http://ieeexplore.ieee.

org/Xplore

Fault Prognosis for Power Electronics Systems

Using Adaptive Parameter Identification
Jason Poon, Student Member, IEEE, Palak Jain, Student Member, IEEE, Costas Spanos, Fellow, IEEE,
Sanjib Kumar Panda, Senior Member, IEEE, Seth R. Sanders, Fellow, IEEE

Abstract—This paper presents the design, implementation, common failure modes of passive components in a power
and experimental validation of a method for fault prognosis electronics systems and the effect that these failures have on
for power electronics systems using an adaptive parameter the resulting parameter value and ESR. The reasons for these
identification approach. The adaptive parameter identifier uses
a generalized gradient descent algorithm to compute real-time failures vary widely, and include manufacturing defects, harsh
estimates of system parameters (e.g. capacitance, inductance, environmental conditions (e.g. temperature and humidity),
parasitic resistance) in arbitrary switching power electronics aging, high voltage stress, insulation failures, interconnection
systems. These estimates can be used to monitor the overall failures, mechanical wear, and vibrations and shocks [4], [6]–
health of a power electronics system, and predict when faults [8]. Moreover, the effect of the failures can be classified as
are more likely to occur. Moreover, the estimates can be used
to tune control loops that rely on the system parameter values. either ‘hard’ or ‘soft’ faults. A hard fault is one that causes
The parameter identification algorithm is general in that it can a sudden and catastrophic effect in the system (e.g. a short
be applied to a broad class of systems based on switching power circuit), while a soft fault is one that causes a gradual effect
converters. We present a real-time experimental validation of the or degradation in the system, generally related to lifetime wear
proposed fault prognosis method on a 3 kW solar photovoltaic or aging.
interleaved boost dc-dc converter system for tracking changes in
passive component values. The proposed fault prognosis method Parameter identification has been investigated previously
enables a flexible and scalable solution for condition monitoring in the context of power electronics systems. One salient
and fault prediction in power electronics systems. application for parameter identification in power electronics
systems has been for estimating the capacitance or equivalent
series resistance (ESR) of a dc-link capacitor [9]–[14]. In
I. I NTRODUCTION
many converters, the dc-link capacitor, particularly electrolytic
Many mission-critical power electronics systems, including capacitors, is one of the primary points of failure in the
renewable energy integration, data center power delivery, and converter. Actively monitoring the capacitance or ESR of
motor drives applications, require high reliability and availabil- the capacitor enables detection and prediction of when these
ity of service [1], [2]. In many of these scenarios, techniques failures will occur.
for fault prognosis are commonly employed, that is, methods In general, most algorithms for parameter identification
for actively monitoring the system condition and predicting compare measurements from the physical system (e.g. voltage,
when failures or faults will occur. A central technology that current, temperature) with a model structure. This model
enables fault prognosis is parameter identification, or identify- structure could be a black-box (e.g. neural networks [15]–
ing the values of system parameters in a real-time and online [18]) or based on a physical model (e.g. Kalman filter, state
manner. By tracking the values of important system parameters observers [10], [19]–[24]).
in real-time, operators can actively monitor the overall health These existing approaches in literature, however, have cer-
of a system and anticipate when maintenance or repairs will tain characteristics that limit that effectiveness and applica-
be needed. Moreover, fault prognosis can be achieved by bility of the methods. Some of these limitations include: (1)
monitoring if estimated parameter values are above or below high computational requirements (e.g. requires an external PC
an accepted tolerance range. or graphical processing unit), (2) custom analog implementa-
The failure modes and mechanisms for power electronics tions, (3) techniques based on heuristic analysis that are only
systems have been widely investigated, for instance in [3]–
[7]. Passive components, such as capacitors and inductors,
are a key failure point. Table I provides an overview of the TABLE I: Typical fault modes for capacitors (electrolytic, ceramic, and film)
and inductors (air inductors, ferrites, and iron-core) in power electronics
systems [4], [7].
This work is supported in part by the NSF through the Graduate Research
Fellowship program, and in part by Republic of Singapore’s National Research
Foundation through a grant to the Berkeley Education Alliance for Research Component Fault modes Fault type Parameter drift direction
in Singapore (BEARS) for the Singapore-Berkeley Building Efficiency and Open circuit (OC) Hard C = 0, ESR = ∞
Sustainability in the Tropics (SinBerBEST) Program. Capacitors Short circuit (SC) Hard C = 0, ESR = 0
J. Poon, C. Spanos, and S. Sanders are with the Department of Electrical Wear out Soft C = ↓, ESR = ↑
Engineering and Computer Sciences at the University of California, Berkeley, Open circuit Hard L = 0, ESR = ∞
CA 94720 USA (email: {jason, spanos, seth.sanders}@berkeley.edu). Full-winding SC Hard L = 0, ESR = 0
Inductors
P. Jain and S. Panda are with the Department of Electrical and Computer Inter-turn SC Soft L = ↓, ESR = ↓
Engineering at the National University of Singapore, Singapore 119077 Core to winding SC Soft L = ↓, ESR = ↑
(email: {palakjain, sanjib.kumar.panda}@nus.edu.sg. Wear out Soft L = ↓, ESR = ↑

Digital Object Identifier 10.1109/TIA.2017.2664052

0093-9994 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
applicable to a specific switching converter, (4) injection of
external voltage or current signals, or (5) need for extensive R1 L1 SW1 SW2
data sets and training.
In this paper, we present a technique for fault prognosis
R2 L2 C
for power electronics systems using an adaptive parameter iload (t)
identifier approach. The fundamental algorithm (first proposed
in [25]) uses a switched linear model of a switching power
vin (t)
converter and a generalized gradient descent algorithm to
dynamically track the values of passive components, such as SW3 SW4
capacitors and inductors, in a power electronics converter. The
algorithm is digitally implemented in real-time on the same Fig. 1: Circuit topology of an interleaved boost dc-dc converter.
embedded processor as the control system, and is used to
monitor when passive component parameter values are above TABLE II: Possible open/close switch positions for the interelaved boost
or below a predefined tolerance range, which would indicate converter in continous conduction mode.
a fault scenario.
The design advantages of the proposed method compared σ(t) 1 2 3 4
s1 0 1 0 1
to state-of-the-art is threefold. First, the modeling and identi- s2 0 0 1 1
fication algorithm is general in that it is (in theory) applicable s3 1 0 1 0
to any switching power converter which can be modeled as s4 1 1 0 0
a switched linear system. Thus, although the focus of this
work is on dc-dc converters for solar photovoltaic applications,
the technique can also be applied in the context of motor devices (e.g. MOSFETs, IGBTs, diodes). Such systems can
drive applications, cable integrity monitoring, among others. be modeled as a switched linear system, that is:
Second, the real-time algorithm is implemented digitally, as
opposed to techniques that require custom analog implemen- ẋ(t) = Aσ(t) x(t) + Bσ(t) u(t) (1)
tations for each converter or estimated parameter. Moreover, y(t) = C x(t) (2)
the algorithm has low computational overhead, which enables
it to be implemented on the same computing platform as where Aσ(t) and Bσ(t) are the collection of linear state space
the control system. Third, the proposed approach requires no models, and σ(t) is the continuous-time switching signal that
additional sensors or computing devices (aside from those indicates the active mode.
already used for closed-loop control purposes), or injection We can build a open-loop switched linear state estimator of
of external signals into the system. the system from (1) as follows:
The remainder of the paper is outlined as follows. Section II
presents the modeling approach and algorithm design of the ż(t) = Aσ(t) z(t) + Bσ(t) u(t) (3)
proposed adaptive parameter identifier. Section III presents a
where z(t) is an estimate of the state vector x(t) from (1) and
simulation of the identifier algorithm for an interleaved boost
Aσ(t) and Bσ(t) are obtained by solving Kirchhoff’s circuit
dc-dc converter. The simulation tests a variety of operating
laws for the ideal circuit topology.
points and parameter perturbations, and provides a simulation
analysis of the robustness of the algorithm to measurement Example 1. Consider the interleaved boost converter topology
noise and converter non-idealities. Section IV describes the in Fig. 1. In the continuous conduction mode of operation,
hardware implementation and experimental testbed. Section V the mode of the system is determined explicitly by the PWM
presents experimental results which validate the real-time applied to SW1 and SW2 . We can construct an open-loop
implementation of the algorithm on a 3 kW dc-dc converter. state estimator of the system as follows:
Section VI concludes the paper. ⎡ ⎤
îL1 (t)    
vin (t) i (t)
II. A DAPTIVE PARAMETER I DENTIFIER D ESIGN z(t) = ⎣îL2 (t)⎦ , u(t) = , y(t) = in
iload (t) vout (t)
In this section, we present the derivation of the proposed v̂C (t)
adaptive parameter identifier algorithm. We first present the
⎡ ⎤ ⎡1 ⎤
modeling approach used to describe the nominal dynamics of a −R
L1
1
0 − Ls11 L1 0
converter. Next, we use this model to analyze the dynamics of Aσ(t) =⎣ 0 −R
L2
2
− Ls22 ⎦ , Bσ(t) = ⎣ L12 0 ⎦
a converter when parameters are perturbed from their nominal s1
C
s2
C 0 0 − C1
values. Finally, we present a gradient descent algorithm that
is used to track perturbed parameter values and reconstruct  
1 1 0
estimates of the actual parameter values. C=
0 0 1
A. Nominal converter model and state estimator The possible values for the switching signal σ(t) are given
Many power electronics systems consist of linear passive in Table II, where sk = 0 indicates switch SWk is open, and
elements (e.g. capacitors, inductors, resistors) and switching sk = 1 indicates switch SWk is closed.
 ⎛ ⎞
B. Converter model and dynamics with unknown parameters −R1 îL1 (t) − s1 v̂C (t) − vin (t)
Wσ(t) (z, u) = diag ⎝−R2 îL2 (t) − s2 v̂C (t) − vin (t)⎠
s1 îL1 (t) + s2 îL2 (t) − iload (t)
Consider the case when Aσ(t) and Bσ(t) are not precisely
known. Let δAσ(t)  Âσ(t) − Aσ(t) where Âσ(t) is an
C. Gradient descent algorithm design and dynamics
estimate of Aσ(t) , and δBσ(t)  B̂σ(t) − Bσ(t) where B̂σ(t)
is an estimate of Bσ(t) . The open-loop state estimator is now The objective of the adaptive parameter identifier is to
given by: perturb the parameter estimate vector θ(t) as to drive the
measured output error γ(t) to zero. In this way, we expect
ż(t) = Âσ(t) z(t) + B̂σ(t) u(t) (4) θ(t) to converge to the actual parameter values θ ∗ .
There are a number of algorithms available for adaptive
The dynamics of the error e(t) = z(t) − x(t) are governed parameter identification [26]. For purposes of simplicity in de-
by: sign and implementation, we propose the use of a generalized
ė(t) = Aσ(t) e(t) + δAσ(t) z(t) + δBσ(t) u(t) (5) gradient descent algorithm [25], [26]. Noting that θ̇(t) = φ̇(t),
the structure of the gradient descent algorithm is as follows:
Let θ ∗ be a vector containing the actual value of the
θ̇(t) = −G HT (t) γ(t) (11)
unknown parameters of interest, and let θ(t) be an estimate
of θ ∗ . In general, a switched linear state space model can be where G  0 is a positive definite matrix gain which is used to
parameterized as follows: control the rate of convergence of the error term γ(t) to zero.
The structure of G is generally of the form diag(1 , ..., n ),
Aσ(t) x(t) + Bσ(t) u(t) = Wσ(t) (x, u)θ ∗ (6)
where n is the number of terms to be identified, and each
where W(x, u) is a time-varying matrix that depends on the  term is parameter update gain for the associated unknown
state and inputs. parameter.
The matrix gain G, in general, influences the rate of
We define φ(t)  θ(t) − θ ∗ as the difference between
convergence and the stability of the gradient descent algorithm.
the estimate and the actual value of the unknown parameters.
The design of G depends on a number of factors (see [25] for
Using this, we can rewrite (5) as:
a mathematical analysis). Among these include:
ė(t) = Aσ(t) e(t) + Wσ(t) (z, u)φ(t) (7) 1) Persistency of excitation of H(t). The gradient descent
algorithm requires that H(t) satisfies a persistency of
We can solve for the time domain evolution of ė(t) as excitation condition [26]. Practically, this implies that
follows: the input u(t), the input voltage and load current in
 t
Examples 1 and 2, are varying with time sufficiently
e(t) = Φ(t0 , t)e(t0 ) + φ Φ(t, τ )Wσ(t) (z, u)dτ (8) such that the dynamics of the system make the parameter
t0
identification possible. This condition is easily satisfied
Note that we assume the term φ(t) evolves slowly (essen- in many practical applications, for instance, in a solar
tially constant) compared to the state dynamics such that they photovoltaic system where both the input voltage and
appear essentially constant; this allows us to bring the term load current depend on time-varying irradiation and
outside of the integral. The corresponding error in the output temperature of the solar panel. Conversely, a purely dc
γ(t)  C (z(t) − x(t)) implies that: steady-state system would not allow the parameters to
 t converge, as there would not be enough information in
γ(t) = C Φ(t0 , t)e(t0 ) + φ Φ(t, τ )Wσ(t) (z, u)dτ the H(t) signal.
t0 2) Number of measurements versus number of unknown
(9) parameters. The number of measurements available
directly influences the excitation of H(t), i.e. more
Finally, the term Φ(t0 , t)e(t0 ) asymptotically decays to
measurements mean higher excitation. Likewise, as the
zero, so the evolution of γ(t) is as follows:
number of unknown parameters increases, the amount
γ(t) = H(t)φ (10) of excitation in H(t) required for the gradient descent
t
algorithm to converge also increases. Thus, it is gen-
where H(t) = C t0
Φ(t, τ )Wσ(t) (z, u)dτ . erally straightforward to achieve convergence when the
number of unknown parameters is less than or equal
Example 2. Again, consider the interleaved boost converter to the number of available measurements, as will be
topology in Fig. 1. The unknown parameters to be estimated shown in Sections III and V. Applying the gradient
are L1 , L2 , and C. The resulting expressions for z(t), θ(t), descent problem to the scenario with more unknown
and Wσ(t) (z, u) are: parameters than available measurements can be an ill-
⎡ ⎤ ⎡ 1 ⎤
posed problem.
îL1 (t) L (t)
⎢ 1 ⎥ 3) Desired convergence rate. Each  is chosen such that
z(t) = ⎣îL2 (t)⎦ , θ(t) = ⎣ L21(t) ⎦ ,
1 the dynamics of θ̇(t) evolve much slower than the state
v̂C (t) C(t) dynamics, in order to satisfy the assumption made in
 TABLE III: Simulation and experiment parameters. 1.5

C(t) [p.u.]
Component parameters 1
R1,2 0.082 Ω C∗
L1,2 (nominal) 5.0 mH
C (nominal) 2.85 mF
0.5
Simulated parasitic parameters (Section III-D)
Resr 38 mΩ ±50% 0
Ron 1 mΩ 0 0.05 0.1 0.15 0.2
Rpar1,2 100 Ω Time [s]
Rpar3 5Ω
Cpar1 1 pF Fig. 2: Simulation of the adaptive parameter identifier tracking a step
Cpar2 1 nF perturbation in capacitance C (t = 0.1s from C to 0.65 C).
Lpar1 5.00 nH
Lpar2 1 nH 1.5
Nominal operating point

L1 (t) [p.u.]
vin (t) 100 V
1
vout (t) 380 V
iload (t) 2.5 A + 1 A, 10 Hz ripple L∗1
Converter switching frequency 10 kHz 0.5
Adaptive parameter identifier gains
1 , 2 (L1 (t), L2 (t)) 1 · 106 0
3 (C(t)) 3 · 106 0 0.1 0.2 0.3 0.4
Embedded computing platform Time [s]
Device Zynq-7000 SoC XC7Z020
Fixed computation time step 500 ns Fig. 3: Simulation of the adaptive parameter identifier tracking a step
Voltage sensor bandwidth 100 kHz perturbation in inductance L1 (t = 0.2s from L1 to 0.5 L1 ).
Current sensor bandwidth 200 kHz

estimated parameter value leaves the acceptable range, a fault

Eq. (8). Moreover, in the case of multiple unknown flag can be raised.
parameters,  can be chosen such that all parameters The output measurements, unknown parameter vector, and
exhibit similar convergence rates so certain parameters gain matrix from Examples 1 and 2 are used in the simulation;
do not converge much faster than others. that is:
Finally, we require an estimate of H(t) as follows:    
 t i (t) 1 1 0
y(t) = in , C=
vout (t) 0 0 1
Ĥ(t) = C Φ̂(t, τ )Wσ(t) (z, u)dτ (12)
t0 ⎡ 1
⎤ ⎡ ⎤
where Φ̂ uses updated parameters from θ̇(t) to provide an L (t) 1 0 0
⎢ 11 ⎥
estimate of the state transition matrix Φ. θ(t) = ⎣ L2 (t) ⎦ , G = ⎣0 2 0⎦
1 0 0 3
The complete adaptive parameter identifier is shown here: C(t)

ż(t) = Âσ(t) z(t) + B̂σ(t) u(t) (13) For simulation and experiments, we estimate either one or
two parameters at a time. As discussed in Section II-C, since
γ(t) = C z(t) − y(t) (14)
only two measurements are made in the simulations, iin (t)
d
Ĥ(t) = Âσ(t) Ĥ(t) + W(z, u) (15) and vout (t), H(t) would not have enough excitation and θ(t)
dt would not converge if there are more than two unknown
θ̇(t) = −G HT (t) γ(t) (16) parameters. For these scenarios, the i associated with the ith
desired parameter(s) to be estimated are set to the value shown
III. S IMULATION AND A NALYSIS in Table III, and the j th parameter(s) not being estimated has
j = 0.
In this section, a variety of MATLAB simulations are used
Moreover, in all simulation cases, a 1 A peak-to-peak ripple
in order to validate the performance and robustness of the
at 10 Hz is added to the 2 A load current iload (t). In addition to
proposed adaptive parameter identifier for a number of test
emulating time-varying load dynamics, this ripple introduces
cases and operating conditions. The interleaved boost dc-dc
the persistency of excitation to H(t) that is necessary for
converter of Fig. 1 is used as the device-under-test. Table III
the convergence of the error term γ(t) to zero. A similar
presents the converter parameters and operating point for the
perturabtion could be added to the input voltage vout (t) that
nominal simulation cases.
would emulate the time-varying photovolatic irradiation.
In order to implement the adaptive parameter identifier in a
simulation environment, Eqs. (13), (14), (15), and (16) are dis-
cretized using the Euler method, and executed in a simulation A. Step perturbation in single parameter
loop with a fixed 500 ns time step. We inject perturbations in One application of the proposed adaptive parameter iden-
the output capacitor and a series input inductor of the plant tifier is the tracking of ‘hard’ faults in passive components,
converter using additional shunt and series elements. For the that is drastic changes in the parameter value indicative of
purposes of fault monitoring, one can define a lower and upper a catastrophic component failure. To simulate these types of
bound of the acceptable range of parameter value. When the faults, we introduced a step perturbation in the parameter
 1.5 1.5

C(t) [p.u.]

L(t) [p.u.]
1 1
C∗ L∗1
0.5 0.5

0 0
0 0.1 0.2 0.3 0.4 0.5 0 1 2 3 4
Time [s] Time [s]
Fig. 4: Simulation of the adaptive parameter identifier tracking a ramp 1.5
perturbation in capacitance C (−166.7 μF/s from C to 0.5 C).

C(t) [p.u.]
1
1.5 C∗
L1 (t) [p.u.]

0.5
1
L∗1 0
0.5 0 1 2 3 4
Time [s]
0
0 0.1 0.2 0.3 0.4 0.5 Fig. 6: Simulation of adaptive parameter identifier tracking simultaneous ramp
Time [s] perturbations in inductance L1 (−1.9 mH/s at t = 1.75 s) and capacitance
(−35.7 μF/s from C to 0.5 C at t = 0.2 s).
Fig. 5: Simulation of the adaptive parameter identifier tracking a ramp
perturbation in inductance L1 (−16.7 mH/s at t = 0.1s from L1 to 0.5 L1 ).
inductor L1 . The rate of change of these parameter values
was set at −1.9 mH/s and −35.7 μF/s, respectively. As shown
value of passive components in the plant. As shown in Fig. 2,
in Fig. 4, the adaptive parameter identifier tracks the correct
when a 35% step perturbation introduced in the capacitance
value of this perturbation with unnoticeable error. Slower
C, the estimate C(t) converges to the correct value in around
perturbations are tracked at equal or better accuracy.
50 ms. Similarly, when a 50% step perturbation introduced in
the series inductance L1 , the estimate L1 (t) converges to the
correct value in around 100 ms. D. Effect of measurement noise and circuit non-idealities
Simulation was also used to test the robustness of the
B. Ramp perturbation in single parameter
proposed adaptive parameter identifier to measurement noise
Another application of the proposed adaptive parameter and parameter and model non-idealities. The modified circuit
identifier is the tracking of slow varying perturbations in the model in Fig. 7a is used to test this scenario. Note that
parameter values of passive components. These types of slow the adaptive parameter identifier still uses the ‘ideal’ circuit
varying changes occur on the time scale of hours to years, model in Fig. 1, while the modified circuit model provides
and represent a type of ‘soft’ fault, that is, gradual changes the measurement vector y(t). As shown, the modified circuit
indicative of lifetime wear or aging. model incorporates switch and passive component parasitics,
We simulated ramp perturbations in the model to emulate including the ESR of the output capacitance Resr,C . The
these types of faults. First, a ramp perturbation of −166.7 μF/s steady state dynamics are shown in Fig. 7b. As shown,
is introduced in the output capacitor C. This is a relatively when compared with the steady state dynamics of the ideal
fast perturbation meant to identify the upper limit of accuracy model, the parasitic model introduces high frequency current
of the adaptive parameter identifier. Slower perturbations are transients of approximately 4 Apk−pk in the series induc-
tracked at equal or better accuracy. As shown in Fig. 4, the tor currents iL (t), and voltage transients of approximately
adaptive parameter identifier tracks the correct value of this 20 Vpk−pk in the output capacitor voltage vC (t).
perturbation with unnoticeable error. Similarly, we introduced We study the impact of the parameter and model non-
a ramp perturbation of −16.7 mH/s in a series input inductor idealities and parasitics on the convergence of the adaptive
L1 . As shown in Fig. 5, the adaptive parameter identifier again parameter identifier. The parameter step perturbation from
tracks the correct value of this perturbation. Section III-A is used as a baseline to compare the ideal and
parasitic models. As shown in Fig. 8, the parasitic model
C. Simultaneous perturbation in multiple parameters causes minimal difference in the dynamics of the adaptive
In general, more than one unknown parameter will be of parameter identifier for parameter step changes in both the
interest in the system. From the analysis in Section II, we output capacitor C and a series inductor L1 . This provides
expect that the convergence rate of θ(t) will be slower than evidence that the algorithm is robust to the non-ideal dynamics
if only one parameter is being estimated. This is due to the introduced by the parasitics.
corresponding increase in the error e(t) and decrease in the A central reason for this is the design of the adaptive param-
excitation in H(t). eter identifier. In (8), we assumed that the φ(t) evolves slowly
To test this scenario, we simulated a simultaneous ramp (essentially constant) compared to the state dynamics. Dynam-
perturbation in both the output capacitor C and a series input ics introduced by parasitics or measurement noise would occur
 IV. H ARDWARE I MPLEMENTATION AND E XPERIMENTAL
T ESTBED
In this section, we present the hardware implementation
Rp,L1 Cp,L1 and experimental testbed for the proposed fault prognosis
algorithm. The complete testbed is shown in Fig. 9. The com-
Cp,S2 Cp,S4 plete specifications for the experimental testbed and computing
Rp,L2 Cp,L2 Rp,S2 Rp,S4 Resr,C platform are given in Table III.
Ron,S2 Ron,S4
Ls,C Rs,C
Ls,S2 Ls,S4
A. Real-time digital implementation
The fault prognosis algorithm is implemented in real-time
(a) Circuit topology of the parasitic simulation model.
on the same computing device as the control system, in this
case, a ZedBoard system-on-chip (SoC) device, a relatively
30
26
low-cost development board for the Xilinx Zynq-7000 SoC
iL1 (t) [A]

20 24
XC7Z020. A central feature of this SoC is the integration
Ideal model
22 programmable logic and a dual-core ARM Cortex-A9 proces-
10 Parasitic model
20 sor. It is worth noting that our implementation only utilized
0 18 11 percent of the programmable logic slices of the device.
0.09 0.1 0.11
Time [s] Moreover, the ARM core was not used in the design. Thus, the
400
algorithm could be implemented on a much lower cost FPGA.
400 380
In general, however, a programmable logic device is preferred
vC (t) [V]

360
for implementation (as opposed to a standard microcontroller)
Ideal model
200 Parasitic model 340
due to the tight constraints on latency and computation time
320
imposed by the fault prognosis algorithm.
0 The adaptive parameter identifier in Eqs. (13), (14), (15),
0.09 0.1 0.11
Time [s]
and (16) is discretized and solved in real-time with a fixed
time step (including input/output latency) of 500 ns. This time
(b) Steady state dynamics of the ideal (from Fig. 1) and parasitic step, which is an order of magnitude faster than the switching
simulation model.
frequency of the converter (10 kHz), is necessary to adequately
Fig. 7: A simulation of an interleaved boost dc-dc converter incorporating model the switching dynamics of the converter.
switch and passive component parasitics and non-idealities is used to validate Moreover, a highlight of the approach and implementation is
the robustness of the proposed adaptive parameter identifier.
that it maintains the flexibility to be reconfigured for different
converters simply by changing the contents of the Âσ(t) ,
1.5 B̂σ(t) , and W(z, u) matrices. The selection of the θ(t) vector
Ideal model
C(t) [p.u.]

Parasitic model
indicates the parameters of interest to be estimated.
1
C∗
0.5 B. Experimental testbed

0
We experimentally validate the proposed fault prognosis
0 0.05 0.1 0.15 0.2 method on a 3 kW interleaved dc-dc converter as shown
Time [s] in Fig. 9. The input of the converter is connected to a
(a) Step perturbation at t = 0.1s from C to 0.5 C. programmable photovoltaic (PV) emulator, and the output is
connected to a programmable dc load. Prior to running the
1.5
experiments, the values of passive components are measured
L(t) [p.u.]

1
Ideal model using an LCR meter. The setup is equipped such that the value
Parasitic model
L∗1 of the output capacitor C and a series inductor L1 can be
0.5 changed in real-time by controlling solid-state switches that
introduce series or parallel elements into the circuit.
0
0 0.1 0.2 0.3 0.4
Time [s] V. E XPERIMENTAL R ESULTS
(b) Step perturbation at t = 0.2s from L1 to 0.5 L1 . This section presents experimental results which validate the
proposed fault prognosis algorithm under a variety of operating
Fig. 8: Comparison of the adaptive parameter identifier convergence when conditions.
applied to the ideal model (Fig. 1) versus parasitic model (Fig. 7a).
We test step changes in the value of the output capacitor C
and a series inductor L1 . Similar to the simulations performed
in Section III-A, the hard fault instantaneously changes the
on a time scale faster than that of the state dynamics, and value of the capacitor and the inductor, and we monitor the
would be ignored by the parameter identification. convergence of the adaptive parameter identifier to the new
 3 kW dc-dc interleaved
Control and boost converter
monitoring PC

Programmable dc load

PV emulator
Zynq-7000 SoC Board

Fig. 9: Photograph of the experimental testbed, including the interleaved dc-dc converter and the Zynq-7000 SoC board.

1.5 f = 2 Hz
1.2 load

f = 5 Hz
C(t) [p.u.]

load

1 f = 10 Hz
C∗
load

f = 15 Hz

C(t) [p.u.]
1 load

0.5

0 0.8 C∗
0 0.05 0.1 0.15 0.2
Time [s]
0.6
Fig. 10: Experiment of adaptive parameter identifier tracking a step perturba-
tion in capacitance C. Step perturbation at t = 0.1 s from C to 0.65 C. 0 2 4 6 8
Time [s]
1.5
Fig. 12: Experiment of adaptive parameter identifier tracking step perturba-
L(t) [p.u.]

tions in capacitance C for a set of iload (t) frequencies: 2 Hz, 5 Hz, 10 Hz,
1
and 15 Hz. Step perturbation at t = 1 s from C to 0.65 C.
L∗
0.5

0 B. Effect of dynamic operating conditions

0 0.1 0.2 0.3 0.4
Next, we varied the frequency of the load current iload (t)
Time [s]
and observed its effect on the convergence of θ(t). We expect
Fig. 11: Experiment of adaptive parameter identifier tracking a step perturba- that a lower iload (t) frequency will result in less excitation
tion in inductance L1 . Step perturbation at t = 0.2 s from L to 0.5 L. in H(t). Thus, with a constant matrix gain G, the rate
of convergence of θ(t) will be slower. For higher iload (t)
frequencies, we expect the converse to be true.
value. Moreover, we test the effect of changing the dynamics We inject a step change in the output capacitance C from
of the load current iload (t) on the speed of convergence of 2.85 mF to 1.85 mF for a set of iload (t) frequencies: 2 Hz,
θ(t). 5 Hz, 10 Hz, and 15 Hz. Fig. 12 shows the results of this test.
For the following experiments, the same output measure- As shown, when the load is perturbed at 15 Hz, convergence
ments, unknown parameter vector, and gain matrix are used as to the new value of C takes approximately 50 ms. At slower
in Section III. Relevant parameter values and operating points perturbation frequencies (e.g. 2 Hz), convergence to the new
can be found in Table III. value of C can take as long as 3 s. These results match
our expectations, and also provide an empirical method for
determining the rate of convergence of θ(t) given a certain
input vector u(t).
A. Step perturbation in C and L1

Fig. 10 shows a fault that causes the output capacitance C VI. C ONCLUSION
to decrease from 2.85 mF to 1.85 mF, a decrease of 35%. This paper presents a method for fault prognosis using
As shown, the adaptive parameter identifier converges to the an adaptive parameter identification approach. The salient
new value of C in approximately 50 ms. Similarly, Fig. 11 advantages of the approach include application flexibility and
shows a fault that causes a series inductor L1 to decrease by the obviated need for additional sensors, computing devices,
50%. Convergence to the new value of L1 takes approximately or injected signals. The proposed technique can be applied
100 ms. These values match the expected time of convergence in domains including motor drive applications (for stator or
from simulations in Section III-A. rotor resistance estimation), cable integrity monitoring, or
other applications where parameter or system identification is [13] A. Wechsler, B. C. Mecrow, D. J. Atkinson, J. W. Bennett, and
a useful technique for fault prognosis or prediction. M. Benarous, “Condition monitoring of dc-link capacitors in aerospace
drives,” IEEE Transactions on Industry Applications, vol. 48, no. 6,
pp. 1866–1874, 2012.
R EFERENCES [14] X. S. Pu, T. H. Nguyen, D. C. Lee, K. B. Lee, and J. M. Kim, “Fault
diagnosis of dc-link capacitors in three-phase ac/dc PWM converters by
[1] J. Poon, P. Jain, I. Konstantakopoulos, C. Spanos, S. Panda, and online estimation of equivalent series resistance,” IEEE Transactions on
S. Sanders, “Model-based fault detection and identification for switching Industrial Electronics, vol. 60, no. 9, pp. 4118–4127, 2013.
power converters,” IEEE Transactions on Power Electronics, vol. 8993,
[15] V. Valdivia, A. Barrado, L. A, zaro, P. Zumel, C. Raga, C. Fern, and
no. c, pp. 1–1, 2016.
ndez, “Simple modeling and identification procedures for black-box
[2] X. Ding, J. Poon, I. Celanovic, and A. D. Dominguez-Garcia, “Fault
behavioral modeling of power converters based on transient response
detection and isolation filters for three-phase ac-dc power electronics
analysis,” IEEE Transactions on Power Electronics, vol. 24, no. 12,
systems,” IEEE Transactions on Circuits and Systems I: Regular Papers,
pp. 2776–2790, 2009.
vol. 60, pp. 1038–1051, apr 2013.
[3] M. Kaufhold, H. Aninger, M. Berth, J. Speck, and M. Eberhardt, [16] R. Leyva, L. Martinez-Salamero, B. Jammes, J. C. Marpinard, and
“Electrical stress and failure mechanism of the winding insulation in F. Guinjoan, “Identification and control of power converters by means
PWM-inverter-fed low-voltage induction motors,” IEEE Transactions on of neural networks,” IEEE Transactions on Circuits and Systems I:
Industrial Electronics, vol. 47, no. 2, pp. 396–402, 2000. Fundamental Theory and Applications, vol. 44, no. 8, pp. 735–742,
[4] D. Shaw, S. Cichanowski, and A. Yializis, “A changing capacitor 1997.
technology-failure mechanisms and design innovations,” IEEE Trans- [17] V. Valdivia, A. Lazaro, A. Barrado, P. Zumel, C. Fernandez, and
actions on Electrical Insulation, no. 5, pp. 399–413, 1981. M. Sanz, “Black-box modeling of three-phase voltage source inverters
[5] V. Smet, F. Forest, J.-J. Huselstein, F. Richardeau, Z. Khatir, S. Lefebvre, for system-level analysis,” IEEE Transactions on Industrial Electronics,
and M. Berkani, “Ageing and failure modes of IGBT modules in high- vol. 59, no. 9, pp. 3648–3662, 2012.
temperature power cycling,” IEEE Transactions on Industrial Electron- [18] B. Karanayil, M. F. Rahman, and C. Grantham, “Online stator and rotor
ics, vol. 58, no. 10, pp. 4931–4941, 2011. resistance estimation scheme using artificial neural networks for vector
[6] H. Wang, M. Liserre, F. Blaabjerg, P. de Place Rimmen, J. B. Jacobsen, controlled speed sensorless induction motor drive,” IEEE Transactions
T. Kvisgaard, and J. Landkildehus, “Transitioning to physics-of-failure on Industrial Electronics, vol. 54, no. 1, pp. 167–176, 2007.
as a reliability driver in power electronics,” IEEE Journal of Emerging [19] D. J. Atkinson, P. P. Acarnley, and J. W. Finch, “Observers for induction
and Selected Topics in Power Electronics, vol. 2, no. 1, pp. 97–114, motor state and parameter estimation,” IEEE Transactions on Industry
2014. Applications, vol. 27, no. 6, pp. 1119–1127, 1991.
[7] S. Yang, D. Xiang, A. Bryant, P. Mawby, L. Ran, and P. Tavner, “Con- [20] O. Jimenez, O. Lucia, I. Urriza, L. A. Barragan, and D. Navarro,
dition monitoring for device reliability in power electronic converters: “Analysis and implementation of FPGA-based online parametric iden-
a review,” IEEE Transactions on Power Electronics, vol. 25, no. 11, tification algorithms for resonant power converters,” IEEE Transactions
pp. 2734–2752, 2010. on Industrial Informatics, vol. 10, no. 2, pp. 1144–1153, 2014.
[8] A. Tanwani, A. D. Domnguez-Garca, and D. Liberzon, “An inversion- [21] M. Hao, M. Xingyun, Z. Ni, and X. Dehong, “Parameter identification
based approach to fault detection and isolation in switching electrical of power electronic circuits based on hybrid model,” in 2005 IEEE 36th
networks,” IEEE Transactions on Control Systems Technology, vol. 19, Power Electronics Specialists Conference, pp. 2855–2860.
no. 5, pp. 1059–1074, 2011.
[22] L. A. Kamas and S. R. Sanders, “Parameter and state estimation in
[9] A. Lahyani, P. Venet, G. Grellet, and P. J. Viverge, “Failure prediction of
power electronic circuits,” IEEE Transactions on Circuits and Systems
electrolytic capacitors during operation of a switchmode power supply,”
I: Fundamental Theory and Applications, vol. 40, no. 12, pp. 920–928,
IEEE Transactions on Power Electronics, vol. 13, no. 6, pp. 1199–1207,
1993.
1998.
[10] K. Abdennadher, P. Venet, G. Rojat, J. M. Retif, and C. Rosset, “A real- [23] A. Kawamura, T. Haneyoshi, and R. G. Hoft, “Deadbeat controlled
time predictive-maintenance system of aluminum electrolytic capacitors PWM inverter with parameter estimation using only voltage sensor,”
used in uninterrupted power supplies,” IEEE Transactions on Industry IEEE Transactions on Power Electronics, vol. 3, no. 2, pp. 118–125,
Applications, vol. 46, no. 4, pp. 1644–1652, 2010. 1988.
[11] G. M. Buiatti, J. A. Martin-Ramos, C. H. R. Garcia, A. M. R. Amaral, [24] G. M. Buiatti, A. M. R. Amaral, and A. J. M. Cardoso, “An online
and A. J. M. Cardoso, “An online and noninvasive technique for the con- technique for estimating the parameters of passive components in non-
dition monitoring of capacitors in boost converters,” IEEE Transactions isolated dc/dc converters,” in 2007 IEEE International Symposium on
on Instrumentation and Measurement, vol. 59, no. 8, pp. 2134–2143, Industrial Electronics, pp. 606–610.
2010. [25] L. Kamas and S. Sanders, “Parameter and State Estimation in Power
[12] K. W. Lee, M. Kim, J. Yoon, S. B. Lee, and J. Y. Yoo, “Condition Electronics Circuits,” IEEE Transactions on Circuits and Systems,
monitoring of dc-link electrolytic capacitors in adjustable-speed drives,” vol. 40, no. 12, pp. 920–928, 1993.
IEEE Transactions on Industry Applications, vol. 44, no. 5, pp. 1606– [26] S. S. Sastry, Nonlinear Systems. New York: Springer-Verlag, 10 ed.,
1613, 2008. 1999.