Unsupervised Adaptive Fleet Battery Pack Fault Detection With Concept Drift Under Evolving Environment

Timely fault detection is critical for ensuring the safety and reliability of electric vehicle battery packs. Capturing the battery’s normal behavior and identifying faults in a fleet operating under dynamic and evolving real-world conditions comes with challenges, including data imbalance, label unavailability, and concept drift. To address these challenges and enhance the robustness of fault detection in evolving environments, we propose an adaptive fleet-based fault detection method. This method comprises two key components. The first component is an OC-aware anomaly detection method, serving as a static model for robust anomaly detection. The second component includes a novel concept drift detection and adaptation mechanism that continuously monitors data distribution and the performance of the anomaly detection model. This mechanism identifies changes in the battery pack’s normal behavior under evolving conditions. Our proposed concept drift detection method reduces false alarms and enhances noise robustness by integrating drift detection and drift isolation within a hierarchical structure. The effectiveness and robustness of the proposed method are validated using real-world in-field data. Note to Practitioners—Existing methods for identifying battery pack faults in electric vehicle (EV) fleets predominantly hinge on static models, deriving from either high-fidelity physics-based battery models or data-driven models trained on extensive labeled datasets. However, the detection performance of static models may degrade when concept drift occurs during real-time fleet monitoring. This paper introduces a novel adaptive fault detection method with an innovative concept drift detection and adaptation mechanism, all without the need for labeled or faulty samples. Furthermore, the method can differentiate real battery faults from concept drift caused by changing operating conditions, significantly reducing false alarms during the model adaptation process. The robustness to unprecedented operating conditions in anomaly detection and the effectiveness of concept detection are validated in case studies using in-field fleet data. In future research, to enhance the efficiency of model retraining and conserve data storage resources, there is a desire to develop an incremental fault detection model that does not require the retention of all historical data.