Data-Driven Fault Detection and Feature Selection from CAN-Bus Diagnostics in Commercial Vehicle

Mitigating unplanned breakdowns in commercial vehicle fleets is critical for operational efficiency, yet challenging due to the complexity of diagnostic data. This study presents a predictive maintenance framework designed for control systems, utilizing fault codes transmitted via the J1939 CAN Bus protocol. We analyzed Diagnostic Message 1 (DM1) structures, specifically Suspect Parameter Numbers (SPNs) and Failure Mode Identifiers (FMIs), to extract interpretable features for real-time applications. The research utilizes a dataset of operational telemetry and diagnostic logs from 120 internal combustion buses over two years. We developed a structured feature engineering pipeline that incorporates temporal alignment, correlation analysis, and XGBoost-based feature weighting to track mechanical fault progression, with a specific focus on engine oil pressure events. To improve model interpretability, we integrated SHAP analysis. Our results identify cumulative runtime and braking system indicators as strong predictors for early fault detection. By converting discrete diagnostic codes into continuous indicators, the proposed framework enables direct integration of predictive intelligence within fleet management and supervisory systems.