Journal of Energy Storage 48 (2022) 103825

Contents lists available at ScienceDirect

Journal of Energy Storage

journal homepage: www.elsevier.com/locate/est

Research papers

A deep belief network approach to remaining capacity estimation for

lithium-ion batteries based on charging process features
Mengda Cao a,b , Tao Zhang a,b , Jia Wang c , Yajie Liu a,b ,∗
a
College of Systems Engineering, National University of Defense Technology, Changsha, Hunan, 410073, PR China
b
Hunan Key Laboratory of Multi-energy System Intelligent Interconnection Technology (HKL-MSI2T), Changsha, Hunan, 410073, PR China
c
Beijing Institute of Tracking and Telecommunication Technology, Beijing, 100094, PR China

ARTICLE INFO ABSTRACT

Keywords: Accurate remaining capacity estimation for lithium batteries can help people understand the working state
Lithium-ion battery of a battery, which ensures the reliability and safety of electric equipment. Since the remaining capacities
Deep belief network (DBN) of lithium-ion batteries are related to internal physicochemical reactions, such as loss of lithium inventory
Health indicator (HI)
and loss of active material, their remaining capacities are usually difficult to directly estimate. Advancements
Sample entropy (SE)
in deep learning have prompted the development of new data-driven approaches to solve this problem that
Capacity estimation
can capture potential degradation correlation information from high-dimensional data via hidden layers of
the deep learning model. This paper presents a novel deep learning method for lithium-ion battery capacity
estimation based on charging process features with the following process. First, multiple health indicators are
analysed and extracted according to different working conditions of batteries. Next, a grey relation analysis is
combined with a cross-validation test, which is employed to eliminate information redundancy and improve
prediction accuracy. Moreover, the optimized health indicators are further extracted through the restricted
Boltzmann machine component of the deep belief network model, and a fully connected layer is adopted for
estimation of lithium-ion battery capacity. A total of 23 battery datasets, including six working condition types,
are employed for verification experiments. The maximum root mean square error of a single battery prediction
is 3%, and that of multiple battery prediction is still within 6%, which confirms the effectiveness and accuracy
of the proposed method.

1. Introduction battery state estimation, battery equalization, and safe and efficient bat-
tery management [6]. The remaining capacity estimation for batteries
As critical components of a power supply system, lithium-ion batter- represents the available battery capacity after it is fully charged, which
ies are widely utilized in consumer electronics, space aircrafts, electric can also be expressed by the remaining useful life (RUL). Remaining
vehicles, and other products due to their merits of high energy density, capacity estimation is important in battery state estimation because it
low self-discharge rate, long cycle life, and lack of memory effect [1,2].
reflects the change in the state and ability to release electrical energy
The health states of lithium-ion batteries usually determine whether a
of a used battery compared with a new battery [7,8]. This approach
power system can operate stably. With repeated charging and discharg-
ing cycling, the available capacities of lithium-ion batteries gradually enables researchers to better understand the states of batteries to a large
decline. Moreover, the degeneration of batteries has great impacts on extent. Therefore, the development of reliable and accurate lithium-ion
the safe and reliable operation of systems, which can increase the battery RUL prediction approaches is necessary.
maintenance cost or even cause catastrophic events [3]. For instance, in To date, a substantial amount of research has been conducted in
2018, the Mars Rover launched by the National Aeronautics and Space the field of remaining capacity estimation for lithium-ion batteries.
Administration (NASA) malfunctioned and was scrapped because the This research can generally be divided into two categories: model-
batteries overheated, induced by overcharging due to a series of human based methods [9,10] and data-driven methods [11]. Model-based
errors [4]. To avoid this situation, a battery management system (BMS)
approaches mainly rely on establishing empirical or mechanism models
is usually installed on electric equipment to properly monitor and
to reflect the internal reaction of the battery and mimic its degradation
manage battery operations [5]. In a BMS, the key technologies involve

∗ Corresponding author at: College of Systems Engineering, National University of Defense Technology, Changsha, Hunan, 410073, PR China.
E-mail address: [email protected] (Y. Liu).

https://doi.org/10.1016/j.est.2021.103825
Received 3 April 2021; Received in revised form 11 December 2021; Accepted 13 December 2021
Available online 13 January 2022
2352-152X/© 2021 Elsevier Ltd. All rights reserved.
M. Cao et al. Journal of Energy Storage 48 (2022) 103825

process, including equivalent circuit model (ECM)-based, electrochem- variations in the length of the input time series. Zheng et al. [35]
ical model (EM)-based, and mathematical model-based methods [12]. proposed a method based on charging curve sections that uses complete
The ECM-based method first identifies the ECM parameters and esti- charging curves as the training data. Moreover, the capacity is esti-
mates the battery capacity based on the identified parameters [13]. mated by calculating the charging capacities during the optimal fixed
Zhang et al. [14] proposed a new decoupled weighted recursive least voltage window determined by particle swarm optimization (PSO). A
squares (DWRLS) method based on battery ECMs, which can con- novel Gaussian process regression (GPR) model based on the charging
currently improve the modelling accuracy and state of charge (SOC) curve is proposed in Ref. [36]. Unlike other studies, where capacity
estimation performance. Electrochemical model-based methods sim- is commonly estimated by cycle life, in this method, four specific
ulate battery ageing by considering the actual electrochemical re- parameters are extracted from charging curves used as the model input
action process, e.g., the pseudo-two-dimensional (P2D) model [15]. instead of cycle numbers. On the other hand, the covariance function
Laue et al. [16] applied a three-step procedure considering quasi-static and similarity measurement of input variables are modified for better
3-electrode measurements of the open-circuit potential, C-rate tests, accuracy, and the performance of the modified GPR model is verified
and electrochemical impedance spectra to parameterize the pseudo-2D by batteries with dynamic discharging profiles. The abovementioned
model. In Ref. [17], a modified reduced-order electrochemical model literature has confirmed that HI extraction is incredibly significant and
based on the single particle (SP) Li-ion battery model was proposed relevant in reflecting the degradation process of lithium-ion batteries.
to achieve higher capacity estimation accuracy than the traditional SP However, several issues in the extraction process are still worthy
model. Mathematical model-based methods translate the influencing of discussion: (1) A clear and precise demonstration of the feature
factors and operating characteristics into mathematical relationships extraction process is lacking. For example, in Refs. [37–39], the HIs are
for analysis [18]. For example, Ossai et al. [19] attempted to add directly given without details of the extraction process. (2) The feature
random effect parameters in a multiphase decay patterned sigmoidal extraction strategy has poor adaptability and robustness. For example,
model to address the uncertainties, and maximum likelihood estima- the HIs applied in Refs. [39–42] are partly extracted from the discharge
tion (MLE) and stochastic approximation expectation maximization phase. However, due to the different discharge strategies, working
(SAEM) algorithms were selected to optimize the random effect param- conditions, and operation methods, the battery data in the discharge
eters, making the predicted results vary within 2% on average. Weng cycle are usually not sufficiently stable and effective for use in evalu-
et al. [20] present an incremental capacity (IC) peak tracking approach ating battery degradation. (3) The influence of thermodynamic factors
based on the incremental capacity analysis (ICA) model for the online is disregarded. These factors not only reflect the thermal stability of
state of health estimation. In general, model-based approaches can battery materials but also have an important influence on the elec-
effectively reflect the physical and electrochemical characteristics of trochemical reaction of battery degradation, as discussed in Refs. [37,
lithium-ion batteries to obtain high prediction accuracy. However, this 39,40]. (4) The HIs are directly utilized, lacking further analysis and
type of method often requires a large amount of prior knowledge about processing. Inadequate extraction usually makes further improvement
a complex system, and it is difficult to know the exact electrochemical of the precision in the performance degradation assessment of lithium-
reaction in practice. As a result, constructing an elaborate and accurate ion batteries difficult. In addition, the intrinsic correlation between HIs
degradation model of lithium-ion batteries is quite difficult. without proper processing may result in information redundancy that
Comparatively, data-driven approaches, i.e., artificial neural net- also affects the accuracy of the estimation model.
works (ANNs) [21], Bayesian networks [22], regression vector ma- Accordingly, in this paper, a deep learning structure based on HIs of
chines (RVMs) [23], and particle filters (PFs) [24], have recently re- the charging process is proposed for capacity estimation of lithium-ion
ceived an increasing amount of attention due to their flexibility and batteries; its prognostic framework is shown in Fig. 1. First, data pre-
easy operation. Data-driven approaches can utilize massive data to processing of the original battery data is conducted for outlier removal.
learn the degradation law of lithium-ion batteries without prior knowl- Second, we extract 17 HIs from the original battery dataset during its
edge of failure mechanisms. Moreover, deep learning methods, in- charging process based on different extraction principles, including the
cluding convolutional neural networks (CNNs) [25], recurrent neural charge capacity, charge time, charge temperature summation, charge
networks (RNNs) [26], deep belief networks (DBNs), generative adver- curve maximum slope, and sample entropy. The charging phase is
sarial networks (GNNs) [27], and autoencoders (AEs) [28], show great relatively stable, which is more suitable for HI extraction, and sample
potential in this field. Deep learning forms a more abstract high-level entropy effectively reflects the thermodynamic variety of the internal
representation attribute category or feature by combining low-level reaction of the batteries. Then, as a common multifactor statistical anal-
features to discover distributed feature representations of data [29]. ysis method, grey relation analysis (GRA) is applied to the extracted HIs
Compared to other deep learning methods, the DBN is notable due to obtain their grey correlation coefficients between HIs and capacity,
to its good scalability and adaptability. The DBN model is composed and the HIs are sorted according to their grey correlation coefficients.
of multiple-layer restricted Boltzmann machines (RBMs) and one layer Fourth, due to high relevance among HIs resulting in information
of a backpropagation neural network (BPNN). The stacked RBMs are redundancy, cross-validation tests are conducted with the sorted HIs to
applied to extract potential information from data, and the BPNN optimize the HIs. Finally, a DBN model is applied to capacity estimation
is employed for prediction. In addition, the health indicators (HIs) with the optimized HIs because the hidden layers of the RBM are able to
extracted by different rules usually influence the performance of data- effectively capture the correlation of high-dimensional data transferred
driven approaches. Apart from choosing algorithms, many researchers from visible layers of the RBM. Based on the superiority of the DBN
have also paid attention to the extraction process of HIs. Cao et al. [30] model, this method attempts to explore HIs with high versatility by
selected the sample entropy (Sampen) and cutoff discharge voltage as charging data and takes the optimized HIs as input to the DBN model.
HIs for state of health (SOH) estimation of batteries. Wu et al. [31] Compared with existing research, the main contributions of this paper
employed importance sampling on the charge curve as the input of are summarized as follows:
a feed forward neural network (FFNN) for RUL estimation. The per- (1) A novel HI extraction approach is proposed based on the charg-
mutation entropy (PE) was extracted from the discharge voltage curve ing process that is suitable for battery operation under different dis-
for analysing battery degradation in Ref. [32]. Ref. [33] proposed an charge scenarios, e.g., discharge current, cutoff voltage and ambient
SOH estimation method based on IC analysis and differential voltage temperature [43].
analysis of the charge process. Li et al. [34] proposed a long short-term (2) The sample entropy, as an indicator of the degree of disorder
memory RNN (LSTM-RNN) framework for online capacity estimation, in a dynamic system, can effectively reflect the changes in the internal
among which voltage-time sensor data from the partial constant current system of lithium-ion batteries. Therefore, sample entropy is added to
phase charging curve are used as an input that has the ability to handle HIs to enhance the correlation between HIs and capacity degradation.

2
M. Cao et al. Journal of Energy Storage 48 (2022) 103825

Fig. 1. Framework of the proposed method.

(3) The extracted HIs are optimized to eliminate information redun- Table 1
dancy by a GRA combined with a cross-validation test and are then Common constituent materials of the basic structure of lithium-ion batteries.

further reprocessed by the RBM layer in the DBN model instead of being Basic structure Common constituent materials
artificially screened, which can avoid the influence of human factors Anode LiCoO2 , LiMn2 O4 , LiFePo4 , LiCo𝑥 Ni𝑦 Mn1−𝑥−𝑦 O2
and improve the prediction accuracy. Cathode Graphite, Li4 Ti5 O12
Positive current collector Al
The superiority and reliability of the proposed model are verified
Negative current collector Cu
with single battery datasets and multiple battery datasets, respectively. Separator polypropylene (PP), polyethylene (PE), ceramics
The structure of this study is as follows: In Section 2, the mathe- Electrolyte solvent ethylene carbonate (EC), propylene carbonate (PC),
matical background of a lithium-ion battery and the related algorithms dimethyl carbonate (DMC)
employed in this paper are introduced. Section 3 details the HI extrac- Electrolyte solute LiPF6
tion and optimization process based on GRA. The experimental results
and discussion are presented in Section 4. The last section discusses the
conclusions.
Taking a lithium iron phosphate (LFP) cathode as an example, the
2. Mathematical background main reaction of the lithium-ion battery cathode is as follows:
discharge(deintercalation)
𝑦Li+ + 𝑦e− + FePO4 ←↽
←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←⇀
←←←←←← Li𝑦 FePO4 (2)
2.1. Working mechanism of Li-ion batteries charge(intercalation)

Lithium-ion batteries are mainly composed of an anode, a cath- 2.2. Degradation mechanism of Li-ion batteries
ode, a separator and an electrolyte. The common materials of these
components are listed in Table 1. As anode and cathode materials are A lithium-ion battery is a dynamic, time-varying, electrochemical
the most important part of lithium-ion batteries, lithium-ion batteries system with nonlinear behaviour and complex internal mechanisms.
are usually named after their distinctive anode or cathode materials. With an increase in the number of charge and discharge cycles, the
During the charging and discharging of lithium batteries, lithium ions performance and life of lithium-ion batteries are gradually degraded
are intercalated and deintercalated in anode and cathode materials. for many reasons, including physical mechanisms (such as thermal
The function of the current collector, which is coated by anode and stress and mechanical stress) and chemical mechanisms (such as side
cathode materials, in a battery is to evenly distribute the current and reactions). Fig. 3 shows a common battery decay mechanism. Differ-
concentrate the conduction of electrons. The electrolyte in a battery is ent attenuation mechanisms cause battery degradation, which can be
equivalent to the carrier of lithium ions, and the lithium ions between divided into two main types: (1) loss of available lithium and (2)
the anode and cathode are conducted through the electrolyte. The loss of active material. The loss of available lithium is mainly caused
anode and cathode are separated by a diaphragm to prevent electrons by the formation and decomposition of a solid electrolyte interface
from passing through but to allow lithium ions to pass through so that (SEI) film, electrolyte decomposition and lithium evolution. Graphite
the battery can form an external working circuit without causing an peeling, binder decomposition, and electrode particle cracking are the
internal short circuit. main reasons for the loss of active material.
Fig. 2 shows the internal working principle of a lithium-ion battery The mechanism of chemical decay is mainly attributed to the de-
during the discharge process. When the battery is discharged, lithium composition and reduction of electrolytes, formation of the SEI, decom-
ions are extracted from the cathode material into the electrolyte, and position of binders, co-intercalation of solvents, dissolution of active
electrons are simultaneously released into the cathode current collector. materials, escape of gases and loss of lithium [44], which have been
In the anode part, lithium ions in the electrolyte are inserted into the extensively investigated. The electrolyte is reduced in the lithiated an-
battery anode material and consume electrons. Due to the difference in ode in the low potential range and oxidized in the lithiated cathode in
the concentration of lithium ions between the anode and the cathode, the high potential range. Impurities in the electrolyte usually catalyze
the lithium ions in the electrolyte move from the cathode to the the occurrence of side reactions, and high temperature also accelerates
anode through the diaphragm. Electrons flow in the external circuit the degradation of battery materials to a certain extent [45]. These
to generate a current, and the battery starts to discharge to drive the side reactions cause some changes in electrode materials. For example,
load. The charging process of a lithium-ion battery is the inverse of its the redox reaction between the electrode and the electrolyte will form
discharging process. The intercalation and deintercalation of lithium an interface layer, which causes an increase in the impedance of the
ions on the anode and cathode may be accompanied by other side battery and a decrease in capacity. In general, the long-term chemical
reactions, resulting in loss of lithium ions or damage to the anode and decay of battery materials may affect the battery life and safety from
cathode active materials, which causes degradation of the lithium-ion the following four aspects: (1) increase in resistance; (2) corrosion or
battery. Considering a graphite anode as an example, the main reaction passivation of active materials by corrosion products; (3) decrease in
of the lithium-ion battery anode is as follows: battery tightness; and (4) electrolyte loss. The DBN model and sample
discharge(intercalation)
+ −
entropy (SE) utilized in the proposed prognostic framework are briefly
Li𝑥 C6 ←↽
←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←←⇀
←←←←←← 𝑥Li + 𝑥e + 6C (1) introduced below.
charge(deintercalation)

3
M. Cao et al. Journal of Energy Storage 48 (2022) 103825

Fig. 2. Schematic of the basic structure and working principle of lithium-ion batteries.

Fig. 3. Main ageing mechanism of lithium-ion batteries [46].

2.3. Related algorithms

2.3.1. DBN
For these machine learning methods in the field of prediction,
the most widely employed method is the BPNN, which calculates the
error between the actual value and the expected value to adjust and
update the network parameters [47]. However, the BPNN has obvious
drawbacks when applying the backpropagation (BP) algorithm to a
deep neural network, such as the necessity of a substantial dataset for
training, slow learning, and inadequate parameter selection techniques
that yield poor local optima. With the development of deep learning, its
potential predictive application value has been demonstrated. As one of
the deep learning approaches, the DBN can overcome the abovemen-
tioned defects of the BPNN due to its powerful linear approximation Fig. 4. Example of the DBN structure as a predictor.
ability. The first emergence of the DBN was proposed by Geoffrey
Hinton [48]; the DBN is a probabilistic generative model that contrasts
with the discriminative nature of traditional neural nets. The related possible is retained to provide better initial parameters for the model.
literature shows that the DBN usually has superior performance in Moreover, the latter period is applied to search for the global optimal
feature extraction and automatic data dimension reduction. The typical solution of model parameters, which can overcome the shortcomings
structure of the DBN is composed of multilayer stacked RBMs and of falling into a local optimum and a long training time.
a single-layer neural network (NN), as shown in Fig. 4 [49]. The
training process of the DBN can generally be divided into two periods: 2.3.2. SE
pretraining and fine-tuning [50]. During pretraining, each layer of the Entropy was originally a thermodynamic concept for measuring the
RBM is trained separately to ensure that when the feature vector is degree of disorder in a thermodynamic system. After the establishment
mapped to different feature spaces, as much feature information as of information theory, the concept and theory of entropy were greatly

4
M. Cao et al. Journal of Energy Storage 48 (2022) 103825

Fig. 5. Charge parameter variation with the ageing of batteries.

developed. As information entropy, SE is used to calculate the com- Table 2

plexity of a time series by measuring the probability of generating new Operating parameters of batteries.

patterns in the signal data [30]. SE has been applied in many fields of Battery Discharge Cutoff Ambient EOL
Label Current Voltage Temperature
time series analysis, for instance, in equipment fault diagnosis, lifetime
(A) (V) (◦ C) (Ahr)
prediction and renewable energy forecasting. The greater the chaos
B0005/B0006 2.7/2.2
of the time series is, the larger the SE. Since the internal complexity 2 1.4
B0007/B0018 2.5/2.2
of lithium-ion batteries varies with age, the SE of battery charging 24
B0025/B0026 0.05 Hz 4 A 2.0/2.2
–
parameters can be considered a reliable HI. B0027/B0028 50% duty 2.5/2.7
The degradation of lithium-ion batteries is partly caused by the syn- B0029/B0030 2.0/2.2
4 43 –
thesis and decomposition of unstable compounds during the charging B0031/B0032 2.5/2.7
B0033/B0034 4 2.0/2.2
and discharging process [30]. The degree of disorder of metal com- B0036 2 2.7
24 1.6
pounds is usually higher than that of metal ions [51], evidencing that B0045/B0046 2.0/2.2
1
the synthesis and decomposition of unstable compounds is related to B0047/B0048 2.5/2.7
4 1.4
the entropy changes of the thermodynamic system inside the batteries. B0053/B0054 2.0/2.2
2
Therefore, as an indicator of the disorder degree of the time series, B0055/B0056 2.5/2.7

the SE of voltage, current, and temperature data sequences can reflect

the influence of thermodynamic system changes on batteries with some
certainty. In this work, 17 HIs are extracted and can be divided into five
groups according to the extraction rules. These five groups of HIs are
3. HI extraction and optimization extracted based on the capacity, temperature, current/voltage drop,
charging time, and SE in the constant current and constant voltage
3.1. HI extraction phases.
(1) Charge capacity-related HIs: C1, C2 and C. Charge capacity-
The lithium-ion battery data used in this paper are sourced from the
related HIs are calculated by the ampere-hour integral method. C1 and
Prognostic Center of Excellence (PCoE) dataset of the Ames Excellence
C2 represent the area under the charged current curve in the constant
Failure Prediction Research Center at NASA. The series of batteries
current phase and constant voltage phase, respectively. C is the total
are all 18,650 lithium-ion batteries. The labels and the working con-
charged capacity, which is the sum of C1 and C2. These three HIs are
dition settings of the battery datasets are shown in Table 2, where the
used to reflect the trend change during the battery charging process.
discharge current indicates the current value in the constant current
(2) Charge time-related HIs: M1, M2 and M1/M2. M1 and M2 are
discharge stage. The cutoff voltage denotes the voltage at the end of
the constant current and constant voltage times, respectively. M1/M2
battery discharge. The ambient temperature is the operating temper-
is the ratio of the constant current time to the constant voltage time.
ature during the charge and discharge cycles. The end of life (EOL)
(3) Charge temperature-related HIs: T1, T2, T, T1/C1, T2/C2, and
represents the remaining capacity of the battery when the battery is
T/C. T1, T2, and 𝑇 represent the area under the temperature curve in
determined to be degraded to failure.
Before HI extraction, the invalid zero values and obvious outliers in the constant current phase, constant voltage phase, and whole charging
the lithium-ion battery dataset are removed. The discharge conditions phase, respectively. T1/C1, T2/C2, and T/C are the ratios of the area
of the batteries, e.g., operating temperature, discharge current, and under the temperature curve to the area under the corresponding
cutoff voltage, usually differ. As a result, the data collected during current curve. These three HIs can indicate the temperature change per
the discharge process are not fixed or sufficiently reliable; thus, the unit charge capacity.
HIs extracted are not suitable for most working scenarios. Therefore, (4) Slope of charge curve-related HIs: S1 and S2. S1 and S2 are the
we extracted the HIs from the charging curve (voltage, current and maximum slope of the voltage curve and the maximum slope of the
temperature), which are shown in Fig. 5. current curve, respectively, which reflect the fluctuation of the voltage
The charging process can be divided into two periods. First, the and current curves.
lithium-ion batteries were charged with a constant current of 1.5 A (5) SE-related HIs: SE1, SE2, and SE3. SE1 and SE2 are the SE of the
until the voltage rose to 4.2 V. Second, the batteries were charged current curve in the constant voltage phase and SE of the voltage curve
at a constant voltage of 4.2 V until the charging current dropped to in the constant current phase, respectively. Fig. 6 shows the variation
0.02 A. Fig. 5 shows that the changes in charge parameter curves in the 17 HIs during the charge/discharge cycle for battery B0005.
present a geometric characteristic with the degradation of batteries.
Due to nonuniform sampling, simple, statistical, and feature extraction 3.2. GRA and HI optimization
methods do not always effectively reflect the degradation process of
lithium-ion batteries. Therefore, certain rules must be formulated to As shown in Fig. 6, the relation of the 17 HIs with the deterioration
extract HIs from these charging curves. of capacity is difficult to show intuitively, so performing a quantitative

5
M. Cao et al. Journal of Energy Storage 48 (2022) 103825

Fig. 6. 17 HIs extracted from the B0005 battery dataset.

analysis is paramount for verifying the effectiveness of HIs. GRA is a sequence. After normalization of the data in the sequence, the grey
{ } { }
common method of multifactor statistical analysis that uses the grey relation coefficient between 𝐱0 = 𝑥0 (𝑘) and 𝐱𝑖 = 𝑥𝑖 (𝑘) is calculated
relation order (GRO) to depict the strength, size, and order of factors. as follows.
{ }
Assume that the sequence of capacity 𝐱0 = 𝑥0 (𝑘) is the reference
{ } min∀𝑖 min∀𝑘 ||𝑥0 (𝑘) − 𝑥𝑖 (𝑘)|| + 𝜌 max∀𝑖 max∀𝑘 ||𝑥0 (𝑘) − 𝑥𝑖 (𝑘)||
𝜉𝑖 (𝑘) = (3)
sequence and that the sequence of HIs 𝐱𝑖 = 𝑥𝑖 (𝑘) is the comparison |𝑥0 (𝑘) − 𝑥𝑖 (𝑘)| + 𝜌 max∀𝑖 max∀𝑘 |𝑥0 (𝑘) − 𝑥𝑖 (𝑘)|
| | | |

6
M. Cao et al. Journal of Energy Storage 48 (2022) 103825

In Eq. (16), 𝜌 is the distinguishing coefficient that determines the 4.2. Single battery dataset
distinction degree of the GRA method. Generally, 𝜌 ∈ [0, 1]. The smaller
𝜌 is, the greater the distinction degree of the GRA method. According First, the proposed method is applied to single battery datasets to
to Ref. [36,42], 𝜌 is set to 0.5. verify its robustness and reliability. In the single battery datasets, the
Because the correlation coefficient is the value of the correlation first battery of each group in different working conditions is selected,
between the comparison sequence and the reference sequence at each including B0005, B0025, B0029, B0045, and B0053. Before capacity
moment, it has more than one value, which makes the information too estimation, a three-fold cross-validation test is performed based on the
training dataset of each battery to determine the optimal number of
scattered to facilitate an overall comparison. Thus, the average value
input HIs. The average RMSE of the DBN model with different numbers
of the correlation coefficient 𝛾𝑖 is taken as the quantitative expression
of HIs is considered the evaluation criterion, in which the number
of the degree of correlation between the comparison sequence and the
corresponding to the minimum RMSE is the optimal value. The average
reference sequence. 𝛾𝑖 is denoted as follows.
RMSE for each battery with different numbers of HIs is shown in
1∑
𝑛 Table 4.
𝛾𝑖 = 𝜉 (𝑘) (4) The bold values in Table 4 are the selected number of HIs for each
𝑛 𝑘=1 𝑖
battery, i.e., 8, 10, 11, 16, and 8 HIs for the B0005, B0025, B0029,
The closer 𝛾𝑖 is to 1, the greater the correlation between the two B0045, and B0053 batteries, respectively. In addition, the accuracy of
sequences. Table 3 shows the correlation coefficients of the 17 HIs the prognostic model is usually related to the amount of data; that is,
related to the capacity sequence, and the average of the grey relational the more data available, the higher the accuracy is. To test the appli-
grades and relation sort results of each HI are shown in the last cability of the model as much as possible, the dataset of a battery is di-
two rows of Table 3. The correlation of the 17 HIs varies, and the vided into three types of data segmentation, i.e., 80%/20%, 70%/30%,
charge parameters in the constant current phase of batteries are highly and 60%/40%, which can be utilized to simulate real lithium-ion
correlated with the ageing of batteries, e.g., C1, M1, and T1; thus, the battery application scenarios. The predicted results are given in Table 5
relevance ranking is 2, 1, and 4, respectively. In contrast, the constant and Fig. 7, in which only the experimental results of 60%/40% are
voltage phase parameters C2, M2, and T2 are sorted as 12, 17, and 16, shown due to space constraints in this paper. In Fig. 7, the image
on the left shows the predicted result of each battery, and the image
respectively, which are less closely related to battery degradation than
on the right presents the relative error between the actual value and
the constant current phase parameters. The SE values of temperature,
the predicted value. Fig. 7 shows that the predicted model adequately
voltage, and current are moderately relevant compared to the strong
fits the degradation trend of the battery. The relative errors of B0005,
correlation of constant current phase parameters and weak correlation
B0025, and B0029 are less than 1.5%, and the relative errors of B0045
of constant voltage phase parameters.
and B0053 are slightly higher but do not exceed 5% and 4%, except for
The HIs extracted from battery operating data are usually directly one predicted value. As shown in Table 5, the accuracy of the method
employed as input of the prognostic model, which causes inadequate improves with an increase in the amount of training data for the first
extraction of potentially useful information in HIs. Artificial interfer- four batteries. The B0053 battery exhibits an anomalous decrease when
ence factors often exist in artificial dimensionality reduction, fusion, comparing the RMSEs in the situations of 70%/30% and 60%/40% data
or selection of features; the RBM layer, as a data mining part of the segmentation. This phenomenon is probably due to the small data size
DBN model, can automatically realize feature extraction and fusion, of only 54 samples and the small increase in the number of training
thus avoiding this problem. samples of only 5 in the different segmentations of 70%/30% data and
However, a high relevance existing among certain HIs may result in 60%/40% data, respectively, for B0053. In conclusion, the RMSEs of
information redundancy that has a negative impact on the prediction B0005, B0025, B0029, and B0045 do not exceed 1.5%, whereas there
results. Therefore, it is necessary to denoise these HIs by reducing their is a certain increase in the RMSE of B0053, which is still within 3%.
number to obtain a better input to the DBN model. The 17 HIs are first
sorted according to their grey correlation coefficient. Then, a series 4.3. Multiple battery datasets
of cross-validation tests are performed by increasing the number of
In the previous section, the effectiveness of the proposed method has
input HIs one by one from high to low correlation degrees. Finally,
been verified with single battery datasets. In certain special application
the optimal number of HIs can be determined with an optimization
scenarios, such as satellite batteries, their capacity data are often diffi-
target of minimizing the prediction error. In this way, the HIs optimized
cult to obtain. To solve this issue, researchers often establish the same
by cross-validation tests are considered as inputs for the RBM layer
working environment and the same tasks for the same types of batteries
of a DBN model for further information extraction to obtain better
used to train a prognostic model based on synchronous experimental
prediction performance. data to predict the RULs of batteries without their capacity data.
Thus, multiple battery dataset experiments are conducted to test the
4. Experimental results and discussion effectiveness of the proposed method in full life cycle prediction in this
section. Due to the lack of predicted battery capacity data, the optimal
number of input HIs cannot be determined by cross-validation tests;
4.1. Experimental results thus, 17 HIs are employed in the multibattery prediction model.
These experiments are based on the same group of battery datasets
Verification experiments are conducted on single battery datasets as in Section 4.2, in which the second and third battery datasets are
and multiple battery datasets of lithium-ion batteries from the PCoE applied to training and the last battery dataset is selected for testing. A
dataset. The root mean square error (RMSE) is applied to evaluate the special experiment is performed on B0033 with the B0034 and B0036
performance of the proposed method, which is presented as follows: datasets considered the training dataset. B0033 is discharged at 4 A,
√ and then, B0034 and B0036 are discharged at 2 A and 4 A, which
√ 𝑛
√1 ∑ can maximize and verify the robustness of the prediction model in the
𝑅𝑀𝑆𝐸 = √ ̄ 2
(𝑥(𝑖) − 𝑥(𝑖)) (5) hybrid operating conditions. The estimated results are given in Fig. 8,
𝑛 𝑖=1
and the estimation errors of multiple battery experiments are listed in
where 𝑥(𝑖) is the 𝑖th actual battery capacity value and 𝑥(𝑖)
̄ is the 𝑖th Table 6. Compared to the single dataset experimental results, the error
predicted battery capacity value. of prediction is much higher. Among them, the relative errors of B0028,

7
M. Cao et al. Journal of Energy Storage 48 (2022) 103825

Table 3
Grey relational grades of HIs.
Battery label C1 C2 C M1 M2 M1/M2 T1 T2 T T1/C1 T2/C2 T/C S1 S2 SE1 SE2 SE3
B0005 0.9789 0.6229 0.9826 0.9790 0.6237 0.9170 0.9605 0.6161 0.6860 0.6774 0.7393 0.6078 0.6529 0.7050 0.7069 0.7258 0.7777
B0006 0.9624 0.6247 0.8576 0.9628 0.6399 0.9117 0.9672 0.6311 0.7013 0.7415 0.7983 0.6559 0.6500 0.6509 0.7538 0.7906 0.7950
B0007 0.8022 0.6567 0.7824 0.8026 0.6504 0.8713 0.8108 0.6428 0.7110 0.7193 0.7006 0.7015 0.7134 0.6220 0.7327 0.8418 0.7842
B0018 0.7443 0.6745 0.7252 0.7424 0.6949 0.7715 0.7582 0.6980 0.7350 0.6928 0.7873 0.7150 0.7558 0.8464 0.7701 0.7522 0.7816
B0025 0.8331 0.7844 0.9021 0.8325 0.8160 0.7197 0.7837 0.8296 0.8264 0.7294 0.7201 0.7705 0.7628 0.7336 0.6762 0.6987 0.7020
B0026 0.6310 0.6767 0.6981 0.6332 0.7303 0.6572 0.6441 0.7207 0.7404 0.7101 0.7033 0.6600 0.5880 0.6425 0.6492 0.5596 0.6352
B0027 0.7500 0.7372 0.8411 0.7484 0.7728 0.7229 0.7476 0.7913 0.8282 0.7893 0.7731 0.7930 0.7482 0.7786 0.6798 0.6873 0.7311
B0028 0.8133 0.8025 0.9398 0.8035 0.8255 0.7136 0.7475 0.8434 0.8347 0.7208 0.7444 0.7714 0.7909 0.8415 0.6566 0.7131 0.7216
B0029 0.9535 0.7131 0.9448 0.9529 0.7609 0.8706 0.8982 0.7620 0.8370 0.7106 0.7923 0.6995 0.7637 0.8485 0.6996 0.7716 0.7006
B0030 0.9534 0.8090 0.9581 0.9516 0.7781 0.8638 0.8820 0.7799 0.8719 0.7066 0.7540 0.7071 0.7903 0.8624 0.6797 0.7469 0.7141
B0031 0.9301 0.7739 0.9352 0.9307 0.7973 0.8351 0.8361 0.7990 0.8419 0.7039 0.8098 0.7509 0.8401 0.8722 0.7149 0.7006 0.7043
B0032 0.9483 0.6857 0.9310 0.9494 0.7648 0.8842 0.9253 0.7657 0.8635 0.7027 0.7809 0.7011 0.8740 0.8485 0.7296 0.7917 0.7105
B0033 0.9029 0.7655 0.9060 0.9040 0.6527 0.7359 0.9157 0.7239 0.7075 0.8003 0.6908 0.7306 0.7650 0.8729 0.8441 0.8065 0.8566
B0034 0.9417 0.6912 0.9510 0.9439 0.6025 0.8029 0.8913 0.6300 0.6338 0.7885 0.6385 0.6570 0.8335 0.8953 0.8057 0.8145 0.8411
B0036 0.8047 0.7038 0.8154 0.8088 0.6097 0.7832 0.8015 0.6424 0.6459 0.7516 0.6869 0.6959 0.8153 0.8528 0.8476 0.6712 0.7736
B0045 0.8831 0.7585 0.7705 0.8833 0.6413 0.8852 0.8726 0.7959 0.8013 0.8060 0.7454 0.7453 0.6758 0.6806 0.5988 0.7173 0.6780
B0046 0.8649 0.6912 0.7102 0.8644 0.6986 0.8731 0.7968 0.7099 0.7309 0.6854 0.7600 0.7626 0.8016 0.8606 0.7895 0.8053 0.7535
B0047 0.9014 0.6512 0.7052 0.9003 0.6461 0.8992 0.8817 0.6789 0.7550 0.7296 0.7853 0.7881 0.7599 0.8436 0.7679 0.7195 0.7819
B0048 0.8961 0.6732 0.7093 0.8962 0.6739 0.9000 0.8668 0.7003 0.7699 0.7328 0.7664 0.7871 0.7859 0.7506 0.7640 0.7935 0.7801
B0053 0.7342 0.7959 0.8698 0.7343 0.7896 0.7379 0.6645 0.7685 0.6654 0.6657 0.6947 0.6878 0.7942 0.7686 0.8186 0.7429 0.7636
B0054 0.7937 0.8678 0.8252 0.7993 0.6111 0.8259 0.7278 0.6168 0.6594 0.7279 0.6928 0.6973 0.8214 0.8412 0.8321 0.8166 0.7941
B0055 0.8794 0.8971 0.8905 0.8805 0.5800 0.8681 0.8158 0.5845 0.5940 0.7220 0.6355 0.6313 0.8762 0.8334 0.7691 0.8860 0.7738
B0056 0.7567 0.8622 0.8566 0.7566 0.6240 0.8270 0.7572 0.6367 0.6154 0.7379 0.6881 0.6771 0.8177 0.8663 0.8283 0.8433 0.7785
Average 0.8548 0.7356 0.8482 0.8548 0.6950 0.8207 0.8240 0.7116 0.7416 0.7284 0.7342 0.7128 0.7686 0.7964 0.7441 0.7564 0.7536
Sort result 2 12 3 1 17 5 4 16 11 14 13 15 7 6 10 8 9

Table 4
Results of the cross-validation test.
RMSE(%) Number of input HIs
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
B0005 1.05 1.86 1.95 1.11 1.73 0.97 1.00 0.82 1.51 1.43 1.83 2.33 3.05 2.68 3.55 3.78 4.14
B0025 3.09 2.79 3.00 2.90 2.83 2.81 2.72 2.80 2.55 2.42 2.53 2.60 2.52 2.74 2.52 2.45 2.43
B0029 7.88 8.50 2.75 3.37 6.33 4.44 3.04 6.37 3.57 2.94 0.79 3.71 3.61 4.31 3.84 3.18 2.08
B0045 1.96 1.90 1.78 1.95 1.97 1.94 2.33 2.30 2.25 2.47 1.92 2.05 2.00 2.07 1.89 1.76 1.92
B0053 3.24 3.24 3.26 3.30 3.26 3.25 3.06 3.05 3.31 3.33 3.32 3.32 3.30 3.33 3.30 3.29 3.30

Table 5 estimation without capacity data or the difficulty in obtaining capacity

Quantitative results of the proposed method on the single lithium-ion battery datasets.
data, such as via application on satellite batteries.
B0005 B0025 B0029 B0045 B0053
However, there are still some challenges to be discussed before this
𝑅𝑀𝑆𝐸80∕20 0.005769 0.007933 0.006831 0.007235 0.021297 method can be widely used in practice. First, it should be noted that
𝑅𝑀𝑆𝐸70∕30 0.005961 0.013023 0.007127 0.012095 0.029890
the batteries worked under stable operating conditions, i.e., the same
𝑅𝑀𝑆𝐸60∕40 0.007323 0.014554 0.007351 0.014266 0.027092
charge–discharge mechanism in each cycle in the above experiments.
However, batteries in practice usually work with multiple discharging
Table 6 schedules, such as the federal urban driving schedule (FUDS) and
Quantitative results of the proposed method for multiple lithium-ion battery datasets.
dynamic stress test (DST). Although the HI extraction method based on
B0018 B0028 B0032 B0048 B0056 B0033
the charging process can alleviate the influence of the variation in the
𝑅𝑀𝑆𝐸 0.035240 0.027565 0.019784 0.011054 0.060549 0.043938 discharging schedule to a certain extent, whether this model is suitable
for unstable conditions is still an open question. The other problem is
that the training of DBN consists of pretraining and fine-tuning, which
B0032, and B0048 do not exceed 5%. Comparably, the relative errors of is time-consuming when used on too much data. Fortunately, we may
B0018, B0056 and B0033 are still within 10% despite several isolated train the DBN model offline or use the sliding window method to solve
points. The multibattery prediction results show that the degradation this problem. In general, the proposed method is fairly promising for
trends of battery capacity are accurately captured by the prediction real-world applications after further improvement.
model, which verifies its robustness and effectiveness.
5. Conclusions
4.4. Discussion
In this work, a deep belief network algorithm is adopted for capacity
As seen in Section 4, the proposed method achieves a satisfactory estimation of lithium-ion batteries. The health indicators applied as
estimation accuracy on the single battery dataset and multiple battery the input of the deep belief network model are extracted from the
datasets with charging phase HIs and trained DBN models. Due to the charging process in the constant current and constant voltage phases.
HIs extracted from the charging phase without complex computation, Before the health indicators are input into the deep belief network
this method is easier to integrate into the BMS of electric vehicles (EVs) model, grey relation analysis is conducted with the HIs to sort them
to monitor the capacity degradation of batteries. In addition, in multi- by the relevance of the extracted health indicators. Moreover, a cross-
ple battery dataset predictions, we tried to utilize battery data of the validation test is performed based on the sorted health indicators
same operation conditions to estimate the battery capacity without its for denoising and optimization of health indicators. The verification
capacity data, and the estimation results showed acceptable accuracy. experiments are performed on the PCoE dataset of NASA. According
Therefore, this method could solve the problem of battery degradation to the different working conditions of batteries, the experiments are

8
M. Cao et al. Journal of Energy Storage 48 (2022) 103825

Fig. 7. RUL prediction results for the single battery datasets.

separated into two types: single datasets and multiple datasets. When mean square error is 2.70%. Although the error of prediction in the
data segmentation is set to 60%/40%, the root mean square error of multiple dataset experiments is higher than that in the single dataset
the single datasets does not exceed 1.5%, except for B0053, whose root experiments, the forecast trends correspond to the actual value. Apart

9
M. Cao et al. Journal of Energy Storage 48 (2022) 103825

The results presented in this paper will facilitate further improve-

ment of the approach. In subsequent studies, we will focus on the
following aspects: (1) The fully connected layer of the deep belief
network can be replaced by other machine learning methods to further
improve the estimation accuracy. (2) Based on existing data, ensemble
learning can be combined with deep learning, which will enhance the
robustness of the prognostic method.

CRediT authorship contribution statement

Mengda Cao: Conceptualization, Software, Writing – original draft

preparation. Tao Zhang: Investigation, Funding acquisition. Jia Wang:
Visualization. Yajie Liu: Conceptualization, Writing – review & editing.

Declaration of competing interest

The authors declare that they have no known competing finan-

cial interests or personal relationships that could have appeared to
influence the work reported in this paper.

Acknowledgement

This work was supported by the National Science Foundation of

China (61973310, 71901210).

References

[1] L. Chang, C. Wang, C. Zhang, L. Xiao, N. Cui, H. Li, J. Qiu, A novel fast capacity
estimation method based on current curves of parallel-connected cells for retired
lithium-ion batteries in second-use applications, J. Power Sources 459 (2020)
227901.
[2] H. Yao, Q. Zhao, Z.-L. Song, Z.-J. Cheng, X. Jia, B. Guo, Lithium-ion battery
capacity estimation based on features selection by grey relational analysis, in:
2018 12th International Conference on Reliability, Maintainability, and Safety,
ICRMS, IEEE, 2018, pp. 312–316.
[3] G. Zhao, G. Zhang, Y. Liu, B. Zhang, C. Hu, Lithium-ion battery remaining useful
life prediction with deep belief network and relevance vector machine, in: 2017
IEEE International Conference on Prognostics and Health Management, ICPHM,
IEEE, 2017, pp. 7–13.
[4] D. Le, X. Tang, Lithium-ion battery state of health estimation using Ah-V
characterization, Annu. Conf. Progn. Health Manag. Soc. 73 (3) (2011) 367–373.
[5] K. Yu, M. Mukai, T. Kawabe, A battery management system using nonlinear
model predictive control for a hybrid electric vehicle, IFAC Proc. Vol. 46 (21)
(2013) 301–306.
[6] Y. Wang, D. Yang, X. Zhang, Z. Chen, Probability based remaining capacity
estimation using data-driven and neural network model, J. Power Sources 315
(2016) 199–208.
[7] X.-S. Si, W. Wang, C.-H. Hu, D.-H. Zhou, Remaining useful life estimation–a
review on the statistical data driven approaches, European J. Oper. Res. 213 (1)
(2011) 1–14.
[8] B. Saha, K. Goebel, J. Christophersen, Comparison of prognostic algorithms for
estimating remaining useful life of batteries, Trans. Inst. Meas. Control 31 (3–4)
(2009) 293–308.
[9] A.V. Virkar, A model for degradation of electrochemical devices based on linear
non-equilibrium thermodynamics and its application to lithium ion batteries, J.
Power Sources 196 (14) (2011) 5970–5984.
[10] Z. Shuzhi, G. Xu, Z. Xiongwen, A novel one-way transmitted co-estimation
framework for capacity and state-of-charge of lithium-ion battery based on
double adaptive extended Kalman filters, J. Energy Storage 33 (2021) 102093.
[11] Z. Wang, J. Ma, L. Zhang, State-of-health estimation for lithium-ion batteries
based on the multi-island genetic algorithm and the Gaussian process regression,
Ieee Access 5 (2017) 21286–21295.
[12] X. Kong, G.L. Plett, M.S. Trimboli, Z. Zhang, D. Qiao, T. Zhao, Y. Zheng, Pseudo-
two-dimensional model and impedance diagnosis of micro internal short circuit
in lithium-ion cells, J. Energy Storage 27 (2020) 101085.
[13] X. Zhang, J. Lu, S. Yuan, J. Yang, X. Zhou, A novel method for identifi-
Fig. 8. RUL prediction results for multiple battery datasets. cation of lithium-ion battery equivalent circuit model parameters considering
electrochemical properties, J. Power Sources 345 (2017) 21–29.
[14] C. Zhang, W. Allafi, Q. Dinh, P. Ascencio, J. Marco, Online estimation of battery
equivalent circuit model parameters and state of charge using decoupled least
from B0056, the root mean square errors of the batteries do not exceed squares technique, Energy 142 (2018) 678–688.
4.5%. The abovementioned prediction results verify the robustness and [15] A. Jokar, B. Rajabloo, M. Désilets, M. Lacroix, Review of simplified pseudo-
two-dimensional models of lithium-ion batteries, J. Power Sources 327 (2016)
effectiveness of the proposed method. 44–55.

10
M. Cao et al. Journal of Energy Storage 48 (2022) 103825

[16] V. Laue, F. Röder, U. Krewer, Practical identifiability of electrochemical P2D [33] Y. Li, M. Abdel-Monem, R. Gopalakrishnan, M. Berecibar, E. Nanini-Maury,
models for lithium-ion batteries, J. Appl. Electrochem. 51 (9) (2021) 1253–1265. N. Omar, P. van den Bossche, J. Van Mierlo, A quick on-line state of health
[17] N. Lotfi, J. Li, R.G. Landers, J. Park, Li-ion battery state of health estima- estimation method for li-ion battery with incremental capacity curves processed
tion based on an improved single particle model, in: 2017 American Control by Gaussian filter, J. Power Sources 373 (2018) 40–53.
Conference, ACC, IEEE, 2017, pp. 86–91. [34] W. Li, N. Sengupta, P. Dechent, D. Howey, A. Annaswamy, D.U. Sauer, Online
[18] M.-F. Ge, Y. Liu, X. Jiang, J. Liu, A review on state of health estimations and capacity estimation of lithium-ion batteries with deep long short-term memory
remaining useful life prognostics of lithium-ion batteries, Measurement (2021) networks, J. Power Sources 482 (2021) 228863.
109057. [35] Y. Zheng, J. Wang, C. Qin, L. Lu, X. Han, M. Ouyang, A novel capacity estimation
[19] C.I. Ossai, N. Raghavan, Statistical characterization of the state-of-health of method based on charging curve sections for lithium-ion batteries in electric
lithium-ion batteries with Weibull distribution function—A consideration of vehicles, Energy 185 (2019) 361–371.
random effect model in charge capacity decay estimation, Batteries 3 (4) (2017) [36] D. Yang, X. Zhang, R. Pan, Y. Wang, Z. Chen, A novel Gaussian process regression
32. model for state-of-health estimation of lithium-ion battery using charging curve,
[20] C. Weng, X. Feng, J. Sun, H. Peng, State-of-health monitoring of lithium-ion J. Power Sources 384 (2018) 387–395.
battery modules and packs via incremental capacity peak tracking, Appl. Energy [37] D. Liu, J. Zhou, H. Liao, Y. Peng, X. Peng, A health indicator extraction
180 (oct.15) (2016) 360–368. and optimization framework for lithium-ion battery degradation modeling and
[21] H. Chaoui, C.C. Ibe-Ekeocha, State of charge and state of health estimation for prognostics, IEEE Trans. Syst. Man Cybern. Syst. 45 (6) (2015) 915–928.
lithium batteries using recurrent neural networks, IEEE Trans. Veh. Technol. 66 [38] C. Lu, L. Tao, H. Fan, Li-ion battery capacity estimation: A geometrical approach,
(10) (2017) 8773–8783. J. Power Sources 261 (2014) 141–147.
[22] K. Khan, T. Hossen, A. Savasci, L. Gauchia, S. Paudyal, Design of a simplified [39] S. Song, C. Fei, H. Xia, Lithium-ion battery SOH estimation based on xgboost
hierarchical Bayesian network for residential energy storage degradation, in: algorithm with accuracy correction, Energies 13 (4) (2020) 812.
2019 IEEE Power & Energy Society General Meeting, PESGM, IEEE, 2019, pp. [40] Y. Li, H. Sheng, Y. Cheng, D.-I. Stroe, R. Teodorescu, State-of-health estimation
1–5. of lithium-ion batteries based on semi-supervised transfer component analysis,
[23] C. Zhang, Y. He, L. Yuan, S. Xiang, Capacity prognostics of lithium-ion bat- Appl. Energy 277 (2020) 115504.
teries using EMD denoising and multiple kernel RVM, IEEE Access 5 (2017) [41] Y. Zhang, B. Guo, Online capacity estimation of lithium-ion batteries based on
12061–12070. novel feature extraction and adaptive multi-kernel relevance vector machine,
[24] A. Guha, A. Patra, State of health estimation of lithium-ion batteries using Energies 8 (11) (2015) 12439–12457.
capacity fade and internal resistance growth models, IEEE Trans. Trans. Electr. [42] P. Guo, Z. Cheng, L. Yang, A data-driven remaining capacity estimation approach
4 (1) (2017) 135–146. for lithium-ion batteries based on charging health feature extraction, J. Power
[25] S. Shen, M. Sadoughi, M. Li, Z. Wang, C. Hu, Deep convolutional neural Sources 412 (2019) 442–450.
networks with ensemble learning and transfer learning for capacity estimation [43] Y. Guo, K. Huang, X. Hu, A state-of-health estimation method of lithium-ion
of lithium-ion batteries, Appl. Energy 260 (2020) 114296. batteries based on multi-feature extracted from constant current charging curve,
[26] M. Savargaonkar, A. Chehade, Z. Shi, A.A. Hussein, A cycle-based recurrent J. Energy Storage 36 (2021) 102372.
neural network for state-of-charge estimation of li-ion battery cells, in: 2020 [44] D.E. Demirocak, B. Bhushan, Probing the aging effects on nanomechanical
IEEE Transportation Electrification Conference & Expo, ITEC, IEEE, 2020, pp. properties of a LiFePO4 cathode in a large format prismatic cell, J. Power Sources
584–587. 280 (2015) 256–262.
[27] H. Zhang, W. Tang, W. Na, P.-Y. Lee, J. Kim, Implementation of generative [45] W. Diao, S. Saxena, M. Pecht, Accelerated cycle life testing and capacity
adversarial network-CLS combined with bidirectional long short-term memory degradation modeling of LiCoO2-graphite cells, J. Power Sources 435 (2019)
for lithium-ion battery state prediction, J. Energy Storage 31 (2020) 101489. 226830.
[28] M. Fasahat, M. Manthouri, State of charge estimation of lithium-ion batteries [46] C.R. Birkl, M.R. Roberts, E. McTurk, P.G. Bruce, D.A. Howey, Degradation
using hybrid autoencoder and long short term memory neural networks, J. Power diagnostics for lithium ion cells, J. Power Sources 341 (2017) 373–386.
Sources 469 (2020) 228375. [47] Q. Wang, P. Wu, J. Lian, SOC estimation algorithm of power lithium battery
[29] H.-I. Suk, S.-W. Lee, D. Shen, A.D.N. Initiative, et al., Hierarchical feature based on AFSA-BP neural network, J. Eng. 2020 (13) (2020) 535–539.
representation and multimodal fusion with deep learning for AD/MCI diagnosis, [48] E.J. Hartman, J.D. Keeler, J.M. Kowalski, Layered neural networks with Gaussian
NeuroImage 101 (2014) 569–582. hidden units as universal approximations, Neural Comput. 2 (2) (1990) 210–215.
[30] M. Cao, T. Zhang, B. Yu, Y. Liu, A method for interval prediction of satel- [49] P. Tamilselvan, P. Wang, Failure diagnosis using deep belief learning based
lite battery state of health based on sample entropy, IEEE Access 7 (2019) health state classification, Reliab. Eng. Syst. Saf. 115 (2013) 124–135.
141549–141561. [50] Y. Bengio, P. Lamblin, D. Popovici, H. Larochelle, Greedy layer-wise training of
[31] J. Wu, C. Zhang, Z. Chen, An online method for lithium-ion battery remaining deep networks, in: Advances in Neural Information Processing Systems, 2007,
useful life estimation using importance sampling and neural networks, Appl. pp. 153–160.
Energy 173 (2016) 134–140. [51] C. CUI, M. YANG, C. YANG, L. TIAN, Y. WEN, Estimation of standard enthalpy of
[32] L. Chen, L. Xu, Y. Zhou, Novel approach for lithium-ion battery on-line remaining Ni-Si eutectic alloy and study on crystal growth mechanism, Hot Work. Technol.
useful life prediction based on permutation entropy, Energies 11 (4) (2018) 820. (2) (2017) 35.

11