利用容量或阻抗的特徵對汰役鋰離子電池進行分類

Abstract

近年來隨著電動車的普及，汰役鋰離子電池的數量大幅增加，處理不當會對環境造成不可逆的傷害和能源的浪費，因此汰役電池的二次利用是目前很受關注的議題。由於汰役電池的容量最高只有初始容量的70 ~ 80%，所以會被應用在容量或功率需求較低的設備上。通常從汰役電池組中淘汰的電池其擁有較低的容量或阻抗一致性，因此找到能正確且快速反應出容量或阻抗一致性的特徵就顯得至關重要。目前利用數據驅動的方法對汰役電池進行分類的文獻都只基於容量進行，這雖然能使分類完的電池具有高容量一致性，但卻無法完全滿足設備對高功率一致性的需求。本研究除了基於容量進行分類之外，首次使用數據驅動的方式對阻抗進行分類，並提出一種容量與功率兼顧的新分類方式，使分類更具靈活性。本文從增量容量分析、電化學阻抗譜以及定電壓充電時的電流曲線中各別提取3個特徵，利用提取出的特徵分別對3種不同的分類方式進行分類，找出不同分類方式適合的特徵，用於未來不同的使用場景。此外我們使用了3種不同的算法包括: catboost、random forest以及one−against−one support vector machine進行分類工作。結果表明不同分類方式有其適合的特徵，使用合適的特徵進行分類，準確率都可以達到93%以上。

In recent years, with the popularity of electric vehicles, the number of retired lithium-ion batteries(LIBs) has significantly increased. Improper handling of these batteries can cause irreversible damage to the environment and lead to energy wastage. Therefore, the secondary use of retired LIBs has become a highly discussed issue. Due to the fact that retired LIBs typically retain only 70 to 80% of their initial capacity, they are often utilized in devices with lower capacity or power requirements. Usually, the batteries retired from battery packs have lower capacity or impedance consistency. Therefore, it is crucial to identify features that can accurately and quickly reflect capacity or impedance consistency. Currently, the literature on data-driven methods for classifying retired batteries is primarily based on their capacity. Although this approach ensures high capacity consistency among classified batteries, it does not fully meet the demand for high power consistency in devices. In this paper, in addition to classification based on capacity, a data-driven approach is being used for the first time to classify retired LIBs based on resistance. Furthermore, a novel classification method is proposed that considers both capacity and power simultaneously, aiming to achieve a more flexible and comprehensive battery classification. In this study, three features were extracted individually from incremental capacity analysis curve, electrochemical impedance spectroscopy curve, and current curves of constant voltage segments. These extracted features are utilized for three different classification methods, enabling us to identify the suitable features for each classification method and their potential use in different application scenarios. Additionally, we have employed three different algorithms, including CatBoost, Random Forest, and one-against-one support vector machine, for the classification task. The results indicate that different classification methods have their suitable features. By using appropriate curve features for classification, the accuracy can be achieved at 93% or higher.