基于1D-CNN与车轨耦合动力学的路基沉降所引起的轨道非协调变形识别

Identification of Uncoordinated Track Deformation Induced by Subgrade Settlement Using 1D-Convolutional Neural Networks Integrated with Vehicle-Track Coupled Dynamics

摘要: 无砟轨道与路基之间的非协调变形是影响列车运行质量和轨道服役性能的关键因素。本文提出了一种基于一维卷积神经网络(1D-CNN)的非协调变形识别方法，通过分析车辆振动响应实现变形程度的实时评估。首先，建立高速车辆–无砟轨道–路基耦合动力学模型，精确模拟车辆运行时轨道与路基层间结构的非协调变形对列车运行状态产生的影响，计算不同程度的非协调变形对车辆振动响应的影响。基于参数敏感性分析，选取轮对加速度作为特征指标，并构建1D-CNN分类模型，从轨道与路基之间的三种协调变形和五种非协调变形工况中选取加速度数据进行训练。该模型能够自动提取振动信号中的深层特征，对非协调变形程度实现端到端的识别。结果表明，在无噪声干扰条件下，模型识别准确率达95.87%；在5%和8%高斯噪声条件下，准确率仍可保持在94.89%和85.66%，展现出较强的鲁棒性。

Abstract: Identifying the uncoordinated deformation between ballastless track and subgrade caused by subgrade settlement has become a key technology to ensure the quality of train operation and the performance of the track. This paper presents a model for identifying the degree of uncoordinated deformation between ballastless track and subgrade. It is based on a one-dimensional convolutional neural network. This study develops a high-speed vehicle-ballastless track-subgrade coupled dynamic theoretical model that efficiently and accurately simulates the impact of uncoordinated deformation in the layered composite structure of track and subgrade on the train’s operational state. The model calculates the vehicle vibration response caused by uncoordinated deformation under different degree of defects. Through analysis, the wheelset acceleration is identified as a sensitive indicator. This study then constructs a comprehensive one-dimensional convolutional neural network model. It is trained using acceleration data extracted from three conditions of coordinated deformation and five conditions of uncoordinated deformation between the track and subgrade. This approach enables real-time processing of wheel vibration response data and the automatic output of the degree of uncoordinated deformation between the ballastless track and the subgrade. The identification accuracy of this model can reach 95.87%. Moreover, the recognition accuracy was 94.89% and 85.66% under 5% and 8% Gaussian noise conditions, respectively, demonstrating strong robustness.