IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 33, NO.
1, JANUARY 2018
Letters
Anti-EMI Noise Digital Filter Design for a 60-kW Five-Level SiC Inverter
Without Fiber Isolation
Lu Wang, Yanjun Shi, and Hui Li
Abstract—Silicon carbide (SiC) power devices are beneficial to
the converters in terms of size reduction and efficiency increase.
Nevertheless, the fast switching of SiC devices results in more seri-
ous electromagnetic interference (EMI) noise issue. Optical fibers
based isolation provides a reliable solution to block the EMI noises
from the power circuit to the control circuit but with additional
cost and size penalty, especially for multilevel converters. This let-
ter aims at the digital filters based solution to suppress the sam-
pling noise caused by EMI without any extra hardware cost. An
improved median filter and a novel magnitude bandpass filter are
proposed. The design tradeoffs among noises filter capability, delay
effect and the computation time have been discussed for proposed
filters. The anti-EMI noise function of proposed filters has been
experimentally verified in a 60-kW five-level SiC inverter.
Index Terms—Anti-electromagnetic interference (EMI) noise,
digital filter, silicon carbide (SiC).
I. INTRODUCTION
ILICON carbide (SiC) device can switch fast to reduce
S switching loss. However, high dv/dt and di/dt generated
with fast switching speed will raise the issue of increased elec-
tromagnetic interference (EMI) [1]–[3]. The EMI noises are
propagated in both the power circuit and the control circuit.
When the voltage/current sampling signals are mixed with the
EMI noises, the control of the SiC converter will be disturbed
or even corrupted.
Optical fibers provide a reliable isolation solution to block
the EMI noises transmission [4]. In [5], the industrial gate drive
for 1.7-kV SiC MOSFET uses optical fibers for the gate signals.
Further, to enhance the gate-drive board power reliability, Zhang
et al. [6] proposes a gate drive with power over fiber-based
isolated power supply for 15-kV SiC MOSFET instead of using
the traditional design based on the isolation transformer. In the
100 kW 100 kHz SiC converter from [7], besides the gate drive
signals, the voltage and current sampling signals also use the
fiber isolation. In this configuration, the control circuit is fully
isolated from the power circuit and the digital controller will
Manuscript received April 10, 2017; revised May 7, 2017; accepted May 23,
2017. Date of publication June 1, 2017; date of current version October 6, 2017.
This work was supported by the Office of Energy Efficiency and Renewable En-
ergy, U.S. Department of Energy, under Award DE-EE0006521 with North Car-
olina State University, PowerAmerica Institute. (Corresponding author: Hui Li.)
The authors are with the Center for Advanced Power Systems, Florida State
University, Tallahassee, FL 32310 USA (e-mail: lw15r@my.fsu.edu; yshi3@
caps.fsu.edu; hli@caps.fsu.edu).
Color versions of one or more of the figures in this letter are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TPEL.2017.2710483
0885-8993 © 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution
requires IEEE permission. See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
13
not be affected by the EMI noises. However, the optical fiber
utilization needs extra space, adds additional cost, and consumes
extra power from the auxiliary power supply, especially when
the system has multiple gate drives and sampling channels such
as multilevel converters.
In this letter, digital filters are applied to suppress the sam-
pling noise for a 60-kW five-level T-type (5LT2 ) SiC inverter
without fiber isolation. There are two main kinds of digi-
tal filters in the time domain and frequency domain, respec-
tively, based on the application requirements [8]. The frequency
domain digital filter is not applicable in this research due to that
the noises in the analog-to-digital (AD) sampling results have
random frequency ranges. The moving average filter is simple
and typically used in the time domain, but the output results
are skewed around the signal edges [9]. On the other hand, me-
dian filter (MF) has been widely used in the image and speech
processing field as an effective nonlinear impulse filter with the
capability to preserve signals edges [10]–[12]. The traditional
MF is time consuming with O(N2 ) time complexity. Therefore,
an improved median filter (IMF) with O(N) time complexity is
proposed in this letter to reduce the digital filter’s computation
time. In addition, a magnitude bandpass filter (MBF) is also
proposed in this letter. The MBF features O(1) time complexity
and much less fixed computation time without introducing time
delay in the control loop. The design tradeoffs among noises fil-
ter capability, delay effect and the computation time have been
discussed for proposed filters. Finally, both proposed digital fil-
ters are implemented in a 60-kW five-level T-type (5LT2 ) SiC
inverter and experimental results demonstrated that the noise
can be suppressed or removed from the AD sampling signals.
II. SAMPLING NOISE ANALYSIS IN THE 60-KW FIVE-LEVEL
T-TYPE (5LT2 ) SIC INVERTER
A. System Description
Fig. 1(a) is the topology of a 60 − kW 5LT2 SiC inverter
where six SiC T-type modules are applied, two of each are par-
allel coupled through an intercell transformers (ICT) and then
connected to the 480 V three-phase three-line grid without any
external filters. The switching frequency is selected as 50 kHz
to achieve both high efficiency (99.2%) and high power density
(27 W/in3 and 3 kW/kg). This 5LT2 topology has taken the full
advantages of SiC device and multilevel waveforms so there is
no external filter required since the leakage inductance of ICT
and grid impedance is enough to serve as ac filter. In addition,
it can suppress the ground leakage current with a small size of
14
Fig. 1. 60 − kW 5LT 2 SiC inverter: (a) topology, and (b) control system
configuration.
common mode choke, therefore it is especially suited for trans-
formerless photovoltaic (PV) string inverter applications. The
detailed operation principles and experimental verifications of
this 5LT2 topology applied as PV converters have been presented
in [13] and [14]. This letter focuses on the digital filters design
to against the noise caused by EMI in the AD sampling signals.
Fig. 1(b) is the control system configuration of this five-level
inverter that consists of TI DSP TMS320F28335 and Spartan-
3AN field-programmable gate array (FPGA) XC3S200AN.
DSP is the main controller and the FPGA is to generate 24
channels’ 50 kHz gate drive signals. Lf is the leakage in-
ductance of the ICT, Lg is the grid impedance, and they
both act as the ac filter. The conventional αβ axis current
control with grid voltage feedforward is implemented in the
5LT2 inverter. The grid side voltage and current sampling
signals are transmitted to the DSP AD ports through signal
conditioning board. Since there is no optical fiber isolation
between the power circuit and the control circuit, the noises
generated by fast switching transient are propagated to the AD
sampling results. The noises in the voltage sampling signals
generate false phase-locked loop angle and feedforward volt-
age magnitude, and the noises in the current sampling signals
produce distorted inputs to the current controller.
B. Sampling Noise Phenomena
The 5LT2 inverter uses Wolfspeed’s 1200 V 100 A full SiC
MOSFET T-type modules. The switching characteristics of the
IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 33, NO. 1, JANUARY 2018
Fig. 2. Switching waveforms at V d c = 720 V, I = 100 A with R g o n =
6.8 Ω and R g o ff = 4.7 Ω: (a) T-type power module, (b) T4 turn-on, and (c) T4
turn-off.
Fig. 3. EMI noises affected AD sampled phase current results from DSP:
waveforms and FFT results.
module have been analyzed in [15]. Fig. 2 shows T4’s turn-on
and turn-off transient at 720 V dc-link voltage and 100-A load
current with 6.8 Ω turn-on and 4.7 Ω turn-off gate resistors.
The dv/dt is 17 V/ns. As a result, the ringing in the voltage
and current waveforms generates severe noises in the whole
system.
Fig. 3 shows four typical phase current AD sampling results
from the DSP. In (a), the noises start at 6 ms and last for nearly
one-fifth of the fundamental period, which generates phase cur-
rent pulses with magnitude between 0 and –110 A. The current
fast Fourier transform (FFT) analysis shows that the noises are
fairly even distributed in the frequency domain below 20 kHz.
In (b), there are two sections affected by noises in one period.
The first section produces large negative current impulse from
100 to –100 A. The second section is concentrated with smaller
magnitude. The FFT results have a peak magnitude around 1
kHz. In (c), when there are only two small current pulses dur-
ing one period, the harmonics magnitude below 20 kHz is less
than 1.5%. In (d), the noises last around one-fourth of the pe-
IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 33, NO. 1, JANUARY 2018
Fig. 4. IMF algorithm: (a) storage arrays, and (b) implementation process.
riod generating both small and large pulses. It can be noticed
that when large current pulse happened, the noises cause the
malfunction of the current controller and produce low frequency
oscillations. The corresponding FFT results show the most
serious harmonics among these four cases, whereby the magni-
tude peak is around 2 and 7 kHz.
From above sampling current waveforms and FFT results,
the sampling noises show random characteristics as nonperiodic
pulses without any predictable frequency domain pattern. With
these sampling noises, the current controller in the DSP will be
disturbed and/or overcurrent protection will be triggered. The
conventional low-pass, bandpass, or high-pass filters cannot re-
move the random noises since they only affect specific frequency
range. The sampling noises discussed in this letter require that
the desired filter should meet the following requirement: ran-
dom pulses suppression with low computation time cost, small
time delay, and no magnitude attenuation to original signals.
III. PROPOSED ANTI-EMI NOISE DIGITAL FILTERS
A. Improved Median Filter
MF is a conventional nonlinear filter used for removing noises
without magnitude attenuation. N points MF algorithm is shown
in (1), where xn , . . . , xn −N −1 are the latest N points input sig-
nals and xm dn is the output signal of the MF. For convenience,
N is considered as odd number in this letter. N points MF is able
to remove (N–1)/2 points noises among the continues N points
signal with (N–1)/2 of control periods delay
xm dn = Median (xn + xn −1 + . . . + xn −N +1 ) .
The algorithm complexity of the conventional MF is O(N2 ),
which is too time-consuming to be used in the microcontroller.
An IMF is proposed in this letter to reduce the computation time,
which keeps the same filter performance of the conventional MF.
As shown Fig. 4(a), there are two storage arrays for the IMF
algorithm: moving data array stores the latest N points sampling
results as the input of IMF, whereas pointer array stores the
pointers that keep N points data sorted from small to large. The
IMF implementation process is shown in Fig. 4(b). When there
is a new data sampled, first the pointer of the old data needs to be
15
Fig. 5 MBF filter algorithm: (a) a unit sinusoidal example signal with 20%
BW, and (b) implementation process.
located in the sorted pointer array. Then, the pointer’s location
needs to be adjusted depending on the new data value through
left side bubble or right side bubble. The output data are always
the median data, which p(N −1)/2 pointed. With the combined
two arrays, the IMF achieves O(N) algorithm complexity.
B. Magnitude Bandpass Filter
An MBF filter is proposed in this letter as another method to
suppress the sample noises. As shown in Fig. 5, the principle is
to remove all the noises points beyond the magnitude bandwidth
(BW) of the baseline signal. When the input signal’s magnitude
is in the effective data band, the signal will be passed through
the MBF without any attenuation and all the frequency domain
information will be preserved. When the input signal’s magni-
tude is beyond the effective data band, the signal will be blocked
from output. The signal baseline can be generated by the ref-
erence or the averaged value. In this letter, the signal baseline
is generated by N points moving average xavg , as shown in (2).
If the sampled new data is beyond the effective data band, it
will be abandoned and replaced by the data from last sampling
period; otherwise the sampled data will be outputted without
any modification
xavg = Mean (xn + xn −1 + . . . + xn −N +1 ) . (2)
The MBF has O(1) algorithm complexity and is able to re-
move unlimited points of EMI noises in the signals without any
magnitude attenuation nor phase delay. The noises points with
magnitude inside the BW cannot be removed.
C. Anti-EMI Noise Filters Design
In order to test the proposed algorithms in a controlled en-
vironment, an M points signal with 0.1M random noise points
is generated by MATLAB as the input of the digital filters.
Both the noise magnitude and noise location of the sampler
is random. The original signal has the constant value of one,
{x1 , x2 , . . . , xM }. The function randperm(M) generates a ran-
dom permutation of the integers from 1 to M inclusive, whereby
(1)
the first 0.1M points are used as the random location of the
noises, {l1 , l2 , . . . , l0.1M }. The noise magnitude is normally dis-
tributed random numbers generated by function randn(0.1M),
{n1 , n2 , . . . , n0.1M }. The following equation shows how the
signal with random noises is generated:
xl i = 1 + ni (i = 1, 2, . . . , 0.1M ) . (3)
Both the number and magnitude of the noises have the impact
on the digital control, so the sum of the noises magnitude is used
to represent the noise and evaluate the digital filter capability,
which is named noise energy in this letter, as shown in (4). In
16
Fig. 6. Delay periods, DSP implementation time, and residue noise energy
percentage with different IMF points.
Fig. 7. DSP implementation time, and residue noise energy percentage with
different BW in the MBF.
the signal sampler, M is 106 and the total noise energy is 7.98e4.
The ratio of the residue noise energy after the digital filter to the
total input noise energy is defined as εnoise in (5) to show the
capability of the digital filters
M
 as the dc source in the experiments. The three-phase output is
Enoise = |xi − 1| (4)
i=1
Enoise residue
εnoise = . (5)
Enoise total
Fig. 6 shows the IMF characteristic with different filter points.
The delay time and DSP implementation time increase nearly
linearly with the IMF points. The IMF does not change the
system magnitude characteristics, while increased IMF points
may decrease the phase margin and reduce the system’s control
BW. Considering the noises removing capability and the impact
to the system control, the IMF points can be selected between
3, 5, and 7.
In the MBF, the computation time is fixed as 0.55 μs with
different moving average points and different data BWs in the
DSP, which is much smaller than that of IMF. There is no time
delay introduced into control loop by the MBF. From Fig. 7,
MBF shows similar noise removing capability as IMF, smaller
BW can remove more noises. However, if the BW is too small,
the ripple in the original signals may be falsely removed. It
can be observed from Fig. 7, if the BW in the MBF is selected
between 10% and 40%, the same attenuation from MBF can be
achieved as that from IMF. Considering there can be up to 30%
high frequency ripple in the sampled data, a 30% BW is chosen
in this research. With 30% BW, a transient 5 V/μs and 1 A/μs
can pass the MBF without being distorted.
IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 33, NO. 1, JANUARY 2018
Fig. 8. Experiment set up of a grid-tied 60-kW SiC inverter prototype.
Fig. 9. Anti-EMI noise capability: (a) phase A current noise, (b) W/ five points
IMF, (c) W/ 30% BW MBF, and (d) W/ fourth Butterworth LPF.
IV. EXPERIMENTAL VERIFICATIONS
The 60 − kW 5LT2 transformerless SiC inverter prototype
is shown in Fig. 8. A 1000 V 150 A dc power supply is used
connected to the grid through a 480 V three-phase 1.5 MW
transformer. The wireless communication between DSP and PC
is able to show the real-time voltage/current sampling signals
in the computer. The control period is 20 μs. The IMF is imple-
mented with five points considering the tradeoff between time
delay and the noise suppression. The MBF has 30% BW with
eight points moving average baseline.
Fig. 9 compares the noise suppression capability of the two
proposed filters and a conventional low-pass filter (LPF). The
LPF is designed as a fourth Butterworth LPF with 9.2 kHz
cutoff frequency, which has a same phase delay at 1 kHz with
the IMF filter. (a) is a 3000-points data extracted from the phase
current AD sampling results, the fundamental component has
been removed for better observation. (b)–(d) show the capability
of IMF, MBF, and LPF, respectively. In zone I, IMF is more
effective since MBF cannot remove the noises within its BW.
In zone II, MBF is more effective since IMF have the points
limitation while removing the continues noises. In zone III,
both IMF and MBF can remove all the noises effectively. Since
LPF is not able to remove the noises under its cutoff frequency,
the noise energy is still large after the LPF in zone I–III.
In the 60 − kW 5LT2 SiC inverter, the grid-tie start-up oper-
ation cannot be realized without digital filters due to the EMI
noises, as shown in Fig. 10(a). With five points IMF, the inverter
is able to operate smoothly from start-up to 60-kW steady state,
as shown in Fig. 10(b). In Fig. 11, with the 30% BW MBF,
IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 33, NO. 1, JANUARY 2018
Fig. 10. EMI noise impact on the 60-kW inverter grid-tie start-up operation:
(a) W/O digital filters, and (b) W/ five points IMF.
Fig. 11. EMI noise impact on the grid-tie steady-state operation: W/ 30% BW
MBF to W/O digital filters.
the grid-tie test is operating normally. Once the MBF is dis-
abled, there are multiple pulses appearing in the grid current
and eventually trigged the overcurrent protection.
The experiment results verify that both IMF and MBF are
able to remove the noises from the AD sampling results and
ensure the system operation. As shown in Fig. 1(b), there are
two current and two voltage signals, which require digital filters
in the inverter system. In the 20 μs control period, four channels
IMF computation already consumes 44% of the DSP interrup-
tion time resources. While in this prototype, MBF is a better
choice, which saves more control time of the inverter.
V. CONCLUSION
This letter presents the noise issue caused by EMI in the
AD sampling signals of the digital control in a multilevel SiC
inverter. The grid-tie experiment results show that the current
control loop will be tripped by overcurrent protection without
fiber isolation. Anti-EMI noise digital filter solution is proposed
to solve this issue. The developed IMF and MBF are imple-
mented and demonstrated to remove the random noises from
the voltage/current sensing signals without any extra hardware
cost. Although in this letter, the proposed methods are imple-
mented in DSP, they can also be implemented in FPGA or SoC
FPGA, which can save the DSP interruption time.
ACKNOWLEDGMENT
The information, data, or work presented herein was funded
in part by the Office of Energy Efficiency and Renewable En-
ergy (EERE), U.S. Department of Energy, under Award Number
DE-EE0006521 with North Carolina State University, Power-
America Institute.
17
DISCLAIMER
The information, data, or work presented herein was funded
in part by an agency of the United States Government. Neither
the United States Government nor any agency thereof, nor any
of their employees, makes any warranty, express or implied,
or assumes any legal liability or responsibility for the accuracy,
completeness, or usefulness of any information, apparatus, prod-
uct, or process disclosed, or represents that its use would not
infringe privately owned rights. Reference herein to any specific
commercial product, process, or service by trade name, trade-
mark, manufacturer, or otherwise does not necessarily constitute
or imply its endorsement, recommendation, or favoring by the
United States Government or any agency thereof. The views and
opinions of authors expressed herein do not necessarily state or
reflect those of the United States Government or any agency
thereof.
REFERENCES
[1] D. Han and B. Sarlioglu, “Study of the switching performance and EMI
signature of SiC MOSFETs under the influence of parasitic inductance in
an automotive DC-DC converter,” in Proc. IEEE Transp. Electrific. Conf.
Expo, Jul. 2015, pp. 1–8.
[2] N. Oswald, P. Anthony, N. McNeill, and B. H. Stark, “An experimental
investigation of the tradeoff between switching losses and EMI generation
with hard-switched all-Si, Si-SiC, and all-SiC device combinations,” IEEE
Trans. Power Electron., vol. 29, no. 5, pp. 2393–2407, May 2014.
[3] B. Sun and R. Burgos, “Assessment of switching frequency impact on
the prediction capability of common-mode EMI emissions of SiC power
converters using unterminated behavioral models,” in Proc IEEE Appl.
Power Electron. Conf. Expo., May 2015, pp. 1153–1160.
[4] Avago Technologies Inc., Building Safe and Reliable Electrical Systems
With Optocouplers, San Jose, CA, USA, Jan. 2012. [Online]. Available:
http://docs.avagotech.com/docs/AV02-3344EN
[5] L. Stevanovic et al., “Industrial readiness of SiC power devices,” in
Proc. CFES Annu. Conf., Feb. 2015. [Online]. Available: https://www.
rpi.edu/dept/cfes/AnnualConference/b2%20Stevanoic%20GE.pdf
[6] X. Zhang et al., “A gate drive with power over fiber-based isolated power
supply and comprehensive protection functions for 15-kV SiC MOSFET,”
IEEE Trans. Power Electron., vol. 4, no. 3, pp. 946–955, Jun. 2016.
[7] D. Boroyevich, “Use of SiC devices in medium-voltage convert-
ers,” in Proc. CPES Workshop WBG Power Electron. Adv. Distrib.
Grids, Apr. 2016. [Online]. Available: https://www.nist.gov/sites/default/
files/documents/pml/high_megawatt/Boroyevich_db-Workshop-on-MV-
WBG-PE-for-ADG-2016.pdf
[8] S. W. Smith, The Scientist and Engineer’s Guide to Digital Signal Pro-
cessing. San Diego, CA, USA: California Tech. Pub., 1999.
[9] C. A. Glasbey and G. W. Horgan, Image Analysis for the Biological
Sciences. Hoboken, NJ, USA: Wiley, 1995.
[10] J. W. Tukey, Exploratory Data Analysis. Reading, MA, USA: Addison-
Wesley, 1970.
[11] R. Kunsoth and M. Biswas, “Modified decision based median filter for
impulse noise removal,” in Proc. IEEE Int. Conf. Wireless Commun.,
Signal Process. Netw., Sep. 2016, pp. 1316–1319.
[12] S. Khatri and H. Kasturiwale, “Quality assessment of median filtering
techniques for impulse noise removal from digital images,” in Proc. IEEE
Int. Conf. Adv. Comput. Commun. Syst., Oct. 2016, pp. 1–4.
[13] Y. J. Shi, L. Wang, R. Xie, Y. X. Shi, and H. Li, “A 60 kW 3 kW/kg
5-level T-type SiC PV inverter with 99.2% peak efficiency,” IEEE Trans.
Ind. Electron., vol. PP, no. 99, p. 1, May 2017.
[14] L. Wang, Y. J. Shi, Y. X. Shi, R. Xie, and H. Li, “Ground leakage current
analysis and suppression in a 60 kW 5-level T-type transformerless SiC PV
inverter,” IEEE Trans. Power Electron., vol. PP, no. 99, p. 1, Mar. 2017.
[15] Y. X. Shi, R. Xie, L. Wang, Y. J. Shi, and H. Li, “Switching characterization
and short-circuit protection of 1200 V SiC MOSFET T-type module in
PV inverter application,” IEEE Trans. Ind. Electron., vol. PP, no. 99, p. 1,
Mar. 2017.