Quality Engineering

ISSN: (Print) (Online) Journal homepage: www.tandfonline.com/journals/lqen20

A reliability prediction method considering

degradation self-acceleration effect in DC-link
electrolytic capacitor

Xuerong Ye, Qisen Sun, Ruishi Lin, Cen Chen, Min Xie, Guofu Zhai & Rui Kang

To cite this article: Xuerong Ye, Qisen Sun, Ruishi Lin, Cen Chen, Min Xie, Guofu
Zhai & Rui Kang (2024) A reliability prediction method considering degradation self-
acceleration effect in DC-link electrolytic capacitor, Quality Engineering, 36:1, 118-130, DOI:
10.1080/08982112.2023.2268703

To link to this article: https://doi.org/10.1080/08982112.2023.2268703

Published online: 19 Oct 2023.

Submit your article to this journal

Article views: 117

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=lqen20
QUALITY ENGINEERING
2024, VOL. 36, NO. 1, 118–130
https://doi.org/10.1080/08982112.2023.2268703

A reliability prediction method considering degradation self-acceleration

effect in DC-link electrolytic capacitor
Xuerong Yea, Qisen Suna,b , Ruishi Linc, Cen Chena, Min Xieb, Guofu Zhaia, and Rui Kangd
a
School of Electrical Engineering and Automation, Harbin Institute of Technology, Harbin, China; bDepartment of Systems
Engineering, City University of Hong Kong, Hong Kong, China; cBeijing Aerospace Automatic Control Institute, Beijing, China; dSchool
of Reliability and Systems Engineering, Beihang University, Beijing, China

ABSTRACT KEYWORDS
The reliability of DC-link electrolytic capacitors is crucial to ensure the quality of power sup­ degradation analysis;
ply systems. The degradation of capacitor parameters may lead to a higher temperature electrolytic capacitor;
and thus accelerate degradation as a self-accelerating effect. In this article, an improved reli­ reliability prediction; self-
acceleration effect
ability prediction method for DC-link electrolytic capacitors is proposed, as existing methods
have not adequately accounted for the self-acceleration effect. The degradation under
dynamic stress is obtained by cumulative computations and the stress is updated according
to the degraded parameters. The degradation models are converted into degradation rate
models to overcome the computational challenges associated with small-step iterations that
may make traditional methods unaffordable. The proposed method for developing the deg­
radation rate model is widely applicable and achieves satisfactory accuracy. To demonstrate
the practicality of the proposed method, a case study of a boost motor drive system is pre­
sented. The appropriate iteration step can be determined by comparing the results of the
lifetime distributions obtained using different iteration steps. Degradation paths considering
self-acceleration effects can be obtained, enabling more precise system quality analysis and
reliability prediction.

1. Introduction further accelerating the degradation. Another example

is bond wire damage in insulated gate bipolar transis­
Electrolytic capacitors are essential components in
tors (IGBT), where damage can cause an increase in
power electronics due to their high capacitance values,
high energy density, and favorable price-performance on-resistance, resulting in higher power dissipation
ratio. They play a crucial role in energy storage and and elevated temperature, consequently exacerbating
filtering, with their reliability being closely linked to the bond wire damage (Qin et al. 2021). The failure of
the quality and reliability of the system. However, electrolytic capacitors is contingent on whether its
among power electronic components, electrolytic parameters meet or exceed the fault threshold, such as
capacitors have the shortest lifespan (Gupta et al. a 10% reduction in capacitance or a 300% increase in
2018), underscoring the significance of accurately pre­ equivalent series resistance (ESR). Parler and Dubilier
dicting their reliability. (2004) presented a simplified model for the lifetime of
Failures in electrolytic capacitors can be categorized electrolytic and film capacitors, where the lifetime is a
into catastrophic and degradation failures. function of applied voltage and temperature. The
Degradation failures are characterized by a drift in Arrhenius model is used to describe the temperature-
capacitance C and equivalent series resistance (ESR) lifetime dependence of electrolytic capacitors (Wang
values. A considerable challenge in predicting and Blaabjerg 2014). To obtain the physical explana­
capacitor reliability is accounting for the degradation tions of the lifetime model variants from different
self-accelerating effect. This effect refers to the phe­ manufacturers, Yang et al. (2018) modeled the influ­
nomenon in which the degradation process is further ence of voltage stress using linear, power law, and
accelerated by its effects. For electrolytic capacitors, exponential equations for low, medium, and high volt­
the degradation of capacitance and ESR leads to age stresses, respectively. Capacitor lifetime under
higher power dissipation, raising the temperature and constant stresses can be easily obtained from

CONTACT Qisen Sun [email protected] Department of Systems Engineering, City University of Hong Kong, 999077 Hong Kong, China.
� 2023 Taylor & Francis Group, LLC
 QUALITY ENGINEERING 119

established models. However, the stresses on capaci­ advantage when performing small-step iterations of
tors in practical applications are often dynamic. The the self-accelerating effect. The article is organized as
damage accumulation method is used to analyze cap­ follows: Section 2 introduces the reliability prediction
acitor lifetime and reliability under dynamic stresses. method considering the degradation self-acceleration
Yao et al. (2020) obtained stresses from electrical and effect. Section 3 presents a reliability prediction case
thermal models and used Miner’s linear damage the­ study of an electrolytic capacitor applied in a boosted
ory (Miner 1945) to calculate the capacitor’s lifetime. motor drive system. Finally, Section 4 provides the
The Miner rule is one of the most widely used linear conclusion.
damage models because of its simplicity. However, its
linearization of the degradation process and ignorance
2. Reliability prediction method considering
of stress sequence effects result in unsatisfactory pre­
degradation self-acceleration effect
diction accuracy. Considering the nonlinear process of
capacitor parameter degradation, nonlinear damage The framework of this method is shown in Figure 1.
models, such as the Manson-Halford model (Manson The initial step involves establishing an electrical
and Halford 1981) have been employed. Wang et al. model, which could be a circuit simulation model or a
(2019) used a nonlinear damage model for the long- mathematical model. This model is used to calculate
term lifetime estimation of DC-Link capacitors and the power loss Ploss of the capacitor under the prede­
achieved more reasonable results. termined mission profile. Subsequently, a thermal
The lifetime models become insufficient when con­ model is established in the second step. This model
sidering the self-acceleration effect, as the stress levels utilizes the previously acquired Ploss as an input par­
require updates based on the degradation parameters. ameter. It is used to compute the core temperature
In contrast, degradation models are considered a Tcore of the capacitor, considering the ambient tem­
superior alternative to lifetime models because they perature Ta. The final step is the development of a
are more informative (Zhang et al. 2023; Hamada time-independent degradation rate model. This model
2005). The capacitor degradation under normal condi­ depends on two inputs: the Tcore obtained from the
tions was quantified by Kulkarni et al. (2010), and the
ESR increase over time was determined using
the Arrhenius model. Sun et al. (2015) performed the
accelerated degradation test on electrolytic capacitors
and developed degradation models for the capacitance
and ESR under temperature stress. Linear or nonlinear
cumulative methods are also employed in calculating
degradation under dynamic stress. Lv et al. (2021) uti­
lized a linear approximation of the capacitor’s degrad­
ation and conducted cumulative calculations
accordingly. When the degradation path is not linear,
performing degradation accumulation requires solving
the nonlinear degradation models. There are various
forms of degradation models, and the cost of accumu­
lation calculation will increase when the inverse func­
tion of the degradation model cannot be expressed in
closed form. As iteration steps become smaller, the
computational cost can become unmanageable.
Convincing prediction results cannot be obtained
without computing the results calculated for different
iteration steps. Consequently, the practical application
of traditional methods in real-world engineering prob­
lems may be limited due to computational expenses.
To address this issue, this article proposes a prac­
tical method for capacitor reliability prediction con­
sidering mission and ambient temperature profiles. Figure 1. Reliability prediction method considering degrad­
This method offers a significant computational speed ation self-acceleration effect.
120 X. YE ET AL.

thermal model and the capacitor’s current degradation

state. This model is used to calculate the degradation
of the capacitor efficiently.
The process of predicting the reliability of the cap­
acitor commences with the extraction of the Ploss
from the electrical model, which is subsequently used
as an input for the thermal model. The thermal model
is then tasked with computing the Tcore of the capaci­
tor derived from the Ploss input. Following this, the
degradation model uses the Tcore as an input to calcu­
late the degradation of capacitor parameters. The cal­ Figure 2. Frequency-dependent ESR curve.
culated degradation is then fed back into the electrical
model, facilitating an update to the capacitor’s Ploss.
This constitutes a cyclical computational process. relationship between Irms at diverse frequencies and
Eventually, this iterative procedure yields the degrad­ the capacitor parameters degradation exhibits substan­
ation path, which can be utilized to determine the tial complexity. Capacitor parameters are not the sole
capacitor’s lifetime based on the predetermined failure determinants of Irms. Instead, they are significantly
threshold. Additionally, the variability inherent in the influenced by the system mission profile and load
capacitor is considered. By applying a Monte Carlo type. The Ploss variation with capacitor degradation
simulation, a distribution of the capacitor’s lifetime under different mission profiles and load types
can be obtained, and reliability prediction can be car­ requires specific circuit simulation analysis.
ried out. The critical parts of the method will be ela­ In general, performing full-scale high-fidelity circuit
borated upon in subsequent subsections. simulations to determine power loss during each iter­
ation is computationally impractical, and thus, surro­
gate models may be employed as an alternative
2.1. Electrical model solution. A surrogate model is a substitute for an
The DC-link capacitors are subjected to stable voltage expensive simulation model by approximating the
stress, and thermal stress becomes the primary factor input-output responses of the original model. There
affecting their lifetime. The presence of ripple current, are many categories of surrogate models, and selecting
which is comprised of higher-order harmonic compo­ an appropriate model that satisfies the accuracy
nents in the current, causes a voltage drop to occur requirements should be based on the specific charac­
across the capacitor as a result of the ESR. This volt­ teristics of the available data (Jiang, Zhou, and Shao
age drop leads to power loss and generates heat. 2020).
Based on the properties of the electrolytic capacitor,
the ESR is a frequency-dependent parameter that dis­ 2.2. Thermal model
plays a considerable decrease with increasing fre­
When determining capacitor’s core temperature Tcore,
quency in the low-frequency range while remaining
computational efficiency is also crucial. Given that
nearly constant in the high-frequency range, as shown
board-level thermal dissipation interactions between
in Figure 2.
the capacitor and other components can be neglected,
At a given frequency, the power loss of a capacitor
the lumped thermal model can be used (Niu, Wang,
is a function of ripple current and ESR, while the total
and Ye 2020), as shown in Figure 3.
power loss is the sum of the power losses over the
The lumped model consists of the core-case ther­
entire frequency range, which can be calculated as
mal resistance Rth1 and case-ambient thermal resist­
(Baburajan et al. 2022):
ance Rth2. Rth1 is related to the capacitor structure and
X
n � � material, which can be determined from the finite
2
Ploss ¼ Irms ðfi Þ � ESRðfi Þ (1)
i¼1
element model or experiment. Rth2 varies with the
capacitor heat dissipation conditions and soldering
where Irms ðfi Þ and ESRðfi Þ are the ripple current root method, which need to be determined by experiments.
mean square value and ESR at frequency fi, respect­ Then, the case temperature Tcase and core temperature
ively. Irms at each frequency can be calculated by Tcore can be calculated by:
Fourier Analysis of the ripple currents obtained from
Tcase ¼ Ta þ Ploss � Rth2 (2)
the electrical model. In practical applications, the
 QUALITY ENGINEERING 121

computation is deemed a reasonable approach and

has been implemented in the lifetime estimation of
DC-link capacitors (Wang et al. 2019). As illustrated
in Figure 5, the computation process involves an itera­
tive loop with the following steps:

1. The parameter reaches P1 under stress I at tc,

Figure 3. Lumped thermal model of electrolytic capacitor. after which the stress is updated to stress II
according to P1.
Tcore ¼ Tcase þ Ploss � Rth1 (3) 2. Determine the ts value corresponding to P1 under
stress II based on the degradation model.
3. Using Dt as the iteration step, the degradation P2
2.3. Degradation cumulative computation method at ts þ Dt is calculated under stress II.

The degradation cumulative computation method Step (2) can be readily accomplished if the degrad­
needs to be discussed separately when the degradation ation model possesses a closed-form inverse function.
path is linear and nonlinear. Under constant stress, if However, not all degenerate models allow for the der­
the degradation path of a capacitor parameter follows ivation of an exact inverse function form. In this case,
a linear pattern, the degradation rate remains con­ it has been noted that the computational cost escalates
stant, indicating that it solely depends on the stress considerably as the number of iterations grows. This
level applied. The degradation increments resulting increase in cost is attributed to the necessity of identi­
from various stress levels can be cumulated directly in fying the zero point of the nonlinear degenerate
this case. In contrast, a nonlinear degradation path model during each iteration. Consequently, the viabil­
signifies that the degradation rate will vary over time. ity of this method for practical applications may be
The degradation rate depends not only on the stress limited, and using the degradation rate model for self-
level applied but also on the degradation state. acceleration calculation is considered more versatile.
The cumulative computation method for degrad­
ation is analogous to the cumulative computation
method for fatigue damage. A comprehensive review 2.4. Degradation rate model development
of the cumulative damage theories development was The degradation rate model is stresses and degrad­
conducted by Yang and Fatemi (1998). Since the ation-state dependent. As illustrated in Figure 6, the
introduction of the damage accumulation concept by degradation can be obtained by integrating the deg­
Palmgren (1924) and the linear damage rule by Miner radation rate k over time and subsequently updating
(1945), various damage models have been developed, the stress based on the new degradation state. The
such as linear damage rules, nonlinear damage curves, computational speed is considerably enhanced due to
and two-stage linearization approaches. The nonlinear the absence of the requirement to solve the nonlinear
damage curve approach can be categorized into two degradation model.
main methods: time-aligned and state-aligned (Betten Establishing degradation models through acceler­
2008; Thiel, Griggio, and Tiku 2022), as shown in ated degradation tests is a prerequisite step in devel­
Figure 4. oping a degradation rate model, which helps
In the time-aligned method, when the stress determine whether degradation over time is linear or
changes from I to II at tc, the forthcoming degrad­ nonlinear. As an illustration, Abdennadher et al.
ation will be calculated in the path of stress II at tc. In (2010) and our prior research (Yuege, Xuerong, and
the state-aligned method, when the stress changes Guofu 2011) on aluminum electrolytic capacitors have
from I to II at tc, the forthcoming degradation will be demonstrated that the C degrades linearly over time,
calculated in the path of stress II starting with the whereas the ESR exhibits nonlinear degradation.
degradation state at tc in stress I. The primary degrad­ Therefore, establishing a degradation rate model is
ation mechanism for electrolytic capacitors is attrib­ only necessary for the ESR. The degradation rate
uted to the evaporation and deterioration of the model, which is stress and degradation-state depend­
electrolyte (Gupta et al. 2018). The decline in C and ent, is modeled as follows:
the rise in ESR directly manifest the electrolyte evap­ � �
DP q
oration and deterioration in electrolytic capacitors. k¼ 1þ � f ðSÞ (4)
P0
Accordingly, the state-aligned method for cumulative
122 X. YE ET AL.

Figure 4. Cumulative computation method. (a) Time-aligned method and (b) state-aligned method.

been used to describe the degradation of lithium bat­

teries (Thomas et al. 2012, 2019). For ESR degradation
under temperature stress, the degradation rate model
is:
� �q
ESRd
kESR ¼ � f ðTÞ (5)
ESR0
� �
b1
f ðTÞ ¼ b0 � exp − (6)
T
where ESRd is the degradation value, ESR0 is the ini­
tial value. f(T) is an Arrhenius form, describing the
Figure 5. An iteration loop computation based on the degen­ effect of temperature on degradation rate, and b0 and
erate model. b1 are fitting parameters.

2.5. Monte Carlo simulation and reliability

prediction
The degradation model offers a defined path for given
parameters and stresses. However, the parameters of
the degradation model are stochastic due to the
intrinsic differences between the capacitors generated
during the manufacturing process. The statistical char­
acteristics of the degenerate model parameters can be
determined through testing or reasonable assump­
tions. It is widely accepted that a normal distribution
with a certain variation range (e.g., 5%) is assumed
for the model parameters representing the variations
in the manufacturing process (Rahimpour et al. 2022;
Figure 6. Iteration computation based on the degenerate rate Chen et al. 2022; Lv et al. 2021). Consequently, the
model.
parameters of the degradation rate model are also sto­
chastic. Monte Carlo simulation is introduced for ran­
where k is the degradation rate, DP the is parameter dom sampling in each distribution, and the parameter
degradation value, P0 is the parameter initial value, combinations are used for self-acceleration effect itera­
and q is a coefficient that provides additional model­ tive calculation.
ing flexibility. The rate either increases (q > 0) or The capacitor lifetime can be determined by estab­
decreases (q < 0) over time for given stresses. S is the lishing failure thresholds for C and ESR. Specifically,
stress vector and f(S) describes the effect of stresses the lifetime can be identified as the time when the C
on the degradation rate. A similar model form has falls below the threshold or the ESR exceeds the
 QUALITY ENGINEERING 123

Figure 7. Boosted motor drive system.

threshold. The distribution type of lifetime data can to drive a permanent magnet synchronous motor
be determined by distribution identification. For (PMSM), as shown in Figure 7. The machine is sub­
example, the lifetime data can be fitted with a distri­ ject to both acceleration and deceleration. It operates
bution, such as the Weibull distribution (Lv et al. in both motoring and generating modes. The speed
2015). The 3-parameter Weibull distribution probabil­ compensator generates a torque set point that serves
ity density function is as follows: as a reference signal for computing the quadrature
� � " � � # current set point for the inverter controller. The boost
b t − c b−1 t−c b circuit is responsible for converting the input voltage
f ðtÞ ¼ exp − (7)
g g g of 200–400 V DC-link output, and the target electro­
where b is the shape parameter, g is the scale param­ lytic capacitor is located at the output side of the
eter, and c is the location parameter. The reliability boost converter.
derived from the probability density function is:
" � � # 3.2. Mission profile and electrical model
t−c b
RðtÞ ¼ exp − (8)
g As shown in Figure 8, the PMSM accelerates and
decelerates according to a set speed profile. During
In addition, normal distributions are often used in acceleration, the bi-directional converter transfers
engineering problems for life analysis. In this case, the energy from the battery to the PMSM. During decel­
reliability can be calculated as follows: eration, the generated energy is transferred from the
� �
t−l PMSM to the battery.
RðtÞ ¼ 1 − U (9) The circuit simulation model can be constructed
r
using software tools, such as SaberTM and PLECSV. By
R

where l is the mean, r is the standard deviation, and incorporating the speed setting curve, the ripple cur­
U is the cumulative distribution function of the stand­ rent of the capacitor can be calculated. Furthermore,
ard normal distribution. When the sample size of the current at each frequency can be obtained through
Monte Carlo is large enough, the reliability can also Fourier Analysis, as depicted in Figure 9.
be calculated as follows: The degradation of the capacitor can alter the cap­
Survival number at t acitor current under the mission profile, resulting in
RðtÞ ¼ (10)
Number of samples n changes in Ploss. The objective is to obtain the Ploss
quickly and accurately based on the values of C and
ESR. Relying on simulation software for each self-
3. Case study on a boosted motor drive accelerating iteration cannot meet the speed require­
system ments for practical applications. Consequently, based
on the simulation results, we develop a surrogate
3.1. Specification of application situation
model that relates the Ploss to the degradation parame­
The system consists of a battery pack that supplies ters. Establishing a surrogate model comprises three
power to the boost converter and inverter converter integral steps: selecting sample points, training the
124 X. YE ET AL.

model, and validating the model’s accuracy. The first Here, we select parameter combinations within the
step involves the selection of a set of input-output failure threshold of C and ESR for simulation to
data points. The goal is to ensure that these data obtain Ploss. The parameter selection range for C is
points adequately represent the system’s behavior. [90%C0, C0], and for ESR is [ESR0, 300%ESR0]. This
Several methods can be employed to facilitate this means that the degraded capacitance exceeds 90% of
selection process. These include uniform sampling, the initial C0 ¼ 1000 lF or the degraded ESR exceeds
where points are distributed evenly across the input 300% of the initial ESR0 ¼ 0.203 X is considered as
space; adaptive sampling, where sample points are failure. Uniform sampling involves selecting sample
added based on the model’s prediction errors. After points at regular intervals in the input space. It is use­
selecting sample points, the next step is to choose an ful when the input space is well-defined and the rela­
appropriate surrogate model type and train it. The tionship between the input parameters and the system
specific application and the data available influence response is relatively simple. Select six distinct capaci­
the selection of a surrogate model. The final step tance values, each at an interval of 2% from the C0. In
involves evaluating the surrogate model’s accuracy. parallel, select six ESR values, each at a 60% interval
This step is achieved by comparing the model’s pre­ from the ESR0. Combining these selected values of
dictions with the actual output values of the system. It capacitance and ESR will yield 36 sets of inputs,
is essential to note that establishing a surrogate model thereby providing a comprehensive initial sample
is iterative. It may necessitate numerous repetitions of space. Subsequently, each of these sets is processed
the steps involving sample point selection, model through an electrical model to derive the correspond­
training, and evaluation. This iterative process is cru­ ing Ploss.
cial to continually refine the model and enhance its Due to the simple relationship between Ploss and
accuracy. capacitor parameters, the following polynomial is
used:
Ploss ðC, ESRÞ ¼ a0 þ a1 � C þ a2 � ESR þ a3 � C2 þ a4 � C
� ESR þ a5 � ESR2
(11)
where a0, a1, a2, a3, a4, and a5 are fitting parameters.
Using the sample data to fit Equation (11), the estab­
lished Ploss model is shown in Figure 10, and the fit­
ting results are shown in Table 1. The established Ploss
model can accurately fit the data with R-Squared ¼
0.99. We compare the results of thirteen simulation
sets outside the fitted data with the established model
to validate the accuracy. The sum squared error (SSE)
Figure 8. Motor speed setting curve. of 6.4�10−6 is obtained, indicating satisfactory accur­
acy. It is imperative to underscore that if the surrogate

Figure 9. Electrical model simulation results. (a) Capacitor current. (b) Fourier analysis results of capacitor current.
 QUALITY ENGINEERING 125

model’s accuracy falls short of the requisite standards, 3.4. Degradation rate model development and
it should be refined by incorporating additional sam­ self-accelerating computation
ple data.
The degradation data and models of electrolytic
capacitors in literature (Sun et al. 2015; Sun et al.
3.3. Ambient temperature profile and thermal 2016) are utilized to demonstrate the process of devel­
model oping the degradation rate model. The linear degrad­
ation of capacitance over time can be established as
Ambient temperature is vital in analyzing thermal
follows:
stress in capacitors, as it significantly affects the reli­
� �
ability prediction. A multi-timescale ambient tempera­ CðtÞ −Ea1
¼ 1 þ A0 � exp �t (13)
ture profile is required to compare the reliability C0 j�T
prediction results using different iteration steps. For
where C0 is the initial capacitance, A0 is the base deg­
instance, an ambient temperature profile for Hong
radation rate, Ea1 is the activation energies, j is the
Kong, which provides minute-level detail, is depicted
Boltzmann constant, t is time, and T is temperature.
in Figure 11.
Without loss of generality, a two-term exponential
The finite element simulation is initially used to
model can be used to describe the nonlinear degrad­
calculate the accurate temperature of the capacitor.
ation of ESR over time:
Once verified by testing, a lumped thermal model can
be constructed. Table 2 presents the thermal proper­ ESRðtÞ
¼ 1 þ exp ðB0 � T B1 � tÞ − exp ðB2 � T B3 � tÞ
ties of the capacitor material necessary to construct ESR0
the finite element model. Subsequently, the tempera­ (14)
ture is calculated through finite element simulation, as where B0, B1, B2, and B3 are fitting parameters.
shown in Figure 12. The capacitance degradation rate kC can be simply
In contrast to power semiconductor devices, the expressed in a time-independent form:
temperature of a capacitor remains relatively stable � �
during operation. As a result, the impact of heat cap­ −Ea1
kC ðTÞ ¼ A0 � exp (15)
acity on the transient temperature is not considered. j�T
The core-to-case thermal resistance Rth1 and the case- To fit the ESR degradation rate model parameters,
to-ambient thermal resistance Rth2 can be calculated degradation rate data corresponding to degradation
as follows: states at different temperatures are generated based on
�
Rth1 ¼ ðTcore − Tcase Þ=Ploss Table 1. Ploss model fitting results.
(12)
Rth2 ¼ ðTcase − Ta Þ=Ploss Fitting parameters Value
a0 −0.2112
According to simulation results, the core-to-case a1 0.000255
a2 0.2836
thermal resistance Rth1 is 29.2 � C/W and the case-to- a3 −1.496�10−8
ambient thermal resistance Rth2 is 72 � C/W. a4 0.001306
a5 −0.5425

Figure 10. Power loss variation versus ESR and capacitance. Figure 11. Annual ambient temperature profile.
126 X. YE ET AL.

Table 2. Electrolytic capacitor material thermal properties. Table 3. Degradation rate model fitting results.
−1 −1
Part Material Thermal conductivity (W�cm �K ) Parameters Value
Can Aluminum 237 b0 800.6
Can surface Phenolic plastic 0.12 b1 5943
Core Electrolyte mixture 1.2 q 0.591
Pins Cu-Sn 109
Rubber seal Rubber 0.16

capacitance degradation paths is 0.02%, and the max­

imum difference between ESR degradation paths is
4.8%. This primarily be attributed to fitting errors
introduced during the curve-fitting process of the ESR
degradation rate model.

3.5. Monte Carlo simulation and reliability

prediction
The statistical characteristics of the degenerate model
coefficients are considered here. In Equation (15) and
Figure 12. Finite element thermal simulation with Equation (5), the variables A0 and b0 are defined.
Ploss ¼ 100 mW and ambient temperature 25 � C. They are assumed to obey a normal distribution,
where the parameter variation is 5%. This implies that
the degradation model. The temperature used for deg­
the range covering �99.9% of the distribution
radation rate model development is first designed in
(equivalent to six standard deviations) is �10% of the
the operating temperature range of the capacitor.
mean value of the distribution. The Monte Carlo
Here 298.15, 318.15, 338.15, and 358.15 K are chosen,
simulation generates 10,000 samples with varying
corresponding to 25, 45, 65, and 85 � C. The degrad­
parameters to explore the possible results. First, the
ation rate model is fitted using Equation (5) and
Equation (6) after calculating the degradation rate for capacitor’s lifetime is calculated without considering
the time corresponding to the ESR failure threshold at the self-acceleration effect, implying that the Ploss
each temperature. Table 3 presents the fitting results, remains constant throughout the degradation process.
and R-Squared ¼ 0.99 and SSE ¼ 7.9�10−9 indicate The calculation is performed for four distinct
that the fitting is almost perfect. scenarios:
After the degradation rate model development, the
algorithm of the self-accelerating computation is � N1: calculate with average temperature of Ta per
shown in Figure 13. The computation starts at ts and minute.
ends at te, Dt is the iteration step. The process begins � N2: calculate with average temperature of Ta per
with the computation of Ploss, as determined by hour.
Equation (11). Following this, the core temperature is � N3: calculate with average temperature of Ta
calculated using Equation (3). Next, the rate of deg­ per day.
radation for both C and ESR is calculated through � N4: calculate with average temperature of Ta per
Equations (15) and (5), respectively. The degradation month.
occurring during Dt is then calculated by multiplying
the degradation rate by Dt. The calculation process is The results without considering the degradation
repeated until ts reaches a predefined end time te. The self-acceleration effect are used for comparison with
Ploss is updated utilizing the degraded C and ESR val­ the subsequent lifetime results considering the degrad­
ues. Degradation leads to a higher capacitor core tem­ ation self-acceleration effect. The lifetime results are
perature and self-acceleration effects are introduced. presented in Figure 14 and Table 4. The distribution
Additionally, a comparative evaluation is conducted type can be ascertained through distribution identifi­
between the degradation paths (degenerate to failure) cation. The common method for individual distribu­
obtained by the proposed and the degradation model- tion identification uses probability plots and
based methods. The results show that the degradation goodness-of-fit tests. Probability plots are used to
paths are fundamentally consistent and the accuracy is visually inspect the data and determine if it follows a
satisfactory. The maximum difference between particular distribution. Taking the normal probability
 QUALITY ENGINEERING 127

Figure 14. Lifetime distribution without considering self-accel­

eration effect.

during temperature fluctuations. This discrepancy sug­

gests that, in our case, temperature fluctuations within
a single day exert a negligible influence on the lifetime
results.
The lifetime calculation, which considers the self-
acceleration effect, is separated into seven distinct
scenarios:

� M1: iteration per minute with Ta the average per

minute.
� M2: iteration per hour with Ta the average per
hour.
� M3: iteration per hour with Ta the average
Figure 13. Algorithm of the self-accelerating computation.
per day.
� M4: iteration per day with Ta the average per day.
plot as a case in point, one should scrutinize the dis­ � M5: iteration every two days with Ta the average
tance between the individual data points and the line every two days.
of best fit, which appears as the central straight line � M6: iteration every five days with Ta the average
on the plot. The degree to which these points align every five days.
closely with the straight line indicates the data’s com­ � M7: iteration every ten days with Ta the average
patibility with a normal distribution. Goodness-of-fit every ten days.
tests, such as the Anderson-Darling test, compare the
These seven scenarios are used to compare the life­
data to a set of candidate distributions and determine
time results calculated under different iteration steps,
which best fits the data. MinitabV is a statistical soft­
R

and the settings of M2 and M3 are used to compare

ware that provides probability plots and goodness-of- the impact of ambient temperature profile granularity
fit tests for individual distribution identification. on the lifetime results. Considering the actual degrad­
Other statistical software and tools also provide simi­ ation situation, computations at smaller steps are not
lar features for individual distribution identification. performed. As shown in Figure 15, the lifetime results
Here, the lifetime results can be accepted as a normal
distribution upon conducting distribution identifica­ Table 4. Lifetime distribution without considering self-acceler­
tion. The lifetime results of N1, N2, and N3 exhibit ation effect.
Cases Mean l l 95% CI Variance r r 95% CI
high similarity. In contrast, the results for scenario N4
N1 581.70 [581.33, 582.07] 18.80 [18.55, 19.07]
deviate considerably from the other scenarios. The N2 581.30 [580.94, 581.66] 18.44 [18.19, 18.70]
overestimated lifetime of N4 can be attributed to dis­ N3 580.92 [580.57, 581.29] 18.38 [18.12, 18.63]
N4 616.16 [615.74, 616.57] 21.13 [20.85, 21.43]
regarding the high temperature’s role in degradation
128 X. YE ET AL.

Figure 15. Lifetime distribution considering self-acceleration effect.

Table 5. Lifetime distribution considering self-acceleration Table 6. Comparison of execution times.

effect. Execution time (s)
Cases Mean l l 95% CI Variance r r 95% CI
100 days 200 days 365 days 1825 days
M1 196.79 [196.68, 196.91] 5.69 [5.61,5.77]
Proposed method 0.550 0.902 1.565 7.017
M2 196.97 [196.86, 197.08] 5.68 [5.60, 5.76]
Degradation 2177.647 8774.014 27832.715 —
M3 196.77 [196.66, 196.87] 5.52 [5.45, 5.60]
model-based method
M4 199.16 [199.05, 199.27] 5.59 [5.51, 5.66]
M5 201.54 [201.43, 201.64] 5.50 [5.38, 5.53]
M6 275.81 [275.64, 275.99] 9.06 [8.94, 9.19]
M7 330.46 [330.21, 330.71] 12.55 [12.38, 12.73] temperature fluctuations within a single day negatively
influence the lifetime results. A comparative analysis
of the results from M4, M5, M6, and M7 reveals a
progressive increase in the average lifetime with an
escalation in the iteration step. These results suggest
that the degradation triggers considerable changes in
Ploss on a temporal scale that transcends a single day.
When considering computational efficiency and cost,
in our case, the M4 is sufficient and satisfactory for
engineering applications. Comparing N3 and M4, the
average lifetime prediction results are overestimated
by 2.9 times. This implies that the degradation self-
accelerating effect is a crucial factor to take into
account in the process of reliability prediction.
To illustrate the computational speed advantage of the
proposed method, a comparison is made between our
Figure 16. Reliability of M1 and M4. method and the degradation model-based method. The
degradation model-based method employs Brent’s
method (Brent 1971) for calculating ts in Figure 5, a com­
can be accepted as a normal distribution upon con­ bination of bisection and secant methods recognized for
ducting distribution identification. with the parame­ its robustness and speed. Self-accelerating degradation cal­
ters shown in Table 5. culations are performed on 10,000 capacitors for different
The lifetime results for M1, M2, M3, and M4 durations, and the codes are run on a multi-core com­
exhibit a substantial degree of congruence, and the puter with an Intel i7-1170 processor (2.50 GHz). The
comparison of reliability curves between M1 and M4 results presented in Table 6 demonstrate the superior
is illustrated in Figure 16. The average lifetime differ­ computational speed of the proposed method.
ence between M1 and M4 is only three days. Such
results show that when considering the self-acceler­
ation effect, degradation occurring within a single day
4. Conclusion
proceeds at a relatively slow pace and does not We propose a reliability prediction method consider­
yield significant alterations in Ploss. Furthermore, ing the degradation self-acceleration effect in DC-Link
 QUALITY ENGINEERING 129

electrolytic capacitors. The degradation cumulative articles. His research interests include quality control and
computation method based on the degradation model process monitoring, robust parameter design, and reliability.
Rui Kang is a professor in School of Reliability and
is first discussed and determined. To address the chal­
Systems Engineering, Beihang University, Beijing, China. He
lenge of high computational costs in the self-accelerat­ is currently serving as the associate editor of IEEE Trans.
ing iterative calculation of the degradation model on Reliability and is the founder of China Prognostics and
without a closed-form inverse function, a degradation Health Management Society. His main research interests
rate model development method is proposed. The include reliability, resilience for complex system, and mod­
eling epistemic uncertainty in reliability and
method proposed for developing the degradation rate maintainability.
model demonstrates broad applicability and satisfac­
tory accuracy. A case study of a boosted motor drive
system demonstrates the practicality of the reliability ORCID
prediction method. Considering the accuracy and Qisen Sun http://orcid.org/0000-0002-7038-2599
computation cost, iterative computation by daily aver­
age temperature in daily steps proved appropriate.
References
The degradation paths considering the self-acceler­
ation effect can be obtained by this method to serve Abdennadher, K., P. Venet, G. Rojat, J. M. Retif, and C.
Rosset. 2010. A real-time predictive-maintenance system
for more accurate system quality analysis and reliabil­
of aluminum electrolytic capacitors used in uninterrupted
ity prediction. Furthermore, it has the potential to be power supplies. IEEE Transactions on Industry
extended to quality analysis and reliability prediction Applications 46 (4):1644–52. doi: 10.1109/TIA.2010.
for other types of capacitors and components consid­ 2049972.
ering the self-acceleration effect. Baburajan, S., H. R. Wang, F. Mandrile, B. Yao, Q. Wang,
D. Kumar, and F. Blaabjerg. 2022. Design of common
DC-link capacitor in multiple-drive system based on
About the authors reduced DC-link current harmonics modulation. IEEE
Transactions on Power Electronics 37 (8):9703–17. doi: 10.
Xuerong Ye is currently a Professor at the Department of 1109/TPEL.2022.3143545.
Electrical Engineering, Harbin Institute of Technology. He Betten, J. 2008. Creep mechanics. Berlin: Springer Science &
received B.S., M.S., and Ph.D. degrees in electrical engineer­ Business Media.
ing from the Harbin Institute of Technology, China. His Brent, R. P. 1971. An algorithm with guaranteed conver­
research interests include robust parameter design and reli­ gence for finding a zero of a function. The Computer
ability prediction. Journal 14 (4):422–5. doi: 10.1093/comjnl/14.4.422.
Qisen Sun is a Ph.D. student in the Department of Chen, Z. G., J. P. Xu, P. Davari, and H. Wang. 2022. A
Systems Engineering at the City University of Hong Kong, mixed conduction mode-controlled bridgeless boost PFC
Hong Kong, China. He received an M.S. degree in electrical converter and its mission profile-based reliability analysis.
engineering from the Harbin Institute of Technology, IEEE Transactions on Power Electronics 37 (8):9674–86.
Harbin, China. His research interests include statistical doi: 10.1109/TPEL.2022.3153558.
engineering and reliability prediction. Gupta, A., O. P. Yadav, D. DeVoto, and J. Major. 2018. A
Ruishi Lin is the Deputy Chief Engineer of the Beijing review of degradation behavior and modeling of capaci­
Institute of Aerospace Automation, Beijing, China. tors. Paper presented at the ASME 2018 International
Technical Conference and Exhibition on Packaging and
Cen Chen is currently an Associate Professor at the
Department of Electrical Engineering, Harbin Institute of Integration of Electronic and Photonic Microsystems.
doi: 10.1115/IPACK2018-8262.
Technology. He received his B.S. and Ph.D. degrees in elec­
Hamada, M. 2005. Using degradation data to assess reliabil­
trical engineering from the Harbin Institute of Technology,
ity. Quality Engineering 17 (4):615–20. doi: 10.1080/
Harbin, China. His research interests include electronic sys­
08982110500225489.
tem reliability prediction, fault diagnosis, and health Jiang, P., Q. Zhou, and X. Shao. 2020. Surrogate model-
management. based engineering design and optimization. Singapore:
Min Xie is a Chair Professor at the City University of Springer.
Hong Kong. Before that, he was a Professor at the National Kulkarni, C., G. Biswas, X. Koutsoukos, J. Celaya, and K.
University of Singapore. He received his undergraduate and Goebel. 2010. Experimental studies of ageing in electro­
postgraduate education in Sweden with a Ph.D. from lytic capacitors. Annual Conference of the PHM Society 2
Linkoping University in 1987. Prof Xie has published over (1). doi: 10.36001/phmconf.2010.v2i1.1727.
300 journal articles and eight books. Prof Xie has been an Lv, C. L., J. J. Liu, Y. Zhang, W. J. Lei, R. Cao, and G. T.
elected fellow of IEEE since 2006 and has been elected to Lv. 2021. Reliability modeling for metallized film capaci­
the European Academy of Sciences and Arts. tors based on time-varying stress mission profile and
Guofu Zhai is currently a Professor in the Department aging of ESR. IEEE Journal of Emerging and Selected
of Electrical Engineering at Harbin Institute of Technology. Topics in Power Electronics 9 (4):4311–9. doi: 10.1109/
He has published more than 40 peer-reviewed journal JESTPE.2020.3001604.
130 X. YE ET AL.

Lv, S. S., Z. W. Niu, L. Qu, S. G. He, and Z. He. 2015. systems. Circuit World 48 (4):425–33. doi: 10.1108/CW-
Reliability modeling of accelerated life tests with both 08-2020-0195.
random effects and nonconstant shape parameters. Thomas, E. V., I. Bloom, J. P. Christophersen, and V. S.
Quality Engineering 27 (3):329–40. doi: 10.1080/ Battaglia. 2012. Rate-based degradation modeling of lith­
08982112.2015.1037393. ium-ion cells. Journal of Power Sources 206:378–82. doi:
Manson, S. S., and G. R. Halford. 1981. Practical implemen­ 10.1016/j.jpowsour.2012.01.106.
tation of the double linear damage rule and damage Thomas, E. V., I. Bloom, J. P. Christophersen, D. C.
curve approach for treating cumulative fatigue damage. Robertson, L. K. Walker, C. D. Ho, and V. S. Battaglia.
International Journal of Fracture 17 (2):169–92. doi: 10. 2019. Modeling memoryless degradation under variable
1007/BF00053519. stress. Journal of Quality Technology 51 (3):284–99. doi:
Miner, M. A. 1945. Cumulative damage in fatigue. 10.1080/00224065.2019.1569963.
Niu, H., S. J. Wang, and X. R. Ye. 2020. Reliability-oriented Wang, H., and F. Blaabjerg. 2014. Reliability of capacitors
optimization of aluminum electrolytic capacitor consider­ for DC-link applications in power electronic converters–
ing uncertain mission profile. Microelectronics Reliability An overview. IEEE Transactions on Industry Applications
114:113817. doi: 10.1016/j.microrel.2020.113817. 50 (5):3569–78. doi: 10.1109/TIA.2014.2308357.
Palmgren, A. 1924. Die Lebensdauer von Kugellagern. Vol. Wang, H. R., P. Davari, H. Wang, D. Kumar, F. Zare, and
68, 339–41. Berlin: Veifahrenstechinik. F. Blaabjerg. 2019. Lifetime estimation of DC-link capaci­
Parler, S. G., and P. C. Dubilier. 2004. Deriving life multi­ tors in adjustable speed drives under grid voltage unbal­
pliers for electrolytic capacitors. IEEE Power Electronics ances. IEEE Transactions on Power Electronics 34 (5):
4064–78. doi: 10.1109/TPEL.2018.2863701.
Society Newsletter 16 (1):11–2.
Yang, and A. Fatemi. 1998. Cumulative fatigue damage
Qin, F., X. R. Bie, T. An, J. R. Dai, Y. W. Dai, and P. Chen.
mechanisms and quantifying parameters: A literature
2021. A lifetime prediction method for IGBT modules
review. Journal of Testing and Evaluation 26 (2):89–100.
considering the self-accelerating effect of bond wire dam­
doi: 10.1520/JTE11978J.
age. IEEE Journal of Emerging and Selected Topics in
Yang, Y., H. Wang, A. Sangwongwanich, and F. Blaabjerg.
Power Electronics 9 (2):2271–84. doi: 10.1109/JESTPE. 2018. Design for reliability of power electronic systems.
2020.2992311. In Power electronics handbook, 1423–40. Oxford:
wRahimpour, S., H. Tarzamni, N. V. Kurdkandi, O. Husev, Butterworth-heinemann.
D. Vinnikov, and F. Tahami. 2022. An overview of lifetime Yao, B., X. L. Ge, H. M. Wang, H. R. Wang, D. Zhou, and
management of power electronic converters. IEEE Access B. Gou. 2020. Multitimescale reliability evaluation of DC-
10:109688–711. doi: 10.1109/ACCESS.2022.3214320. link capacitor banks in metro traction drive system. IEEE
Sun, B., X. J. Fan, C. Qian, and G. Q. Zhang. 2016. PoF-simula­ Transactions on Transportation Electrification 6 (1):213–
tion-assisted reliability prediction for electrolytic capacitor in 27. doi: 10.1109/TTE.2020.2974182.
LED drivers. IEEE Transactions on Industrial Electronics 63 Yuege, Z., Y. Xuerong, and Z. Guofu. 2011. Degradation
(11):6726–35. doi: 10.1109/TIE.2016.2581156. model and maintenance strategy of the electrolytic
Sun, B., X. Fan, C. A. Yuan, Q. Cheng, and G. Zhang. 2015. capacitors for electronics applications. Paper presented at
A degradation model of aluminum electrolytic capacitors the 2011 Prognostics and System Health Management
for LED drivers. Paper Presented at the 2015 16th Conference, 24–25 May 2011.
International Conference on Thermal, Mechanical and Zhang, J. P., Y. J. Zhang, L. Sun, J. Fu, Q. Zhou, J. Hu, C.
Multi-Physics Simulation and Experiments in Bao, and S. Y. Chen. 2023. A life extrapolation model
Microelectronics and Microsystems, 19–22 April 2015. using accelerated equivalent luminance and its applica­
Thiel, G., F. Griggio, and S. Tiku. 2022. Cumulative degrad­ tion. Quality Engineering 35 (2):258–66. doi: 10.1080/
ation methodology to predict reliability of electronic 08982112.2022.2119409.