel of rail flaw detection car,the generation principle of B-scan data and the classification of rail flaw,the differences between the rail flaw identification by rail flaw detection carand the common image recognition are analyzed.The detection data are regardedas the images superimposed by the binary matrices of 16 channels.A deep learning structure is designed which includes 1 input layer,3 convolutional layers,3pooling layers,2 fully connected layers and 1 output layer.By noise and channel preprocessing,the“object detection”problem of rail flaw is converted to the“classification”problem.The artificial rail flaw detection data of some placeare expanded and used as training sets to obtain the intelligent recognition model of rail flaw based on deep learning.The artificial rail flaw detection dataof another place are used as test data to evaluate the recognition effect of themodel.Comparing the model identification results with those obtained from theexisting system of rail flaw detection car and manual analysis,the comparison results show that the intelligent recognition model of rail flaw based on deep learning is superior to the existing system of rail flaw detection car in the accuracy rate and false alarm rate.The model meets the index requirements of manualanalysis and improves the accuracy.