Implementing predictive maintenance with AI for Moroccan railway freight networks: a case study on rails and critical components
Taoufiq El Moussaoui, Alaa Eddine El MoussaouiPurpose
This study develops and evaluates an artificial intelligence–driven predictive maintenance framework tailored to the operational challenges of Moroccan railway freight networks. These challenges include heterogeneous infrastructure conditions, coexistence of legacy and modern systems, limited real-time monitoring capabilities and increasing freight demand along strategic logistics corridors. The objective is to enhance infrastructure reliability, improve safety, and reduce maintenance costs by anticipating failures in rails, switches, and critical components before severe degradation occurs.
Design/methodology/approach
A multimodal hybrid predictive architecture was developed by integrating structured gradient boosting models (XGBoost) for operational and maintenance data, long short-term memory (LSTM) networks for temporal degradation analysis and convolutional neural networks (CNN) for automated defect detection from rail imagery. Model outputs were combined using a calibrated probabilistic fusion mechanism to generate reliable failure risk estimates. The framework was trained using a five-year dataset from 127 high-traffic Moroccan rail segments, incorporating maintenance logs, IoT sensor streams, environmental indicators, operational loads and annotated inspection images.
Findings
The proposed multimodal ensemble achieved strong predictive performance, reaching an AUC-ROC of 0.963 and recall of 0.921, outperforming baseline statistical and single-modality machine learning models. Deployment simulations demonstrated measurable operational benefits, including a 31% reduction in unscheduled downtime, a 28% decrease in emergency repairs, and a 38% increase in mean time between failures. Economic analysis further indicated an 18% reduction in total maintenance costs. The system maintained stable performance under increased sensor noise and traffic loads, confirming robustness under realistic railway operating conditions.
Originality/value
This study presents a context-adapted predictive maintenance framework specifically designed for emerging railway environments characterized by infrastructure variability and partial digitalization. The originality lies in integrating XGBoost, LSTM and CNN models within a probabilistically calibrated multimodal fusion architecture capable of producing reliable risk estimates for safety-critical operations. By explicitly addressing Moroccan railway constraints and demonstrating quantifiable operational and economic improvements, the research provides both methodological innovation and practical guidance for implementing AI-driven maintenance strategies in developing freight rail systems.