基于1D-CNN的无砟轨道CA砂浆脱空识别

CA mortar void identification for ballastless track using 1D-CNN

为解决水泥乳化沥青砂浆(CA砂浆)脱空位置识别的难题,提出考虑CA砂浆脱空的无砟轨道模型建模与计算方案。首先,基于CRTS I型板式无砟轨道“梁-板-板”建模理论,使用非线性弹簧模拟CA砂浆脱空病害,输入扣件支点反力作为模型的激励,得到监测点位的垂向加速度数据集。其次,通过对数据集缩放、添加噪声的方式进行数据增强,增强数据集的差异,以增强识别模型的泛化性,获得更加准确的识别结果。最后,建立脱空病害的识别方法,搭建一维卷积神经网络(1D-CNN)进行CA砂浆的脱空位置识别,构建准确率、查准率和查全率等评价指标评估识别结果,使用t-SNE降维算法对结果进行可视化。对比其他3种神经网络的表现,探讨1D-CNN应用于CA砂浆脱空位置识别的优势。研究结果表明:“梁-板-板”模型反映了无砟轨道的力学本征,用于CA砂浆脱空问题的仿真模拟是可行的;通过对数据集信号缩放、添加噪声的方式进行数据增强可提高深度学习模型的性能;从深度学习的角度来看,CA砂浆脱空发生在板中造成的危害较小,应重点关注板端CA砂浆脱空的养护与维修问题;1D-CNN相比其他3种模型运算时间短,在CA砂浆脱空位置识别数据集上优势更大,识别准确率可达95%以上。研究结果为进一步提高无砟轨道结构监测的自动化和智能化水平提供参考。

In order to solve the challenge of identifying the location of void in Cement-emulsified Asphalt mortar(CA mortar),a modeling and calculation scheme was proposed for ballastless tracks considering the CA mortar void.Based on the CRTS I slab ballastless track beam-shell modeling theory,a non-linear spring was used to simulate CA mortar void damage,and the fastener forces were input as the excitation of the model to obtain the vertical acceleration dataset at the monitoring points.Second,data augmentation was performed by scaling the sequences and adding noise to enhance the differences in the dataset.These approaches aimed to improve the recognition model to obtain more accurate recognition results.Finally,a method for the identification of voids was established.A one-dimensional Convolutional Neural Network(1D-CNN)was built to identify the location of CA mortar voids.Evaluation metrics such as accuracy,recall,and precision were constructed to assess the identification results and visualize the results using the t-SNE dimensionality reduction algorithm.The advantages of 1D-CNN applied to CA mortar void recognition were discussed by comparing the performance of the other 3 neural networks.The results show that the beam-shell model reflects the mechanical intricacies of the ballastless track and is feasible for the simulation of the CA mortar void problem.Data augmentation by scaling the dataset signal and adding noise can improve the performance of the deep learning model.From a deep learning perspective,the CA mortar void in the middle of the slab is less damaging and should focus on maintaining the CA mortar void at the slab end.1D-CNN has a shorter running time compared with the other 3 models and has a greater advantage in the CA mortar void location identification dataset,with an identification accuracy of more than 95%.The results of the study provide a reference for further improving the automation and intelligence of ballastless track structure monitoring.