基于智能图像识别的轨道交通钢轨焊接接头识别

Identification of Rail Transit Rail Weld Joints Based on Intelligent Image Recognition

目的:轨道交通钢轨焊接接头作为钢轨上最为薄弱的结构,易发展为钢轨焊接接头不平顺,从而加剧轮轨间作用力,影响乘客乘坐舒适性,因此有必要对钢轨焊接接头进行识别。方法:介绍了钢轨焊接接头识别中所采用的AlexNet CNN(卷积神经网络)模型和小波变换的原理,以及轨道冲击指数的计算方法。提取北京地铁19号线某运营列车的轴箱振动加速度信号,通过小波变换提取其时域及频域特征,并采用AlexNet CNN进行钢轨焊接接头的识别。对其钢轨表面不平顺进行检测,计算其滑动峰峰平均值,并与同一里程处根据列车轴箱振动加速度计算得到的轨道冲击指数进行对比,采用线性拟合的方式分析钢轨焊接接头对轨道冲击指数的影响。结果及结论:钢轨焊接接头在列车轴箱振动加速度信号中存在较为明显的空间域分布特征与时频特征:空间域分布特征表现在每两个钢轨焊接接头信号之间的间距为25 m,且每个信号由间距为2.2 m的两个尖峰所构成;时频特征表现为钢轨焊接接头所在里程附近有一明一暗两个亮点。这种清晰的特征使AlexNet CNN高效地对钢轨焊接接头进行识别,准确率在92.98%左右。对钢轨表面不平顺实测值进行30~100 mm带通滤波后可发现与钢轨焊接接头空间分布特征相一致的特征峰值,且轨道冲击指数与钢轨焊接接头对应的钢轨表面不平顺滑动峰峰平均值的峰值具有相似的空间分布特征;钢轨焊接接头处的轨道冲击指数随钢轨焊接接头不平顺的滑动峰峰平均值增大呈现先减小后增大的趋势,二者并非简单的正相关关系。

Objective:RWJ(rail weld joints)in rail transit systems are the weakest structural points on rails,which are prone to develop into RWJ irregularities,exacerbating wheel-rail interaction forces and affecting passenger comfort.Therefore,it is necessary to conduct identification of RWJ.Method:The principles of the AlexNet CNN(convolutional neural network)model and wavelet transform used in the identification of RWJ are introduced,along with the calculation method of rail impact index.The axle box vibration acceleration signals of a train operating on Beijing Subway Line 19 are extracted,then time-domain and frequency-domain features are extracted through wavelet transform,and RWJ are identified using AlexNet CNN.The rail surface irregularities are detected,its sliding peak-to-peak average value is calculated,and it is compared with the rail impact index obtained from the train axle box vibration acceleration at the same mileage.The impact of RWJ on the rail impact index is analyzed using linear fitting.Result&Conclusion:RWJ exhibit distinct spatial domain distribution and time-frequency features in the train axle box vibration acceleration signals:the spatial domain distribution is characterized by a spacing of 25 meters between consecutive rail weld joint signals,each signal is composed of two peaks with a spacing of 2.2 meters.The time-frequency features manifest as two bright spots near the mileage of the RWJ,one brighter and one dimmer.These clear features enable efficient identification of rail weld joints using AlexNet CNN,achieving an accuracy of around 92.98%.After applying 30-100 mm band-pass filtering to the measured rail surface irregularities,it is observed that characteristic peaks are consistent with the spatial distribution features of all RWJ,and the rail impact index and the sliding peak-to-peak average value of rail surface irregularities at corresponding RWJ share similar spatial distribution features.The rail impact index at RWJ exhibits a trend of decreasing followed by increasing as the sliding peak-to-peak average value of RWJ irregularities increase,indicating a complex relationship between the two factors rather than simple positive correlation.