Elsevier

Microelectronics Reliability

Volumes 88–90, September 2018, Pages 1189-1194
Microelectronics Reliability

Remaining useful life prediction for lithium-ion batteries based on an integrated health indicator

https://doi.org/10.1016/j.microrel.2018.07.047Get rights and content

Highlights

  • An integrated health indicator is developed using capacity, resistance, and CCCT.

  • The constant voltage charge time is not a good health indicator.

  • A threshold of 0.85 is recommended as the end-of-life criterion.

  • Battery remaining useful life is predicted using a particle filter algorithm.

Abstract

State of health estimation and remaining useful life prediction of lithium-ion batteries is challenging due to various health indicators characterizing battery degradation. This paper develops an integrated health indicator to predict remaining useful life by incorporating capacitance, resistance, and constant current charge time with the help of a beta distribution function, based on the correlation analysis between parameter variations and aging mechanisms. A three-order polynomial model is employed to fit the battery health degradation process, remaining useful life is predicted using a particle filter algorithm, and the probability density function for the battery remaining useful life is then provided. A case study is conducted to validate the health degradation model and battery remaining useful life prediction. The results show that the constant voltage charge time is not a good health indicator, and a threshold of 0.85 is recommended as the end-of-life criterion based on the integrated health indicator. The developed method provides a reference for battery remaining useful life prediction when sufficient energy and power are required.

Introduction

Lithium-ion batteries are regarded as the most promising energy candidate in the automotive industry [1]. Since battery performance inevitably degrades with cycling, information about the state of health (SOH) of lithium-ion batteries is critical for safe and reliable operation [2]. Generally, the remaining useful life (RUL) is predicted to give users an estimate of battery cycle life, so that decisions on battery replacement can be made. However, controversies exist for how the SOH indicators are determined. This paper develops a new integrated health indicator, from the point of view of capacitance and power, to represent battery SOH and predict battery RUL.

Generally, capacity and internal resistance are employed to represent battery SOH [3]. For example, a capacity lower than 80% is considered as the criterion for end-of-life (EOL) [4], and electrochemical impedance spectroscopy is used to measure a battery's internal resistance and characterize the varying aging and fault processes [5]. Because batteries in plug-in hybrid vehicles (PHEVs) require sufficient energy and power [6], and power fades when resistance increases as the battery ages [7], both the capacity and internal resistance should be integrated to estimate battery SOH. In addition, Eddahech et al. [8] explored an SOH determination method based on constant voltage charge time (CVCT). Williard et al. [9] introduced the beta distribution function to combine capacitance, resistance, constant current charge time (CCCT), and CVCT to produce a new SOH indicator. However, they did not validate the effectiveness through aging mechanism analysis on the CCCT or CVCT. Nor did they focus on RUL prediction based on the fused indicator or a single indicator. Therefore, this paper correlates the parameter variations of CCCT and CVCT to the aging mechanisms and develops a new integrated SOH indicator.

Estimating battery SOH and predicting RUL is difficult because batteries are sophisticated nonlinear systems with elusive chemical reactions [10]. Thus, many data training algorithms have been applied to predict battery RUL. Terzimehic et al. [11] applied support vector machine (SVM) to predict RUL using load collectives as the training and test data. Jiang et al. [12] used genetic algorithm (GA) to estimate the battery model parameters, and then determined the battery RUL using the identified diffusion capacitance. Because the particle filter (PF) is able to provide the probability distribution, it has become one of the most effective important methods to predict RUL. The PF was used for RUL prediction based on a state-space model [13]. Dempster–Shafer (DS) theory and the PF method were combined to predict battery RUL [14]. This paper uses a three-order polynomial model and the PF algorithm to predict battery RUL based on an integrated health indicator.

The remainder of this paper is organized as follows. Section 2 deals with the identification of the integrated SOH indicator. Section 3 describes the RUL prediction method in detail. Section 4 validates the model and discusses the prediction results. Section 5 presents the conclusions.

Section snippets

Identification of the integrated SOH indicator

Three battery cells with a rated capacity of 1.1 Ah underwent cycle life tests using Arbin BT2000 test equipment under room temperature. The testing profile is shown in Fig. 1. Capacitance, resistance, CCCT, and CVCT were measured as shown in Fig. 2. The capacitance and CCCT decreased with cycles, whereas the resistance and CVCT increased with cycles. The peak point observed in Fig. 2d could indicate specific internal degradation mechanisms that are not as apparent in the discharge capacity in

RUL prediction based on PF algorithm

A three-order polynomial model was employed to fit the battery degradation process with a trade-off between model accuracy and calculation complexity,S=a1×k3+a2×k2+a3×k+a4where S is the health state of the battery; and a1, a2, a3, a4 are model parameters. The parameter estimates can be computed with the help of Matlab. The state transfer equation of a lithium-ion battery isxk=a1,ka2,ka3,ka4,kTpxk/xk1where ar,k = ar,k−1 + var(k), var(k) ~ N(0,δv) are Gaussian noise with zero mean and standard

Case study

The degradation model is assessed by goodness-of-fit. The parameters of the degradation model are estimated using the fused testing data of cells CS2-35, CS2-36, and CS2-37, as shown in Table 1. The fitted curves are presented in Fig. 5. The goodness-of-fit for the degradation model is estimated and shown in Table 2. The results represent the effectiveness of the degradation model, and the model can be employed to predict battery RUL.

RUL prediction results at cycle 370 for cell CS2-35 based on

Conclusions

This paper developed an RUL prediction approach using an integrated SOH indicator and a particle filter algorithm. CCCT and CVCT aging mechanisms and incidence relationships with capacitance and resistance were analyzed. The CVCT is not a good SOH indicator, and the capacitance, resistance, and CCCT should be integrated to produce a new SOH indicator from the point of view of capacity and power. Consequently, a method to integrate this new SOH indicator was developed using the beta distribution

Acknowledgements

This research was sponsored by Natural Science Foundation of Heilongjiang Province (QC2016068), and University Nursing Program for Young Scholars with Creative Talents in Heilongjiang Province (UNPYSCT-2017087). The authors thank Cheryl Wurzbacher for editing and comments to improve the paper's quality.

Cited by (50)

  • Multivariable Fractional Polynomials for lithium-ion batteries degradation models under dynamic conditions

    2022, Journal of Energy Storage
    Citation Excerpt :

    The powers of included variable have been determined by the MFP algorithm (Section 3.1). It should be emphasized that the rationale for the selection of such polynomials is not of a physico-chemical nature: the transformations are suggested by the data themselves, hence are purely data-driven and relevant for the specific dataset rather than universally valid as it is the case, for example, of the health indicators resarchead by Zhou et al. [60] and Sun et al. [61]. This also comes as a consequence of the fact that the model is multivariable: it does not account for the marginal effect of each input, but rather it accounts for each variable in combination with the other variables in the model.

View all citing articles on Scopus
View full text