Deformation of high-speed railway subgrades, due to low temperatures, is a common phenomenon in cold regions. In winter, the uneven frost heave of subgrade soil would cause hazards to train safety. It is therefore necessary to estimate and predict the subgrade properties. Since the variation of frost heave is non-stationary over time, traditional time series analyses have difficulties where complex physical parameters are not available. In this study, we introduce two models based on deep learning technology to predict frost heave deformation of railway subgrade. These include the artificial neural network (ANN) and long-short term memory (LSTM) network, where we used data of four sections to build the ANN and LSTM. The experimental results of the LSTM model provided lower MAE and RMSE with different datasets. The prediction of three deep deformations for the K1959 + 580 and K1962 + 618 section with slight fluctuation in the data and the performance of the ANN with MAE is 0.0090‐–0.0660 and 0.0069‐–0.0201 of the LSTM models. In the K2005 + 948 and K2029 + 829 section, ANN and LSTM estimated the frost heave deformation with MAE of 0.0061‐–0.0681 and 0.0054‐–0.0309 for a more intense fluctuation on the deformation. Our findings suggest that the network topology of the LSTM model with 12 hidden neurons performs best with fewer parameters, with an average RMSE of 0.0210 mm and MAE of 0.0138 for all the training samples, indicating that the deep learning model has high precision in this scenario.

低温导致的高速铁路路基变形是寒冷地区的常见现象。在冬季，路基土壤冻胀不均会危害火车安全。因此，有必要估计和预测路基的性质。由于霜冻随时间的变化是不平稳的，因此传统的时间序列分析在无法获得复杂物理参数的情况下会遇到困难。在这项研究中，我们介绍了两种基于深度学习技术的模型来预测铁路路基的冻胀变形。这些包括人工神经网络（ANN）和长期短期记忆（LSTM）网络，我们使用四个部分的数据来构建ANN和LSTM。LSTM模型的实验结果为不同的数据集提供了较低的MAE和RMSE。L1E模型的K1959 + 580和K1962 + 618截面的三个深度变形的预测随着数据的轻微波动和具有MAE的ANN的性能分别为0.0090-0.0660和0.0069-0.0201。在K2005 + 948和K2029 + 829部分中，ANN和LSTM估计霜冻变形的MAE为0.0061-0.0681和0.0054-0.0309，以使变形波动更大。我们的发现表明，具有12个隐藏神经元的LSTM模型的网络拓扑在参数较少的情况下表现最佳，所有训练样本的平均RMSE为0.0210 mm，MAE为0.0138，这表明深度学习模型在这种情况下具有较高的精度。LSTM模型的0201。在K2005 + 948和K2029 + 829部分中，ANN和LSTM估计霜冻变形的MAE为0.0061-0.0681和0.0054-0.0309，以使变形波动更大。我们的发现表明，具有12个隐藏神经元的LSTM模型的网络拓扑在参数较少的情况下表现最佳，所有训练样本的平均RMSE为0.0210 mm，MAE为0.0138，这表明深度学习模型在这种情况下具有较高的精度。LSTM模型的0201。在K2005 + 948和K2029 + 829部分中，ANN和LSTM估计霜冻变形的MAE为0.0061-0.0681和0.0054-0.0309，以使变形波动更大。我们的发现表明，具有12个隐藏神经元的LSTM模型的网络拓扑在参数较少的情况下表现最佳，所有训练样本的平均RMSE为0.0210 mm，MAE为0.0138，这表明深度学习模型在这种情况下具有较高的精度。