IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 69, NO.

7, JULY 2022 7019

A Fast Impedance Calculation-Based Battery

State-of-Health Estimation Method
Yumeng Fu, Jun Xu , Member, IEEE, Mingjie Shi, and Xuesong Mei

Abstract—State-of-health (SOH) is crucial to the main- (SOH) is generally defined as the ratio of the current maximum
tenance of various kinds of energy storage systems, in- capacity to the nominal capacity [1], [2]. Accurately estimating
cluding power batteries. Relevant research articles are SOH is one of the core functions of the battery management
mostly based on battery external information, such as
current, voltage, and temperature, which are suscepti- system (BMS) [3], [4]. High-quality SOH estimation helps to
ble to fluctuation and ultimately affects the SOH estima- improve battery energy utilization, ensure operational safety,
tion accuracy. In this article, to solve these problems, a promote the maintenance efficiency of the battery system, and
fast impedance calculation-based battery SOH estimation optimizes the energy management of BMS [5], [6]. However,
method for lithium-ion battery is proposed from the per- power batteries represented by lithium-ion batteries inherently
spective of electrochemical impedance spectroscopy (EIS).
The relationship between EIS and state of charge and that have some unfriendly characteristics, including limited num-
between EIS and degraded capacity is first studied by ex- ber of measurable parameters, intense nonlinearity, and strong
perimental tests. Some impedance features called health time-varying property. Besides, batteries used in electric vehi-
factors effectively indicating battery aging states are se- cles usually face complicated and variable environments under
lected. Second, an improved fast Fourier transform (FFT) various operating conditions [7]. Therefore, it is challenging to
utilizing the conversion relationship between the real and
complex signals is proposed to realize online fast EIS ac- obtain high-precision SOH estimation results for these adverse
quisition. Compared with ordinary FFT, such treatments re- factors.
duce computational complexity. Then, the SOH evaluation There have been many SOH estimation methods reported,
model is built by the extreme learning machine with regu- which can be generally classified into the experimental methods
larization mechanism, further reducing the computational and the model-based methods [8]. The most direct experimental
burden. The relationship between the health factors and
aging capacity of batteries is established. Finally, an exper- method is to calibrate battery capacity by Ampere-hour counting
imental bench is established. The results indicate that the [9]. However, this method is time consuming since it has to
estimated SOH can be obtained within 35 s for a four-cell fully charge and discharge the battery and not available online.
series-connected battery pack and the estimation errors are Besides, it is vulnerable to accumulative errors caused by sensor
less than 2%. accuracy. Computing internal resistance using current pulse and
Index Terms—Electrochemical impedance spectroscopy voltage drop is another feasible approach, but this method can
(EIS), extreme learning machine (ELM), fast Fourier trans- only make a rough SOH evaluation since the internal resistance
form (FFT), fast impedance calculation, state-of-health
(SOH) estimation.
is greatly affected by the current amplitude, state of charge
(SOC), temperature, etc. The charging curve is also used to
I. INTRODUCTION demonstrate SOH for it may change as the battery degradation
processes [10], but to obtain the complete charging curve is
ONSIDERED as a quantitative indicator describing the
C degree of battery aging and usage, the state-of-health
usually time consuming. Furthermore, some researchers pro-
posed incremental capacity analysis (ICA) [11] and differential
voltage analysis (DVA) [12] to get regular featured peaks in
Manuscript received February 1, 2021; revised April 16, 2021 and
June 11, 2021; accepted June 27, 2021. Date of publication July 21, ICA or DVA curves to assess SOH but they were almost only
2021; date of current version February 9, 2022. This work was sup- suitable for small current charging/discharging conditions. In
ported in part by the National Natural Science Foundation of China a word, the experimental methods are logically intuitive and
under Grant 52075420, in part by the National Key Research and De-
velopment Program of China under Grant 2020YFB1708400, in part by but not suitable for online applications for the time consum-
the Shaanxi Key Research and Development Plan Industry Innovation ing or the requirements of charging/discharging conditions’
Chain Project under Grant 2019ZDLGY15-04-01. (Corresponding au- problems.
thor: Jun Xu.)
The authors are with the State Key Laboratory for Manufacturing The model-based methods mainly consist of electrochemical
System Engineering, Xi’an Jiaotong University, Xi’an 710049, China, model-based, equivalent circuit model-based, and data-driven
and also with the Shaanxi Key Laboratory of Intelligent Robots, methods. Xiong et al. [8] used electrochemical models, pseudo-
Xi’an Jiaotong University, Xi’an 710049, China (e-mail: fym19961003@
stu.xjtu.edu.cn; [email protected]; [email protected]; two-dimensional model, to reasonably describe the battery in-
[email protected]). ternal electrochemical mechanism with some model parameters
Color versions of one or more figures in this article are available at reflecting the battery decay. However, it is extremely high com-
https://doi.org/10.1109/TIE.2021.3097668.
Digital Object Identifier 10.1109/TIE.2021.3097668 putational to identify model parameters and solve related partial

0278-0046 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: VISVESVARAYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on February 10,2023 at 09:28:41 UTC from IEEE Xplore. Restrictions apply.
 7020 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 69, NO. 7, JULY 2022

TABLE I the corresponding relationship between the ohmic resistance in

MAIN CHARACTERISTICS OF THE EXISTING SOH ESTIMATION METHODS
the high-frequency range of EIS and the aging capacity [23].
Jiang et al. [24] presented the relation trend between the EIS
phase at a specific frequency point and the aging capacity of the
lithium battery, and the SOH evaluation model is established.
But generally there exist two main problems in these methods.
First, the EIS frequency range selected is narrow and the number
of EIS features used for modeling is quite small. Second, these
methods are difficult to realize online because the traditional EIS
test approach takes long time to obtain battery impedance infor-
mation by sweeping frequency on expensive specialized precise
instruments [25].
To quickly obtain the EIS of lithium-ion batteries, researchers
have proposed some calculation methods using time-domain
signal processing techniques [26]. The time-domain methods
apply a transient signal to the battery. Then, the input and re-
sponse signals are mathematically transformed to the frequency
domain. EIS is obtained from the quotient of the response
signal spectrum and the input signal spectrum. Compared with
differential equations. Equivalent circuit models are much sim- the EIS calculated by sweeping frequency, the measurement
pler than electrochemical models and usually utilized together duration can be effectively reduced. Arora et al. [27] applied
with adaptive filtering methods. Most of these methods derive a step signal to the equivalent circuit model and transformed the
from Kalman filtering (KF), such as extended KF, unscented response into s domain through Laplace transform to obtain the
KF, adaptive extended KF, and so on [8]. In addition, parti- model impedance. Katayama and Kogoshi [28] applied mixed
cle filtering [13], proportional–integral observer, and H-infinity sinusoidal current signals of the same amplitude, different fre-
observer are also applied. Such methods are relatively easy to quencies and phases to a lithium-ion battery, and then obtained
implement online application, but strict requirements are placed the EIS through Fourier transform (FT) method. Gomez-Luna et
on the accuracy of model parameters identification. Data-driven al. [29] proposed the use of continuous wavelet transform to di-
methods [14] are popular recently in SOH estimation-related agnose battery impedance based on transient response. Debenjak
research field, because these methods utilize statistical learning et al. [30] used a pseudorandom binary sequence as a disturbance
techniques [15], such as machine learning [16] to mine potential signal and successfully estimated the battery cell impedance
laws from battery historical data to establish data-based models using Morlet continuous wavelet transform. However, there
[17] for SOH estimation, instead of studying on complicated is still room for improvement in the simplicity of impedance
knowledge and sophisticated battery internal mechanisms. How- calculation.
ever, due to heavy calculational burden and high algorithm To solve these problems, a fast EIS calculation method based
complexity, data-driven methods are difficult to be applied on- on improved fast FT (FFT) is introduced to obtain the EIS of each
line. Main characteristics of these existing methods are listed in cell in a series pack in this article. Utilizing the decomposition
Table I. and synthesis relationship between the complex signals and real
The above-mentioned methods utilize easily measured ex- signals, this approach further reduces the calculation cost of
ternal signals, including current, voltage, and temperature, to signal spectrum processing effectively. This method, along with
describe and analyze the characteristics of the battery. However, an EIS-feature-based SOH estimation model using regularized
such signals are normally affected by environmental condi- extreme learning machine (ELM), constitutes the whole SOH
tions, equipment performances, and measurement methods. By evaluation framework. As a nonlinear data model, ELM runs
contrast, the battery internal characteristics, such as electro- much quicker than ordinary neural networks because its training
chemical impedance spectroscopy (EIS), can reflect the decay process is based on matrix algebra operations instead of recur-
degree and health state of the battery better [18]–[20]. As sive iterative calculation. It makes the proposed method suitable
an important electrochemical analysis technique, EIS displays for online applications.
battery internal impedance information in a wide frequency This article is organized as follows. In Section II, the ex-
range, demonstrating some interior reactions, such as interface perimental platform is established to study on the relation-
reaction, charge transfer, and diffusion [21]. Generally, under ships among EIS, SOC, and aging capacity. In Section III,
a steady external environment, such as constant temperature, first, the relevant EIS features are explored and the SOH
the EIS characteristics of a cell can usually remain stable. This evaluation model based on regularized ELM is established.
makes SOH estimation based on EIS related features more Then, the improved FFT approach is introduced and sim-
advantageous. De Sutter et al. [22] found the change rule of EIS ulated on the fractional-order model. In Section IV, the
curve with the increase of aging cycles for lithium-ion batteries. whole framework is validated by joint experiments. Finally,
Galeotti et al. [23] proposed an SOH estimation method using Section V concludes this article.

Authorized licensed use limited to: VISVESVARAYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on February 10,2023 at 09:28:41 UTC from IEEE Xplore. Restrictions apply.
 FU et al.: FAST IMPEDANCE CALCULATION-BASED BATTERY STATE-OF-HEALTH ESTIMATION METHOD 7021

TABLE II
BASIC PARAMETERS FOR NCR18650A BATTERY

Fig. 2. EIS at five different SOCs.

selected cell is charged to 4.2 V, namely 100% SOC. Next, the

cell rests for at least 3 h to obtain a stable internal state before
EIS measurement. Then, it is clamped on the electrochemical
workstation to start measuring EIS with a frequency range
Fig. 1. The maximum available capacity and capacity differences of from 0.001 to 1000 Hz. After the accomplishment, the cell is
the batteries tested.
discharged with constant current for decreasing SOC level to
80%, 60%, 40%, and 20% in sequence, thus EIS is measured
at different SOCs. A rest for 3 h is always necessary between
II. EXPERIMENTAL TEST FOR THE EIS AND every discharge and EIS measurement.
RELATIONSHIP ANALYSIS The measurement results at five different SOCs for this cell
To realize the health state assessment of lithium-ion batter- (2.905 Ah) is displayed in Fig. 2. Totally 61 data points are
ies from the perspective of EIS, the relationship between the collected with 10 recorded every decade. The horizontal and
impedance characteristics and main battery parameters should vertical axes represent real and imaginary part of EIS, respec-
be explored first. In this part, a batch of lithium-ion batteries tively. The measurement result curves show a relatively obvious
with different aging degrees is selected as samples. A test bench trend in terms of SOC. Each curve has an inflection point
is set up to achieve EIS measurement. The relationship between about the imaginary part near the real part of 0.2 Ω, which here is
SOC and EIS and that between the capacity of batteries with called the low-frequency inflection point. All five low-frequency
different aging degrees and EIS under the same SOC level is inflection points correspond to the frequency of 0.0215 Hz. The
investigated. inflection point divides each curve into two parts. In the left
part, where the frequency is between 0.0215 and 1000 Hz, both
real and imaginary parts of EIS vary small with the change
A. Experimental Test Platform
of SOC when frequency is fixed. However, in the right area,
An experimental platform for battery EIS measurement is where the frequency is between 0.001 and 0.0215 Hz, EIS
established to study the relationship between EIS and battery curves shift downward with SOC decreasing, and the real part
aging capacity. It includes an electrochemical workstation, a almost remains unchanged, while the imaginary part gradually
thermal chamber, a battery test system, and a host computer. The decreases as SOC reduces.
electrochemical workstation is used to obtain battery EIS. The
thermal chamber is to keep the battery in a constant temperature C. Study on EIS–Capacity Relationship
surrounding, specifically 25 °C in this study. The battery test
system is utilized to charge and discharge to keep the battery To explore how the aging capacity influences EIS, all 25
at any SOC. The NCR18650A cylindrical lithium-ion battery sample batteries are fixed at the same SOC level to eliminate
is studied. Its basic parameters are given in Table II. Total the effect from SOC. All samples are charged to 4.2 V first to
25 batteries with good capacity gradient are chosen from the get 100% SOC. Then, their EIS are measured after fully rested.
laboratory battery set as the experimental samples. All samples The frequency range here is from 0.01 to 1000 Hz. The result is
have gone through different working cycles and they are in displayed in Fig. 3. Intuitively it can be seen that these curves
various aging states. The current maximum available capacity of are more concentrated at high-frequency range. Otherwise in
the samples and the absolute capacity differences between each the middle- and low-frequency areas, as the capacity decreases,
sample and the maximum are shown in Fig. 1. the curves overall move to upper right and the elliptical arc
contour in the middle-frequency area gradually become larger. In
addition, the low-frequency inflection point only appears on the
B. EIS–SOC Relationship Analysis
curves of Samples 1–7, which are in relatively low aging states.
The EIS characteristics at various SOC levels are explored. According to above analysis, the following conclusions can be
The experimental procedure is described as follows. First, the drawn. The SOC primarily affects the EIS in the low-frequency

Authorized licensed use limited to: VISVESVARAYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on February 10,2023 at 09:28:41 UTC from IEEE Xplore. Restrictions apply.
 7022 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 69, NO. 7, JULY 2022

part (HF2), and the amplitude at the peak of the imaginary part
(HF3); the imaginary part (HF4) and the amplitude (HF5) at
0.1Hz; the phase peak value (HF6). The selected HFs are shown
in Fig. 5. The length of the red dashed line segment and the blue
dashed line segment represent HF3 and HF5, respectively. The
angle between the green dashed line segment and the real axis
represents HF6. To fully consider the effectiveness of selected
HF, the correlation degree between HFs and aging capacity is
tested using Spearman correlation coefficient. The calculation
method of the correlation coefficient is
Fig. 3. EIS of 25 batteries at 100% SOC. n

Cspearman = 1 − 6 {d2 /[n(n2 − 1)]} (1)
i=1
region below the inflection point frequency. When the frequency where d is the difference between the ranks of two variables,
is fixed, the imaginary part of EIS decreases with the decrease and n is the length of the data sequence. The closer the result is
of SOC, but the real part is almost not affected. Differently, the to 1 or −1, the stronger the positive or negative correlation is.
aging capacity mainly influences EIS both in the middle- and Besides, the p-value is calculated to denote the probability that
low-frequency regions. Along with the capacity reduction both the current correlation coefficient fails to correctly describe the
the real and imaginary parts of EIS show an apparent increasing true correlation. In statistics, it is generally believed a p-value
tendency. Except for seven batteries on a higher capacity level, below 5% means a reliable correlation coefficient. Table III lists
the inflection point frequency of most batteries is lower than the correlation evaluation results. It can be seen from Table III
0.01 Hz. that the Spearman coefficients of selected HFs are all less than
−0.9, even some very close to −1, which can be strong proof of
III. ANALYSIS OF THE PROPOSED SOH intensive negative correlation between these six HFs and aging
ESTIMATION METHOD capacity.
In this section, the proposed SOH estimation method is The ELM algorithm is used to formulate the relation between
explained and analyzed in detail. The overall framework of HFs and capacity. The ELM is an algorithm framework in the
the SOH estimation method is shown in Fig. 4, which mainly form of a three-layer feedforward neural network and has the
includes offline construction of the ELM-based SOH estimation same structure as the basic feedforward neural network. But
model and the online fast EIS calculation and SOH estimation. it is completely different from the traditional neural network
This part is organized by two sections. In Section III-A, the using the error back propagation algorithm. First, the connection
selection of relevant EIS features based on correlation analysis weights between the input layer and the hidden layer and the
is illustrated first. Next, the ELM-based SOH estimation model hidden layer thresholds are randomly set at the beginning, and
with regularization mechanism is established using the selected no longer change during calculation. Second, the connection
EIS features. In Section III-B, the methodology of the fast EIS weights between the hidden layer and the output layer are
calculation utilizing improved FFT is explained mathematically determined by solving algebraic equations instead of repeated
and validated by simulation on the fractional-order impedance iterative computational steps. These characteristics ensure a
model. simpler calculation method and a higher training speed of ELM
algorithm. Naturally, with the six selected HFs as input variables
A. Offline Construction of ELM-Based SOH and the capacity of 25 sample batteries as output variables, an
Estimation Model ELM-based data-driven model for SOH estimation is estab-
lished. The schematic diagram of ELM model is displayed in
To estimate SOH using relevant characteristics of EIS, the in- Fig. 6. The model input sequence is x = [x1 , x2 , . . . , xn ]T and
fluence of SOC on EIS must be considered. But it is worth noting output sequence is y = [y1 , y2 , . . . , ym ]T . The number of hidden
that, in some frequency ranges, the impact of SOC difference neurons is k. The connection weight matrix between the input
on EIS is negligible. Therefore, selecting related EIS features in layer and the hidden layer and that between the hidden layer and
such a frequency range may realize SOH estimation without the output layer are shown as
knowing the current SOC. The EIS feature is called health ⎛ ⎞ ⎛ ⎞
factor (HF) in this study. Several principles should be followed α11 α12 · · · α1n β11 β12 · · · β1m
in the selection of HF. First, the EIS frequency range should ⎜ α21 α22 · · · α2n ⎟ ⎜ β21 β22 · · · β2m ⎟
⎜ ⎟ ⎜ ⎟
be carefully chosen such that HFs in this range are strongly α=⎜ . .. .. .. ⎟ , β = ⎜ .. .. .. .. ⎟ (2)
⎝ .. . . . ⎠ ⎝ . . . . ⎠
correlated to SOH but almost unchanged with the variation of
αk1 αk2 · · · αkn βk1 βk2 · · · βkm
SOC. Second, try to avoid choosing HFs whose frequencies are
far too low, because the acquisition of low-frequency impedance where αji represents the weight between the ith neuron in the
is usually time consuming. Following these selection methods, input layer and the jth neuron in the hidden layer, and βji
the selection consists of six HFs, namely: the peak value of the represents the weight between the jth neuron in the hidden layer
imaginary part (HF1), the real part at the peak of the imaginary and the ith neuron in the output layer. The threshold sequence

Authorized licensed use limited to: VISVESVARAYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on February 10,2023 at 09:28:41 UTC from IEEE Xplore. Restrictions apply.
 FU et al.: FAST IMPEDANCE CALCULATION-BASED BATTERY STATE-OF-HEALTH ESTIMATION METHOD 7023

Fig. 4. Proposed SOH estimation framework.

TABLE III
SPEARMAN COEFFICIENTS BETWEEN HFS AND CAPACITY

The activation function is g(). Then, the output matrix of the

network can be expressed as

Fig. 5. Schematic diagram of selected HFs. T = [t1 , t2 , . . . , tP ]T (4)

and Eqs. (5) and (6) shown at the bottom of next page.
The calculation of the ELM network can be simplified as

in the hidden layer is b = [b1 , b2 , . . . , bk ]T . Suppose the input Hβ = T T . (7)

feature matrix and output matrix of P samples are, respectively
H is the output matrix of the hidden layer and it can be
⎛ ⎞ ⎛ ⎞ specifically expressed as
x11 x12 · · · x1P y11 y12 · · · y1P ⎛ ⎞
⎜ x21 x22 · · · x2P ⎟ ⎜ y21 y22 · · · y2P ⎟ g(α1 x1 + b1 ) g(α2 x1 + b2 ) · · · g(αk x1 + bk )
⎜ ⎟ ⎜ ⎟
X=⎜ . .. .. .. ⎟ , Y = ⎜ .. .. .. .. ⎟ . ⎜ g(α1 x2 + b1 ) g(α2 x2 + b2 ) · · · g(αk x1 + bk ) ⎟
⎝ .. . . . ⎠ ⎝ . . . . ⎠ ⎜ ⎟
H=⎜ .. .. .. .. ⎟.
xn1 xn2 · · · xnP ym1 ym2 · · · ymP ⎝ . . . . ⎠
(3) g(α1 xP + b1 ) g(α2 xP + b2 ) · · · g(αk xP + bk )
(8)

Authorized licensed use limited to: VISVESVARAYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on February 10,2023 at 09:28:41 UTC from IEEE Xplore. Restrictions apply.
 7024 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 69, NO. 7, JULY 2022

Fig. 6. Schematic diagram of the ELM.

The objective function of the network training is obj =

Fig. 7. The training and test error performances when C = 100.
min Hβ − T T . To improve generalization ability, an L2 reg-
β
ularization term is added to the current objective function. The
objective function is formulated as
2
obj = min(C Hβ − T T /2 + β2 /2) (9)
β

where H is the model parameter matrix, T is the target out-

put, and C is the weight coefficient used to control the de-
gree of regularization. The required value of weight β can be
expressed as

T
H (I/C + HH T )−1 T T , P ≤ k
β= (10)
(I/C + H T H)−1 H T T T , P > k

where P is the number of samples and k is the number of

nodes in the hidden layer. After the acquisition of β, all the
model parameters are determined. This means one round of
model training is completed. The data training set and test set
are generated according to the leave-one-out cross-validation
Fig. 8. Capacity estimation results of samples.
method. Specifically, 24 samples compose the training set and
the remaining one individually composes a test set in each round
of training. Every sample becomes a test dataset after totally 25
rounds of training. When the regularization coefficient C is 100, B. Online Fast EIS Calculation
the training and test error performances are shown in Fig. 7. It can FFT is the classic fast algorithm, which is commonly utilized
be clearly seen that the percentage error of all training samples to find the frequency spectrum of the aperiodic discrete-time
in each round of training remains within 2%, and the error of signal. The calculation expression of FFT is displayed as
each test sample does not exceed 2.5%. Besides, the training
error is roughly equivalent to the test error level, indicating that X(k) = FFT[x(n)]
there is no common overfitting problem.

X(k) = X1 (k) + WNkn X2 (k)

The above results show that the aging battery capacity estima- = (11)
X(k + N /2) = X1 (k) − WNkn X2 (k)
tion model based on regularized ELM has acceptable estimation
accuracy and generalization ability. All 25 battery samples are where X1 (k) is the spectrum of even sequence x1 and X2 (k) is
2π
utilized to establish a complete capacity estimation model. The that of odd sequence x2 , WNkn = e−jk N n is the basis function
estimation results, as shown in Fig. 8, show that the SOH and N is the number of sampling points. In fact, the computa-
estimation error based on this model does not exceed 2%. tional cost can be further decreased using conjugate symmetry

T
 T 
k 
k 
k
tj = t1j , t2j , . . . , tmj = βi1 g(αi xj + bi ), βi2 g(αi xj + bi ), . . . , βim g(αi xj + bi ) (5)
i=1 i=1 i=1

αi = [αi1 , αi2 , . . . , αin ], xj = [x1j , x2j , . . . , xnj ]T (6)

Authorized licensed use limited to: VISVESVARAYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on February 10,2023 at 09:28:41 UTC from IEEE Xplore. Restrictions apply.
 FU et al.: FAST IMPEDANCE CALCULATION-BASED BATTERY STATE-OF-HEALTH ESTIMATION METHOD 7025

Fig. 10. The excitation waveform of general step signal with the ampli-
tude of 0.2 A.

Fig. 9. Test bench and its schematic diagram.

of its basis function. The current and voltage sequences can

compose a complex sequence
x(n) = i(n) + ju(n) (12)
where j is an imaginary number unit. Suppose the conjugation
sequence of x(n) is x∗ (n). If the spectrum calculation result of Fig. 11. Amplitude spectrum of voltage response.
x(n) is X(k), the spectrum of x∗ (n) can be calculated utilizing
conjugate symmetry, shown as
N −1 N −1
2.755 Ah (95.01%), 2.669 Ah (92.04%), and 2.562 Ah (88.34%),
  −(N −k)n respectively. A dc power supply and an electronic load compose
DFT[x∗ (n)] = x∗ (n)WNnk = x∗ (n)WN
n=0 n=0
a charge and discharge system to apply excitation signal to
batteries. The loading resolution of the power supply and load
N
 −1
(N −k)n ∗ is 1 mA. The battery signal acquisition platform mainly consists
= x(n)[WN ] = X ∗ (N − k). (13) of a current sensor with high accuracy, a voltage acquisition
n=0
board, a series battery pack, and a main control board. The sensor
Formula (11) tells that the spectrum of x∗ (n) is just equivalent resolution is 1 mA and the total error of the voltage acquisition
to the conjugate of the inverse spectrum sequence of x(n). The board for 12 cells is about 1.3 mV. The core of the main control
real and imaginary parts of x(n) can be expressed by x(n) and board is STM32F103, which is an ARM Cortex-M3 MCU based
x∗ (n) as microcontroller. A LabVIEW-based software is equipped on the
host PC to transmit charge or discharge instructions to the power
i(n) = [x(n) + x∗ (n)]/2, u(n) = [x(n) − x∗ (n)]/(2j). (14)
supply and electronic load. The experimental input current and
Therefore, the spectrum of i(n) and u(n) can be represented output voltage data recorded by the signal acquisition platform
by X(k) and X ∗ (N − k) as can be returned to the host PC for further investigation. The
trained ELM-based model and the signal spectrum calculation
I(k) = [X(k) + X ∗ (N − k)]/2, U (k)
method based on improved FFT is accomplished in MATLAB
= − j[X(k) − X ∗ (N − k)]/2. (15) R2020a. The computer processor used is Intel(R) Core(TM)
i5-8250U CPU with 8GB RAM. The algorithm program of SOH
Formula (13) indicates that an N-point FFT of a complex
evaluation framework is written in MATLAB GUI.
signal composed by the current and voltage can simultaneously
A 30 s general step signal with the amplitude of 0.2 A is
obtain the spectrum of two real signals of length N, which causes
input and the step begins at the fifth second. The excitation
that the computational burden further reduces to nearly half.
waveform is shown in Fig. 10. The amplitude spectrum of
the derivative of the terminal voltage increment is shown in
IV. EXPERIMENTAL VALIDATION AND DISCUSSION Fig. 11. Here the frequency is from 0.1 to 10 Hz because the
To test and verify the proposed fast EIS calculation and frequencies of the selected battery HFs are all in this frequency
SOH estimation method, an experimental bench is built, as range. Fig. 11 shows that in the entire frequency range, the
shown in Fig. 9. A series battery pack, including four cells, spectral amplitudes of the voltage impulse responses of the
is equipped, which are additionally selected and do not belong four cells used for testing all attenuate with the increase in
to the previous 25 battery samples. The measured capacities frequency, but the rate of amplitude decrease gradually slows
of the four cells used in the experiment are 2.815 Ah (97.05%), down. The minimum amplitude at 10 Hz of Cell 2 is about

Authorized licensed use limited to: VISVESVARAYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on February 10,2023 at 09:28:41 UTC from IEEE Xplore. Restrictions apply.
 7026 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 69, NO. 7, JULY 2022

2.529 Ah, respectively. Correspondingly, the estimated SOH

values are, respectively, 96.28%, 96.01%, 92.59%, and 87.22%.
Thus, the experimental results show that the maximum SOH
estimation error (Cell 4) and the minimum error (Cell 3) of the
tested cells are 1.12% and 0.55%, respectively, which means
the proposed battery SOH estimation method based on fast
impedance calculation has achieved relatively good results.
In addition, the timeliness and efficiency should also be
discussed. The total time consumption of the entire online SOH
estimation for the four-cell series-connected battery pack mainly
comes from three parts, including battery signal acquisition,
Fig. 12. The EIS calculation results of the 4 battery cells with the fast
calculation.
EIS calculation, and the computation of SOH estimation model.
Due the high collection speed of the voltage acquisition board,
the collection of battery signals and the excitation of the input
TABLE IV current are almost simultaneous, consuming nearly 31 s. The last
HFS OF THE FOUR BATTERIES
two parts are accomplished in MATLAB. Using the tic–toc (a
MATLAB command used to calculate program running time),
the running time of EIS calculation and estimation model com-
putation are 1.89 s and 1.39 s, respectively. The total consumed
time is smaller than 35 s. In a word, using the proposed SOH es-
timation framework, the SOH of four cells in a series-connected
battery pack is relatively accurately obtained with the maximum
error of 1.12% in less than 35 s.

V. CONCLUSION
An SOH evaluation method based on fast impedance calcula-
tion for series-connected batteries was proposed in this article.
This method consisted of the fast EIS acquisition with improved
FFT method and the EIS-feature-based SOH estimation model
with regularized ELM. Utilizing the conjugate property of the
ordinary FT and the decomposition relationship between the
real signals and complex signals, the improved FFT method
further reduced the computational burden of signal spectrum.
This method with optimized signal parameters was verified by
simulation on the battery fractional-order impedance model. The
Fig. 13. The SOH estimation results by the SOH estimation model relationship between battery EIS and SOC and that between EIS
utilizing the obtained HFs. and degraded capacity were studied through EIS experimental
test for 25 battery samples in various SOH. Based on these
relationships, six EIS features indicating the battery degradation
40 dB, indicating that the intensity of the voltage response signal degree were selected. The ELM was used to build the aging
of all the cells is relatively ideal. Therefore, the reliability of the capacity or SOH estimation model based on selected HFs, and
EIS calculation results can be further guaranteed. Meanwhile, the regularization mechanism was introduced to improve the
the capacity difference of the four cells is also shown on the model generalization ability. The leave-one-out cross-validation
amplitude–frequency characteristic curve. In the entire range of results showed that the training and test errors on all 25 samples
1 to 10 Hz and the area close to 0.1 Hz, the four curves display were both less than 2%. An experimental bench was established
obvious distinction. After fast calculation, the EIS results of four to validate the proposed framework. Total four series-connected
cells are displayed in Fig. 12. The figure records 21 calculated lithium-ion cells were tested. An appropriate step excitation
EIS values for each cell. The corresponding values of six HFs current amplitude was determined by the experiments. For step
can be obtained from the calculated EIS, as given in Table IV. current input, the test results for four cells were acquired in 35 s
It can be seen from any row in the table that the calculated in total. Both the capacity estimation error and SOH estimation
value of each HF increases as the battery capacity decreases. deviation were less than 1.5%. The experimental results revealed
This is in line with the strong negative correlation between the that this framework may provide the possibility to accurately
selected HFs and battery capacity obtained in Section III. The obtain the health state of the series battery pack in a short
results of the estimated SOH by the SOH estimation model time.
utilizing the obtained HFs are shown in Fig. 13. The online It should be noted that the rapid SOH assessment method
calculated capacities of the four batteries output by the trained proposed in this article was applicable to a constant temperature
regularized ELM-based model are 2.792, 2.784, 2.685, and of 25 °C. This makes the application of this method have certain

Authorized licensed use limited to: VISVESVARAYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on February 10,2023 at 09:28:41 UTC from IEEE Xplore. Restrictions apply.
 FU et al.: FAST IMPEDANCE CALCULATION-BASED BATTERY STATE-OF-HEALTH ESTIMATION METHOD 7027

limitations. In fact, temperature is also an important factor [21] R. Al Nazer, V. Cattin, P. Granjon, M. Montaru, and M. Ranieri, “Broad-
affecting the impedance spectrum of the battery. Therefore, the band identification of battery electrical impedance for HEVs,” IEEE Trans.
Veh. Technol., vol. 62, no. 7, pp. 2896–2905, Sep. 2013.
future research direction of this article is to try to establish a [22] L. De Sutter, Y. Firouz, J. De Hoog, N. Omar, and J. van Mierlo, “Battery
battery SOH evaluation model that is suitable for a variety of aging assessment and parametric study of lithium-ion batteries by means of
temperatures or a certain temperature range. a fractional differential model,” Electrochimica Acta, vol. 305, pp. 24–36,
2019.
[23] M. Galeotti, L. Cinà, C. Giammanco, S. Cordiner, and A. Di Carlo,
“Performance analysis and SOH (state of health) evaluation of lithium
polymer batteries through electrochemical impedance spectroscopy,” En-
ergy, vol. 89, pp. 678–686, 2015.
REFERENCES [24] J. Jiang, Z. Lin, Q. Ju, Z. Ma, C. Zheng, and Z. Wang, “Electrochemical
impedance spectra for lithium-ion battery ageing considering the rate of
[1] J. Tian, R. Xiong, and Q. Yu, “Fractional-order model-based incremental
discharge ability,” Energy Procedia, vol. 105, pp. 844–849, 2017.
capacity analysis for degradation state recognition of lithium-ion batter-
[25] C. de Beer, P. S. Barendse, and P. Pillay, “Fuel cell condition monitoring
ies,” IEEE Trans. Ind. Electron., vol. 66, no. 2, pp. 1576–1584, Feb. 2019.
using optimized broadband impedance spectroscopy,” IEEE Trans. Ind.
[2] B. Jiang, H. Dai, and X. Wei, “Incremental capacity analysis based
Electron., vol. 62, no. 8, pp. 5306–5316, Aug. 2015.
adaptive capacity estimation for lithium-ion battery considering charging
[26] J. Huang, Z. Li, B. Y. Liaw, and J. Zhang, “Graphical analysis of electro-
condition,” Appl. Energy, vol. 269, 2020, Art. no. 115074.
chemical impedance spectroscopy data in Bode and Nyquist representa-
[3] J. Xu, X. Mei, X. Wang, Y. Fu, Y. Zhao, and J. Wang, “A relative state
tions,” J. Power Sources, vol. 309, pp. 82–98, 2016.
of health estimation method based on wavelet analysis for lithium-ion
[27] S. Arora, W. Shen, and A. Kapoor, “Review of mechanical design and
battery cells,” IEEE Trans. Ind. Electron., vol. 68, no. 8, pp. 6973–6981,
strategic placement technique of a robust battery pack for electric vehicles,”
Aug. 2021.
Renewable Sustain. Energy Rev., vol. 60, pp. 1319–1331, 2016.
[4] R. Xiong, Y. Zhang, J. Wang, H. He, S. Peng, and M. Pecht, “Lithium-ion
[28] N. Katayama and S. Kogoshi, “Mixed-signal Fourier transform for elec-
battery health prognosis based on a real battery management system
trochemical impedance spectroscopy,” J. Fuel Cell Sci. Technol., vol. 10,
used in electric vehicles,” IEEE Trans. Veh. Technol., vol. 68, no. 5,
no. 1, 2013, Art. no. 011006.
pp. 4110–4121, May 2019.
[29] E. Gomez-Luna, D. Silva, G. Aponte, J. G. Pleite, and D. Hinestroza,
[5] J. Xu, X. Mei, H. Wang, and J. Wang, “A hybrid self-heating method for
“Obtaining the electrical impedance using wavelet transform from the
batteries used at low temperature,” IEEE Trans. Ind. Informat., vol. 17,
time response,” IEEE Trans. Power Del., vol. 28, no. 2, pp. 1242–1244,
no. 7, pp. 4714–4723, Jul. 2021.
Apr. 2013.
[6] H. Wang, T. Tao, J. Xu, X. Mei, X. Liu, and P. Gou, “Cooling capacity
[30] A. Debenjak, P. Boškoski, B. Musizza, J. Petrovčič, and Đ. Juričić, “Fast
of a novel modular liquid-cooled battery thermal management system for
measurement of proton exchange membrane fuel cell impedance based
cylindrical lithium ion batteries,” Appl. Thermal Eng., vol. 178, 2020,
on pseudo-random binary sequence perturbation signals and continuous
Art. no. 115591.
wavelet transform,” J. Power Sources, vol. 254, pp. 112–118, 2014.
[7] Y. Li, J. Xu, X. Mei, and J. Wang, “A unitized multiwinding transformer-
[31] J. Xu, C. C. Mi, B. Cao, and J. Cao, “A new method to estimate the state
based equalization method for series-connected battery strings,” IEEE
of charge of lithium-ion batteries based on the battery impedance model,”
Trans. Power Electron., vol. 34, no. 12, pp. 11981–11989, Dec. 2019.
J. Power Sources, vol. 233, pp. 277–284, 2013.
[8] R. Xiong, L. Li, and J. Tian, “Towards a smarter battery management
system: A critical review on battery state of health monitoring methods,”
J. Power Sources, vol. 405, pp. 18–29, 2018.
[9] R. Xiong, J. Tian, H. Mu, and C. Wang, “A systematic model-based degra- Yumeng Fu was born in Xianyang, Shaanxi,
dation behavior recognition and health monitoring method for lithium-ion China, in 1996. He received the B.S. degree
batteries,” Appl. Energy, vol. 207, pp. 372–383, 2017. in vehicle engineering and the M.S. degree
[10] X. Han, M. Ouyang, L. Lu, and J. Li, “A comparative study of commercial in mechanical engineering from Xi’an Jiaotong
lithium ion battery cycle life in electric vehicle: Capacity loss estimation,” University, Xi’an, China, in 2018 and 2021,
J. Power Sources, vol. 268, pp. 658–669, 2014. respectively.
[11] X. Li, C. Yuan, X. Li, and Z. Wang, “State of health estimation for His research interest includes state estima-
Li-ion battery using incremental capacity analysis and Gaussian process tion and life prediction of power lithium-ion bat-
regression,” Energy, vol. 190, 2020, Art. no. 116467. tery for electric vehicles.
[12] M. Berecibar, M. Garmendia, I. Gandiaga, J. Crego, and I. Villarreal,
“State of health estimation algorithm of lifePO4 battery packs based on
differential voltage curves for battery management system application,”
Energy, vol. 103, pp. 784–796, 2016. Jun Xu (Member, IEEE) received the B.S. and
[13] J. Li, C. Lyu, L. Wang, L. Zhang, and C. Li, “Remaining capacity estima- Ph.D. degrees in mechanical engineering from
tion of Li-ion batteries based on temperature sample entropy and particle Xi’an Jiaotong University, Xi’an, China, in 2009
filter,” J. Power Sources, vol. 268, pp. 895–903, 2014. and 2013, respectively.
[14] K. A. Severson et al., “Data-driven prediction of battery cycle life before From 2011 to 2013, he was a Joint Ph.D. Stu-
capacity degradation,” Nat. Energy, vol. 4, no. 5, pp. 383–391, 2019. dent with the U.S. Department of Energy GATE
[15] J. Qu, F. Liu, Y. Ma, and J. Fan, “A neural-network-based method for Center for Electric Drive Transportation, College
RUL prediction and SOH monitoring of lithium-ion battery,” IEEE Access, of Electrical and Computer Science, University
vol. 7, pp. 87178–87191, 2019. of Michigan, Dearborn, MI, USA. He is currently
[16] Z. Ma, R. Yang, and Z. Wang, “A novel data-model fusion state-of-health an Associate Professor with the School of Me-
estimation approach for lithium-ion batteries,” Appl. Energy, vol. 237, chanical Engineering, Xi’an Jiaotong University.
pp. 836–847, 2019. He has conducted extensive research and authored more than 60 peer-
[17] Y. Deng et al., “Feature parameter extraction and intelligent estimation of reviewed papers. He holds 20 patents. His research interests include
the state-of-health of lithium-ion batteries,” Energy, vol. 176, pp. 91–102, design, modeling, control of battery systems, electric vehicles, renew-
2019. able energy systems, and robots.
[18] K. Qian, B. Huang, A. Ran, Y.-B. He, B. Li, and F. Kang, “State-of- Dr. Xu was the recipient of the innovation award of China Industry-
health (SOH) evaluation on lithium-ion battery by simulating the voltage University-Research Cooperation Innovation Award in 2018, the first
relaxation curves,” Electrochimica Acta, vol. 303, pp. 183–191, 2019. prize of Shaanxi Province University Science and Technology Award
[19] Q. Yang, J. Xu, X. Li, D. Xu, and B. Cao, “State-of-health estimation in 2018, the second prize of Shaanxi Province Teaching Achievement
of lithium-ion battery based on fractional impedance model and interval Award in 2019, and the special award of Xi’an Jiaotong University Teach-
capacity,” Int. J. Elect. Power Energy Syst., vol. 119, 2020, Art. no. 105883. ing Achievement Award in 2019. He was the conference organizing
[20] Q. Yang, J. Xu, B. Cao, X. Li, and B. Wang, “State-of-health estimation of committee Co-Chair of the 2017 International Symposium on Electric
Lithium-ion battery based on interval capacity,” Energy Procedia, vol. 105, Vehicles, Stockholm, Sweden.
pp. 2342–2347, 2017.

Authorized licensed use limited to: VISVESVARAYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on February 10,2023 at 09:28:41 UTC from IEEE Xplore. Restrictions apply.
 7028 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 69, NO. 7, JULY 2022

Mingjie Shi received the B.S. degree in Xuesong Mei received the Ph.D. degree in me-
mechanical engineering in 2019 from Xi’an chanical engineering from Xi’an Jiaotong Uni-
Jiaotong University, Xi’an, China, where he is versity, Xi’an, China, in 1991.
currently working toward the M.S. degree in me- He is a Full Professor with the School of
chanical engineering. Mechanical Engineering and the Director of the
His research interests include battery mod- Shaanxi Key Laboratory of Intelligent Robots,
eling, state of charge, and state of health Xi’an Jiaotong University. His research interests
estimation. include intelligent manufacturing, robotics, and
precision laser processing.

Authorized licensed use limited to: VISVESVARAYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on February 10,2023 at 09:28:41 UTC from IEEE Xplore. Restrictions apply.