AI-driven predictive maintenance for connected vehicles using environmental context integration evaluated through simulation benchmarking and field validation

Abstract Predictive maintenance for connected vehicles can reduce unexpected breakdowns and lower maintenance costs, but most existing systems rely solely on internal diagnostic signals and are validated on simulated or industrial data. This paper presents a contextual data fusion framework that combines vehicle-internal sensor streams with external environmental signals–road quality, weather, traffic density, and driver behaviour–acquired via third-party APIs and Vehicle-to-Everything (V2X) communication, with inference at the vehicle edge. The framework is evaluated across four complementary layers. A feature group ablation study on a physics-informed synthetic dataset shows contextual features contribute a 2.6-point F1 improvement (macro F1: 0.855 vs. 0.807 internal-only). Benchmarking on the AI4I 2020 dataset (10,000 samples) yields LightGBM AUC-ROC of 0.973 under 5-fold stratified cross-validation with SMOTE confined to training folds. A noise sensitivity analysis shows macro F1 remains above 0.88 under moderate noise, degrading to 0.74 at high noise. Field validation on five heterogeneous vehicles across three countries (India, Germany, Brazil), comprising 992 trips and 11 evaluable service events, demonstrates 100% detection of six wear-driven events (mean MAE 12.2 days; SD $$= 9.8$$ days; range: 4.2−30.5 days); per-vehicle fine-tuning reduces MAE from 25.9 to 12.2 days while binary detection is already present in the base model. SHAP analysis identifies API-sourced contextual features among the top 15 predictors; however, their marginal contribution in real field data could not be independently quantified from 11 events–the SHAP evidence is derived from the synthetic dataset and constitutes a simulation-domain finding. Edge-based inference reduces estimated latency from 3.5 to under 1.0 s.