Multisource Domain Metalearning Network for Battery State-of-Health Estimation Under Multitarget Working Conditions

Abstract

State of health (SOH) estimation is important in battery health prognostics. The discrepancy in working conditions gives rise to the phenomenon of domain shift, which presents a significant obstacle to the accurate estimation of battery SOH. Domain adaptation (DA) has been widely used to solve the domain shift problem in battery SOH estimation by minimizing the distribution discrepancy between different domains. However, the multitarget domain shift problem is not be addressed well for existing methods. Most existing DAs only consider a single target domain, which means that the model needs to be retrained when new scenarios emerge, which differ from the original target working condition. In this article, multisource domain metalearning network (MSDMLN) is proposed for battery SOH estimation under multiple target working conditions. A novel metalearning method, multisource domain metalearning (MSDML) strategy is developed for enhancing the generalization of the network by diversifying battery health degradation features based on meta fusion block. Empirical mode decomposition-densely connected recurrent-convolution network is developed to extract global degradation tendency and local fluctuation features of Li-ion batteries. The effectiveness of MSDMLN is verified on the combined Li-ion battery dataset. The results demonstrate that MSDMLN achieved low mean absolute error (i.e., 0.0474) and normalized root mean square error (i.e., 0.2187) for three different target operating conditions, which is better than other state-of-the-art methods. This illustrates the outperformance of MSDMLN in generalizing the estimation of battery SOH under multitarget working conditions, due to the integration of MSDML and the meta fusion block.