多视图特征融合与SVM集成：锂离子电池健康预测新方法

65 浏览量更新于2024-08-26 收藏 2.28MB PDF 举报

本文探讨了基于多视图特征融合和支持向量回归集成的方法来预测锂离子电池的健康状态（State of Health, SOH）。锂离子电池在众多电子产品系统中扮演着关键角色，准确估计其SOH对于确保其安全性和可靠性至关重要。机器学习方法在SOH预测中占据主导地位，但面临两个主要挑战：一是如何提取具有区分度的特征；二是如何有效地利用这些特征。作者Chao Ma、Xu Zhai等人提出了一种解决方案，他们采用了滑动窗口基特征提取技术（Sliding Window-Based Feature Extraction, SWBFE），这是一种创新的方法，旨在解决这两个问题。SWBFE通过分析电池在不同时间窗口内的数据，动态捕捉电池性能随时间变化的趋势，从而提取出反映电池健康状况的关键特征。这种方法考虑了电池行为的时空连续性，提高了特征的代表性和解释性。支持向量回归（Support Vector Regression, SVR）集成则是另一个核心组成部分。通常，SVR以其强大的泛化能力和小样本学习能力被用于预测任务。通过将多个独立的SVR模型进行集成（如Bagging或Boosting等集成策略），可以进一步提高预测精度，减少过拟合风险，并增强对复杂关系的理解。文章还强调了多视图特征融合的重要性。在实际应用中，电池数据可能来源于不同的传感器或模态，如电流、电压、温度等，这些数据可以从不同的角度反映电池状态。通过融合这些多源信息，研究者能够获取更全面的电池健康特征，避免单个视角的局限性，提升预测模型的鲁棒性和准确性。该研究工作于2017年11月14日接收并最终于2018年8月7日接受发表在《国际机器学习与网络计算》杂志上，doi号为10.1007/s13042-018-0865-y。研究结果表明，结合多视图特征融合和支持向量回归集成的技术，能够有效地提高锂离子电池健康状态预测的精度，为电池管理系统的优化和电池寿命的延长提供了有力的支持。

International Journal of Machine Learning and Cybernetics

1 3

namely, Ada-TargetSOH and Sta-TargetSOH, for robust pre-

diction of lithium-ion battery SOH. To evaluate the eﬃcacy

of the two proposed predictors, we implemented Ada-Tar-

getSOH and Sta-TargetSOH on three types of lithium-ion

battery datasets. The experimental results have shown that

our predictors outperform other existing lithium-ion battery

SOH predictors.

2 Materials andmethods

2.1 Benchmark datasets

The lithium-ion battery experimental data that were used

to evaluate the proposed methods in this research were

obtained from National Aeronautics and Space Adminis-

tration (NASA) Ames Prognostics Center of Excellence

(PCoE) [52]. There were various kinds of lithium-ion battery

datasets in the NASA database, and we chose three types of

lithium-ion batteries (005, 006, and 007) as benchmarks.

For each type, three diﬀerent operational proﬁles, namely,

charge process, discharge process and impedance measure-

ment process, were involved as following: ﬁrstly, charge was

performed at a constant current level of 1.5A until the bat-

tery voltage reached 4.2V and then continued in a constant

voltage level until the charge current fell to 20mA; next, dis-

charge was carried out in a constant current mode with 2A

until the battery voltage respectively decreased to 2.7, 2.5

and 2.2V for batteries 005, 006 and 007; ﬁnally, impedance

measurement was performed by an electrochemical imped-

ance spectroscopy (EIS) frequency sweep from 0.1Hz to

5kHz. In this paper, we focused on the battery discharge

process. For each type of battery, the corresponding dis-

charge process consisted of 168 discharge cycles in constant

current mode. In each cycle, various battery measurements

were listed, such as battery voltage, battery current, and bat-

tery capacity. Among all measurements, the most important

was battery capacity. According to the battery capacity, we

can easily calculate the SOH value of a battery with the fol-

lowing formula:

where

is the initial battery capacity and

is the capac-

ity at the

i-th

discharge cycle. Figure1 illustrates the SOH

variation curves of batteries 005, 006, and 007 as functions

of the discharge cycle number. From Fig.1, we can observe

that the SOH shows the degradation trend during discharge

cycles for each type of battery. In addition, it can’t escape

from our notice that there are several spikes in some cycles.

This phenomenon can be explained as follows: in charge/

discharge proﬁles, the side reactions, which occur between

the electrodes and electrolyte of the battery, make the elec-

trochemical performance of battery decrease; therefore, the

(1)

SOH = C

∕C

value of SOH gradually drops on the whole; however, when

the battery rests during charge or discharge processes, the

reaction products may dissipate, which leads to that the elec-

trochemical performance has an opportunity to recover; as a

result, the SOH in current cycle is increased compared to the

former cycle. The above phenomenon is called regeneration

or self-recharge, introduced and analyzed in [53].

In this work, to validate the eﬃcacy of the proposed

methods, we designed two strategies, which are denoted as

DStrategy1 and DStrategy2, for constructing the training

dataset and test dataset for each type of battery. In DStrat-

egy1, the data from cycle 1 to cycle 80 were used as the

training dataset, and the data from the remaining cycles were

selected as the test dataset. In DStrategy2, we chose the data

from cycle 1 to cycle 100 for constructing the prediction

model and utilized the remaining data to evaluate the trained

model.

2.2 Sliding‑window‑based feature extraction

In the discharge process, ﬁve measurements, namely, battery

voltage (BV), battery current (BC), load voltage (LV), load

current (LC), and battery temperature (BT), were recorded

for each type of battery in the database. Since the discharge

process is based on constant current mode, all the recorded

values of BC and LC are approximately equal. Therefore,

only BT, BV, and LV are considered in constructing the fea-

ture set in this work. For each type of measurement, numer-

ous values were recorded in each cycle, and the number of

recorded values in the current cycle is diﬀerent from the

number in the next cycle. Hence, it is diﬃcult to use these

recorded values as features directly.

To solve this problem, the following steps are executed.

Let

i,j

={M

i,j

, M

i,j

, … , M

i,j

, … , M

i,j

}

be the

Fig. 1 The SOH variation curves of three types of batteries vs. dis-

charge cycle number

剩余13页未读，继续阅读

weixin_38503483

粉丝: 8
资源: 942

多视图特征融合与SVM集成：锂离子电池健康预测新方法

基于非并行支持向量机的多视图学习

Android仿MIUI周视图月视图切换日历

Oracle10g视图的神图

新能源汽车电池故障预测大数据平台架构详解

多视图聚类学习：一种新型 proximity 学习方法

Fritzing电路设计软件全面指南：多视图同步与操作详解

多视图三维重建：彩色点云技术在Matlab中的实现

基于分类算法的SSM项目学业警示预测系统

Vue解决数组和对象变化不触发视图更新的方法

互联网路由算法详解：距离向量法

最新资源