Professional Documents
Culture Documents
Abstract—This paper presents the design, analysis, and exper- Few works have focused on a generalized model-based
imental validation of a fault detection and identification (FDI) methodologies for fault prognosis and diagnosis which are
scheme for dc-dc power electronic converters. The FDI scheme converter topology- and fault-agnostic [4], [10], [11]. How-
includes two new classes of fault filters: (1) a linear-switched
fault detection (FD) filter and (2) a bank of linear-switched fault ever, the FDI schemes based on parity space methods in [4],
identification (FI) filters. Both the FD filter and the bank of FI [10] lack design for robustness, and are limited in identifying
filters have a structure similar to that of Luenberger observer. faults (for instance, the fault signatures used to identify dif-
The FD filter detects a fault event and the bank of FI filters ferent faults are indistinguishable). Similarly, the FDI scheme
identify a faulted converter component, for instance, failure in based on parameter estimation in [11] is promising for passive
switching or passive components. We present simulation and
experimental results for a prototype 2 kW solar photovoltaic component degradation, however it does not account for
(PV) boost dc-dc converter system to demonstrate the efficacy of failures in switching components.
the proposed FDI scheme. The experimental results demonstrate In this paper, we present a robust model-based FDI scheme
accurate fault detection and identification for a collection of based on linear-switched Luenberger observer for dc-dc power
catastrophic component faults in PV dc-dc power converters. electronic converters. The proposed scheme includes a FD
Index Terms—Fault detection, fault diagnosis, fault location,
converters, state estimation, photovoltaic (PV) systems filter and a bank of FI filters. The gain matrix for the FD
filter is appropriately designed such that its residual function
converges to near-zero in absence of faults and is robust
I. I NTRODUCTION
in the presence of noise and other non-idealities. In a fault
The dc power electronic systems are increasingly becoming event, the FD filter residual function exceeds a pre-defined
complex and their applications in mission-critical- and safety- threshold and the bank of FI filters are actuated to determine
systems (e.g. avionics, space-transport, electric vehicles), in- the faulted component. The bank includes a number of FI
formation and communication systems (e.g. data centres), filters designed for each converter component, such that the
renewable energy technologies (e.g. photovoltaics systems), residual function for FI filter which represents the faulted
zero-energy buidings and intergrid [1] require a high level of component is minimum of all residual functions generated by
dependability and reliabilty [2], [3]. For instance, converter other FI filters which represent healthy components.
component faults may degrade system performance and en- Both the proposed FDI scheme, and the converter control
danger safety and security of system operation. Thus, it is system are digitally implemented on a single all-programmable
necessary, in many cases, to quickly detect and accurately system-on-chip (SoC) device. In order to illustrate the perfor-
identify the location of any fault in these systems. This is mance of the proposed scheme, We present a simulation and
the purpose of fault detection and identification (FDI) scheme experimental validation for a 2 kW dc-dc boost converter used
whose design becomes a mandatory complement to the system in solar PV applications.
design process. In general, FDI schemes fall into two broad The remainder of the paper is organized as follows. Sec-
categories: model-based or model-free [4]. tion II presents the design of the proposed FDI scheme.
Model-based FDI schemes leverage analytical redundancy Section III presents the simulation study. The experimental
which includes signal processing techniques, state estimators implementation and results that validate the proposed scheme
(e.g. Luenberger observers and Kalman filters), parity space are presented in Section IV. Section V concludes the paper.
methods, and parameter estimation [5]. These FDI schemes
have been investigated in the literature before [4], [6]–[11] II. A NATOMY OF P ROPOSED FDI S CHEME
for dc-dc converters. However, most of the earlier works In this section, we describe the modeling framework for the
are fault-specific, for instance in [6], [7] only switch faults dc-dc power electronic converters under nominal and faulted
are investigated using signal processing techniques for dc-dc operating conditions, design methodology for the FD and
converters. the FI filters, and develop the proposed FDI scheme for a
l-)))
A power electronics system converter captured by the state vector in (1) through additive
Power stage: dc-dc deviations ΔA(t) and ΔB(t) in the nominal Aσ(t) and Bσ(t) ,
solar PV system
iL(t) Inductor current
respectively. Hence, the generalized component fault model is
iload(t) Load current
vpv(t) PV voltage iload(t) vc(t) Capacitor voltage as follows:
SW1
ipv(t) PV current R L +
iL(t)
vC(t)
C
ẋ(t) = (Aσ(t) + ΔA(t))x(t) + (Bσ(t) + ΔB(t))u(t) (5)
-
vpv (t)
Load ẋ(t) = Aσ(t) x(t) + Bσ(t) u(t) + Hfi (t) (6)
SW2 current
PV module
where H is an identity matrix; i = 1, ..., I, I is the number of
Half-bridge leg
converter component parameters in the linear-switched state
u(t)
Switching signals (σ (t)) y(t) space model; and fi (t) is the component fault vector which
Computing Platform represents fault severity and a unique direction in fault vector
MPPT controller Sensing and
space. Table II represents the components, their parameters
communication and the component fault vectors for common failure modes in
Fault Diagnosis Scheme devices the dc-dc boost converter.
Fault Indicator
σ (t) σ (t)
A bank of fault
Solar PV u(t) identification (FI)
dc-dc converter
source
Fault y(t) filters r^θ1(t) Fault
detection FI filter 1 identification FI flag
u(t) y(t)
logic r^θ(t) logic
FI filter 2 2 J(r(t))
^θfault
Fault detection (FD) filter rθ^(t)
r(t) FD flag FI filter 3 3
TABLE II: Components, their parameters, their common failure modes, and fault vectors (i.e. fi (t)) for a dc-dc boost converter. A general fault vector f (t)
captures the fault dynamics both in magnitude and direction in the fault vector space. Note that f (t) ∈ F where F is the set of all the possible fault vectors
in a power electronic converter. For instance, fi (t) here represents component fault vectors.
for a FI filter design. This follows from the fact that if power from the fact that Aσ(t) , Bσ(t) and θi are not precisely known
electronic converter is stable during its operation in a fault during a fault in component parameter θi .
event, the parameter θi will always lie in the certain range θi The FI filter gain Lθ̂i is chosen such that (14) under arbitrary
¯
and θ̄i where θi represents lower fault bound and θ̄i represents switching is asymptotically stable which is determined by
¯
upper fault bound, that is: a common Lyapunov function Vθ̂i (t, rθ̂i ) = rTθ̂ (t)Pθ̂i rθ̂i (t)
i
θi < θi < θ̄i , ∀ θi > 0, θ̄i < ∞ (10) where Vθ̂i ≥ 0, Pθ̂i > 0 and V̇θ̂i ≤ 0.
¯ ¯
The linear-switched FI filter for the converter component The estimated parameter θ̂i is determined by using (10) and
parameter θi is given by following set of equations: (13) where binary function γ(w) is defined as [18]:
żθ̂i (t) = Â(σ(t),θ̂i ) zθ̂i (t) + B̂(σ(t),θ̂i ) u(t) + Lθ̂i rθ̂i (11) 1, if w ≥ 0
γ(w) = (16)
0, if w < 0
rθ̂i = y(t) − Czθ̂i (t) (12)
θ̂i = θi + (θ̄i − θi )γ(w) (13) where w = rTθ̂ (t)Pθ̂i Ψ(zθ̂i (t), u(t)).
i
¯ ¯
where θ̂i is the estimated parameter θi in the fault event; zθ̂i (t)
C. FDI scheme for solar PV power electronic converter
is the state estimation vector for θ̂i ; Â(σ(t),θ̂i ) and B̂(σ(t),θ̂i )
are estimates of Aσ(t) and Bσ(t) for θ̂i respectively; Lθ̂i is We develop and design the proposed FDI scheme for the
solar PV dc-dc boost converter shown in the Fig. 1. The
the FI filter gain for θ̂i ; rθ̂i is the corresponding filter residual
parameters of the dc-dc boost converter for which FD and
vector; and γ(w) is a binary function which takes value either
FI filters are designed are given in Table III. An overview of
one or zero to estimate θi (see (16)).
the proposed FDI scheme is shown in Fig. 2. A FD filter for
The dynamics of FI filter residual rθ̂i for the component the proposed system can be described using (7), and (8) where
parameter θi is given by: FD filter gain L is designed using (9).
Fault detection logic: A norm-based residual evaluation
ṙθ̂i = (Aσ(t) − Lθ̂i C)rθ̂i + Ψ(zθ̂i (t), u(t))δθi (14)
function J(r(t)) is generated to make a binary decision on a
Ψ(zθ̂i (t), u(t))δθi = δA(σ(t),δθi ) żθ̂i (t) + δB(σ(t),δθi ) u(t) fault event. The FD filter residual evaluation function is given
(15) below [18]:
t
where δθi = θi − θˆi , δA(σ(t),δθi ) = Aσ(t) − Â(σ(t),θ̂i ) , and
J(r(t)) = c1 r(t)2 + c2 e−(t−τ ) T r(τ )2 dτ (17)
δB(σ(t),δθi ) = Bσ(t) − B̂(σ(t),θ̂i ) respectively. These follow 0
0.5 TABLE III: Simulation and experimental parameters
Fault injected
J(r(t))
Γ PV emulator ratings 150 V, 40 A, 2 kW
0
Programmable ac/dc
−0.05 0 0.05 0.1 0.15 0.2 electronic load 300 V, 36 A, 3.3 kW
Time [s]
Boost converter parameters 400 V, 40 A, IGBT devices
1 Fault detected L 5 mH ±20%
F Dflag (t)
R 1 Ω ± 20%
0.5 C 2.85 mF ±20%
0 Operating point
−0.05 0 0.05 0.1 0.15 0.2 vpv (t) 120 V
Time [s]
vC (t) 240 V
(a) Residual evaluation function J(r(t)) for FD filter (top) iload (t) 2A
and Fault detection logic (bottom). Sensing and computing platform
3
Model Xilinx ZynqTM-7000(AP SoC)
Parameter, i
i = 1, C Residual
2 i = 2, L Residual Fault identification begins only after a fault is detected.
i = 3, SW Residual
0
Since fault identification is a classification problem, a template
−0.05 0 0.05 0.1 0.15 0.2
Time [s] matching algorithm is used. A bank of fault identification
(b) Fault identification logic (top) and FI filter residual evalu-
filters (which is analogous to a library of fault templates)
ation functions J(rθi (t)) where i ∈ {1, 2, 3} (bottom). is designed using (14) and (15) for distinguishing between
various component faults fi , as shown in Table II. Each FI
Fig. 3: Simulation results for step decrease in capacitance C by ΔC at t = 0.
filter can be described using (11), (12), (13), and (16). We
0.5
design three FI filters for the proposed system which correlates
Fault injected to faults in inductor, capacitor and half bridge leg respectively.
J(r(t))
2 − (t−τ )
0.5
J(rθ̂i (t)) = cθ̂i 1 rθ̂i (t) + cθ̂i 2 e θ̂i rθ̂i (τ )2 dτ
0
0 (18)
−0.05 0 0.05 0.1 0.15 0.2
Time [s]
where cθ̂i 1 > 0 and cθ̂i 2 > 0 are the weighted constants for
(a) Residual evaluation function J(r(t)) for FD filter (top)
and Fault detection logic (bottom).
the instantaneous and past values of L2 norm of filter residual
3 vector, respectively; and θ̂i > 0 is a forgetting factor.
Parameter, i
Fault identified, i = 2
2 Next, the moving window average of each FI filter residual
1 evaluation function is computed. Lastly, the averaged residual
0
−0.05 0 0.05 0.1 0.15 0.2 evaluation function of all the FI filters are compared with each
Time [s]
other to find the minimum of all in the bank, that is:
2 F Ifilter : t t
J(rθi (t))
i = 1, C Residual
1 i = 2, L Residual θf ault = min( J(rθ̂1 (τ ))dτ, J(rθ̂2 (τ ))dτ,
i = 3, SW Residual
t−W t−W
0
−0.05 0 0.05 0.1 0.15 0.2
t
Time [s]
J(rθ̂3 (τ ))dτ ) (19)
(b) Fault identification logic (top) and FI filter residual evalu- t−W
ation functions J(rθi (t)) where i ∈ {1, 2, 3}(bottom).
where W is the window size for moving average. Note that
Fig. 4: Simulation results for step decrease in inductance L by ΔL at t = 0. the residual evaluation function of ith FI filter correlates to
a potential fault in ith converter component. Thus, whenever
the fault occurs in the corresponding component, the averaged
where c1 > 0 and c2 > 0 are the weighted constants for the residual evaluation functions of that FI filter converges to a
instantaneous and past values of L2 norm of filter residual minimum value identifying the faulted component.
vector, respectively; and > 0 is a forgetting factor.
III. S IMULATION STUDY
Fault detection is achieved by monitoring the residual
evaluation function J(r(t)) at each time step and comparing Simulation results are presented for detecting and identify-
it with a empirically calculated fault detection threshold Γ as ing three different kinds of converter component faults which
follows: J(r(t)) > Γ. include step decrease in capacitance C by 20% (i.e. Fig. 3),
0.5
Fault injected fault which is detected in 34 μs. After fault detection, the
J(r(t))
Γ fault identification logic is activated and the FI filter residual
0 evaluation functions are computed, as shown in Fig. 4b. Since
−0.05 0 0.05 0.1 0.15 0.2
Time [s] J(rθ2 (t)) is minimum of all residual evaluation functions (i.e.
1 Fault detected J(rθ1 (t)) and J(rθ3 (t))), fault identification logic correctly
F Dflag (t)
500 i = 1, C Residual
i = 2, L Residual The proposed FDI scheme is implemented in FPGA fabric of
i = 3, SW Residual
0 a single, low-cost all programmable SoC device and experi-
−0.05 0 0.05 0.1 0.15 0.2
Time [s] mentally validated on a prototype solar PV power electronic
(b) Fault identification logic (top) and FI filter residual evalu- system. The complete specifications of the digital platform and
ation functions J(rθi (t)) where i ∈ {1, 2, 3} (bottom). the prototype testbed setup are presented in Table III.
Fig. 5: Simulation results for an open switch fault in SW2 at t = 0. A. Digital implementation of the proposed scheme
The proposed FDI scheme is implemented on an all pro-
Output grammable SoC, a ZedBoard XC7Z020clg484 ZYNQ device.
Input filters filter
This device combines a dual core ARM Cortex-A9 processor
Fault diagnosis terminal
Programmable dc with a field-programmable gate array (FPGA). It enables
load
design flexibility, low latency, and high computation speed,
which are necessary attributes for performing FDI and control
on power electronic systems. Fault detection and identification
FDI and control on
Xilinx Zynq-7000 filters are discretized and executed with a fixed 500 ns time
dc-dc boost SoC device
PV emulator converter step on FPGA. Outputs from the PV plant (i.e. Vpv and Ipv )
and the switching power converter (i.e y(t)) are sampled
through ADCs using 100 kHz bandwidth sensors for control
Fig. 6: Validation of the proposed FDI scheme on a prototype solar PV power and fault diagnosis purposes.
electronic system.
B. Solar PV dc-dc boost converter prototype
The programmed SoC device is experimentally tested on
step decrease in inductance L by 20% (i.e. Fig. 4) and open a prototype solar PV power electronic system. The complete
switch fault in SW2 (i.e. Fig. 5). experimental testbed setup is shown in Fig. 6. It consists of
Fig. 3a presents residual evaluation function J(r(t)) for FD (1) a CHROMA 62150H-600S PV emulator, (2) a dc-dc boost
filter, which is constantly compared against threshold Γ to converter, and (3) a programmable dc load. Table III presents
detect a fault event. Capacitance fault is injected at t = 0 additional details of the experimental setup.
which is detected in 54.7 ms, as shown in fault detection plot.
Fig. 3b presents the fault identification logic and the residual C. Experimental results
evaluation functions for all three FI filters which are computed Experimental results are presented for the proposed FDI
after the fault is detected. As shown, J(rθ1 (t)) converges scheme which detects and identifies three different kinds of
to a minimum value while residual evaluation functions for component faults in the dc-dc boost converter (see Fig. 7).
inductor (J(rθ2 (t))) and half-bridge leg (J(rθ3 (t))) struggle These three faults are: (1) step decrease in capacitance C by
to achieve a stable minimum value. To find the minimum 66%; (2) step decrease in inductance L by 66%; and (3) open
residual evaluation function which indicates the faulty element, switch fault in IGBT.
fault identification logic computes the moving window average Fig. 7a presents fault detection (FD) flag and residual
of the residual evaluation functions which are compared with evaluation functions J(rθi (t)), ∀ i ∈ {1, 2, 3} where rθi (t) is
each other. The capacitance fault is identified within 54.7 ms. generated by ith FI filter in the bank, for capacitance fault. As
Similarly, Fig. 4a presents FD filter residual evaluation shown in fault-free condition, both the FD flag and the FI filter
function J(r(t)) and fault detection logic for inductance residual evaluation functions are zero. When a step decrease in
FD flag
FD flag FD flag
Fault detected
Fault detected Fault detected
FI filter residuals
FI filter residuals FI filter residuals ‘C’ residual
‘C’ residual ‘C’ residual ‘L’ residual
‘L’ residual ‘L’ residual ‘SW’ residual
‘SW’ residual ‘SW’ residual
(a) Step decrease in capacitance C by ΔC. (b) Step decrease in inductance L by ΔL. (c) Open switch fault in IGBT SW2 .