A Bayesian deep learning pipeline for lithium‐ion battery SOH estimation with uncertainty quantification

• Published in: Quality and reliability engineering international 2024-02, Vol.40 (1), p.406-427
• Main Authors: Ke, Yuqi, Long, Mingzhu, Yang, Fangfang, Peng, Weiwen
• Format: Article
• Language: English
• Subjects: Algorithms; automatic feature selection; Bayesian analysis; Bayesian neural network; Charging; Datasets; Deep learning; Electric potential; Feature extraction; Lithium-ion batteries; Machine learning; Neural networks; SOH estimation; Uncertainty; uncertainty quantification; Voltage; Working conditions
• ISSN: 0748-8017; 1099-1638
• DOI: 10.1002/qre.3424

In recent years, deep learning (DL) methods for state of health (SOH) estimation of lithium‐ion (Li‐ion) batteries have attracted great attention. However, most existing DL‐based methods for SOH estimation were designed using specific battery dataset. To gain better performance, the procedure can be labor‐intensive, involving designing intricate features and fine‐tuning complex DL model, which significantly limits the applications of these methods. Moreover, uncertainty quantification, that is how much uncertainty these DL‐based methods can have on their SOH estimation, cannot be addressed with only point estimation. To this end, this paper proposes a Bayesian DL pipeline for Li‐ion battery SOH estimation targeting at automatic feature extraction and uncertainty‐aware DL‐model developing for general application to batteries under different working conditions. The proposed pipeline is composed of three modules accounting for common feature extraction, automatic feature selection, and Bayesian neural network (BNN) model construction. Specifically, a frequently‐used feature set is designed using the voltage, current and time data of constant current and constant voltage (CC‐CV) charging mode, while the recursive feature elimination (RFE) with cross validation (RFECV) algorithm is used to automatically select features based on relevance. Based on the automatically selected features, DL model is automatically constructed without too much human intervention. The uncertainty of SOH estimation is quantified by BNN model using Monte Carlo (MC) dropout method, which is subsequently used to calculate the confidence of the estimation results. Four battery datasets, covering regular charging, fast charging, and second‐use applications, are used to demonstrate the performance of the proposed method. The proposed pipeline shows superior performance to state‐of‐art DL‐based methods in terms of application capability to different working conditions and uncertainty quantification for SOH estimation.