基于物理模型‒数据驱动混合的钢轨滚动接触疲劳裂纹发展预测

Prediction of Rail Rolling Contact Fatigue Crack Growth Based on Physical Model-data Driven Hybrid

摘要:

受轮轨反复作用，钢轨滚动接触疲劳裂纹沿轨面发展，呈现物理萌生、随机扩散和扩展的分布过程。为预测裂纹发展随机过程，提出一种物理模型与数据驱动混合的裂纹发展预测方法。将裂纹发展过程的萌生、扩散中的随机参数引入裂纹发展物理模型，采用近似贝叶斯推断算法估计上述参数，并结合神经网络进行数据训练，最后通过物理机制与随机数据驱动的融合建立裂纹扩散与线路和运营条件的关系。结果表明，在曲线半径R一定时，裂纹扩散率λ与日通过总重G_d呈正相关，且R<650 m时G_d对λ影响显著；G_d对初始裂纹倾角θ影响微弱，而R与θ呈负相关；R减小或G_d增大均会显著提前钢轨裂纹扩散起始时间T₀。

Abstract:

Under repeated wheel-rail loading， rolling contact fatigue cracks in rails propagate along the rail surface， exhibiting a distributed evolution process involving physical initiation， stochastic diffusion， and subsequent growth. To predict the stochastic process of crack evolution， a hybrid crack development prediction method integrating physics-based modelling and data-driven approaches is proposed in this paper. Random parameters associated with crack initiation and diffusion are incorporated into the physical model of crack evolution. These parameters are estimated using an approximate Bayesian inference algorithm， and neural networks are further employed for data training. On this basis， a relationship between crack diffusion and track operational conditions is established by integrating physical mechanisms with stochastic data-driven modelling. The results show that， for a given curve radius R， the crack diffusion rate λ is positively correlated with the daily gross tonnage Gd， and the effect of Gd on λ becomes significant when R<650 m. In contrast， Gd has only a weak effect on the initial crack inclination angle θ， whereas R is negatively correlated with θ. Moreover， a decrease in R or an increase in Gd significantly advances initial crack diffuse time T0.