The ballast bed constitutes the cornerstone of the ballasted track. A fouled ballast bed poses a significant threat to its performance, potentially resulting in severe consequences. In recent years, studies have shown that infrared thermography (IRT) technology has emerged as a promising method for detecting the fouled ballast bed. The surface temperatures of clean and fouled ballast beds differ because of their distinct thermodynamic properties. To effectively utilize temperature for identifying the fouled ballast bed, it is essential to accurately predict surface temperatures for two threshold levels of fouling in real-time. To address the issue, this paper proposes an improved BiGRU model (CBGA), and the main contributions are as follows. First, a formula for the surface heat flux density of the ballast bed was derived. In conjunction with existing research findings, the key factors affecting its surface temperature were identified as inputs to the neural network, including the solar radiation intensity, air temperature, wind speed, and humidity. Next, thermodynamic finite element models were established based on a field experiment, which can be utilized to expand the sample library. Leveraging this groundwork, 430 days of temperature data and meteorological data were acquired to train the neural network. Before inputting data into the BiGRU model, CNN and Attention mechanisms were employed to extract local and significant features. Furthermore, a residual network was introduced to ensure the model’s performance. It exhibits superior performance compared to other models. Subsequently, the CBGA model was used to study the impact of different time steps on the prediction accuracy. It was found that a time step of 660 minutes resulted in the best predictive performance. At this time, the evaluation indicators on the testing dataset were: MSE = 0.057, RMSE = 0.238, MAE = 0.168, and MAPE = 0.008. Finally, the reliability and feasibility of the CBGA model were validated using experimental data. These findings demonstrate that the proposed method can achieve real-time prediction of the ballast surface temperature, laying a solid foundation for the practical application of IRT technology in railway maintenance.

中文翻译：

---

使用红外热成像 （IRT） 观察和识别结垢的压载床：基于增强型 BiGRU 模型的实时温度预测研究


有砟床构成了有砟轨道的基石。结垢的道砟床对其性能构成重大威胁，可能导致严重后果。近年来，研究表明，红外热成像 （IRT） 技术已成为检测污染镇流床的一种很有前途的方法。清洁和结垢的压载床的表面温度不同，因为它们具有不同的热力学特性。为了有效地利用温度来识别结垢的压载床，必须实时准确预测两个阈值水平的表面温度。针对这一问题，本文提出了一种改进的BiGRU模型（CBGA），主要贡献如下。首先，推导了镇流床表面热通量密度的公式。结合现有的研究结果，确定了影响其表面温度的关键因素作为神经网络的输入，包括太阳辐射强度、空气温度、风速和湿度。接下来，基于现场实验建立了热力学有限元模型，可用于扩展样品库。利用这一基础，我们获得了 430 天的温度数据和气象数据来训练神经网络。在将数据输入到 BiGRU 模型之前，采用 CNN 和 Attention 机制来提取局部和重要特征。此外，还引入了残差网络以确保模型的性能。与其他型号相比，它表现出卓越的性能。随后，使用 CBGA 模型研究不同时间步长对预测精度的影响。结果发现，660 分钟的时间步长产生了最佳的预测性能。 此时，测试数据集上的评价指标为：MSE = 0.057、RMSE = 0.238、MAE = 0.168 和 MAPE = 0.008。最后，使用实验数据验证了 CBGA 模型的可靠性和可行性。这些结果表明，所提方法可以实现道砟表面温度的实时预测，为 IRT 技术在铁路养护中的实际应用奠定了坚实的基础。