Session: 12-06-01: Scientific Machine Learning (SciML) for Characterization, Modeling, and Design of Structures and Materials

Paper Number: 150125

150125 - Physics-Based Machine Learning Model for Predicting Rail Failures Due to Internal Defects 

This research, sponsored by the Federal Railroad Administration, details the development of physics-based Machine Learning (ML) predictive model of rail failures due to internal defect growth. Many rail breaks occur every year in the US railroad network due to internal defects. The railroads adopted many strategies to minimize occurrence of these failures. Strategies include increased rail testing, rail grinding and improved rail steels. These strategies are resource intensive. The goal of this research was to develop a comprehensive predictive method that considers multiple external variables and underlying physical mechanisms for internal defect growth to improve effectiveness of these strategies. Such a model would allow improvements in preventive maintenance and would have direct effect on railroad safety.

This research investigated physical contributors such as rail age with external factors such as rail temperature difference, traffic density as well as track and rail condition to predict time to failure from known detail fracture defect size. ENSCO, Inc. (ENSCO) successfully simulated the fracture behavior of detail fractures and contributing factors using Ansys and FRANC3D modeling software. The stress intensity factors, and the J-integral were used to compare the results of different contributing factors for the static simulation. Remaining lives were calculated based on the Paris' law for detail fracture growth analysis.

Furthermore, this research investigated suitability of various ML models and developed a multi-variant predictive model for detail fracture defect growth. ENSCO partnered with a Class I railroad that provided rail flaw inspection data with precise geolocation and the associated foot-by-foot track geometry and rail profile measurements to develop the approach. Regarding model selection, regression-type models were preferred to classification models due to the need for continuous estimations of rail detail fracture growth rate. Neural network models were also considered but classical machine learning models generalized better for the available small dataset. Linear models performed better compared to non-linear ML models. Lasso Regression with Cross-Validation (LassoCV) model was selected for the final life predictions due to the lower errors, less overfitting (using L1 regularization) and more generalization.

Both the multi-variant ML model and fracture mechanics model predicted similar trends in predicted life for various contributing factors like rail temperature difference and track curvature, confirming previous findings in the literature. The results revealed that the rail temperature difference parameter was the most significant contributor among the investigated factors. However, both models overestimated the remaining life in MGTs compared to previous research due to assumptions and limitations in the available dataset. The available dataset did not include repeated measurements of defect size at the same defect locations or inspection frequencies. The performance of the multi-variant model would be vastly improved if additional rail defect dataset would become available. Including residual stress in the fracture mechanics model in addition to a more detailed rail model would increase the accuracy of the physical model.

Presenting Author: Mohamad Ghodrati ENSCO

Presenting Author Biography: Mohamad Ghodrati is a Senior Engineer at ENSCO. He earned his PhD in Mechanical Engineering from Virginia Tech in 2020 and joined ENSCO after graduation. He primarily works on railroad R&D projects focused on structural analysis and data science. His areas of expertise are finite element analysis, mechanical fatigue, and data science. He lives in Portland Metro Area, so feel free to ask for recommendations of places to visit!

Authors:

Mohamad Ghodrati ENSCO
Rakan Alturk ENSCO