Scholarlink

<p>To accurately predict the effects of freeze–thaw cycles on the crushing strength of different types of railway ballast, this study proposes a high-precision prediction framework based on the Tabular Prior-Data Fitted Network (TabPFN) model, which integrates a Bayesian inference mechanism within a Transformer architecture. Additionally, we utilize the SHapley Additive exPlanations (SHAP) method to analyze model interpretability, providing insights into the key influencing factors and the model’s decision-making process. We trained and validated the model using a multidimensional feature set comprising freeze–thaw cycle count, geometric dimensions, shape indices, and mechanical strength parameters, based on a limited sample dataset. The results demonstrate that the TabPFN model achieved excellent performance, with a mean squared error (MSE) of 0.066, a mean absolute error (MAE) of 0.033, and a coefficient of determination (R<sup>2</sup>) of 0.993 on the training set. On the testing set, it maintained strong generalization, yielding an MSE of 0.240, an MAE of 0.118, and an R<sup>2</sup> of 0.968—significantly outperforming conventional models such as Random Forest (R<sup>2</sup> = 0.935), Extreme Gradient Boosting (R<sup>2</sup> = 0.947), and Support Vector Regression (R<sup>2</sup> = 0.728). Furthermore, the SHAP analysis revealed that the number of freeze–thaw cycles, needle-like index, and flakiness index are the dominant factors influencing ballast crushing strength.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Machine Learning Prediction of Ballast Strength Under Freeze–Thaw Effects

  • 期刊论文

  • 发表日期：

    2025年8月21日

查看全文
  • Xiaohang Tang,
  • Wenrui Bian,
  • Yundong Ma,
  • Zhongchang Wang

摘要

To accurately predict the effects of freeze–thaw cycles on the crushing strength of different types of railway ballast, this study proposes a high-precision prediction framework based on the Tabular Prior-Data Fitted Network (TabPFN) model, which integrates a Bayesian inference mechanism within a Transformer architecture. Additionally, we utilize the SHapley Additive exPlanations (SHAP) method to analyze model interpretability, providing insights into the key influencing factors and the model’s decision-making process. We trained and validated the model using a multidimensional feature set comprising freeze–thaw cycle count, geometric dimensions, shape indices, and mechanical strength parameters, based on a limited sample dataset. The results demonstrate that the TabPFN model achieved excellent performance, with a mean squared error (MSE) of 0.066, a mean absolute error (MAE) of 0.033, and a coefficient of determination (R2) of 0.993 on the training set. On the testing set, it maintained strong generalization, yielding an MSE of 0.240, an MAE of 0.118, and an R2 of 0.968—significantly outperforming conventional models such as Random Forest (R2 = 0.935), Extreme Gradient Boosting (R2 = 0.947), and Support Vector Regression (R2 = 0.728). Furthermore, the SHAP analysis revealed that the number of freeze–thaw cycles, needle-like index, and flakiness index are the dominant factors influencing ballast crushing strength.