Variational autoencoder-driven adversarial SVDD for power battery anomaly detection on real industrial data

Automotive power batteries are integral to the operation of electric vehicles (EVs) and hybrids, playing a critical role in ensuring their safety and performance. Obtaining sufficient high-quality fault data is challenging, and prior data annotation often relies on expert knowledge, which is both costly and prone to errors. Developing a battery condition detection scheme that does not require labeled data remains a persistent challenge for the industry. In recent years, the Deep Support Vector Data Description method, which is not affected by data distribution, has gained significant attention. However, this approach faces the risk of hypersphere collapse. In response, this paper introduces Deep Variational AutoEncoder-Based Support Vector Data Description with Adversarial Learning (DVAA-SVDD), a specialized one-class classification model tailored for power battery anomaly detection. This approach employs a variational autoencoder to regularize the feature distribution of normal samples, preventing hypersphere collapse caused by deterministic outputs. Additionally, adversarial learning is used to incorporate the quality of model generation into anomaly detection. Experiments conducted on a dataset containing 5 million real-world automotive battery operation data points demonstrate that DVAA-SVDD exhibits outstanding performance in vehicle anomaly detection, making it a robust solution for ensuring the reliability and safety of automotive power batteries.