Forecasting cross-tie condition based on the dynamic adjacent support using a theory-guided neural network model

Cross-ties represent a key infrastructure asset of the railroad industry. Recent research has shown that the cross-tie life is not only affected by the traditionally defined load and track design parameters but also by support condition, and in particular, support condition as represented by the condition of adjacent cross-ties. This paper builds upon the recent research and is focused on predicting a cross-tie’s future condition as a function of the changing condition of the surrounding cross-ties. As accurate cross-tie condition information becomes available from automated inspection systems, this data allows for the development of a theoretical framework for predicting cross-tie degradation and corresponding cross-tie life. This theoretical framework allows for the modeling of the interactions between adjacent cross-ties as a complex and dynamic system. Thus, the objective of this paper is to develop a model that uses theory guided machine learning framework as supported by well-defined railroad engineering relationships, such as the Beam on Elastic Foundation theory, to forecast the cross-tie condition as a function of its adjacent cross-ties and their corresponding degradation rates. The resulting model outperformed a more conventional traditional neural network model. The theory guided machine learning model showed very good correlation with actual data exhibiting an R² of 88.6% and an a²⁰-index of 91% suggesting that the incorporation of domain knowledge into the machine learning model leads to demonstrably better cross-tie life prediction results.