 |
International Society of Science and Applied Technologies |
| A Hybrid Generative Framework for Overvoltage Fault Detection in Power Batteries of Electric Vehicles | ||||
| Author |
Zhijie Rong
|
|||
| Co-Author(s) |
Yahan Sun; Jiayou Chen; Jiawei Xiong; Miaomiao Zhao; Jian Zhou
|
|||
| Abstract | As electric vehicles (EVs) become increasingly popular, overvoltage faults present a significant safety risk for lithium iron phosphate (LFP) batteries in EVs. Therefore, reliable early-warning systems are necessary. However, a severe scarcity of real-world fault data limits their predictive capabilities. To address this issue, this paper proposes a hybrid generative framework to synthesize overvoltage fault data. We employ a Random Forest baseline model to capture macroscopic electrochemical trends within the battery. Then, a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) is developed to model the standardized residuals separating the actual operating data from the baseline predictions. This captures the complex, nonlinear characteristics of overvoltage events. By applying K-Means clustering, the framework introduces state-specific constraints that reflect the varying driving conditions of EVs, ensuring the generated sequences remain physically plausible. Experimental validation on actual EV datasets shows that the synthesized fault data closely matches real-world conditions. Compared with the baseline model trained solely on the original dataset, integrating the synthetic data improves the Area Under the Curve (AUC) of downstream diagnostic models from 0.9937 to 0.9971, while reducing the false negative rate by 77.7%.
|
|||
| Keywords | Power Battery, Overvoltage Fault, Fault Early-Warning, WGAN-GP, Data Augmentation | |||
| Article #: RQD2026-182 | ||||
Proceedings of 31st ISSAT International Conference on Reliability & Quality in Design |