Machine Learning-Based Prediction of Diagnostic Trouble Codes in Electric Vehicle Batteries : A Multi-Temporal Analysis

Sammanfattning

Lithium-ion battery (LiB) dominates the traction battery market for battery electric vehicles (BEVs), in no small part due to its high energy density and long life-cycle. But the safety risks involved with the use of LiBs include overheating of battery cells that can potentially lead to fire incidents. This highlights the need for early fault prediction systems. Moreover, with the adoption of cloud technology, an increasing number of data-driven approaches have been enabled for the fault prediction of BEV LiBs. In particular, machine learning (ML) approaches have found success within this field, but there are some key challenges it faces such as limited fault data and the need for timely detection. This thesis proposes leveraging diagnostic trouble codes (DTCs) as proxy fault labels to predict critical BEV LiB faults using ML. DTCs are alerts generated by a vehicle’s self-diagnosis system and indicate a wide range of battery issues, but they are inherently reactive. By predicting critical DTCs using real-world BEV monitoring data, this work can transform these reactive alerts into proactive tools for fault prediction. Traditional supervised ML models—Decision Tree, Random Forest, and XGBoost—were trained on battery signals such as voltage, current, temperature, and state of charge (SoC) to predict critical DTCs. Short-term (120 minutes) and long-term (30 days) prediction frameworks were explored to explore the influence of temporal differences. The results demonstrates that ML models have robust performance when predicting critical DTCs. Random Forest achieved a F1-score of 0.9722 when predicting critical DTCs 45- minutes ahead while Decision Tree had a F1-score of 0.833 when predicting 3-days ahead on the hold-out test set. Long-term predictions benefited from XGBoost’s ability to handle non-linear feature interactions, with extended prediction windows yielding consistent improvements. Feature importance analysis revealed current-based signals as the most predictive, while SoC emerged as a key factor in identifying slower-evolving faults. This study highlights the potential of DTCs as proxy fault labels for ML-based BEV LiB fault prediction, advancing research for BEV predictive maintenance.