Received April 21, 2021, accepted May 7, 2021, date of publication May 11, 2021, date of current version

May 24, 2021.

Digital Object Identifier 10.1109/ACCESS.2021.3079332

A New Method of Insulation Detection on Electric

Vehicles Based on a Variable Forgetting Factor
Recursive Least Squares Algorithm
ZHENBIN CHEN 1 , WEIYA CUI 1, XIANGYU CUI 2, HUIMIN QIAO 3,

HAO LU 1 , AND NA QIU 1

1 Mechanical and Electrical Engineering College, Hainan University, Haikou 570228, China
2 Key Laboratory of Electrochemical Energy Storage and Energy Conversion of Hainan Province, College of Physics and Electronic Engineering, Hainan Normal
University, Haikou 571158, China
3 Haima Automobile Company Ltd., Zhengzhou 450000, China

Corresponding author: Xiangyu Cui (mymq0002@[Link])

This work was supported by the Finance Science and Technology Project of Hainan Province of China under Grant ZDYF2020032.

ABSTRACT The lithium-ion batteries of an electric vehicle belong to a high-voltage direct-current system.
The high-voltage insulation performance of electric vehicles is very important for their safe operation.
To solve the problems of slow response and the poor estimation accuracy of the insulation resistance
under complex vehicle working conditions, a real-time insulation resistance detection method based on the
variable forgetting factor least squares algorithm is proposed in this paper. Based on the low-frequency
signal injection method and considering the influence of the Y capacitor, the corresponding circuit model
and the mathematical model of the reflected wave voltage are established, and the mathematical model is
linearized by a first-order Taylor expansion. By analyzing the influence of the forgetting factor on model
parameter identification and setting appropriate shutdown criteria, the least squares algorithm with a variable
forgetting factor is designed to quickly and accurately estimate the insulation resistance and Y capacitance.
The experimental test results show that the proposed method can quickly track the changes in the insulation
resistance and Y capacitance under the condition of noise interference and that the root mean square error
of the estimation resistor is within 0.012.

INDEX TERMS Electric vehicle (EV), embedded micro-control unit, insulation detection, lithium-ion
batteries, variable forgetting factor recursive least squares (VFFRLS).

NOMENCLATURE SYMBOLS
ABBREVIATIONS a1 Steady-state component of response
DC direct current a2 Response gain
EKF extended Kalman filter a3 Response time constant
EV electric vehicle a1 (k − 1)
b Estimated value of a1 at time k − 1
HVDC high-voltage direct-current a2 (k − 1)
b Estimated value of a2 at time k − 1
MCU micro-control unit a3 (k − 1)
b Estimated value of a3 at time k − 1
PC personal computer A Parameters a1 , a2 , a3
RC resistor capacitor A(k − 1)
b Estimated values of parameters at time k − 1
RLS recursive least squares b Scaling factor
RMSE root mean square error Cn Negative Y capacitance
UKF unscented Kalman filter Cp Positive Y capacitance
VFFRLS variable forgetting factor recursive Cp //Cn Equivalent Y capacitance value
least squares Cp //Cn Equivalent Y capacitance measurement result
e(i) Difference between the observed value and
The associate editor coordinating the review of this manuscript and the estimated value at time i
approving it for publication was Wei Xu . Ea1 (k) Mean value of a1 in the sliding window

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see [Link]
73590 VOLUME 9, 2021
Z. Chen et al.: New Method of Insulation Detection on Electric Vehicles Based on a VFFRLS Algorithm

Ea3 (k) Mean value of a3 in the sliding window battery belongs to high-voltage direct current (DC) systems,
H Observation matrix the high-voltage insulation performance of automobiles is
I1 Current flowing through R1 of great significance for the safe driving of electric vehi-
I2 Current flowing through R2 cles [10]–[12]. The operating conditions and operating envi-
I3 Current flowing through Rp ronment of electric vehicles, including high temperature,
I4 Current flowing through Rn high humidity, and high salt spray operating environments,
If Current flowing through Rf are complex [13]–[18]. High-voltage electrical components
If + Current flowing through Rf in the exhibit high-voltage insulation changes or even high-voltage
positive half period insulation failure. The China national standard file GB
J Cost function 18384-2020 [19] and the International Organization for Stan-
K Gain of RLS dardization standard file ISO 6469-1: 2019 [20] stipulated
M Maximum voltage value of the reflected wave that the insulation resistance of the high voltage systems of
P Error covariance matrix electric vehicles must not be less than 100 V/ohm. Therefore,
Q Root mean square noise value the real-time monitoring of electric vehicle high-voltage insu-
R1 Positive current limiting resistor lation performance changes and responding to the changes in
R2 Negative current limiting resistor the insulation resistance, represent important guarantees for
Rf Sampling resistor the safe operation of electric vehicles. Because it is difficult
Rn Negative insulation resistance to directly measure the high-voltage insulation of electric
Rn Negative insulation resistance vehicles online, researchers have invented a variety of meth-
measurement result ods to detect the high-voltage insulation resistance [21]–[24],
Rp Positive insulation resistance but most of them face problems such as a slow response
Rp Positive insulation resistance time, insufficient estimation accuracy, and poor reliability.
measurement result Therefore, it is necessary to study a new insulation resistance
U Power battery voltage detection method with a rapid response, high estimation accu-
Uf Reflected wave voltage racy and high reliability.
Uf + Reflected wave voltage in the positive
half period A. REVIEW OF INSULATION RESISTANCE
Uf − Reflected wave voltage in the negative DETECTION APPROACHES
half period In the laboratory, a high-precision voltmeter is usually used
Un0 Initial voltage value of Cn to measure the voltage of the high-voltage direct-current
Up0 Initial voltage value of Cp (HVDC) system supplied to the electrical chassis in order to
Us Voltage value generated by the signal generator calculate the insulation resistance of the positive and negative
Us+ Voltage value generated by the signal generator poles of the HVDC system with respect to the automobile
in the positive half period electrical chassis [19], [20]. This method requires a stopped
Us− Voltage value generated by the signal generator electric vehicle for static testing, so it does not employ actual
in the negative half period driving conditions to measure the insulation resistance of
V One-dimensional random observation noise the HVDC system; therefore, the value of the promotion
Y Constant error term of the electric vehicle is not utilized. The balanced bridge
β Parameter for calculating the forgetting factor method [25] has been the most commonly used high-voltage
δ Relative error insulation resistance measurement method for electric vehi-
1t Sampling interval cles due to its advantages of a simple measurement circuit
λ Forgetting factor and low cost. This method can accurately measure the insu-
λ0 Parameter for calculating the forgetting factor lation resistance of an HVDC system when the different
λ1 Parameter for calculating the forgetting factor insulation resistance value exists between the positive and
σa1 (k) Standard deviation of a1 in the sliding window negative poles of the HVDC system and the automobile elec-
σa3 (k) Standard deviation of a3 in the sliding window trical chassis. In the case of same insulation resistances, this
method has an obvious measurement error, and in the case of
high temperature and high humidity, the measurement error
I. INTRODUCTION will be enhanced: the balanced bridge method has thus been
As automobile pollutant emissions become increasingly eliminated by mainstream electric vehicle manufacturers.
serious [1], electric vehicles are becoming increasingly The unbalanced bridge method can overcome the shortcom-
popular [2]–[6]. Lithium-ion batteries have high energy den- ings of the balanced bridge method but will greatly increase
sity and power density, and their high voltage can signifi- the cost of the measurement circuit. Additionally, this method
cantly improve the energy utilization efficiency. Therefore, cannot respond to the rapid change in the insulation resistance
most electric vehicles use high-voltage lithium-ion batteries when the system has Y capacitance, so it is gradually replaced
as power batteries [7]–[9]. As the high-voltage lithium-ion by other new methods in current electric vehicles.

VOLUME 9, 2021 73591

 Z. Chen et al.: New Method of Insulation Detection on Electric Vehicles Based on a VFFRLS Algorithm

In view of the influence of the Y capacitance on the mea- variable forgetting factor recursive least squares (VFFRLS)
surement system, Mathsyaraja Aravind [26] proposed two algorithm to estimate the insulation resistance of an HVDC
new insulation resistance measurement methods. Method 1 system. Specifically, based on the low-frequency signal injec-
designs the Y capacitor discharge circuit and uses the resistor tion method, the influence of system Y capacitance is consid-
capacitor (RC) characteristics of the HVDC system to esti- ered, the corresponding circuit model and the mathematical
mate the insulation resistance of the system. This method model of the reflected wave voltage are established, and the
can effectively measure the insulation resistance of the high mathematical model is linearized by using a first-order Taylor
voltage system when the Y capacitor is large enough, but expansion. By analyzing the influences of the forgetting fac-
when the Y capacitor is small, the discharge current is small, tor on model parameter identification and setting appropriate
and the current measurement error will significantly affect shutdown criteria, the least squares algorithm with a vari-
the measurement result. Method 2 injects a high voltage able forgetting factor is designed to quickly and accurately
into the HVDC system and calculates the insulation resis- estimate the insulation resistance and Y capacitance values.
tance of the HVDC system by observing the charging process This method has the advantages of a fast response speed (the
of the Y capacitor. This method can quickly and accurately average response time is 3 s), high robustness, high estimation
calculate the insulation resistance of the HVDC system but accuracy (the root mean square error (RMSE) is less than
may cause damage to the HVDC systems of electric vehi- 0.012) and strong anti-interference ability, which is suitable
cles. Moreover, there is a risk that the measurement error for promotion in electric vehicles.
will increase or even reach failure under the high frequency
interference of the high voltage system. Due to its safety, C. ORGANIZATION OF THIS PAPER
reliability and low cost, the low-voltage and low-frequency The sections organized in the remainder of the paper are as
voltage injection method to measure insulation resistance follows: In Section 2, the calculation model of insulation
has gradually become the mainstream test method for the resistance and the reflected wave voltage model affected by
insulation resistance of the high voltage systems of electric the Y capacitor are described. The VFFRLS algorithm for
vehicles [27]. The method injects a ± 34 V square wave insulation resistance estimation is proposed in Section 3.
with a frequency of 0.1 Hz into the high voltage system Simulations and experimental configurations are presented in
of electric vehicles, returns to the electric chassis after the Section 4. Three experiments were performed on the simula-
voltage divider, and collects feedback from the electrical tion and test bench to verify the proposed method in Section 5.
chassis. In addition, the Kalman filter algorithm [28] has been The main conclusions are drawn in Section 6.
applied to filter noise, and the recursive least squares (RLS)
algorithm [29] has been applied to solve the problem of the II. MODEL DERIVATION BASED ON LOW-FREQUENCY
fluctuation of the insulation resistance detection results of a SIGNAL INJECTION METHOD
high voltage system caused by the unstable voltage of a power A. INSULATION RESISTANCE CALCULATION MODEL
battery under complex working conditions. This method In this paper, when the circuit model for measuring insulation
needs to calculate both the Kalman filter algorithm and RLS resistance by the low-frequency signal injection method is
algorithm at the same time, which leads to an increase in established, the power battery voltage is taken into account,
computational complexity and the failure to run smoothly on and the calculation formula of the insulation resistance under
a low-cost on-board microprocessor, which limits the applica- the joint action of the pulse voltage and high voltage system is
tion of this method in electric vehicles. Chuanxue Song [30] derived. Fig. 1 shows the insulation detection circuit model,
proposed an equivalent circuit model for HVDC insulation in which the power battery voltage is U , the pulse signal
resistance measurements based on the method of low voltage generator is Us , the equivalent insulation resistances of the
and low-frequency voltage injection and used the extended positive and negative electrodes are Rp and Rn , the current
Kalman filter (EKF) algorithm to estimate the insulation limiting resistors of the detection equipment are R1 , R2 , and
resistance of an HVDC system. This method can accurately R1 = R2 = R, and the voltage at both ends of the sampling
estimate the insulation resistance of the HVDC system, but resistor Rf is Uf . The voltage values generated by the pulse
due to the large rounding error in the nonlinear process of the signal generator in the positive and negative half periods are
extended Kalman algorithm [31], the estimation of insulation set as Us+ and Us− , respectively.
resistance cannot be further improved. At the same time, In the positive half period, the pulse signal generator gener-
in the case of strong noise interference, this method poses the ates a step voltage with value Us+ , and the currents flowing on
problem of algorithm divergence [32]. R1 , R2 , Rp , Rn , and Rf are I1 , I2 , I3 , I4 , and If + , respectively.
U and Rp , Rn form loop II, U and R1 , R2 form loop I, Us and
B. CONTRIBUTIONS OF THIS STUDY Rp , R1 , Rf form loop III. It is assumed that the power battery
To overcome the problems of poor accuracy, slow response, voltage remains constant within one pulse period. According
low robustness and weak anti-interference ability of the above to Kirchhoff’s law:
methods, based on the low-voltage low-frequency injection
method, this paper analyzes the reflected wave, establishes If + = I1 + I2 (1)
a new insulation detection circuit model, and proposes the If + = I3 + I4 (2)

73592 VOLUME 9, 2021

Z. Chen et al.: New Method of Insulation Detection on Electric Vehicles Based on a VFFRLS Algorithm

FIGURE 2. Insulation detection circuit model with the Y capacitor.

FIGURE 1. Insulation detection circuit model based on the low-frequency
signal injection method.

capacitance is constructed based on the low-frequency signal

U = Rn I4 − Rp I3 (3) injection method, as shown in Fig. 2, and the step response of
U = R1 I1 − R2 I2 (4) the circuit model with respect to the pulse injection signal is
analyzed.
Us+ = Rp I3 + R1 I1 + Rf If + (5)
In Fig. 2, U is the power battery voltage, Us is the pulse
According to (1)-(5), the voltage Uf + across the sampling signal generator, Rp and Rn are the power battery positive
resistor Rf in the positive half period can be obtained as: and negative insulation resistors, respectively, Cp and Cn are
positive and negative Y capacitors, respectively, R1 and R2
U (Rp − Rn ) + 2Us+ (Rp + Rn )
Uf + = Rf (6) are current limiting resistors, R1 = R2 = R, and Rf is the
(R + 2Rf )(Rp + Rn ) + 2Rp Rn sampling resistance. When the pulse generator generates a
Similarly, in the negative half period, the voltage Uf − at step voltage with an amplitude of Us , the currents flowing
both ends of the sampling resistor Rf is: over R1 , R2 , and Rf are set to I1 , I2 , and If , respectively. The
sum of the current flowing to Rp and Cp is I3 , the sum of
U (Rp − Rn ) + 2Us− (Rp + Rn )
Uf − = Rf (7) the current flowing to Rn and Cn is I4 , and the initial voltage
(R + 2Rf )(Rp + Rn ) + 2Rp Rn values of capacitors Cp and Cn are Up0 and Un0 , respectively.
When the insulation resistances Rp and Rn remain U and Rp , Rn form loop I, U and R1 , R2 form loop II, and Us
unchanged in a pulse period, according to (6) and (7), respec- and Rp , R1 , Rf form loop III. According to Kirchhoff’s Law,
tively, the insulation resistances Rp and Rn of the two poles the formula for s-domain is as follows:
of the power battery relative to the vehicle chassis can be
If (s) = I1 (s) + I2 (s) (10)
obtained as follows:
If (s) = I3 (s) + I4 (s) (11)
(R + 2Rf )(Uf + U − Uf − U ) + 2Rf U (Us− − Us+ )
Rp = Rn (I4 (s) + Cn Un0 ) Rp (I3 (s) + Cp Up0 )
−2Uf + Us− + 2Uf − Us+ + U (Uf − − Uf + ) U (s) = −
1 + sCn Rn 1 + sCp Rp
(8)
(R + 2Rf )(Uf + U − Uf − U ) + 2Rf U (Us− − Us+ ) (12)
Rn = U (s) = R1 I1 (s) − R2 I2 (s) (13)
2Uf + Us− − 2Uf − Us+ + U (Uf − − Uf + )
(9) Rp (I3 (s) + Cp Up0 )
Us (s) = + R1 I1 (s) + Rf If (s) (14)
1 + sCp Rp
The voltages Us+ and Us− generated by the pulse signal
generator in the positive and negative cycles are set val- Sorting out (10)-(14), the reflected wave voltage Uf across
ues. The power battery voltage U can be measured by the the sampling resistor Rf can be obtained, and the inverse
high-voltage measuring circuit, and the voltages Uf + and Laplace transform can be performed to obtain:
Uf − on the sampling resistor can be measured by the sam- Uf (t) = a1 + a2 exp(−t/a3 ) (15)
pling circuit in the positive and negative cycles. Therefore,
as long as the corresponding Uf + , Uf − , and U are measured Equation (15) is a function of the voltage of the sampling
in a signal period, the corresponding insulation resistances resistance varying with time in a half period, where t is a time
Rp and Rn in the period can be calculated by (8) and (9), variable and the function parameter expression is:
respectively. Rf (U (Rp − Rn ) + 2Us (Rp + Rn ))
a1 = (16)
(Rp + Rn )(R + 2Rf ) + 2Rp Rn
B. REFLECTED WAVE VOLTAGE MODEL AFFECTED
(Cp + Cn )Rp Rn (R + 2Rf )
BY Y CAPACITOR a3 = (17)
(Rp + Rn )(R + 2Rf ) + 2Rp Rn
Considering the influence of the Y capacitance between the
positive and negative poles of the power battery and the where a1 is the steady-state component of the response, which
chassis of the vehicle on the measurement of the insulation is the stationary value of the reflected wave voltage. a2 is a
resistance, an insulation detection circuit model with the Y response gain, which is independent of the calculation of the

VOLUME 9, 2021 73593

 Z. Chen et al.: New Method of Insulation Detection on Electric Vehicles Based on a VFFRLS Algorithm

resistance and capacitance values and is not considered in this

paper. a3 is the response time constant of the equivalent Y
capacitance.
The equivalent Y capacitance value can be obtained
by (17), that is, the parallel value of capacitance Cp and Cn :
(Rp + Rn )(R + 2Rf ) + 2Rp Rn
Cp //Cn = a3 (18)
Rp Rn (R + 2Rf )
where Cp //Cn represents the equivalent Y capacitance value.
In summary, the insulation resistances Rp and Rn can be
calculated by the identified a1 value and (8) and (9), and
the equivalent Y capacitance value can be calculated by the
identified a3 value and (18).

III. PARAMETER IDENTIFICATION METHOD BASED ON FIGURE 3. Positive half-period reflected wave voltage when Rp = Rn =
THE VFFRLS ALGORITHM 2000 k and Cp = Cn = 0.2 µF.

A. FORGETTING FACTOR RLS ALGORITHM

It can be found from (15) that the reflected wave voltage Uf
where Â(k − 1) represents the parameter estimation value at
is a nonlinear model under t. Considering the computational
time k − 1. The above formula can be written as a regression
complexity and estimation accuracy, the first-order Taylor
equation of least squares, and its expression is:
series expansion is used to linearize it.
According to (15), the discrete time function of Uf can be Uf (k) = H (k)A(k) + Y (k) + V (k) (22)
obtained, and its expression is as follows: where
k1t
Uf (k) = a1 (k) + a2 (k) exp(−k1t/a3 (k)) −
− â k1t â2 (k − 1)k1te â3 (k−1)
H (k) = [1 e 3 (k−1) ] (23)
= F(A(k), k) (19) â23 (k − 1)
k1t
where 1t is the sampling interval. â3 (k − 1)â2 (k − 1)k1te
− â
3 (k−1)
Y (k) = − (24)
A(k) = [a1 (k) a2 (k) a3 (k)]> (20) â23 (k − 1)
Equation (22) can be regarded as a model with constant
The linearization method of (19) is as follows. Expand
error term Y (k) in the observation equation. H (k) is the obser-
Uf (k) around the estimated value Â(k − 1) at time k − 1 into a
vation matrix, V (k) is one-dimensional random observation
Taylor series, omitting the quadratic and above terms, so that
noise, which is a zero mean and independent white Gaussian
the equation can be reduced to:
noise sequence.
Uf (k) ≈ F(Â(k − 1), k) The forgetting factor RLS algorithm is used to identify
∂F the model parameters. The implementation process of the
+ (A(k) − Â(k − 1)) (21) algorithm is shown in Table 1.
∂A> A(k)=Â(k−1)

TABLE 1. The implementation process of the forgetting factor RLS algorithm.

73594 VOLUME 9, 2021

Z. Chen et al.: New Method of Insulation Detection on Electric Vehicles Based on a VFFRLS Algorithm

FIGURE 4. The identification process when the forgetting factor is 0.99: (a) the identification process of a1 and (b) the identification process of
a3 .

FIGURE 5. The identification process when the forgetting factor is 1: (a) the identification process of a1 and (b) the identification process of a3 .

B. VARIABLE FORGETTING FACTOR and the estimation error of the parameter at convergence is
In the RLS algorithm, the forgetting factor λ is mainly used to also larger [33], [34].
increase the weight of the current data. The idea is to add the To obtain the influence of different λ values on the con-
forgetting factor to the old data and attenuate the influence vergence speed and parameter estimation, set the insula-
of the old data on identification in the form of exponential tion resistance Rp = Rn = 2000 k, the Y capacitance
weighting. The cost function of the forgetting factor RLS Cp = Cn = 0.2 µF, the power battery voltage U =
algorithm is 300 V, and the pulse injection voltage amplitude Us =
X 40 V. The positive half-period reflected wave voltage data
J (n) = λn−i e2 (i) (29) shown in Fig. 3 are obtained through simulation, and the
data noise variance is 0.001. Due to the charging process
where i = 1, 2 . . . , n, 0 < λ ≤ 1, e(i) represents the difference of the Y capacitor, the reflected wave voltage decreases
between the observed value and the estimated value at time i. exponentially, and as time increases, the voltage gradu-
By analyzing the recursive formula in Table 1, it is found ally stabilizes, which conforms to the description of (15).
that the larger the λ value is, the smaller the gain K (k), By identifying with the forgetting factor RLS algorithm,
which reduces the correction effect of the observation data the identifying processes with the forgetting factor equal to
on the parameter estimation and thus leads to the slower 0.99 and 1 are obtained, as shown in Figs. 4 and 5, respec-
convergence speed of the algorithm. However, it is not sen- tively. Figs. 4(a) and 5(a) show the identification process of
sitive to noise. The estimation error of the parameter is also parameter a1 , and Figs. 4(b) and 5(b) show the identification
smaller. In the same way, if the value of λ is smaller, then the process of parameter a3 . Through comparison, it can be found
convergence speed is faster, but it is more sensitive to noise, that when λ = 0.99, the identification process converges

VOLUME 9, 2021 73595

 Z. Chen et al.: New Method of Insulation Detection on Electric Vehicles Based on a VFFRLS Algorithm

FIGURE 6. The identification process when the variable forgetting factor: (a) the identification process of a1 and (b) the identification process
of a3 .

faster, starting to converge in approximately 0.6 s, but the con- that the variable forgetting factor RLS algorithm is better than
vergence accuracy is relatively low, and the parameters a1 and the fixed forgetting factor in terms of the convergence speed
a3 fluctuate greatly. When λ = 1, the convergence accuracy and convergence accuracy.
is relatively high, and the fluctuations of parameters a1 and a3
are also relatively small, but the convergence speed is slower
C. ALGORITHM IMPLEMENTATION
and starts to converge in approximately 1.0 s. In addition,
in Fig. 4(b), since λ < 1 and the reflected wave voltage In the process of using the VFFRLS algorithm to identify
gradually stabilized after 1.0 s, the identification error of a3 model parameters, the stopping condition of the algorithm
begins to increase after 1.0 s. Therefore, when designing the is one of the key factors affecting the effectiveness of the
forgetting factor RLS algorithm, the selection of the λ value is algorithm. If it stops prematurely, then the algorithm does
very important, and it should be considered as a compromise not converge at this time, and the error of the identification
between the convergence speed and stability of the algorithm. result is large. If the shutdown occurs too late, then the mea-
In view of the characteristics of the system reflected wave surement period will increase, which reduces the response
voltage signal, it is hoped that when the injected square wave speed of the measurement system, wastes substantial com-
signal changes between positive and negative values, λ takes puting resources, and may also increase the number of errors.
a smaller value so that the system can quickly track the signal As shown in Fig. 4(b), in the process of identifying a3 when
trend and improve the dynamic response speed. When the λ = 0.99, the reflected wave voltage tends to be stable in
algorithm starts to converge, a larger λ value is used to obtain the later stage, causing the error to increase again. Therefore,
a smaller steady-state error of the system and improve the it is necessary to determine whether the algorithm converges
accuracy of convergence. Since the fixed forgetting factor on the basis of weighing the identification accuracy and the
cannot take into account the requirements of accuracy and measurement period to establish an appropriate shutdown
convergence speed [35], this paper proposes a variable forget- criterion. The specific process of judging whether the algo-
ting factor method to achieve both accuracy and convergence rithm has converged is shown in Fig. 7. When the standard
speed. The variable forgetting factor is defined as follows: deviation of a1 and a3 in the sliding window is less than ε at
the same time, the algorithm is considered to have converged
λ(k) = λ1 − (λ1 − λ0 ) exp(−βk1t) (30) and the iteration is exited. Using the form of the standard
deviation within the sliding window can effectively determine
where λ0 , λ1 , and β are the parameters for calculating the whether the algorithm has converged, control the stopping of
forgetting factor. The parameter values used in this article are the algorithm, and avoid misjudgment of the convergence.
λ0 = 0.96, λ1 = 1, and β = 40. It can be determined from Fig. 8 is the implementation flow chart of the insulation
(30) that after the injected signal changes between positive resistance measurement method proposed in this paper. First,
and negative values, λ will gradually increase from a smaller the initial value and parameters of the algorithm are set. Then,
value of 0.96 to 1 until the algorithm reaches the set conver- the VFFRLS algorithm is used to identify the model parame-
gence accuracy. By simulating the VFFRLS algorithm, the ters. When the convergence accuracy is reached, the average
results shown in Fig. 6 are obtained. It can be found from values of a1 and a3 in the sliding window are used as the result
the figure that the algorithm starts to converge in 0.4 s, the of parameter identification. Finally, the insulation resistance
convergence accuracy is relatively high, and the fluctuations value is calculated by (8) and (9), and the equivalent Y
of parameters a1 and a3 are relatively small. The results show capacitance value is calculated by (18).

73596 VOLUME 9, 2021

Z. Chen et al.: New Method of Insulation Detection on Electric Vehicles Based on a VFFRLS Algorithm

TABLE 2. The parameters of the insulation detector system.

rapidity and accuracy of the identification algorithm conver-

gence. Therefore, in the experiment, the insulation resistance
value and the Y capacitance value were changed artificially
to cause a sudden change.

A. SIMULATION CONFIGURATION
The simulation model of the system is established using
MATLAB software, as shown in Fig. 9. Fig. 9(a) depicts a
circuit simulation model that generates the required data by
simulating the high-voltage circuit and insulation detection
circuit of electric vehicles. Rp and Rn are the insulation
FIGURE 7. The flow chart of judging whether the algorithm has resistances, Cp and Cn are the Y capacitors, R1 and R2 are
converged. the current limiting resistors, Rf is a sampling resistor, Us
is an injection voltage signal, and the power battery voltage
is 300 V. Fig. 9(b) shows the VFFRLS model. The input of
The calculation formulas for the mean values Ea1 (k) and the model is the data obtained by the simulation in Fig. 9(a),
Ea3 (k) and the standard deviations σa1 (k) and σa3 (k) are as and the output is the identified a1 and a3 . The model mainly
follows: consists of a time module, an observation matrix module, a
1
k
X forgetting factor module, a gain module, an estimated value
Ea1 (k) = ab1 (i) (31) update module, an error variance update module, and a judg-
w
i=k−w+1 ment convergence module. In the simulation experiment, the
1 Xk fixed-step calculation mode is adopted, and the simulation
Ea3 (k) = ab3 (i) (32) step is set to 0.001 s according to a sampling frequency
w
i=k−w+1 of 1 kHz.
v
u
u1 X k
σa1 (k) = t a1 (i) − Ea1 (k))2
(b (33) B. EXPERIMENTAL CONFIGURATION
w
i=k−w+1 The structure of the test bench is shown in Fig. 10. The test
v
u k bench mainly consists of a high-voltage DC power supply,
u1 X
σa3 (k) = t a3 (i) − Ea3 (k))2
(b (34) an insulation detector, a low-voltage DC power supply, two
w DC resistance boxes, two decimal adjustable capacitor boxes,
i=k−w+1
a CAN monitor, a downloader and a personal computer (PC).
IV. SIMULATION AND EXPERIMENTAL CONFIGURATION The high-voltage DC power system is used to simulate the
To verify the electric vehicle insulation detection method power battery voltage and is set to 300 V. The low-voltage DC
based on the VFFRLS algorithm proposed in this paper, sim- power system supplies power to the insulation detector, and
ulation experiments and bench experiments were carried out the power supply voltage of the insulation detector is 12 V.
under different insulation resistances and Y capacitances. The The two DC resistance boxes simulate the insulation resis-
parameters of the insulation detector system in the process of tance of the positive and negative electrodes. The two capac-
the simulation experiment and bench experiment are shown itor boxes simulate positive and negative Y capacitors. The
in Table 2. In the actual use of electric vehicles, a long PC downloads the software project file to the micro-control
transition time is required for the insulation resistance and Y unit (MCU) of the insulation detector through the downloader
capacitance values to change. However, considering that the and collects experimental data such as the power battery volt-
model parameters change slowly, it is difficult to verify the age, injection voltage signal, reflected wave voltage signal,

VOLUME 9, 2021 73597

 Z. Chen et al.: New Method of Insulation Detection on Electric Vehicles Based on a VFFRLS Algorithm

FIGURE 8. Flowchart of the insulation resistance estimation based on VFFRLS.

and insulation resistance measurement value through a CAN TABLE 3. Insulation resistance and Y capacitance value of Four cycles.
monitor.

V. RESULTS AND DISCUSSION

A test bench for the simulation with the configuration shown
in Fig. 9 was built using MATLAB software. Three tests
were executed to test the performance of the VFFRLS algo-
rithm using the test bench for the simulation. The system
cost should be lower for the electrical vehicle, so the hard-
ware of insulation detection was designed for the lower cost.
Freescale’s KEA128 was used as an embedded MCU to test resistance and Y capacitance values of the four cycles are
the performance of the proposed method with respect to the shown in Table 3.
system cost. Hardware was designed with the selected MCU, Using the reflected wave voltage data without noise for
a test bench was built, and three tests were executed to test the the simulation, the measurement results shown in Table 4 are
performance of the proposed method with the configuration obtained, where Rp and Rn are the measurement results of
of Fig. 10. the positive and negative insulation resistances, respectively,
Cp //Cn is the measurement result of the equivalent Y capac-
A. EXPERIMENT 1: SIMULATIONS WITHOUT/WITH itance, and δ is the relative error. It can be found from the
MEASURING NOISE table that the relative error of measurement is within 1% and
The reflected wave voltage data, as shown in Fig. 11, are the measurement time T is within 0.3 s without noise. The
obtained through Simulink simulation. The injection signal insulation resistance and Y capacitance can be accurately and
cycle is 3 s, the simulation time is 12 s, and a total of 4 cycles quickly measured.
are simulated. The blue curve represents the reflected wave The simulation is carried out using reflected wave voltage
voltage data without noise. To simulate the actual reflected data with a noise variance of 0.01. Each set of data is simu-
wave voltage value, Gaussian white noise with a variance lated ten times, the average values are taken as the measure-
of 0.01 is superimposed on the signal sampling end to obtain ment results of the insulation resistance and the Y capacitance
the red curve in Fig. 11. The model parameters exhibit value, and the measurement results shown in Table 5 are
sudden change at the end of each cycle, and the insulation obtained. It can be determined from the table that in the

73598 VOLUME 9, 2021

Z. Chen et al.: New Method of Insulation Detection on Electric Vehicles Based on a VFFRLS Algorithm

FIGURE 9. Simulation model of the insulation detection system: (a) the insulation detection circuit model and (b) the VFFRLS model.

TABLE 4. Measurement results without noise.

presence of noise interference, the relative measurement error three algorithms adopt the same initial value and convergence
of insulation resistance is within 1.8%, and the relative mea- condition. When using noisy data for simulation, it was found
surement error of the Y capacitor is within 1.9%. VFFRLS that EKF and UKF diverged. Therefore, the data without
effectively reduces noise interference. noise in the second cycle is selected for simulation, and
Through simulation, compare EKF, unscented Kalman fil- the results are shown in Table 6. The results show that in
ter (UKF) and the algorithm proposed in this paper. The the absence of noise, the convergence speed of EKF and

VOLUME 9, 2021 73599

 Z. Chen et al.: New Method of Insulation Detection on Electric Vehicles Based on a VFFRLS Algorithm

FIGURE 10. Insulation resistance test bench.

FIGURE 11. Reflected wave voltage data.

TABLE 5. Measurement results with a noise variance of 0.01.

UKF is faster than the VFFRLS algorithm, and the mea- that although UKF is better than VFFRLS, when there is
surement accuracy of the VFFRLS is higher than EKF, noise, UKF and EKF will diverge, and VFFRLS can run
but slightly lower than UKF. The simulation results show smoothly.

73600 VOLUME 9, 2021

Z. Chen et al.: New Method of Insulation Detection on Electric Vehicles Based on a VFFRLS Algorithm

TABLE 6. Measurement results of different algorithms.

FIGURE 12. Reflected wave voltage data without measuring noise and with measuring noise of reflected wave: (a) without
measuring noise and (b) with 2.5% measuring noise.

B. EXPERIMENT 2: COMPUTING COMPLEXITY TEST TABLE 7. Computing complexity of different algorithms.

To verify the computing complexity of the proposed algo-

rithm, a contrast test with an EKF, an UKF and the proposed
algorithm is executed. The bus frequency is set to 48 MHz,
the programs are executed once every 3 ms, the period of the
reflected wave is 3 s, and the results are shown in Table 7.
It is shown that the computing complexity of the proposed
algorithm is the lowest among the three algorithms. The load
rate of the KEA128 embedded MCU is lower than 70% when requires a total of 300 s and is divided into four time periods.
running the program of the proposed algorithm. This result The parameter settings of the insulation resistance and Y
shows that the VFFRLS algorithm for insulation detection capacitor in different time periods are shown in Table 8. The
can run smoothly on the KEA128 embedded MCU. period of the wave is 3 s, so the half period of the reflected
wave is 1.5 s.
Electric vehicles have many components with high elec-
C. EXPERIMENT 3: TESTS WITHOUT AND WITH tromagnetic interference (including a relay, solenoid valve,
MEASURING NOISE motor control, motor, etc.) and the operating environment
To verify the performance of the VFFRLS algorithm in an and working conditions are complex. The signal sampling
actual demo with the test bench shown in Fig. 10, the test circuit is easily interfered with, and the insulation detection

VOLUME 9, 2021 73601

 Z. Chen et al.: New Method of Insulation Detection on Electric Vehicles Based on a VFFRLS Algorithm

FIGURE 13. Results of a1 and a3 without measuring the noise of the reflected wave: (a) estimated a1 and (b) estimated a3 .

TABLE 8. Working conditions of the tests. Fig. 18 displays the error distribution of the estimation results
of Rp , Rn , and Cp //Cn .
When the changes in the capacitance or resistance before
and after the switching point are small, the fluctuation of the
estimated results of a1 and a3 is minimal; in contrast, the fluc-
tuation of the estimated results of a1 and a3 is large. The
results of a1 can converge to the correct value in one reflected
wave period, while a3 can converge to the correct value in two
reflected wave periods. According to the method described
in this paper, the values of Rp , Rn , and Cp //Cn are further
algorithm must be useful, with strong anti-interference ability obtained from the estimated values of a1 and a3 . At the same
and good robustness. To test the anti-interference perfor- time, the calculation of Rp , Rn , and Cp //Cn values requires a
mance and robustness of the proposed algorithm, 2.5% sam- complete reflected wave period, so the calculated fluctuation
pling signal noise was applied to the reflected wave sampling period of Rp , Rn , and Cp //Cn values will be further extended,
circuit, and the noise was described as follows: and the maximum fluctuation time is approximately 8 s in
experiment 3. According to the error distribution diagram,
Q = b∗ M /3 (35)
with or without the reflected wave measurement error, except
where Q is the root mean square noise value, b represents for the error during initialization and near the switching point,
the scaling factor, and M is the maximum voltage value of the relative error of the Rp , Rn , and Cp //Cn parameters in the
the reflected wave. The measurement noise is loaded with whole process is less than 2%, and the RMSE is less than
2.5% of the actual signal. The measured signal of the reflected 0.012. In practical engineering, limit filtering can be carried
wave is shown in Fig. 12 without/with measuring noise. out at the switching point to further smooth the fluctuation
Fig. 13 shows the results of the estimated parameters of a1 error caused by large parameter changes during the switching
and a3 achieved by the VFFRLS algorithm without measur- to further improve the estimation accuracy of the whole pro-
ing noise, and the results of the two parameters are shown cess. Considering that drastic changes in the capacitance and
in Fig. 14 with measuring noise. Figs. 15-17 are the results of resistance at 0.003 s are rare under actual work conditions,
Rp , Rn , and Cp //Cn without and with measuring noise, and the maximum fluctuation of the estimated value is approx-

73602 VOLUME 9, 2021

Z. Chen et al.: New Method of Insulation Detection on Electric Vehicles Based on a VFFRLS Algorithm

FIGURE 14. Results of a1 and a3 with 2.5% measuring noise of reflected wave: (a) estimated a1 and (b) estimated a3 .

FIGURE 15. Results of Rp without and with 2.5% measuring noise of reflected wave: (a) Rp without measuring noise
and (b) Rp with 2.5% measuring noise.

imately 3 s (that is, convergence can be completed in one Therefore, in the cases of different Y capacitors and insu-
reflected wave cycle) when the change in working conditions lation resistors, the detection accuracy of the insulation resis-
is small. tance is still very high (the RMSE is within 0.012) after

VOLUME 9, 2021 73603

 Z. Chen et al.: New Method of Insulation Detection on Electric Vehicles Based on a VFFRLS Algorithm

FIGURE 16. Results of Rn without and with 2.5% measuring noise of reflected wave: (a) Rn without measuring noise and
(b) Rn with 2.5% measuring noise.

FIGURE 17. Results of Cp //Cn without and with 2.5% measuring noise of reflected wave: (a) Cp //Cn without
measuring noise and (b) Cp //Cn with 2.5% measuring noise.

loading the measurement noise with 2.5%, the response speed maintained after running continuously. The results show that
is very short (the average response time is within 3 s), and the proposed method exhibits high anti-interference ability
the convergence property of the continuous algorithm is still and robustness.

73604 VOLUME 9, 2021

Z. Chen et al.: New Method of Insulation Detection on Electric Vehicles Based on a VFFRLS Algorithm

FIGURE 18. Distribution of the relative error of the estimation results: (a) Rp relative error without measuring noise;
(b) Rn relative error without measuring noise; (c) Cp //Cn relative error without measuring noise; (d) Rp relative error
with measuring noise; (e) Rn relative error with measuring noise; and (f) Cp //Cn relative error with measuring noise.

VI. CONCLUSION ACKNOWLEDGMENT

Based on the circuit model of low-frequency signal injec- The authors would also like to thank Hainan Normal Univer-
tion to measure the insulation resistance, the principle of sity for the support.
insulation detection is described in detail, and the calcu-
lation formula of insulation resistance is derived. Consid- REFERENCES
ering the influence of the Y capacitance of the system, [1] U. Chong, S. H. L. Yim, S. R. H. Barrett, and A. M. Boies, ‘‘Air quality
and climate impacts of alternative bus technologies in Greater London,’’
a nonlinear reflected wave voltage model is established Environ. Sci. Technol., vol. 48, no. 8, pp. 4613–4622, Apr. 2014, doi: 10.
according to the step response of the injected signal. After 1021/es4055274.
linearizing it through a first-order Taylor expansion, the [2] J. Remmlinger, M. Buchholz, M. Meiler, P. Bernreuter, and K. Dietmayer,
‘‘State-of-health monitoring of lithium-ion batteries in electric vehi-
model parameters are identified by the variable forgetting cles by on-board internal resistance estimation,’’ J. Power Sources,
factor recursive least squares algorithm, and the insulation vol. 196, no. 12, pp. 5357–5363, Jun. 2011, doi: 10.1016/[Link].2010.
resistance value and Y capacitance value are calculated 08.035.
according to the model parameters. The experimental test [3] S. J. Gerssen-Gondelach and A. P. C. Faaij, ‘‘Performance of batteries for
electric vehicles on short and longer term,’’ J. Power Sources, vol. 212,
results show that the proposed method can quickly track pp. 111–129, Aug. 2012, doi: 10.1016/[Link].2012.03.085.
the changes in the insulation resistance and Y capacitance [4] L. Lu, X. Han, J. Li, J. Hua, and M. Ouyang, ‘‘A review on the key issues
under the condition of noise interference. The root mean for lithium-ion battery management in electric vehicles,’’ J. Power Sources,
vol. 226, pp. 272–288, Mar. 2013, doi: 10.1016/[Link].2012.10.060.
square error is within 0.012, and the average response [5] R. Xiong, F. Sun, X. Gong, and H. He, ‘‘Adaptive state of charge estimator
time is within 3 s. Simulation experiments and bench tests for lithium-ion cells series battery pack in electric vehicles,’’ J. Power
prove that this method exhibits the advantages of a fast Sources, vol. 242, pp. 699–713, Nov. 2013, doi: 10.1016/[Link].
2013.05.071.
response speed, high robustness, high estimation accuracy
[6] R. Adany, D. Aurbach, and S. Kraus, ‘‘Switching algorithms for extending
and strong anti-interference ability. At present, this method battery life in electric vehicles,’’ J. Power Sources, vol. 231, pp. 50–59,
has been tested on a bench in the laboratory. In future Jun. 2013, doi: 10.1016/[Link].2012.12.075.
work, the proposed method shall be applied to actual elec- [7] X. Wang, H. Hao, J. Liu, T. Huang, and A. Yu, ‘‘A novel method for
preparation of macroposous lithium nickel manganese oxygen as cathode
tric vehicles to further verify its performance in realistic material for lithium ion batteries,’’ Electrochimica Acta, vol. 56, no. 11,
environments. pp. 4065–4069, Apr. 2011, doi: 10.1016/[Link].2010.12.108.

VOLUME 9, 2021 73605

 Z. Chen et al.: New Method of Insulation Detection on Electric Vehicles Based on a VFFRLS Algorithm

[8] Z. Xie, S. Ellis, W. Xu, D. Dye, J. Zhao, and Y. Wang, ‘‘A novel [29] X. Zhang, Y. Wang, C. Liu, and Z. Chen, ‘‘A novel approach of remaining
preparation of core–shell electrode materials via evaporation-induced self- discharge energy prediction for large format lithium-ion battery pack,’’
assembly of nanoparticles for advanced Li-ion batteries,’’ (Cambridge, J. Power Sources, vol. 343, pp. 216–225, Mar. 2017, doi: 10.1016/j.
England), Chem. Commun., vol. 51, no. 81, pp. 15000–15003, 2015, jpowsour.2017.01.054.
doi: 10.1039/c5cc05577f. [30] C. Song, Y. Shao, S. Song, S. Peng, F. Zhou, C. Chang, and D. Wang,
[9] Y. K. Zeng, T. S. Zhao, X. L. Zhou, L. Zeng, and L. Wei, ‘‘The effects of ‘‘Insulation resistance monitoring algorithm for battery pack in electric
design parameters on the charge-discharge performance of iron-chromium vehicle based on extended Kalman filtering,’’ Energies, vol. 10, no. 5,
redox flow batteries,’’ Appl. Energy, vol. 182, pp. 204–209, Nov. 2016, p. 714, May 2017, doi: 10.3390/en10050714.
doi: 10.1016/[Link].2016.08.135. [31] X. Cui, Z. Jing, M. Luo, Y. Guo, and H. Qiao, ‘‘A new method for
[10] J. Jiang and H. Ji, ‘‘Study of insulation monitoring device for DC system state of charge estimation of lithium-ion batteries using square root cuba-
based on multi-switch combination,’’ in Proc. 2nd Int. Symp. Comput. ture Kalman filter,’’ Energies, vol. 11, no. 1, p. 209, Jan. 2018, doi: 10.
Intell. Design, Changsha, China, Dec. 2009, pp. 429–433. 3390/en11010209.
[11] C. Li, S. Luo, C. Cole, and M. Spiryagin, ‘‘An overview: Modern tech- [32] X. Cui, Z. He, E. Li, A. Cheng, M. Luo, and Y. Guo, ‘‘State-of-charge
niques for railway vehicle on-board health monitoring systems,’’ Vehi- estimation of power lithium-ion batteries based on an embedded micro
cle Syst. Dyn., vol. 55, no. 7, pp. 1045–1070, Jul. 2017, doi: 10.1080/ control unit using a square root cubature Kalman filter at various ambi-
00423114.2017.1296963. ent temperatures,’’ Int. J. Energy Res., vol. 43, no. 8, pp. 3561–3577,
[12] Y.-H. Chiang, W.-Y. Sean, C.-Y. Huang, and L.-H. Chiang Hsieh, ‘‘Adap- Apr. 2019, doi: 10.1002/er.4503.
tive control for estimating insulation resistance of high-voltage battery [33] M. Beza and M. Bongiorno, ‘‘Application of recursive least squares algo-
system in electric vehicles,’’ Environ. Prog. Sustain. Energy, vol. 36, no. 6, rithm with variable forgetting factor for frequency component estima-
pp. 1882–1887, May 2017, doi: 10.1002/ep.12644. tion in a generic input signal,’’ IEEE Trans. Ind. Appl., vol. 50, no. 2,
[13] H. Li, Z. Li, F. Lin, Y. Chen, D. Liu, and B. Wang, ‘‘Measurement pp. 1168–1176, Mar. 2014, doi: 10.1109/TIA.2013.2279195.
and analysis of insulation resistance of metallized polypropylene film [34] X. Wei, M. Yimin, and Z. Feng, ‘‘Identification of parameters in Li-ion
capacitor under high electric field,’’ in Proc. IEEE Int. Power Modula- battery model by least squares method with variable forgetting factor,’’
tor High Voltage Conf. (IPMHVC), San Diego, CA, USA, Jun. 2012, Int. J. Comput. Methods, vol. 17, no. 7, Sep. 2020, Art. no. 1950027,
pp. 617–619. doi: 10.1142/S0219876219500270.
[14] H. Torkaman and F. Karimi, ‘‘Measurement variations of insulation resis- [35] Q. Song, Y. Mi, and W. Lai, ‘‘A novel variable forgetting factor recursive
tance/polarization index during utilizing time in HV electrical machines— least square algorithm to improve the anti-interference ability of battery
A survey,’’ Measurement, vol. 59, pp. 21–29, Jan. 2015, doi: 10. model parameters identification,’’ IEEE Access, vol. 7, pp. 61548–61557,
1016/[Link].2014.09.034. 2019, doi: 10.1109/ACCESS.2019.2903625.
[15] E. David, R. Soltani, and L. Lamarre, ‘‘PDC measurements to assess
machine insulation,’’ IEEE Trans. Dielectr. Electr. Insul., vol. 17, no. 5,
pp. 1461–1469, Oct. 2010, doi: 10.1109/TDEI.2010.5595547.
[16] R. Rui and I. Cotton, ‘‘Impact of low pressure aerospace environment
on machine winding insulation,’’ in Proc. IEEE Int. Symp. Electr. Insul.,
San Diego, CA, USA, Jun. 2010, pp. 1–5.
[17] B. X. Du and Y. Yamano, ‘‘Effects of atmospheric pressure on DC
resistance to tracking of polymer insulating materials,’’ IEEE Trans. ZHENBIN CHEN was born in 1968. He received
Dielectr. Electr. Insul., vol. 12, no. 6, pp. 1162–1171, Dec. 2005, doi: 10. the B.E. degree in agricultural mechanization and
1109/TDEI.2005.1561796.
the M.E. degree in agricultural extension from
[18] G. Mitic and G. Lefranc, ‘‘Localization of electrical-insulation and partial-
Hainan University, Haikou, China, in 1990 and
discharge failures of IGBT modules,’’ IEEE Trans. Ind. Appl., vol. 38,
2007, respectively, and the Ph.D. degree in power
no. 1, pp. 175–180, Jan./Feb. 2002, doi: 10.1109/28.980373.
machinery and engineering from Tongji Univer-
[19] Electric Vehicles Safety Requirements, Standard GB 18384-2020,
May 2020. sity, Shanghai, in 2015.
[20] Electrically Propelled Road Vehicles-Safety Specifications-Part 3: Electri- From 1990 to 1996, he worked as an Engineer
cal Safety, Standard ISO 6469-3:2018, Oct. 2018. with the Hainan Agricultural Machinery Research
[21] W. Zhen-Jun and W. Li-Fang, ‘‘A novel insulation resistance monitoring Institute. Since 2004, he has been working as a
device for hybrid electric vehicle,’’ in Proc. IEEE Vehicle Power Propuls. Lecturer with Hainan University. Since 2015, he has been working as
Conf., Harbin, China, Sep. 2008, pp. 1–4. the Vice President of the Mechanical and Electrical Engineering College.
[22] W. Xuezhe, B. Lu, and S. Zechang, ‘‘A method of insulation failure In the past five years, he has published more than 20 academic articles
detection on electric vehicle based on FPGA,’’ in Proc. IEEE Vehicle Power and two monographs. His research interests include energy-saving and new
Propuls. Conf., Harbin, China, Sep. 2008, pp. 1–5. energy vehicles, advanced manufacturing and automation, and intelligent
[23] C. H. Piao and T. Cong, ‘‘Study on isolation monitoring of high-voltage connected vehicles. He received the Hainan Provincial Science and Technol-
battery system,’’ Appl. Mech. Mater., vols. 44–47, pp. 571–575, Dec. 2010, ogy Progress Award. He currently serves as a reviewer for several journals,
doi: 10.4028/[Link]/AMM.44-47.571. and a reviewer and organizer for national and international conferences.
[24] J. Li, Z. Wu, Y. Fan, Y. Wang, and J. Jiang, ‘‘Research on insu-
lation resistance on-line monitoring for electric vehicle,’’ in Proc.
Int. Conf. Electr. Mach. Syst., Nanjing, China, vol. 1, Sep. 2005,
pp. 814–817.
[25] J. S. Simons and M. T. Richards, ‘‘Non-destructive electrical test methods
for evaluating high-voltage stator insulation,’’ Proc. IEE A, Power Eng.,
vol. 109, no. 3, pp. 71–79, Feb. 2010, doi: 10.1049/pi-a.1962.0013.
[26] A. Mathsyaraja, ‘‘Ground fault detection for flexible high voltage power
WEIYA CUI was born in 1995. He received the
systems,’’ SAE Int. J. Commercial Vehicles, vol. 4, no. 1, pp. 185–197,
Sep. 2011, doi: 10.4271/2011-01-2252. B.E. degree from the Anyang Institute of Tech-
[27] J. Tian, Y. Wang, D. Yang, X. Zhang, and Z. Chen, ‘‘A real-time insu- nology, Anyang, China, in 2014. He is currently
lation detection method for battery packs used in electric vehicles,’’ pursuing the M.E. degree with the Mechanical and
J. Power Sources, vol. 385, pp. 1–9, May 2018, doi: 10.1016/[Link]. Electrical Engineering College, Hainan Univer-
2018.03.018. sity, Haikou, China. His research interests include
[28] Y. Wang, C. Zhang, and Z. Chen, ‘‘An adaptive remaining energy new energy electric vehicles, power battery insula-
prediction approach for lithium-ion batteries in electric vehicles,’’ tion testing, least squares algorithm, and Kalman
J. Power Sources, vol. 305, pp. 80–88, Feb. 2016, doi: 10.1016/[Link]. filter algorithm.
2015.11.087.

73606 VOLUME 9, 2021

Z. Chen et al.: New Method of Insulation Detection on Electric Vehicles Based on a VFFRLS Algorithm

XIANGYU CUI was born in 1980. He received HAO LU was born in 1993. He received the B.E.
the M.E. degree in vehicle engineering from the degree in vehicle engineering from Hainan Univer-
Wuhan University of Technology, Wuhan, China, sity, Haikou, China, in 2014, where he is currently
in 2011, and the Ph.D. degree in vehicle engi- pursuing the M.E. degree with the Mechanical
neering from Hunan University, Changsha, China, and Electrical Engineering College. His research
in 2019. He is currently a Lecturer with the interests include new energy electric vehicles and
Key Laboratory of Electrochemical Energy Stor- automotive electronics.
age and Energy Conversion of Hainan Province,
College of Physics and Electronic Engineering,
Hainan Normal University. He has published three
articles and holds nine patents. His research interests include power battery
modeling and simulation on electric vehicles, design, and control theory of
the hybrid power train. He received the first prize for Scientific or Technical
Awards from the Haima Automobile Group, in 2011.

HUIMIN QIAO was born in 1983. She received NA QIU received the Ph.D. degree in mechanical
the B.S. degree in automation and the M.S. degree engineering from Tongji University. She was a
in control theory and control engineering from Joint-Ph.D. Student of the University of Florida.
Jilin University, Changchun, China, in 2005 and She is currently a Lecturer with the Mechani-
2008, respectively. She is currently pursuing the cal and Electrical Engineering College, Hainan
Ph.D. degree in vehicle engineering with the University, Haikou, China. Her research interests
College of Mechanical and Vehicle Engineer- include data mining, artificial intelligence, addi-
ing, Hunan University, Changsha, China. Her tive manufacturing, crashworthiness optimization,
research on subject with Haima Motor Corpo- automotive structure design, optimization and
ration. Her research interests include automobile lightweight design.
software/architecture development innovative subject.

VOLUME 9, 2021 73607