Automatically constructing a health indicator for lithium-ion battery state-of-health estimation via adversarial and compound staked autoencoder

Precisely assessing the state of health (SOH) has emerged as a critical approach to ensuring the safety and dependability of lithium-ion batteries. One of the primary issues faced by SOH estimate methods is their susceptibility to the influence of noise in the observed variables. Moreover, we prefer to automatically extract explicit features for data-driven methods in certain circumstances. In light of these considerations, this paper proposes an adversarial and compound stacked autoencoder for automatically constructing the SOH estimation health indicator. The compound stacked autoencoder consists of two parts. The first one is a denoising autoencoder that learns three different denoising behaviors. The second is a feature-extracting autoencoder that employs adversarial learning to improve generalization ability. The experimental results show that the proposed compound stacked autoencoder can not only get explainable explicit features but also can achieve accurate SOH estimation results compared with its rivals. In addition, the results with transfer learning demonstrate that the proposed method not only can provide high generalization ability but also be easily transferred to a new SOH estimation task.