The reliability of Drive Unit (DU) oil pumps is critical to the performance and safety of electric vehicles, as these pumps provide essential lubrication and thermal management. However, direct physical sensors for real-time pump health monitoring are absent in current architectures, making early fault detection particularly challenging. To address this limitation, we present a novel, data-driven anomaly detection framework that leverages large-scale customer fleet telemetry and advanced machine learning to identify incipient pump degradation that traditional diagnostic methods often fail to capture.
The core of the approach is an XGBoost regression model trained on time-series features—including commanded pump speed, oil temperature, and historical pump current—to predict expected current behavior under nominal conditions. Deviations are quantified using the Mean Absolute Percentage Error (MAPE) between predicted and actual currents, providing a continuous and interpretable measure of anomaly severity. To reduce false positives, anomaly scores are aggregated across operational sessions for each vehicle over a rolling three-week window, producing a stable vehicle-level anomaly score. This score is benchmarked against calibrated thresholds and classified into low, medium, or high severity alerts. A fully automated pipeline ingests daily telemetry, performs session segmentation, executes predictive modeling, and records anomaly outcomes in backend databases for continuous monitoring and engineering review.
The proposed framework enables continuous, fleet-wide predictive maintenance of DU oil pumps. It improves early detection of degradation, reduces vehicle downtime, enhances safety, and increases customer satisfaction. More broadly, it highlights the potential of large-scale data analytics and machine learning to advance predictive maintenance and reliability in electric vehicle systems.