6432 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 64, NO.

8, AUGUST 2017

A New FXLMS Algorithm With Offline and Online

Secondary-Path Modeling Scheme for Active
Noise Control of Power Transformers
Tong Zhao, Member, IEEE, Jiabi Liang, Liang Zou, and Li Zhang

Abstract—In this study, the active noise control (ANC) propagation path, is difficult to analyze quantitatively due to
method was used to suppress high-decibel and low- its engineering experience, and it is difficult to realize larger
frequency power transformer noise. An appropriate ANC noise reduction because of the restriction on the environment
system was selected based on the transformer noise char-
acteristics and experimental condition. A new filter-X least and cost [4]. Comparatively, based on the destructive inter-
mean square (FXLMS) adaptive ANC algorithm based on ference of two sound waves, the active noise control (ANC)
offline and online secondary-path modeling was proposed method [5]–[8], which generates a secondary sound signal with
to realize faster and more stable secondary-path online the opposite phase to offset the primary noise, can reduce low-
modeling than that of the random white-noise FXLMS al- frequency noise, such as transformer noise, more effectively,
gorithm and to ensure the convergence, stability, and re-
duction in transformer noise control. Moreover, the genetic with better controllability, easier installation, and lower cost.
algorithm is adopted to optimize the convergence coeffi- The transformer ANC system consists of electroacoustic de-
cient, while the effect of the convergence coefficient on the vices and a controller. In this system, the reference and the
algorithm was analyzed using simulation and theory. In ad- error microphones collect the reference and the error signal of
dition, the transformer noise online monitoring and active the transformer noise, respectively, and the controller based on
control system was designed including software and hard-
ware, and the hardware devices were selected based on the the adaptive filter algorithm can calculate the secondary sound
noise feature. In the 50 000 KVA transformer noise reduc- signal to reduce the transformer noise through the secondary
tion experiment, the system achieved a noise reduction of sound source. Obviously, the adaptive filter algorithm has im-
8–15 dB and an 84.10–96.86% decrease in average sound portant impact on the noise reduction performance of the sys-
energy density in a certain area.
tem. For example, the least mean square (LMS) algorithm is
Index Terms—Active noise control (ANC), convergence a traditional ANC adaptive algorithm [9], but it is easily af-
coefficient, filter-X least mean square (FXLMS), genetic fected by the changed secondary path (SP, the sound path be-
algorithm (GA), power transformer noise, secondary-path tween the error microphone and the secondary sound source) and
(SP) online modeling.
diverges.
I. INTRODUCTION Thus, the filter-X least mean square (FXLMS) algorithm is
used to solve this divergence problem with modeling the SP
ransformer noise is mainly produced by the vibration of the
T stator core from the silicon steel sheet magnetostriction.
Because the magnetostrictive period is half of the current and
[10]–[14]. As changing slowly with the environment, the SP
needs to model rapidly, smoothly, and accurately online. At
present, the random white-noise method is widely used in this
the magnetic circuits of the stator core vary in lengths, the case, which needs the random noise generator to send the Gaus-
noise mainly includes low harmonic frequency of 100 Hz when sian white noise into the SP. Based on the LMS algorithm, it
the power frequency is 50 Hz [1], [2]. This low-frequency noise regards the random noise signal as the reference signal and the
sound level of nearly 50–90 dB may cause human chronic injury error signal as the desire signal, and then the SP modeling can
and result in neurological diseases [3]. Therefore, it is necessary be obtained in the online modeling filter real timely. However,
to reduce transformer noise. because the error signal needs to adjust the noise reduction filter,
The traditional passive noise control method, which includes if this signal mixed with the random noise signal, it will result in
vibration damping of the transformer and insulation of the the instability of the whole ANC system. Additionally, with the
Manuscript received October 30, 2016; revised January 21, 2017; ac-
characteristic of randomness and fluctuation, the random noise
cepted February 11, 2017. Date of publication March 14, 2017; date will also lead to model the SP online unstably.
of current version July 10, 2017. This work was supported by the Two approaches so far have been developed to overcome
National Natural Science Foundation of China under Grant 51207084.
(Corresponding author: Tong Zhao.)
these shortcomings in a certain extent. The first approach uses
The authors are with the School of Electrical Engineering, Shan- the modified structure FXLMS algorithm [10], [11] or increases
dong University, Jinan 250061, China (e-mail: [email protected]; the number of the adaptive filters [12], [13] to eliminate the
[email protected]; [email protected]; [email protected]).
Color versions of one or more of the figures in this paper are available
random noise signal from the error signal as much as possible
online at http://ieeexplore.ieee.org. before the error signal adjusts the noise reduction filter. The
Digital Object Identifier 10.1109/TIE.2017.2682043 second approach utilizes the power of the reference signal and
0278-0046 © 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.

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on December 23,2022 at 03:54:40 UTC from IEEE Xplore. Restrictions apply.
ZHAO et al.: NEW FXLMS ALGORITHM WITH OFFLINE AND ONLINE SECONDARY-PATH MODELING SCHEME FOR ACTIVE NOISE CONTROL 6433

Fig. 1. Time-domain analysis of transformer noise. Fig. 2. Frequency-domain analysis of transformer noise.

the error signal [14] to decrease the amplitude of the random

noise signal when the FXLMS algorithm converges. However, it
is still inevitable to increase the computational complexity and
limit the noise reduction.
This paper proposes a new FXLMS algorithm with offline and
online SP modeling. Comparing with the random white-noise
FXLMS algorithm, this algorithm has three advantages. First, it
costs less in hardware because it does not need to add a random
noise generator. Second, with only two adaptive filters, it can Fig. 3. Single-channel feedforward adaptive ANC system.
reduce the computational complexity and ensure the real-time
performance. Third, without the effect of the random noise,
it can rapidly and smoothly model the SP online and realize
a stable and large reduction in transformer noise. In addition,
it uses the offline SP modeling as the initial value, so it can
increase the speed of modeling the SP online. Moreover, after
the SP changes, it also can quickly and stably model the new
SP, reconverge the system and reach the same noise reduction
performance as before.
The convergence coefficient for the noise reduction filter is Fig. 4. Active noise controller.
the most important control parameter in the proposed algorithm.
Based on the derivation and experience, the range of this param-
eter is provided to guarantee the convergence of the algorithm. noise is stable, periodic, and low in frequency. Its frequency
Additionally, in order to obtain the optimal convergence co- mainly consists of harmonics of 100 Hz. Therefore, it is benefi-
efficient and realize the best noise reduction performance, the cial to use the ANC to reduce the transformer noise.
genetic algorithm (GA) is used in optimizing this parameter The transformer ANC system is designed to be single-
[15], [16]. With global search ability, GA based on the pro- channel, feedforward, and adaptive to attain a stable and large
vided convergence range can quickly optimize this parameter noise reduction, as shown in Fig. 3, considering the transformer
with less populations and generations. In addition, the effect of noise characteristics and noise reduction equipment expense.
this parameter on the performance of the algorithm is analyzed The controller is the most important part of the ANC system.
using simulation and theory, which can prove the correction of It uses the adaptive filter algorithm, which includes the filter
the optimal value calculated by GA. called the noise reduction filter and the adaptive algorithm, as
Based on the proposed algorithm, the transformer ANC sys- shown in Fig. 4. The filter is used to filter the reference signal,
tem is designed including software platform and hardware whereas the adaptive algorithm adjusts the weight coefficient of
equipment [17], [18]. Meanwhile, the hardware devices are se- the filter.
lected by the transformer noise feature. In the 50 000 KVA In Figs. 3 and 4, x(n) collected by the reference microphone
transformer noise reduction experiment, when the error micro- is the reference signal of the transformer noise, e(n) is the error
phone is placed in the semicircular area with the secondary signal collected by the error microphone, y(n) is the secondary
sound source as the center and a radius of 2 m, the system can sound signal from the secondary sound source, p(n) is the pri-
stably and quickly model the SP online, and realize a noise re- mary sound field signal, and s(n) is the secondary sound field
duction of 8–15 dB and an 84.10–96.86% decrease in average signal.
sound energy density in this semicircular area. The filter uses the finite impulse response transversal structure
to maintain the stability of the ANC system. Thus, y(n) can be
II. POWER TRANSFORMER ANC SYSTEM expressed as follows:
Figs. 1 and 2 show the analyses of the indoor oil-immersed

Lw
self-cooling 50 000 KVA power transformer noise in the time y (n) = wl (n)x (n − l + 1) = W (n)T X w (n) (1)
and frequency domains, respectively. The transformer body l=1

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on December 23,2022 at 03:54:40 UTC from IEEE Xplore. Restrictions apply.
6434 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 64, NO. 8, AUGUST 2017

Fig. 6. Part 2: Primary-path offline modeling. (a) Primary-path offline

modeling method block diagram. (b) Primary-path offline modeling sys-
tem in the transformer ANC system.

Fig. 5. Part 1: SP offline modeling. (a) SP offline modeling method

block diagram. (b) SP offline modeling system in the transformer ANC
system.

where Lw is the filter order for W (n), wl (n)(l = 1, 2, . . . , Lw )

is the weight coefficient, W (n) = [w1 (n), w2 (n), . . . ,
wL w (n)]T is the weight vector and X w (n) = [x(n), x(n − 1),
. . . , x(n − Lw + 1)]T is the reference signal vector.
The adaptive algorithm should be suitable for a realistic trans-
former ANC system because it has a strong effect on the con-
vergence, stability, and reduction. Although the traditional LMS
algorithm can achieve large noise reduction, it is easily affected
by the easily changed SP, which will cause worse accuracy and
timeliness of the secondary source signal and result in system
divergence. Moreover, the system cannot reconverge regardless Fig. 7. Part 3: Primary sound field signal cancellation FXLMS
algorithm.
of how the noise control parameters are adjusted.
Consequently, the FXLMS algorithm can be used to solve the TABLE I
divergence problem that results from the existence of the SP, but MEANING OF THE SIGNAL VARIABLE IN EACH PART
it must rapidly, smoothly, and accurately model the SP online.
Part 1 Part 2 Part 3

III. NEW FXLMS ADAPTIVE ANC ALGORITHM WITH Reference signal y 1 (n ) x(n) x(n)
OFFLINE AND ONLINE SP MODELING Error signal e 1 (n ) e 2 (n ) e(n )/e  (n )
Primary sound signal x(n) x(n) x(n)
A. New FXLMS Algorithm Secondary sound signal y 1 (n ) y(n)
Primary sound field signal p(n) p(n) p(n)
The new FXLMS algorithm with offline and online SP mod- Primary sound field estimate signal p (n) p (n)
eling consists of three parts, as shown in Figs. 5–7. Parts 1 and Secondary acoustic field signal s 1 (n ) s(n)
Secondary acoustic field estimate signal s 1 (n ) s  (n )
2 separately model the SP and primary path offline, and Part 3 Filter-x signal r(n)
models the SP online and reduces the transformer noise, and it
also adopts GA to optimize the convergence coefficient for the
TABLE II
noise reduction filter. Table I provides the meaning of the signal SIGNAL COLLECTED OR OUTPUT THROUGH ELECTROACOUSTIC DEVICES IN
variables in Figs. 5–7. Table II provides the signals collected EACH PART
or output by the electroacoustic devices in the ANC system.
Table III shows the flowchart of the proposed algorithm. Part 3 Part 1 Part 2 Part 3
is the main part of the proposed algorithm, whereas Parts 1 and
Reference microphone x(n) x(n)
2 provide assistance. Error microphone s 1 (n ) p(n) e(n)
In Figs. 5–7, Hp (z) is the transfer function of the primary Secondary sound source y 1 (n ) y(n)
path, Hs (z) is the transfer function of the SP, Ĥp (z) is the

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on December 23,2022 at 03:54:40 UTC from IEEE Xplore. Restrictions apply.
ZHAO et al.: NEW FXLMS ALGORITHM WITH OFFLINE AND ONLINE SECONDARY-PATH MODELING SCHEME FOR ACTIVE NOISE CONTROL 6435

L
TABLE III
FLOW OF THE PROPOSED ALGORITHM
where Ĥ p (n) = [ĥ1p (n), ĥ2p (n), . . . , ĥp p (n)]T is the impulse
response of Ĥp (z), X p (n) = [x(n), x(n − 1), . . . , x(n −
Parameter: L s = filter order of Ĥ s (z ) Lp + 1)]T is the reference signal vector, μp is the convergence
μ s = convergence coefficient for Ĥ s (z ) coefficient for Ĥp (z), and Lp is the filter order of Ĥp (z).
Computation: n = 0, 1, . . .
Part 1 s 1 (n ) = Ĥ Ts (n )Y 1 (n ) Part 3 uses the primary sound field signal cancellation method
e 1 (n ) = s 1 (n ) − s 1 (n ) FXLMS algorithm to model the SP online and reduce trans-
Ĥ s (n + 1) = Ĥ s (n ) + 2μ s e 1 (n )Y 1 (n ) former noise, and it also adopts GA to optimize the convergence
Parameter: L p = the filter order of Ĥ p (z ) coefficient for the noise reduction filter. The entire transformer
μ p = convergence coefficient for Ĥ p (z )
Computation: n = 0, 1, . . .
ANC system with the offline modeling result of Parts 1 and 2
Part 2 p  (n ) = Ĥ Tp (n )X p (n ) can rapidly and stably model the SP online and ensure the con-
e 2 (n ) = p(n ) − p  (n ) vergence, stability, and reduction of transformer noise control.
Ĥ p (n + 1) = Ĥ p (n ) + 2μ p e 2 (n )X p (n )
Consequently, the SP online modeling formula can be expressed
Parameter: L s = filter order of Ĥ  (z ) 
μ s = convergence coefficient for Ĥ s (z ) Ĥ  s (0) = Ĥ s
L w = filter order of W(z)
μ w = convergence coefficient for W(z) Ĥ  s (n + 1) = Ĥ  s (n) + 2μ s e (n) Y (n) n = 0, 1, . . .
Computation: n = 0, 1, . . . (4)
Ĥ  (n ) = Ĥ s n = 0  L S
Part 3 { s
Ĥ s (n ) = Ĥ  s (n − 1) + 2μ  s e  (n )Y (n )n = 1, . . . where Ĥ s (n) = [ĥ1s (n), ĥ 2
s (n), . . . , ĥ s (n)]T
is the impulse
y (n ) = W T (n )X w (n ) response of Ĥs (z), Ĥ s is the SP offline model, Y (n) =
p  (n ) = Ĥ Tp (n )X p (n ) [y(n), y(n − 1),... , y(n − Ls + 1)]T is the secondary sound
s  (n ) = Ĥ T s (n )Y (n )
s  (n ) = Ĥ T s (n )Y (n )
signal vector, e (n) is the error signal of the SP online mod-
e  (n ) = e(n ) − p  (n ) − s  (n ) eling, and μs is the convergence coefficient for Ĥs (z), and Ls
GA optimizes μ w (n) is the filter order of Ĥs (z). The weight vector W (n) iterative
W (n + 1) = W (n ) + 2μ w (n )e(n )R(n )
formula of the new FXLMS algorithm is
W (n + 1) = W (n) + 2μw (n) e (n) R (n) (5)
offline estimate of the transfer function of the primary path, where R(n) = [r(n), r(n − 1), . . . , r(n − Lw + 1)]T is the
Ĥs (z) is the offline estimate of the transfer function of the SP, filter-x signal vector, e(n) is the error signal of the FXLMS
and Ĥs (z) is the online estimate of the transfer function of the algorithm, μw (n) is the convergence coefficient for the noise
SP. reduction filter W (z), and Lw is the filter order of W (z).
Part 1 uses the random white-noise method based on the LMS The convergence coefficient μw (n) for the noise reduction
algorithm to model the SP offline. The modeling result is used filter is the most important control parameter in the proposed
as the initial value of the online modeling in Part 3 to make it fast algorithm, as it significantly affects the convergence speed, sta-
and stable. Without changing the built transformer ANC system, bility, and reduction of the transformer noise control. According
Part 1 uses the secondary sound source to generate random white to the simulation and theoretical analysis below, there is a cer-
noise and regards it as the reference signal while collecting the tain range of μw (n) that makes the proposed algorithm converge
secondary sound field signal through the error microphone. The in the ANC system, which is called the convergence range of
effect of the primary sound field signal can be ignored in this the proposed algorithm. In the convergence range, there is an
process. Thus, the SP offline modeling formula is optimal μw (n) that makes the algorithm obtain the maximum
Ĥ s (n + 1) = Ĥ s (n) + 2μs e1 (n) Y 1 (n) (2) noise reduction and the fastest convergence speed. Thus, the
proposed algorithm uses GA to optimize μw (n).
where Ĥ s (n) = [ĥ1s (n), ĥ2s (n), . . . , ĥLs S (n)]T is the impulse Based on the deduction and experience, the convergence
response of Ĥs (z), Y 1 (n) = [y1 (n), y1 (n − 1), . . . , y1 (n − range of μw (n) is provided below. Mentioned in [6], the con-
Ls + 1)]T is the reference signal vector, μs is the convergence vergence range of the FXLMS algorithm is
coefficient for Ĥs (z), and Ls is the filter order of Ĥs (z).
1
Part 2 uses the two-microphone transfer function method 0 < μw (n) < L w (6)
based on the LMS algorithm to model the primary path of- n=1 E [r2 (n)] ks
fline. The modeling result is a part of Part 3. The frequency of 
L w
the transformer noise is considered sufficiently abundant, and where E[r2 (n)] is the total input power of r(n) and ks is
n=1
Part 3 does not require a high accuracy for the primary-path of- the time delay of the SP.
fline model. Therefore, without changing the built transformer As for the proposed algorithm, the base value of μw (n) is
ANC system, Part 2 uses the reference microphone to collect defined as follows:
the transformer noise as the reference signal and the error mi-
1
crophone to collect the primary sound field signal. In this case, μw b = L w  2  k s (7)

the primary-path offline modeling formula can be expressed as n = 1 E r (n) 2
follows:
where r (n) = Ĥ Ts (n)X s (n), Ĥs (z) is the offline model of
Ĥ p (n + 1) = Ĥ p (n) + 2μp e2 (n) X p (n) (3) the SP, which can be obtained in Part 1, k s can be obtained

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on December 23,2022 at 03:54:40 UTC from IEEE Xplore. Restrictions apply.
6436 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 64, NO. 8, AUGUST 2017

by observing Ĥs (z) and X s (n) = [x(n), x(n − 1), . . . , x(n − TABLE IV
FREQUENCY COMPONENTS OF THE PRIMARY SOUND SOURCE
Ls + 1)]T is the reference signal vector.
Then, based on experience, the convergence range of μw (0)
Frequency (Hz) 100 200 300 400 500 600 700
can be set as Amplitude (V) 0.9 0.9 1.0 1.0 0.5 0.9 0.9

μw b × 10−2 < μw (0) < μw b . (8)

TABLE V
After considering the changed SP, the range is determined as CONTROL PARAMETER SETTINGS

1
μw b × 10−2 < μw (n) < L w . (9) Part 1 Ls μs

n=1 E [r2 (n)] k2s 300 0.001

Proposed algorithm Part 2 Lp μp
200 1 × 10–4
In order to obtain the optimal μw (n) and realize the fastest Part 3 L s μ s Lw
convergence speed and the maximum noise reduction for the 200 0.003 214
transformer ANC system, the proposed algorithm adopts the Random white noise algorithm L s μs Lw μw
GA to optimize the μw (n). 200 0.003 214 8 × 10–7

Constraint can be set based on (8) and (9), so the fitness

function can be set as
N −1
1  2
min f (μw (n)) = e (i)
N i=0
N −1
1 
= [d (i) + s (i)]2
N i=0

1  T 2
N −1
= Ĥ p (i) X p (i)+ Ĥ T s (i) Y (i)
N i=0
Fig. 8. Primary-path modeling.
y (i) = W T (i) X w (i)
⎧
⎨W (n) , i=0 In addition, the transfer function of the primary path is
W (i) = W (i−1) + 2μw (n) Hp (z) = 0.05z −6 − 0.05z −7 + 0.01z −8 + 0.14z −9
⎩
× e (i) R (i) , i = 1, . . . , N−1
 + 0.1z −10 − 0.045z −11 − 0.02z −12
μw b ×10−2 < μw (n) < μw b , n=0
s.t. μ ×10−2 < μ (n) <  1
, n = 1, . . . + 0.05z −13 − 0.01z −14 (11)
wb w Lw
E [r 2 (n )] k2s
n=1

(10) and the transfer function of the SP is

where N is the iterations in the fitness function, and Hs (z) = 0.01z −3 + 0.01z −4 + 0.9z −5 − 0.01z −7 − 0.75z −8 .
the initial values should be X p (0) = X p (n), X s (0) = (12)
X s (n), X w (0) = X w (n), Y (0) = Y (n), R(0) = R(n), The control parameters of the proposed algorithm and random
 
Ĥ p (i) = Ĥ p (n), and Ĥ s (i) = Ĥ s (n). white-noise algorithm are set in Table V.
The proposed algorithm consists of Parts 1, 2, and 3. The
B. Simulation Analysis offline SP modeling result of Part 1 is used as the initial value of
the online SP modeling in Part 3 to make the online modeling
In the following simulation, the proposed new FXLMS faster, while the offline primary-path modeling of Part 2 is one
algorithm is compared with the random white-noise FXLMS part of Part 3. Therefore, it is necessary to analyze the modeling
algorithm in the performance which includes the convergence, performance of Parts 1 and 2. The relative error ΔS(n) is usually
stability and reduction of transformer noise, as well as the sta- regarded as the evaluation criterion for modeling
bility, convergence speed and accuracy of online SP model- ⎡ 2 ⎤
L  i
ing. Meanwhile, the optimization of the convergence coefficient
⎢ i = 0 h s (n) − ĥ i
s (n) ⎥
μw (n) with GA is analyzed. In addition, the performance and ΔS (n) = 10log10 ⎣ L ⎦ (13)
{h i (n)}2
ability to deal with the changed SP of the proposed algorithm is i=0 s
also discussed.
The transformer noise is periodic, and its frequency mainly where L represents the filter order of Ĥs (z), Ĥp (z) or Ĥs (z).
consists of harmonics of 100 Hz. Hence, the primary sound The primary-path and the SP modeling results are shown in
source is set to be the superposition of sinusoids with each Figs. 8 and 9 respectively, while the modeling relative errors
frequency component in Table IV. are shown in Fig. 10. Obviously, Part 1 can accurately offline

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on December 23,2022 at 03:54:40 UTC from IEEE Xplore. Restrictions apply.
ZHAO et al.: NEW FXLMS ALGORITHM WITH OFFLINE AND ONLINE SECONDARY-PATH MODELING SCHEME FOR ACTIVE NOISE CONTROL 6437

Fig. 9. SP modeling.

Fig. 11. Noise reduction performance of the proposed algorithm and

the random white-noise FXLMS algorithm. (a) Error signal. (b) MSE
Fig. 10. Modeling relative error. learning curve.

model the SP. It is beneficial to make the online modeling in

Part 3 faster and more stable. Additionally, Part 2 can effectively
model the primary path offline. Although this modeling does not
have high precision, it satisfies the requirement of Part 3.
The modeling relative error ΔS(n) curves shown in Fig. 10
can intuitively compare the performance of online SP mod-
eling between Part 3 and the random white-noise method. The
ΔS(n) curve of Part 3 is smooth, and it can realize stability after
Fig. 12. Optimization of convergence coefficient μ w (n).
2000 iterations. Comparatively, the ΔS(n) curve of the random
white-noise method has obvious twist and fluctuation, which
seems that the curve is difficult to stabilize and the method will rithm (about 0 dB). Apparently, the proposed algorithm can
diverge at any time. Apparently, without the effect of random realize much more noise reduction.
noise, the modeling process of Part 3 is relatively smooth and Corresponding to the noise reduction of the proposed algo-
stable. rithm in Fig. 11, Fig. 12 shows the iterative process of conver-
In addition, the proposed algorithm requires only 2000 itera- gence coefficient μw (n) optimized by the GA. The optimization
tions to make the ΔS(n) curve reach –6 dB, while the random of μw (n) has slight fluctuation around 8.3 × 10–7 , but it still
white-noise method needs almost 3600 iterations. Obviously, can make the proposed algorithm obtain the maximum noise
the proposed algorithm uses the offline modeling result of Part 2 reduction and the fastest convergence speed.
as the initial value of Part 3, which can rapidly realize online In order to analyze the processing capacity of the proposed
modeling. algorithm when the SP changes, the iteration number is set
The noise reduction of the proposed algorithm and the ran- as 10 000, and the transfer function of the SP changes from
dom white-noise FXLMS algorithm is shown in Fig. 11. In Hs (z) to 0.5Hs (z) after 5000 iterations. Then, the simulation
Fig. 11(a), the error signal e(n) of the proposed algorithm can results are shown in Fig. 13. After the SP changes, the proposed
converge to 0 and maintain as 0 stably after 1000 iterations. By algorithm still can guarantee that the ΔS(n) curve of online
contrast, with the effect of the random noise, e(n) of the random SP modeling can smoothly and stably reach –6 dB again after
white-noise algorithm cannot converge to 0, and it also has high the next 2000 iterations, which indicates that the proposed al-
volatility, which seems that the algorithm will diverge at any gorithm can online model the changed SP quickly, stably and
time. Obviously, the proposed algorithm can realize more stable accurately. In addition, the optimal value of μw (n) optimized
and faster convergence. by the GA changes from 8.3 × 10–7 to 3.4 × 10–6 . Although
Additionally, in Fig. 11(b), the noise reduction upper limit the optimization process has slight fluctuation around the opti-
of the proposed algorithm is approximately –20 dB, which is mal value, it can ensure that the error signal e(n) reconverges to
considerably lower than that of the random white-noise algo- 0 and maintains as 0 stably after the next 1000 iterations, and

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on December 23,2022 at 03:54:40 UTC from IEEE Xplore. Restrictions apply.
6438 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 64, NO. 8, AUGUST 2017

Fig. 13. Noise reduction performance of the proposed algorithm with

consideration of the SP changing.

the noise reduction upper limit of the proposed algorithm is still

approximately –20 dB. It is shown that after the SP changing,
the proposed algorithm can quickly and stably model the new
SP, realize reconvergence and reach the same noise reduction as
before.

C. Convergence Coefficient
The effect of convergence coefficient μw (n) on the conver-
gence speed, stability and reduction of the transformer ANC Fig. 14. Simulation analysis of convergence coefficient influence.
(a) Error signal. (b) MSE learning curve.
system is analyzed in simulation and theory, which can prove
the correction of the optimal value calculated by GA.
In the simulation, the primary sound source signal, Additionally, in the convergence process, the weight vector
Hp (z), Hs (z) and other parameters are set as Table IV, (11), randomly fluctuates near the optimal weight vector W o , which
(12) and Table V, respectively. The error signal and MSE learn- makes the objective function J∞ unequal to its minimum Jm in .
ing curve are represented by their upper envelopes are shown in Thus, the misadjustment can be defined as
Fig. 14. J∞ − Jm in
In Fig. 14, μw is increased from 1 × 10–7 to 1 × 10–6 , and the δ= . (15)
Jm in
other parameters remain unchanged. When μw increases from
The misadjustment δ is relative to μw and λp
1 × 10–7 to 6 × 10–7 , the convergence speed accelerates, and
the noise reduction increases. With a continuous rising of μw to 
LW
1 × 10–6 , both the speed of convergence and the noise reduction δ = μw λp . (16)
decrease, and the system tends to diverge. The system diverges i=1
when μw exceeds 1 × 10−6 . According to (14)–(16), when μw increases from a small
Therefore, in the simulation, it can conclude that when the value with other parameters unchanged, the shortened decay
convergence coefficient μw is less than or equal to 1 × 10–6 , the time τp plays a leading role in the noise reduction process.
proposed algorithm can realize converge. Moreover, the optimal Hence, the system converges more rapidly, which makes the
value of μw is in 6 × 10–7 –9 × 10–7 , which can make the al- algorithm more able to perform in real time and increases the
gorithm obtain the fastest convergence speed and the maximum noise reduction. However, the misadjustment δ also increases,
noise reduction. Being consistent with this simulation result, which makes J∞ diverge from Jm in . Hence, when μw increases
the optimal value of μw (n) optimized by the GA above is about to a sufficiently large value, the misadjustment δ is dominant and
8.3 × 10–7 , which indicates that the GA can optimize μw (n) decreases the stability and noise reduction. Finally, the system
reliably in the proposed algorithm. diverges.
The effect of μw on the convergence speed, stability, and
reduction in the simulation is theoretically analyzed as follows. IV. TRANSFORMER NOISE ONLINE MONITORING AND ACTIVE
The decay time τp of the pth sample in the MSE learning curve CONTROL SYSTEM
is inversely proportional to the pth eigenvalue λp of matrix Q
(Lw order autocorrelation matrix of R(n)), The transformer noise online monitoring and active control
system which includes software platform and hardware equip-
1 ment is designed and constructed by the devised transformer
τp = . (14)
2μw λp ANC system and the new FXLMS algorithm. It consists of

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on December 23,2022 at 03:54:40 UTC from IEEE Xplore. Restrictions apply.
ZHAO et al.: NEW FXLMS ALGORITHM WITH OFFLINE AND ONLINE SECONDARY-PATH MODELING SCHEME FOR ACTIVE NOISE CONTROL 6439

Fig. 15. Function module of the transformer noise online monitoring

and active control system.

three function modules, as shown in Fig. 15. The data acqui-

sition, access and output module can acquire and access the
transformer noise reference signal and error signal and output
the secondary source signal with the DSP data acquisition card.
The noise measurement and analysis module can analyze the
transformer noise in the time and frequency domains and dis-
play the noise reduction in real time. The ANC module mainly
includes the new FXLMS algorithm with the introduced offline
and online SP modeling. These three modules play important
roles in the transformer noise control, and no module should be
missing.

A. Software System
The software system is installed in the computer, and its inter-
face is shown in Fig. 16. According to the proposed algorithm,
the software interface divides into two parts. The Interface 1
shown in Fig. 16(a) corresponding to the Parts 1 and 2 in the
proposed algorithm can achieve to model the primary path and
the SP offline, and calculate μwb based on (7). Besides, the Inter-
face 2 shown in Fig. 16(b) corresponding to Part 3 can achieve
to model the SP online, optimize convergence coefficient μw (n)
by the GA and reduce the transformer noise.
Fig. 16. Software interface. (a) Primary path and SP offline model soft-
B. Hardware System ware interface. (b) Online monitoring and active control for transformer
noise software interface.
The hardware system, which was constructed based on the
transformer ANC system in Fig. 3, consists of a reference mi-
crophone, error microphone, power amplifier, loudspeaker, DSP
data acquisition card and computer, as shown in Fig. 17. In ad-
dition, the DSP data acquisition is NI PCI-4461, which can
connect the computer host through card slot. It has two analog
signal input interfaces to connect the reference and error mi-
crophone respectively, and it also has two analog signal output
interfaces to connect the power amplifier with loudspeaker. All
of these interfaces are parallel. After connecting the hardware
equipment, the hardware system is established in accordance
with the layout of the hardware equipment in Fig. 17. Addition-
ally, the ANC system aims at reducing the power transformer
noise by approximately 70–90 dB with a harmonic frequency Fig. 17. Hardware system.
of 100 Hz. Thus, the selected devices should collect or emit
low-frequency and high-decibel signals.
humidity is 46%, and westerly is slight. The hardware equipment
of the transformer ANC system is installed in the experimental
V. TRANSFORMER NOISE ACTIVE CONTROL EXPERIMENT
field as shown in Fig. 18, while the relative position of each
In the realistic 50 000 KVA transformer noise reduction device is shown in Fig. 19. As is nearly symmetrical, the trans-
experiment, the temperature of the day is 3 °C, the relative former sound field can be analyzed by arranging the measuring

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on December 23,2022 at 03:54:40 UTC from IEEE Xplore. Restrictions apply.
6440 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 64, NO. 8, AUGUST 2017

Fig. 18. Active transformer noise control experiment.

Fig. 20. Online SP modeling results when the error microphone placing
at 2, 3, and 6.

TABLE VI
NOISE SOUND PRESSURE LEVEL CHANGING OF EACH POSITION WHEN THE
ERROR MICROPHONE PLACING AT 2, 3, AND 6

Position ANC OFF (dB) ANC ON (dB)

3 2 6

1 62.64 72.56 70.35 67.05

2 66.33 54.64 56.16 58.67
3 72.32 58.48 61.45 63.63
4 61.93 69.25 67.25 64.78
5 64.90 74.60 73.84 70.93
6 69.37 56.65 58.59 60.82
7 60.27 66.93 64.99 62.86
8 62.63 70.47 68.33 66.79
9 63.76 75.50 74.93 71.78

Fig. 19. Positions of the hardware devices.

positions on the one side of the symmetrical axis like the po- is closer to the secondary sound source, the noise pressure level
sition numbers 1–9 in Fig. 19. In addition, when the distance will increase larger.
of the error microphone from the secondary sound source is in- In Fig. 20, when the error microphone places in different
creased more than 2 m, the SP will increase the significant delay positions, all the SP online modeling results are smooth. In
on the secondary sound signal, which will narrow the conver- addition, during the ANC experiment, the modeling process has
gence range of the proposed algorithm. This not only will easily no fluctuation and oscillation. Therefore, the proposed algorithm
lead to divergence of the transformer ANC system, but also will can model the SP online steadily. Because the environment is
decrease the optimal convergence coefficient, and reduce the relatively stable, the modeling results always remain unchanged
convergence speed and the noise reduction. Therefore, in this as is shown in Fig. 20.
experiment, the error microphones are only placed at the po- It can be seen from Table VI when the error microphone is
sitions 2, 3, and 6, and the corresponding SP online modeling in the position 3, the transformer noise is reduced nearly 13 dB
results are shown in Fig. 20, while the corresponding changes in the noise reduction region, which means that the positions 2
of the noise sound pressure level in each position are shown in and 6 can realize the same noise reduction as the position 3
Table VI. When the error microphone is in the position 3, the at this time. The reason for this is that the similar SPs can
noise changes in positions 1–9 are shown in Fig. 21. result in similar attenuations and delays to the secondary sound
As is shown in Table VI, when the error microphone is located signal, and then the signal will have nearly the same changes
at position 2, 3 or 6, the transformer ANC system based on the in amplitude and phase. It is shown in Fig. 20 that the SPs
proposed algorithm can reduce the transformer noise of the of the positions 2, 3, and 6 have high degree of similarity in
positions 2, 3, and 6, while the noise pressure levels of other shape. Therefore, when the error microphone places at a certain
positions are increased. In this way, the noise reduction region position in the noise reduction region, the other positions in
is a semicircular area with the secondary sound source as the this region can have the same noise reduction. As for the non-
center and a radius of 2 m, and the rest region can be regarded noise reduction region, there are obvious differences on the SPs
as the non-noise reduction region where the measuring position between this region and the noise reduction region, so the noise

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on December 23,2022 at 03:54:40 UTC from IEEE Xplore. Restrictions apply.
ZHAO et al.: NEW FXLMS ALGORITHM WITH OFFLINE AND ONLINE SECONDARY-PATH MODELING SCHEME FOR ACTIVE NOISE CONTROL 6441

that in the noise reduction region, when the error microphone is

closer to the secondary sound source, the noise reduction will be
larger. The reason for this is that if the SP is shorter, it will have
smaller delay to the secondary sound signal, and then the con-
vergence rang of the proposed algorithm will be larger. With the
larger optimal convergence coefficient, the system can realize
much more noise reduction.
When the error microphone is placed at the position 3, the
noise changes of the ANC OFF and ON in each position is shown
in Fig. 21. In the noise reduction region, the transformer ANC
system based on the proposed algorithm can reduce the trans-
former noise focusing on low harmonic frequency of 100 Hz,
and the noise sound pressure level of other frequency is nearly
unchanged, which indicates that the propose algorithm can re-
duce the transformer noise stably. As for the non-noise reduction
region, the increasing of transformer noise also mainly focuses
on the low harmonic frequency of 100 Hz.
After several experiments, when the error microphone is
placed in the semicircular area with the secondary sound source
as the center and a radius of 2 m, the transformer ANC system
based on the proposed algorithm can stably model the SP on-
line, and realize an 8–15 dB reduction in total sound level and an
84.10–96.86% decrease in average sound energy density in that
semicircular area. However, there is a non-noise reduction re-
gion, where the position is closer to the secondary sound source,
and the noise sound pressure level is increased larger. Therefore,
the designed transformer ANC system now is mainly used in a
small area in the transformer station or in the residence near the
station which really needs to reduce the transformer noise.

VI. CONCLUSION
This paper proposed a new FXLMS adaptive ANC algorithm
with offline and online SP modeling. In the simulation analy-
sis, compared with the random white-noise FXLMS algorithm,
it not only costs less in hardware, but also can realize bet-
ter performance in SP online modeling and transformer noise
reduction.
The proposed algorithm adopted the GA to optimize the con-
Fig. 21. Noise changing of each position when the error microphone vergence coefficient, and realized the fastest convergence speed
placing at position 3. (a) Noise changing of the positions 1–3. (b) Noise
changing of the positions 4–6. (c) Noise changing of the positions 7–9. and the maximum noise reduction. Additionally, the effect of
the convergence coefficient on the algorithm was analyzed with
simulation and theory, which also verified the correction of the
sound pressure level in this region will increase when the error result optimized by GA.
microphone is in the noise reduction region. Moreover, when The transformer noise online monitoring and active control
the measuring position is closer to the secondary sound source, system including software and hardware was designed. In the
the noise sound pressure level is increased larger. realistic 50 000 KVA transformer noise reduction experiment,
When the error microphone places in positions 2 and 6, re- when the error microphone is placed in the semicircular area
spectively, the corresponding noise reduction is 10 and 8 dB in with the secondary sound source as the center and a radius of
the noise reduction region obviously lower than 13 dB, which 2 m, the system can stably and quickly model the SP online,
indicates that the position of the error microphone will affect and realize an 8–15 dB reduction in total sound level and an
the noise reduction. It can be seen from Fig. 19 that both posi- 84.10–96.86% decrease in average sound energy density in this
tions 3 and 6 point right to the secondary sound source, which semicircular area.
means that the deviation angle of the position between the error
microphone and the secondary sound source has no effect on the REFERENCES
noise reduction. Apparently, because the distance of the posi-
[1] Y. Champoux, B. Gosselin, and J. Nicolas, “Application of the intensity
tion 3 from the secondary sound source is 1 m smaller than that technique to the characterization of transformer noise,” IEEE Trans. Power
of position 2 (1.8 m) and position 6 (2 m), it can be concluded Del., vol. 3, no. 4, pp. 1802–1808, Oct. 1988.

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on December 23,2022 at 03:54:40 UTC from IEEE Xplore. Restrictions apply.
6442 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 64, NO. 8, AUGUST 2017

[2] A. C. Binojkumar, B. Saritha, and G. Narayanan, “Acoustic noise Tong Zhao (M’16) received the B.Sc. and
characterization of space-vector modulated induction motor drives— Ph.D. degrees from Shandong University, Jinan,
An experimental approach,” IEEE Trans. Ind. Electron., vol. 63, no. 6, China, in 2002 and 2008, respectively, both in
pp. 3362–3371, Jun. 2015. electrical engineering.
[3] I. Alimohammadi, S. Sandrock, and M. R. Gohari, “The effects of low He is currently an Associate Professor of elec-
frequency noise on mental performance and annoyance,” Environ. Monit. trical engineering at Shandong University. He
Assessment, vol. 185, no. 8, pp. 7043–7051, Aug. 2013. completed his postdoctoral research in electri-
[4] T. Wen and X. W. Zhang, “Investigate and control of power transformer cal engineering at Tsinghua University, China,
noise,” in Proc. Asia-Pacific Power Energy Eng. Conf., Mar. 2009, pp. 1–4. from 2008 to 2010. His special fields of interest
[5] P. Lueg, “Process of silencing sound oscillations,” U.S. Patent 2043416, include high-voltage engineering, condition
Jun. 9, 1936. monitoring and fault diagnostics, etc.
[6] S. M. Kuo and D. R. Morgan, Active Noise Control Systems: Algorithms
and DSP Implementation. New York, NY, USA: Wiley, 1996.
[7] K. Chen, “Active noise controller,” in Active Noise Control, 1st ed. Beijing,
China: Nat. Defense Ind. Press, 2003.
[8] C. Y. Chang and S. T. Li, “Active noise control in headsets by using Jiabi Liang received the Bachelor’s degree in
a low-cost microcontroller,” IEEE Trans. Ind. Electron., vol. 58, no. 5, electrical engineering from Shandong Univer-
pp. 1936–1942, May 2011. sity, Jinan, China, in 2015, where she is currently
[9] B. Widrow and S. D. Stearns, Adaptive Signal Processing. Englewood working toward the Ph.D. degree in electrical
Cliffs, NJ, USA: Prentice-Hall, 1985. engineering.
[10] M. T. Akhtar, M. Abe, and M. Kawamata, “A new variable step size LMS
algorithm-based method for improved online secondary path modeling
in active noise control systems,” IEEE Trans. Audio, Speech, Language
Process., vol. 14, no. 2, pp. 720–726, Mar. 2006.
[11] P. A. C. Lopes and J. A. B. Gerald, “Auxiliary noise power scheduling
algorithm for active noise control with online secondary path model-
ing and sudden changes,” IEEE Signal Process. Lett., vol. 22, no. 10,
pp. 1590–1594, Oct. 2015. Liang Zou received the B.Sc., M.Sc., and Ph.D.
[12] M. Zhang, H. Lan, and W. Ser, “Cross-updated active noise control sys- degrees in electrical engineering from Shan-
tem with online secondary path modeling,” IEEE Trans. Speech, Audio, dong University, Jinan, China, in 2004, 2007,
Process., vol. 9, no. 5, pp. 598–602, Jul. 2001. and 2011, respectively.
[13] D. Chang and F. Chu, “Feedforward active noise control with a new He is currently an Associate Professor in the
variable tap-length and step-size filtered-x LMS algorithm,” IEEE Trans. School of Electrical Engineering, Shandong Uni-
Audio, Speech, Language Process., vol. 22, no. 2, pp. 542–555, Feb. 2014. versity, with a broad research interest covering
[14] P. Davari and H. Hassanpour, “Designing a new robust on-line secondary condition monitoring and reliability analysis of
path modeling technique for feedforward active noise control systems,” electrical equipment, fault current limiting tech-
Signal Process., vol. 89, no. 6, pp. 1195–1204, Jun. 2009. nology, etc.
[15] X. Du, K. K. K. Htet, and K. K. Tan, “Development of genetic-algorithm-
based nonlinear model predictive control scheme on velocity and steering
of autonomous driving vehicles,” IEEE Trans. Ind. Electron., vol. 63,
no. 11, pp. 6970–6977, Jun. 2016.
[16] Y. Du, J. Fang, and C. Miao, “Frequency-domain system identification of
Li Zhang received the B.Sc., M.Sc., and Ph.D.
an unmanned helicopter based on an adaptive genetic algorithm,” IEEE
degrees in electrical engineering from Shan-
Trans. Ind. Electron., vol. 61, no. 2, pp. 870–881, Apr. 2013.
dong University, Jinan, China, in 2001, 2005,
[17] A. Rosado-Munoz, M. Bataller-Mompean, E. Soria-Olivas, C. Scarante, and 2009, respectively.
and J. F. Guerrero-Martı́nez, “FPGA implementation of an adaptive filter
He is currently an Associate Professor in
robust to impulsive noise: Two approaches,” IEEE Trans. Ind. Electron.,
the School of Electrical Engineering, Shandong
vol. 58, no. 3, pp. 860–870, Mar. 2011.
University, with a broad research interest cover-
[18] A. Melkonyan, A. Gampe, M. Pontual, G. Huang, and D. Akopian, “Fa- ing power systems electromagnetic compatibil-
cilitating remote laboratory deployments using a relay gateway server
ity, condition monitoring, and reliability analysis
architecture,” IEEE Trans. Ind. Electron., vol. 61, no. 1, pp. 477–485,
of electrical equipment.
Jan. 2013.

Authorized licensed use limited to: National Cheng Kung Univ.. Downloaded on December 23,2022 at 03:54:40 UTC from IEEE Xplore. Restrictions apply.