Predictive Maintenance for Heavy-Duty Truck Fleets: An Explainable Machine Learning Approach on the SCANIA Component X Dataset

Abstract

Component failures in heavy-duty truck fleets generate substantial operational costs and safety risks. This study develops a cost-optimized predictive maintenance system using real-world heavy-duty truck telemetry. A calibrated gradient boosting model achieves a 45% reduction in operational maintenance cost and 86% failure recall. The primary contribution lies in configuration-stratified explainability analysis using a game-theoretic attribution framework. We demonstrate that vehicle specification groups exhibit mechanistically distinct failure pathways, with feature importance rankings diverging significantly across categories. Global predictors include sensor variance and usage intensity metrics, but stratified analysis reveals specification-dependent degradation signatures previously unreported. The observed precision ceiling reflects inherent data constraints rather than methodological limitations. These findings challenge the homogeneous fleet assumption and suggest that specification-aware alerting could substantially improve maintenance efficiency. This work establishes stratified explainability as a foundation for knowledge-driven, targeted predictive maintenance in industrial fleet management.