PSI - Issue 57

Olivier Vo Van et al. / Procedia Structural Integrity 57 (2024) 104–111 O. VoVan / Fatigue Design 2023 00 (2023) 000–000

110

7

that it deteriorates the model, in this case, the variable describing acceleration. However, one notable limitation of this method is that if two variables are entirely redundant, the importance score will be the same for both variables in models that include either one. Consequently, the importance of both variables will be reduced to zero. The compar ative importance of tonnage and temperature damage is questioning. Classical size for stress in the wheel / rail contact patch is of 500MPa while the stress induced by temperature daily cycles rarely reach more than 50MPa. The high im portance of temperature could be explained by a combined e ff ect of both phenomena, only attributed to temperature damage. SHAP analysis provides numerous possibilities beyond the scope of this presentation, including the assessment of interactions between variables and targets. However, accurately calculating these interactions is known to be challeng ing, and the TreeShap method serves as an approximation for this purpose.

Fig. 6: Individual explanation for SHAP importance. In this figure, each point represent one rail segment while the color stands for the value of feature. The abscissa axis is the impact of the feature in the prediction. The reading of this plot shows that temperature damage has a positive impact on prediction, while the cant has a negative, yet significant, impact on the prediction.

4. Conclusions

In this paper, a mechanically representative model was built to consider fatigue damage due to temperature vari ations in a data based model. While taking special precautions in the sensitivity analysis steps, results showed two remarkable points. First, stress cycles with amplitude lower than the fatigue limit contribute significantly to the RCF. Muhamedsalih [16] observed in his paper that recent rails have lower squat rates than older ones. Several explanations were proposed and the temperature damage can now appears to be another significant one. Then, daily variations have more impact than seasonal variations. For expert knowledge, this result is counter intuitive because rail breaks are more frequently observed during very cold days of winter. These result show that crack initiation is not sensitive the same sizes than crack propagation, the latest being the phenomenon leading to rail break.

5. Perspectives

While strong hypotheses were taken for computing temperature damage, the proposed methodology combining physical and data models is a promising way to unveil weak signals. The next step will be to add data from thermal imager as shown in [5] or refine thermal models as proposed by [15] and implement it in the machine learning scheme. Physical model can even be perfected by combining thermal and mechanical stresses and computing the resulting Von Mises stress on which one applies a damage model.