An adaptive mixture prior in Bayesian convolutional autoencoder for early detecting anomalous degradation behaviors in lithium-ion batteries

Abstract

Accurate and timely detection of anomalies in lithium-ion batteries is crucial for ensuring their reliability and safety. Complex degradation patterns and limited availability of labeled data pose significant challenges in identifying abnormal behaviors in battery usage. This paper proposes an unsupervised adaptive mixture distribution-based Bayesian convolutional autoencoder (AMDBCAE) method for detecting anomalous degradation behaviors in lithium-ion batteries at earlier cycles of reliability test. As the prior for the model parameters, we propose a mixture of the Laplace and Student’s $t$ distributions by taking uncertainties in the weights of the convolutional network and their heavy-tailed characteristics into account. Using a modified form of the Bayes by backprop algorithm, the parameter of mixture proportion is adaptively updated to capture diverse and complex degradation patterns in battery degradation data more efficiently. Extracted latent features are then processed through unsupervised clustering algorithms to identify abnormal degradation behaviors of lithium-ion batteries. The analyses of two real-world lithium-ion battery datasets demonstrate the efficiency and accuracy of the proposed unsupervised framework with limited number of testing data. The proposed method addresses the limitations of manual feature extraction and the need for extensive experimental knowledge by leveraging the adaptive BCAE model to automatically extract latent features as a virtual health indicator in sparse data environments.