Lithium-ion battery anomaly detection with variational autoencoder based on feature representation learning
Journal of Energy Storage, vol. 152, pp. 120545
Abstract
Battery failure is an increasingly critical issue in lithium-ion battery applications, primarily arising from manufacturing defects and operational abuse. Non-invasive battery anomaly detection based on monitoring data is of significant importance as disassembling a battery pack is impractical. A fault diagnosis approach employing Gaussian Mixture Model and Variational Autoencoder is proposed to effectively extract anomalous features from voltage, temperature, current, and State of Charge. The latent features are utilized for binary and multi-categories anomaly diagnosis through clustering and Random Forest, respectively, including Internal Short Circuit, overcharge, and over-temperature scenarios. The cluster compactness of latent variable distribution is improved by incorporating Maximum Mean Discrepancy into the training loss. The model is validated with both simulation and real battery data, which yields well-separated latent features and outstanding performance with a total accuracy of 98 % for binary classification task. A total accuracy of 99 % is achieved for multi-label classification using self-supervised learning with 15 % fault data and 5% contamination in training set.
Authors 3
-
Affiliation as printed
Chair of Thermodynamics of Mobile Energy Conversion Systems, RWTH Aachen University, Forckenbeckstraße 4, Aachen, 52074, Germany
-
Affiliation as printed
Chair of Thermodynamics of Mobile Energy Conversion Systems, RWTH Aachen University, Forckenbeckstraße 4, Aachen, 52074, Germany
-
Affiliation as printed
Chair of Thermodynamics of Mobile Energy Conversion Systems, RWTH Aachen University, Forckenbeckstraße 4, Aachen, 52074, Germany
Cited by 5 stored of 5
5 results
No patents citing this paper on Lens.org (checked 2026-10-06).