Invariant learning based multi-stage identification for Lithium-ion battery performance degradation

TL;DR

This paper introduces an invariant learning-based multi-stage identification method for analyzing lithium-ion battery degradation, revealing multiple degradation behaviors and improving understanding of degradation mechanisms from data.

Contribution

It proposes a novel multi-stage division strategy using invariant learning to identify different degradation behaviors in battery performance data.

Findings

Effective in identifying multiple degradation stages

Provides insights into battery degradation mechanisms

Validated on benchmark datasets

Abstract

By informing accurate performance (e.g., capacity), health state management plays a significant role in safeguarding battery and its powered system. While most current approaches are primarily based on data-driven methods, lacking in-depth analysis of battery performance degradation mechanism may discount their performances. To fill in the research gap about data-driven battery performance degradation analysis, an invariant learning based method is proposed to investigate whether the battery performance degradation follows a fixed behavior. First, to unfold the hidden dynamics of cycling battery data, measurements are reconstructed in phase subspace. Next, a novel multi-stage division strategy is put forward to judge the existent of multiple degradation behaviors. Then the whole aging procedure is sequentially divided into several segments, among which cycling data with consistent degradation speed are assigned in the same stage. Simulations on a well-know benchmark verify the efficacy of the proposed multi-stages identification strategy. The proposed method not only enables insights into degradation mechanism from data perspective, but also will be helpful to related topics, such as stage of health.