Multi-level contrastive self-supervised learning with dynamic spectral-temporal embedding for state-of-health estimation of lithium-ion battery with limited labeled data

Abstract

Insufficient labeled data is a common issue in state-of-health (SOH) estimation of Lithium-Ion battery. Self-supervised learning method provide a feasible direction to solve the problem of labeled data scarcity by setting up pretext task to fully exploit the intrinsic features of unlabeled data. However, the existing self-supervised methods do not fully consider the physical temporal dependence of signal data, and cannot capture the multi-level key features of the signal, resulting in unsatisfactory SOH estimation results. In order to solve this problem, a multi-level contrastive self-supervised learning with dynamic embedding method (MCSSL-DE) is proposed in this paper. This method fully considers the temporal dependence of the sequence and designs a multi-scale dynamic patch embedding method with local dependence. Specifically, the frequency domain representation of the signal is used to guide the temporal signal to obtain a local comprehensive embedding at the subsequence level, thereby capturing point-level unavailable context information. On this basis, a new multi-level feature extraction model for prediction tasks is designed, which adopts a self-supervised learning method combining contrastive learning and mask modeling to learn the intrinsic multi-level features on the easily available unlabeled data. Furthermore, the multi-layer features are adaptively aggregated through the frequency domain information, and the features are input into the projection network to achieve high-precision SOH estimation by using limited labeled data. Finally, the SOH estimation performance of MCSSL-DE has been verified on the lithium-ion battery dataset NASA, and the interpretability analysis and visual analysis of the model are carried out based on the experimental results, which proves the superiority and practicability of our method.