PROGRAM | Civil Engineering

By: Kenza Soufiane Chair: Allan M. Zaremsbki

ABSTRACT

Understanding the deterioration behavior of crossties (also known as railroad ties or sleepers) is of paramount importance due to the major safety concerns associated with their failure. Studies have been conducted to assess ties degradation and their maintenance planning considering different parameters: tonnage, climate conditions, type of tracks, frequency of trains, and degree of curvature, among others. However, very few studies determined the relation between a crosstie’s degradation rate or failure and that of its adjacent ties.

Degradation rates of adjacent ties are interconnected and interact in a dynamic manner over time. On track, ties have different conditions which leads to an imbalanced load distribution: as crossties deteriorate, more load is exerted on their adjacent ties, causing them to degrade faster. Consequently, as these adjacent ties deteriorate as well, their capacity to withstand additional loads diminishes, resulting in the transfer of more load to nearby ties, accelerating their degradation. The work performed in this PhD thesis investigates this complex dynamic interaction between changing crosstie condition over time (through degradation and replacement) and the support it provides to adjacent crossties. In fact, the objective is to use a theory-driven machine learning framework incorporating real data to understand, model, and predict cross-tie degradation behavior and the corresponding cross-tie life under actual in-field conditions.

Using different machine learning techniques, tie degradation trends were investigated to not only determine the dynamic impact of adjacent support condition on a center tie, but to also quantify the load distribution effect over time. Then, using automated tie inspection data, theory was integrated into the machine learning framework using three key approaches: constructing the model’s architecture in accordance with a comprehensive mathematical formulation to ensure it adheres to tie degradation mechanisms; designing a loss function that aligns with the underlying tie degradation trends, in addition to theory-guided calibration of the tie condition input data. The established framework allowed for the use of real tie data, supported by well-defined railroad engineering relationships, to forecast the tie condition as a function of its adjacent ties and their corresponding degradation rates over time.

Theory Guided Machine learning models were developed with the domain knowledge from an engineering-based Beam-On-Elastic-Foundation track model with variable local stiffness. The models were able to determine the load variation and associated rate of degradation on a given cross-tie, based on the condition of both the cross-tie itself and that of the adjacent ones. Using three years of railroad cross-ties inspection data, the resulting models forecast  the fourth-year condition.

The models showed excellent correlation with the test data set, exhibiting strong performance indicators, and outperforming more conventional traditional neural networks. They also accurately represented tie degradation behavior and effectively captured the complex dynamics introduced by tie replacements, suggesting  that the incorporation of domain knowledge into the machine learning model leads to demonstrably better tie life prediction results.

Back >