Machine Learning to Predict Railway Infrastructure Defects

The rapid advancement of the digital revolution has propelled machine learning (ML) to the forefront of transformative technologies within the transportation sector. However, exploring its potential for predictive maintenance in railway infrastructure is still burgeoning. This research paper delves into this nascent field, probing the trajectory of ML applications in railway maintenance and the effectiveness of various ML models in asset and equipment condition prediction. Employing a Literature Review (LR) and Systematic Literature Review (SLR)—which incorporates both the Delphi method and expert inputs alongside PRISMA-guided, AI-enhanced screening—this study aims to distill the influence of data quality, diversity, and specificity on ML efficacy in this context. The findings indicate a positive trend towards integrating ML, particularly highlighting the increasing reliance on artificial neural networks (ANN) and LSTM models for asset condition forecasting. Models like Random Forest and Gradient Boosting remain prevalent, underscoring the necessity to tailor approaches to the unique requirements of each railway component. The study underscores that while tracks are the primary focus of current research, significant potential lies in extending these methodologies to other critical assets, including signaling systems and catenaries. Data quality emerges as a pivotal factor; subpar data significantly impairs predictive accuracies, while data tailored to specific networks poses challenges to model generalizability. Identified gaps span data variability, model validation on actual datasets, parameterization, fault categorization, integration into maintenance planning, operational safety, and cost analysis. These insights call for concerted efforts to bridge these gaps, paving the way for robust, context-adaptable ML solutions in railway maintenance.