Accepted Manuscript

A narrowband active noise control system with a frequency

estimation algorithm based on parallel adaptive notch filter

Han Wang , Hongling Sun , Yunping Sun , Ming Wu , Jun Yang

PII: S0165-1684(18)30273-1
DOI: https://doi.org/10.1016/j.sigpro.2018.08.012
Reference: SIGPRO 6904

To appear in: Signal Processing

Received date: 5 March 2018

Revised date: 10 July 2018
Accepted date: 21 August 2018

Please cite this article as: Han Wang , Hongling Sun , Yunping Sun , Ming Wu , Jun Yang , A nar-
rowband active noise control system with a frequency estimation algorithm based on parallel adaptive
notch filter, Signal Processing (2018), doi: https://doi.org/10.1016/j.sigpro.2018.08.012

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

Highlights:
 adaptive notch filters are often used to estimate frequency and obtain reference
signal in narrowband active noise control systems
 conventional cascaded structures of the adaptive notch filter suffer from several
drawbacks
 the proposed parallel adaptive notch filter algorithm has good performance and
low complexity

T
IP
CR
US
AN
M
ED
PT
CE
AC

1
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

A narrowband active noise control system with a frequency estimation algorithm based on

T
parallel adaptive notch filter

IP
Han Wanga,b, Hongling Suna, Yunping Suna,b, Ming Wua,b, and

CR
Jun Yanga,b

US
AN
M
ED
PT

a
Key Laboratory of Noise and Vibration Research, Institute of Acoustics, Chinese
CE

Academy of Sciences, No. 21 North 4th Ring Road West, Beijing 100190, China,

Email: [email protected]
AC

b
University of Chinese Academy of Sciences, Beijing, China,

2
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

A narrowband active noise control system with a frequency estimation algorithm based on

parallel adaptive notch filter

Han Wanga,b, Hongling Sunb*, Yunping Suna,b, Ming Wua,b and Jun Yanga,b

a
Key Laboratory of Noise and Vibration Research, Institute of Acoustics, Chinese Academy of

T
Sciences, No. 21 North 4th Ring Road West, Beijing 100190, China

IP
b
University of Chinese Academy of Sciences, Beijing, China

CR
Abstract

US
Active noise control (ANC) instead of traditional passive method shows good low-frequency

performance especially for periodic noise signals. To obtain accurate reference signal and
AN
remove the secondary feedback effect, frequency estimation algorithms are often used in

narrowband ANC systems. In this work, a parallel adaptive notch filter (parallel ANF, PANF)
M

algorithm is proposed to obtain accurate estimation of frequency which is used for generating

reference signal in a sinusoidal form. In the proposed PANF algorithm, multiple tones are
ED

separated in the feedback reconstruction process to reduce the frequency-interfering effect. As a

PT

result, the estimation of multiple frequencies has high accuracy as well as the triangular

cascaded ANF (TCANF) algorithm and the computation burden is greatly reduced to the level
CE

of the linear cascaded ANF (LCANF) algorithm. Detailed performance analysis is conducted to

acquire the deviation of frequency estimation algorithm and the mean square error (MSE) of the
AC

ANC system. Finally, extensive simulations and narrowband ANC experiments are carried out

in a pipe system to verify theoretical results.

Keywords: Narrowband active noise control, Frequency mismatch, Frequency estimation,

Adaptive notch filter (ANF)

*
Corresponding author. Tel.: +86 010 82547500; fax: +86 010 82547500.
E-mail address: [email protected]

3
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

1. Introduction

Noise threatens human’s health both physically and psychologically. On the basis of

insulation and absorption of sound, traditional passive method sustains poor performance in low

frequency. By comparison, active noise control (ANC) technology shows excellent low-

frequency properties and is widely investigated [1-6]. In the case of narrowband active noise

control where periodic signal dominates the primary source, non-acoustic reference signals such

T
as revolving speed, acceleration signal and generated tonal signal are widely used to remove the

IP
secondary path feedback effect [7,8]. The generated tonal signal is effective, provided the

CR
primary signal is stable with fixed frequencies. However, unstable primary sources often lead to

frequency mismatch (FM) especially for those with time-varying frequencies, which may

US
decrease the noise reduction and even make the system unable to work properly in some

extreme circumstances [9]. Although the revolving speed and acceleration signal provide time-
AN
varying reference signal, these two suffer from temperature fluctuation, time-delay effect and

other factors, which may also lead to FM.

M

In order to eliminate the ill-effect caused by FM, some improvements have been made in the

narrowband ANC system involving variable step-size (VSS) algorithms [10] and frequency
ED

estimation algorithms [11-14]. The former method given in [10], suffering from unstable step-
PT

size, results in poor robustness. In [13], a minimum variance distortionless response (MVDR)

iteration algorithm is adopted to estimate the frequency of the primary signal and generate the
CE

reference signal. Although those frequency spectrum based algorithms converge fast, most of

them depend on autocorrelation matrix or its inverse which means high computation complexity
AC

for multi-frequency tracking. Another wonderful frequency tracking algorithm is based on the

extended Kalman filter [15-18]. This algorithm has fast convergence rate and provides unbiased

estimation results. However, it seems that the computation complexity of the extended Kalman

filter tracker is still somewhat high for multi-tone active control in the on-line DSP system. An

easy and widely used frequency-estimation technique is the adaptive notch filter (ANF)

algorithm [19-29]. A narrowband ANC system with a linear cascaded adaptive notch filter

4
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

(LCANF) algorithm has been adopted to reduce the harmonic noise in [11]. Besides, the

LCANF algorithm has been used in [12] to estimate multiple independent time-varying

frequencies and generate the reference signal. In the LCANF algorithm, the ANF blocks are

cascaded in a linear structure to estimate and remove frequencies one by one. However, the

estimation of multiple frequencies given by the cascaded ANF with a linear structure has been

proven to be biased [26]. For a certain ANF block which is tracking a ‘target’ frequency, some

T
other ‘unwanted’ frequency components often lead to interference. In the LCANF algorithm, the

IP
target frequency of a latter ANF block is one of the unwanted frequencies of the former ANF

CR
block, which leads to biased results. In order to estimate multiple frequencies accurately,

numerous additional notch filters are used to remove the unwanted frequencies in front of the

US
frequency-estimation ANF [23], and this method is triangular cascaded ANF (TCANF)

algorithm. Although the TCANF algorithm tracks multiple tones accurately, it suffers from
AN
complicated additional notch filters.

In this paper, a narrowband ANC system with a parallel ANF (PANF) algorithm is proposed
M

to reduce a tonal signal with multiple time-varying frequencies. Similarly to the TCANF

algorithm, all the unwanted frequencies are removed before the real tracking process. However,
ED

it is the ANC and reconstruction process rather than numerous additional notch filters that
PT

decrease these unwanted tones. The reference signal is then generated in a sinusoidal form while

the narrowband ANC method is conducted using a conventional FxLMS algorithm. Considering
CE

the entire narrowband ANC system, the PANF algorithm shows better performance than the

LCANF algorithm while the complexity is quite lower than the TCANF algorithm. This means
AC

that, on the same level of DSP system, the sampling frequency can be set higher in the PANF

algorithm than TCANF algorithm. A system with a higher sampling frequency can provide a

faster convergence rate, a potentially wider control bandwidth and a shorter time delay.

The rest of this paper is organized as follows. In section 2, a new narrowband ANC system

with a PANF algorithm is presented. Section 3 analyzes the performance of the proposed

narrowband ANC system involving the reconstruction effect, the frequency estimation results

5
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

and the noise reduction behavior. Simulations are performed in section 4 to verify the analysis

of PANF algorithm. In section 5, results of ANC experiments in a pipe system are given as a

verification of performance of the proposed narrowband ANC system with a PANF algorithm.

2. A Narrowband ANC System with a PANF Algorithm

In this section, two narrowband ANC systems both with a frequency estimation algorithm are

given. Firstly, a basic narrowband ANC system with a frequency estimation algorithm based on

T
cascaded ANF is introduced in detail. Then, a new narrowband ANC system with a PANF

IP
algorithm is presented.

CR
2.1 A Narrowband ANC System with a Cascaded ANF Algorithm

One of the narrowband ANC systems with a cascaded ANF algorithm can be illustrated in

US
Fig.1 [11][12][14]. In this paper, the primary source signal is assumed as d (n)

N
d ( n )   d i ( n )  v ( n)
AN
(1)
i 1

where di (n) represents the i-th component of the primary signal dominated by frequency f i ,
M

v(n) is the system noise which is often assumed as a zero-mean white noise. Considering that
ED

the acoustic signals have real values in a real-world system, the noise with a frequency f i is

assumed to be the superposition of two complex tonal signals with equivalent amplitudes but
PT

opposite phases

di (n)  Cdi [e j (i ndi )  e j (i ndi ) ]

CE

(2)

where Cdi is the amplitude, di represents the phase and i is a dimensionless normalized
AC

angular frequency which can be written as

fi
i  2 (3)
fs

where the sampling frequency is defined as f s . The reference signals are generated in the form

of two complex tonal signals xai and xbi with unity amplitude and  / 2 phase shift

6
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

xai (n)  e( jˆi n ) (4)

xbi (n)  e j (ˆi n /2) (5)

where the estimated angular frequency of the primary signal ˆ i is obtained by the frequency

estimation algorithm. The secondary path at each frequency is expressed as

H si  Chi e jhi (6)

T
where Chi and hi are the parameters of amplitude and phase. In this paper, xai (n) and xbi (n)

IP
are used to express the reference signal filtered by the estimated secondary path parameter Hˆ si

CR
Hˆ si  Cˆ hi e jhi
ˆ
(7)

US
xai (n)  Hˆ si xai (n)

xbi (n)  Hˆ si xbi (n)

(8)

(9)
AN
The adaptive filter iteration process is updated using the conventional FxLMS algorithm
wai (n  1)  wai (n)   xai (n)e(n) (10)
M

wbi (n  1)  wbi (n)   xbi (n)e(n) (11)

ED

yi (n)  wai (n) xai (n)  wbi (n) xbi (n) (12)

where wai (n) and wbi (n) are the filter weights,  is the step size, the error signal e(n) is
PT

obtained by the error sensor and the output signal in each frequency channel is represented as
CE

yi (n) . The total acoustic output at the error point can be expressed as

N
y(n)   yi(n) (13)
AC

i 1

where yi(n) is the output signal yi (n) affected by H si

So far, the narrowband ANC system has been introduced except the frequency estimation

method. In addition, a frequency estimation algorithm based on the cascaded ANF algorithm

7
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

will be given in detail. First, the estimation of primary signal is reconstructed in a feedback

method

N
dˆ (n)   yˆi(n)  e(n) (14)
i 1

where yˆi(n) is the estimated secondary-path output which has been filtered by Hˆ si . By

considering that the filtered output signal yi(n) has a real value in a real-world application, in

T
IP
this paper, only the real component of the estimated signal yˆi(n) is employed. Then, the ANF

CR
based frequency estimation algorithm can be given as follows, where the estimated primary

signal dˆ (n) is used as the input signal u (n) of frequency estimation process. First, u (n) is

US
enhanced by an all-pole section D(q 1 ) and enhanced tonal signal can be written as xh (n)

xh (n)  u(n) D(q 1 ) (15)

AN

where q 1 is a time delay operation (e.g. x(n)  q 1  x(n  1) ). Then xh (n) is filtered by an
M

all-zero section N (q 1 ) to remove this sinusoidal signal. This method is called notch filter and

the total transfer function can be written as

ED

1 N (q 1 )
H (q )  (16)
D(q 1 )
PT

In this paper, two types of notch filter are introduced, namely, a direct form and a simplified
CE

lattice form [21, 22, 24, 26]. The notch filter with a direct form can be expressed as

N (q 1 ) 1  2k0 q 1  q 2
H D (q 1 )   (17)
D(q 1 ) 1  2 k0 q 1   2 q 2
AC

where the pole zero contraction factor  is a positive value which is smaller than unity, and k0

is defined as the negative of the cosine of the target angular frequency

k0   cos( ) (18)

Similarly, the transfer function of the lattice notch filter is

8
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

1 N (q 1 ) 1  2k0 q 1  q 2
H L (q )   (19)
D(q 1 ) 1  (1   )k0 q 1   q 2

where  and k0 follow the same definition as in (17).

In order to estimate the frequency, an adaptive frequency estimation algorithm is appended to

the notch process, and this algorithm is also called an adaptive notch filter (ANF). In [24, 25,

27], several adaptive frequency estimation algorithms have been presented based on gradient or

T
Gauss-Newton methods. In this paper, an adaptive frequency estimation algorithm based on the

IP
Burg algorithm is adopted [19, 26]. First, two intermediate variables A(n) and B(n) are

CR
employed with given original values

B(n)   B(n  1)  (1   ) xh2 (n  1) (20)

US
A(n)   A(n  1)  (1   ) xh (n  1)[ xh (n)  xh (n  2)] (21)
AN
where xh (n) is the enhanced tonal signal and the forgetting factor  is a between 0 and 1.

After that, the estimated angular frequency ̂ is obtained as follows

M

A(n)
k0 ( n)   (22)
B ( n)   2
ED

kˆ0 (n)   kˆ0 (n  1)  (1   )k0 (n) (23)

PT

ˆ  arccos[kˆ0 (n) / 2] (24)

where k0 (n) is the original estimation while kˆ0 (n) is the smooth value. The smoothing factor
CE

 is a positive value and is smaller than 1, and  is a parameter with a small value to ensure
AC

that the denominator is positive.

To estimate multiple frequencies simultaneously, higher-order ANF and cascaded ANF have

been used in [20]. The latter implementation has several advantages over the former, so the

cascaded structure is adopted more often. Fig. 2 shows a linear cascaded ANF algorithm, where

the ANF sections are cascaded in a linear structure. To be precise, the output signal of the

9
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

former i-th ANF is used as input signal to the later (i+1)-th ANF, and the sinusoidal signals are

estimated and removed one by one.

The target frequency in the (i+1)-th ANF section is one of the unwanted frequencies of the i-

th section which leads to inaccurate tracking results (the detailed analysis of bias is given in

section 3).The deviation of frequency estimation results leads to FM, which severely influences

the ANC performance. To obtain accurate estimation of multiple frequencies, a TCANF

T
algorithm is given in [23]. In the TCANF algorithm, the input signal is first filtered by

IP
additional notch filter sections with previously estimated frequencies to remove the unwanted

CR
frequencies before the frequency estimation method is conducted. As an example, Fig. 3

illustrates a TCANF algorithm with N=4. By removing the influence of multi-tonal interference,

US
the estimation results of TCANF are accurate. However, the computational burden of TCANF is

greatly increased due to the complex triangular structure.

AN
2.2 New Narrowband ANC System with a PANF Algorithm

In this paper, we introduce a new narrowband ANC system with a frequency estimation
M

technique based on the PANF algorithm. As shown in Fig.4, the proposed narrowband ANC

system is similar to the conventional narrowband ANC system, only differs in the PANF
ED

algorithm. Rather than one single estimated primary signal, the proposed PANF algorithm

reconstructs N estimated signals, namely, d̂1 … dˆN . Fig. 5 illustrates the block diagram of
PT

proposed PANF algorithm where the estimated signals are obtained as the superposition of the
CE

estimated secondary-path signal yˆ i and the filtered error signal e i 1 (n) (where e 0 (n) is set as
AC

the original e(n) )

dˆi (n)  yˆi(n)  e i 1 (n) (25)

In order to avoid the drawback that multiple ANF sections converge at the same frequency, the

error signal is filtered by the notch filter in a linear structure to remove the tonal signal one by

one which is similar to the structure in LCANF

10
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

ei (n)  ei 1 (n)[ Ni (q 1 ) Di (q 1 )] (26)

where Ni (q 1 ) is the i-th all-zero section and Di (q 1 ) is the i-th all-pole section. Both of

Ni (q 1 ) and Di (q 1 ) are set using the estimated frequency given by the previous ANF

sections. Then the input of i-th ANF section is just adopted as the i-th estimated signal

u i (n)  dˆi (n) (27)

T
So, at the beginning of the system convergence process, the proposed PANF algorithm works in

IP
the similar way as the LCANF algorithm. In the steady state, however, its performance is close

CR
to the TCANF algorithm. The complex structure of TCANF has been simplified by taking full

advantage of the narrowband ANC system. Rather than removing unwanted frequencies by the

US
additional notch filters (e.g. the way of TCANF algorithm), in the proposed system, the

unwanted frequencies in the error are sharply reduced by ANC method in the steady state and
AN
estimated output signals with other frequencies will not be added into each estimated signal

dˆi (n) . Benefitting from this independent frequency estimation process, each ANF section
M

tracks a target frequency without the influence from others, and as a result, the estimation of
ED

frequency is more accurate.

3. Performance Analysis
PT

In this section, detailed performance analysis of new narrowband ANC system with a PANF

algorithm is given along with comparison with narrowband ANC system based on the cascaded
CE

ANF algorithms. The analysis is given in three parts, involving the reconstruction of the

estimated signal, the deviation of frequency estimation and noise reduction of the narrowband
AC

ANC system. Before the total analysis, we shall make the following assumptions:

Assumption 1:

The secondary-path modeling has been precisely completed in the off-line method.

Assumption 2:

11
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

The frequencies of multi-tone noise have enough distance (e.g. over 3Hz) and change slowly

so that the conventional narrowband ANC system works.

3.1 Analysis of Reconstruction

First, the reconstruction method of narrowband ANC system with a cascaded ANF algorithm

is analyzed. From (2), (4-9), (12) and (14) the reconstructed primary signal can be expressed as

N
dˆ (n)   yˆi(n)  e(n)

T
i 1

IP
N
  [ yˆi(n)  yi(n)  yi(n)]  e(n)
i 1

CR
N
 d (n)   [ yˆi(n)  yi(n)]
(28)
i 1
N
 d (n)   [Cˆ hi wai (n)e j (ˆi n hi )  Cˆ hi wbi (n)e j (ˆi n hi  /2)  Chi wai (n)e j (ˆi n hi )  Chi wbi (n)e j (ˆi n hi  /2) ]
ˆ ˆ

i 1
N
US
 d (n)   {[ wai (n)  wbi (n)e j /2 ] [Cˆ hi e jhi  Chi e jhi ]  e jˆi n }
i 1
ˆ
AN
Considering the acoustic signal is a real-valued one, it is the real part of filtered output

signal that works in the physical system. As a result, (28) can be rewritten as
M

N
dˆ (n)  d (n)  Re{{[ wai (n)  wbi (n)e j /2 ][Cˆ hi e jhi  Chi e jhi ]  e jˆi n }}
ˆ
(29)
i 1
ED

Cˆ hi e jhi  Chi e jhi  0 and the

ˆ
By considering the assumption 1, it can be found that
PT

{[w (n)  wbi (n)e j /2 ][Cˆ hi e jhi  Chi e jhi ]  e jˆi n })  0 . Then the
 ˆ
superposition term Re( ai
i 1
CE

primary signal can be approximated as its estimation which is also used as the input of

frequency estimation algorithm

AC

u(n)  dˆ (n)  d (n) (30)

Obviously, the frequency estimation result based on the reconstructed signal is reliable in

LCANF algorithm and TCANF algorithm.

Similarly, in the reconstruction process of the proposed narrowband ANC system with a

PANF algorithm

12
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

ui (n) dˆi (n)

 yˆi(n)  ei 1 (n)
(31)
 yˆi(n)  yi(n)  yi(n)  ei 1 (n)
 yi(n)  ei 1 (n)

So, the reconstruction procedure ensures that the power of output signal yi(n) is estimated

correctly (assumption 1). Moreover, as the secondary signals with unwanted frequencies have

T
not been added into dˆi (n) , the power of unwanted tones are significantly reduced in ui (n)

IP
provided the unwanted tones have been removed by ANC process in the error signal in the

CR
steady state (assumption 2), which benefits accurate estimation (detailed analysis given in

section 3.2 ).

US
3.2 Discussion of Deviation of frequency estimation Process Based on ANF Algorithms

In [26], it has been shown that the MSE of frequency estimation result based on ANF has the
AN
following form

MSEk =  2 +  2 (32)
M

where the  2 is the mean square of evaluated error which has been shown to be determined by
ED

(1   ) and the signal noise ratio (SNR) in the system, and the  2 represents the bias term

which is attributed by interference. Considering the target frequency is a time-varying one, (32)
PT

has to be rewritten as

MSE k =  2 +  2
CE

(33)

where deviation term  is defined as

AC

 =  + (34)

where  represents the time delay deviation. In this paper, we make the assumption that the

terms (1   ) and (1   ) are small so that the frequency-estimation error caused by variance is

small [26]. Then the bias and delay are analyzed in detail as follows.

3.2.1 Bias

13
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

From the analysis in [30], it has been shown that (20)(21)(22) can be rewritten in the form of

B(n)   B(n  1)  2(1   ) xh2 (n 1) (35)

yh (n)  xh (n)  2k0 (n  1) xh (n  1)  xh (n  2) (36)

k0 (n)  k0 (n  1)  [(1   ) B(n)]xh (n  1) yh (n) (37)

where yh (n) is the result of all-zero section N (q 1 ) , and it can be expressed that

T
yh (n)  N (q 1 ) xh (n) (38)

IP
Obviously, (35-37) is similar to the update approach based on the normalized gradient. In the

CR
case that there is only one frequency which changes slowly (assumption 2), and then the

US
estimation of k0 (n) converges in the following condition

E[ xh (n  1) yh (n)]  0 (39)
AN
by putting (17) (or (19)) and (38) (where k0 is used as kˆ0 (n  1) ) into (39), and making the

assumption that k0 ()  kˆ0 ()  kˆ0 in steady state, it can be rewritten as
M

E{xh (n  1)[ xh (n)  2kˆ0 xh (n  1)  xh (n  2)]}  0 (40)

ED

In steady state, (40) can also be expressed as

Rx (0)kˆ0  Rx (1)  0 (41)

PT

where Rx (l )  E[ xh (n) xh (n  l )] , from[2], Rx (0) , Rx (1) can be derived as

CE

Cs2 (1   2 ) 2
Rx (0)   (42)
(1   2 )[(1   2 ) 2  4 2 kˆ02 ]
2
2 D(e j )
AC

Cs2 cos( ) 2 2 kˆ0

Rx (1)   (43)
(1   2 )[(1   2 )2  4 2 kˆ02 ]
2
2 D(e j )

for the ANF algorithm with a direct form, and

Cs2 2
Rx (0)   (44)
(1   2 )(1  kˆ02 )
2
2 D(e j )

14
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

Cs2 cos( )  2 kˆ0

Rx (1)   (45)
(1   2 )(1  kˆ02 )
2
2 D(e j )

for the ANF algorithm with a lattice form, where  2 is the variance of system noise which is

usually assumed to be a Gaussian white one and Cs2 represents the variance of sinusoidal signal

with angular frequency  . Then, further derivation can be conducted by putting Rx (0) and

T
Rx (1) into (41). In the direct algorithm, it can be obtained that

IP
Cs2 [kˆ0  cos( )] kˆ0 (1   ) 2
 0 (46)

CR
(1   )[(1   2 ) 2  4 2 (1  kˆ02 )]
2
2 D(e j )

where the autocorrelation function follows expression (42) and (43).

US
In addition, for the case that the input signal is with multiple frequencies (e.g. i=1, 2, …P),

such as the former ANF sections of the LCANC algorithm, (46) has to be rewritten as
AN
P
Csi2 [kˆ0  cos(i )] kˆ0 (1   ) 2
{ 2
} 
(1   )[(1   2 )2  4 2 (1  kˆ02 )]
0 (47)
i 1 2 D(e ji )
M

where Csi2 is the variance of tone with angular frequency i . On the basis of analysis and
ED

experiment, in those applications where the system has high SNR, the influence of the second

term in (47) can be neglected. Then we have the approximated equation

PT

P
Csi2 [kˆ0  cos(i )]
 i 1
ji 2
0 (48)
D (e )
CE

In a system with low SNR, the noise term can never be ignored, and this case is analyzed in
AC

appendix A. By contrast, the equation of k̂0 in the lattice ANF algorithm can be obtained in the

same form of (48) without any approximation as the noise term has been offset in the process of

putting the form of autocorrelation function (44)(45) into equation (41). Obviously, the results

in (48) suggest that the estimation k̂0 is not equal to any one of  cos(i ) . The estimation of

frequency is close to one of the target frequency but always biased. Further, as the form of

15
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

formulation (48) is a higher-order one, the explicit solution of bias is harder to obtain and some

approximation should be adopted for further research. In this paper, we make the assumption

that the solution is close to  cos(i )

kˆ0   cos(i )  k (49)

where k is much smaller than unity ( cos(i ) O(1) , k 1 ) and for simplicity, i is assumed

T
as 1. Then (48) can be rewritten as

IP
P
Csi2 [cos(i )  cos(1 )  k ]
 2
0 (50)
D(e ji )

CR
i 1

In this paper, the ANF algorithm with a direct form is taken as an example (the derivation of the

US
lattice algorithm is easier). Then the denominator of each component can be expressed as

D(e ji )  1  2kˆ0 e ji   2e2 ji

2 2
AN
(51)
2
 Di  2 e ji i
M

where

Di  1  2cos(i ) e ji   2e2 ji (52)

ED

i  cos(i )  cos(1 )  k (53)

PT

From (51), after some straightforward derivation, we obtain

2
D(e ji )  Di  4 2 i2  4 (1   )2 cos(i )i
2
(54)
CE

In addition, (54) can be simplified as

AC

2
D(e j1 )  D1  4 2 k 2  4 (1   )2 cos(1 ) k
2
(55)

for i  1 , where D1  (1   )2 [(1   )2  4 cos 2 (1 )] and

2

2
D(e ji )  (1   )2 [(1   ) 2  4 cos 2 (i )]  4 2 i2  4 (1   ) 2 cos(i )i (56)

for i  1 . For small (1   ) , we make the following assumption

16
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

(1   )2 1 (57)

By considering that k 1 , we assume that k and (1   )2 have small value so that

 k2 k

 k (1   ) 2
k (58)
 k (1   ) 2 (1   ) 2


Applying the relation of (58), by putting (55)(56) into (48) and neglecting the higher-order

T
terms with k 2 and (1   )2 k , we have the approximation expression in (48)

IP
P P

CR
[(1   )2  4 cos 2 (1 )](1   ) 2  {Csi2 [cos(i )  cos(1 )]  4 2 [cos(l )  cos(1 )]2 }
i 2 l  i ,l 1
(59)
P
C 2
k  4 [cos(i )  cos(1 )]  0
2 2

US
s1
i 2

Then the expression of k can be obtained after some straightforward derivation

AN
P P

{Csi2 [cos(i )  cos(1 )]  4 2[cos(l )  cos(1 )]2}

k  [(1   ) 2  4 cos 2 (1 )](1   ) 2  i  2 P
l  i ,l 1

Cs21  4 2 [cos(i )  cos(1 )]2 (60

M

i 2

[(1   ) 2  4 cos 2 (1 )](1   ) 2 P Csi2


4 2
 
i  2 Cs1[cos(i )  cos(1 )]
2
ED

)
PT

For small (1   ) , (60) leads to the approximated form of bias

 1 P
Csi2
   k  (1   )2 sin 2 (1 )  
CE

(61)
i  2 Cs1[cos(i )  cos(1 )]
2

From (61), after some analysis, it can be found that k O[(1   )2 ] , which meets the
AC

assumption (58).

The expression of k in an ANF algorithm with a lattice form can be derived in the similar way

as the direct form

(1   )2 sin 2 (1 ) P Csi2

k 
(1   )2
 
i  2 Cs1[cos(i )  cos(1 )]
2
(62)

17
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

For small (1   ) , we have the approximated form of bias in the ANF algorithm with a lattice

form

 1 (1   )2 sin 2 (1 ) P Csi2

  k   2 (63)
4 i  2 Cs1[cos(i )  cos(1 )]

Until now, the bias forms of both direct algorithm and lattice algorithm for a system with

interference have been given in detail. Further, the bias of a certain section (k-th) in the

T
cascaded ANF algorithm or the PANF algorithm can be obtained as

IP
 N 2
Cski
   2 2
  

CR
(1 ) sin ( ) , direct form

k
i 1, C 2
skk [cos( i )  cos( k )]
ik
k   (64)
 (1   ) sin (k ) 
2 2 N 2
Cski
  , lattice form

2
where Cski

4
ik
US
i 1, Cskk [cos(i )  cos(k )]
2

represents the variance of tones with i in section k .

AN

In the LCANF algorithm, we assume that the detection process is in the order of 1 to  N .
M

Although the former k-1 target tones have been removed in the k-th ANF section, there are still

N-k unwanted frequencies

ED

 0, i  k

2
Cski 2
Cskk  2 2 (65)
ˆ ˆ
Cdi Cdk , i  k

PT

So the bias form of k-th ANF section can be simplified as

 Cˆ di2
CE

 (1   ) sin (k )   ˆ 2
2 2
, direct form
 i  k 1 Cdk [cos(i )  cos(k )]
k   (66)
 (1   ) sin (k ) 
2 2 N
Cˆ di2
 ˆ2
AC

 , lattice form
 4 i  k 1 Cdk [cos(i )  cos(k )]

where k=1,2…N-1, and the last section (k=N) keeps unbiased. Obviously, from the formation of

bias, the value of bias can be adjusted by setting  . However, the value of this parameter can

never be set too close to unity otherwise it leads to slow convergence. So, the deviation caused

by bias can never be completely eliminated in a cascaded ANF algorithm with a linear structure.

18
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

In TCANF algorithm, as the power of unwanted frequencies has been reduced by the

additional notch filters, we have

2
Cski 2
Cskk  0, i  k (67)

and then the frequency estimation result becomes unbiased as each term in (64) has a value of

zero. On the basis of assumption 1 and assumption 2, the functions of these additional notch

filters are realized in the reconstruction process of the proposed system in steady state. More

T
precisely, it is assumed that yˆ k is dominated by ˆ k and the variance of tonal signal has been

IP
greatly reduced to a low level in the error signal, which means that (67) is also satisfied in input

CR
ui (n) of PANF sections. Therefore, in steady state the proposed algorithm can be seen as a

US
simplified TCANF algorithm where the additional notch filtering process has been implemented

in the FxLMS algorithm and reconstruction procedure.

AN
3.2.2 Delay

From (48), in steady system with a constant frequency, it is clear that the estimation of k0 has
M

the form of

kˆ0   cos( )
ED

(68)

However, the time delay deviation can never be neglected in the system with a time-varying
PT

frequency. In this paper, we make the following assumption

kˆ0   cos(   ) (69)

CE

where  is the real frequency, and  is the time delay angular frequency caused by varying
AC

tone. Further, from (20) and (21), (69) is assumed to be

 i
cos(  iv )
cos(   )  i 1

(70)

i 1
i

19
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

where v is the change rate of the angular frequency which is assumed to be a constant [13]. For

small v (we can set sampling frequency to make v small), the first order approximation of

Taylor series expansion of (70) can be expressed as

 

  i cos( )   iv sin( )

i

cos(   )  i 1

 i 1


 i
 i
(71)

T
i 1 i 1

v  sin( )
 cos( ) 

IP
(1   )

CR
Moreover, the time-delay deviation  can be obtained

  k0  kˆ0

US
  cos( )  cos(   )
v  sin( )
 
(1   )
(72)
AN

So far, the deviation of k̂0 can be given in (34)(64)(72), and then the error of frequency
M

estimation can be expressed as

fs
f  arccos[cos( )  ( + )]  f
ED

(73)
2

where f s is the sampling frequency, and f is the target frequency.

PT

3.3 Analysis of Noise Reduction

I n [9], it has been derived that the noise reduction in the narrowband ANC system with FM
CE

effect can be expressed as

AC

Cˆ h2Ch2 cos( h )  cos(   h ) 2 2 h  

NR  [  ]  cos 2 ( ) (74)
2  2cos(  ) Cˆ hCh 2

where NR represents the noise reduction value in the mean square,  is the deviation of

f
angular frequency   2 , and h is the phase error of the estimated secondary path ˆhi .
fs

Although form (74) is detailed and accurate, some parameters such as h and Ch2 can hardly be

20
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

acquired in a real-world application. Considering that in the system with small FM, accurate

secondary-path modeling and large  [9], it can be derived that

cos( h )  cos(   h )  sin( h )

  (75)
Cˆ hCh Cˆ hCh

and

2 h   Cˆ h2Ch2  2

T
2
cos ( ) 1 (76)
2 2  2cos(  )

IP
In this paper, we make the assumption that in the system with small FM and accurate

CR
secondary-path modeling, the noise reduction performance of a certain narrowband ANC

system with constant large step size  follows the approximated form,

NR  US f
1  cos(  )
(77)
AN
where  f is a const value at the frequency f, which can be obtained in experiment. Furthermore,

by putting (73) into (77), it can be derived that

M

f
NR  (78)
1  cos{arccos[cos( )  ( + )]  }
ED

In a multi-tone ANC system with a frequency estimation algorithm based on the TCANF
PT

algorithm or the PANF algorithm, the deviation caused by unwanted frequencies has been

reduced while the delay term can never be completely removed. In order to reduce time-delay
CE

effect, the sampling frequency has to be set as high as possible. So, the computation complexity

of frequency estimation algorithm has to be analyzed. From Fig.2, it is obvious that there are N
AC

pole sections D(q 1 ) and N-1 zeros sections N (q 1 ) in LCANF algorithm. As shown in Fig.5,

these numbers have increased to (2N-1) and (N-1) in the proposed PANF algorithm. It shows

that compared to the LCANF algorithm, the computation burden of the PANF algorithm

increases slightly and keeps in the level of N order. By contrast, as shown in Fig. 3, the TCANF

algorithm has (N(N+3)/2-1) pole sections D(q 1 ) and (N(N+1)/2-1) zeros sections N (q 1 ) ,

21
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

which means that complexity is in the level of N2 order. Therefore, under the same computation

condition, the proposed narrowband ANC system meets a higher sampling frequency and works

better than system with the TCANF algorithm on multi-tone noise reduction especially for those

primary signals with numerous target frequencies (e.g. over 10 frequencies).

4. Simulations

T
In this section, computer simulations are conducted as the verification of performance

IP
analysis of frequency estimation process based on ANF algorithms. In the first simulation, the

CR
theoretical bias of ANF both with direct form and lattice form are compared with the

simulations. The second simulation is conducted to verify the analysis of time delay deviation.

US
Then the third and last simulations are given as a comparison between the proposed PANF

algorithm and other two multi-frequency estimation algorithms based on ANF algorithm,
AN
namely the LCANF algorithm and the TCANF algorithm. The sampling frequency is set as 5

kHz in each simulation.

M

As shown in Fig.6 and Fig.7, the first simulation is given and compared with the theoretical

bias. More specifically, the target tone is a constant one with frequency 100Hz while an
ED

unwanted frequency varies from 125Hz to 500Hz. The power of the target tone and the
PT

unwanted tone is set to be equal. So, this simulation is conducted to verify the analysis of the

bias caused by an unwanted tone. In Fig.6, two types of ANF sections (the direct form and the
CE

lattice form) are compared in the same figure both with two different pole zero contraction

factors. The forgetting factor  is 0.999 in each situation which satisfies the condition that
AC

(1   ) is small. From Fig.6, it is obvious that the theoretical bias agrees well with the

simulations. In addition, as shown in Fig.6, the value of bias changes dramatically if the gap is

less than 100Hz while this gradient becomes gentle at large gap (over 200Hz). In order to show

the influence of SNR, this value ranges from 5dB to 20dB in Fig.7. From the diagram of Fig.7,

the SNR value has little impact on the frequency estimation performance, which meets the

22
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

approximation in section 3. In general, the bias in a direct form ANF is around 4 times of the

lattice one under the same condition.

The second simulation is carried out to illustrate the analysis of time delay deviation. In this

simulation, a sinusoidal signal with time-varying frequency is tracked by the ANF algorithm. In

Fig.8, simulations are given and compared with theoretical value of the delay deviation, where

the pole zero contraction factor is chosen as   0.99 while the forgetting factor  is set

T
between 0.99 and 0.998. The change rate of frequency varies from 0.5Hz/s to 8Hz/s in this

IP
simulation, which means that the value of v is between 1.2566 107 rad/sample and

CR
2.0106 106 rad/sample. From Fig.8, it is obvious that theoretical prediction matches well

US
with the simulations only with a little difference at high change rate. It is also shown that the

deviation value almost increases linearly with the change rate.

AN
The third simulation is conducted as an overall tracking-behavior comparison of three ANF

structures (LCANF, TCANF and the proposed PANF). The basic ANF form is chosen as a

direct one with parameters   0.993 ,   0.997 and   0.995 in each algorithm. The
M

parameters have been adjusted carefully to keep fast convergence as well as accurate
ED

estimation. As illustrated in the time-frequency diagram Fig.9 (a), all these three algorithms

show fast convergence as well as small frequency-estimation variance. In addition, detailed

PT

tracking-behavior is given in Fig.9 (b). The TCANF algorithm and the proposed algorithm show
CE

good performance where the tiny error mainly arises from the time-delay effect. As a

comparison, the LCANF algorithm has lager deviation because of the frequency-estimation bias.
AC

The last simulation is a test of the estimation performance in more challenging conditions. In

Fig.10 (a), two close tones with the same amplitude are tracked and separated where the

minimum gap is only 4 Hz. The ANF section is adopted as a direct form. By considering that

the freuqency gap is very small, the pole zero contraction factor  is chosen as 0.999 to

provide an accurate distinguishability. The system is added with zero-mean white noise to keep

the system SNR=20dB. As shown in Fig.10 (a), all three algorithms distinguish two frequencies

23
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

well. Further, the proposed algorithm and the TCANF algorithm show better estimation

accuracy than the LCANF algorithm especially when the tones are close. Another comparison is

given in Fig.10 (b), which shows the tracking performance in a two-tone tracking system where

the tones have very different amplitudes (10:1). The parameters are chosen as   0.995 ,

  0.99 and   0.9 in the 1-st ANF section and   0.995 ,   0.995 and   0.9 in the

2-nd ANF section for all three algorithms, and system SNR is set as 30dB. All the factors are

T
set carefully to ensure that the ANF sections converge fast with small estimation

IP
variance. Obviously, all these three algorithms converge fast. Moreover, it seems that the

CR
estimation error of the proposed PANF algorithm is slightly smaller than the LCANF algorithm

and the TCANF algorithm in steady state.

5. Experiments US
AN
In this section, experimental results in a pipe noise reduction system are given to verify the

theoretical analysis and compare the performance of the proposed narrowband ANC system and
M

the narrowband ANC system with a LCANF algorithm. First, experiments are conducted to

verify approximated form of the noise reduction effect with FM given in section 3. Next, some
ED

experimental results of multi-tone noise reduction with three constant frequencies are given and

compared with theoretical value. Then, similar work is carried to verify the analysis of time
PT

delay effect. Ultimately, control experiments are conducted to compare the performance of

narrowband ANC system with a LCANF algorithm and proposed narrowband ANC system
CE

where the primary signal is multiple tones with three time-varying frequencies.

As shown in Fig.11, all of the experiments are completed in a circular pipe system with a
AC

compound muffler. The muffler is an active-passive hybrid one with 50mm inner radius and

135mm outer radius which is the same one as our previous work [4]. The length of the muffler

is 500mm while the whole system has a length of 1500mm which is long enough to make sure

that the acoustic wave is in the form of plane wave. There are two acoustic sources in the

experiment system. The primary source is set at the front-end of the pipe system while the

24
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

secondary source at the after-end of the muffler. The acoustic signal is acquired in the form of

sound pressure by the microphone. The microphone is chosen as a B&K 4189-021 which is

located at the after-end of the muffler just close to the secondary source. Signal processing is

performed online by a controller with a development board Omap-L137 based on tms320c-

6747. For convenience, the sampling frequency is set as 5000Hz and the system step size  is

fixed at 0.05 in each experiment. The secondary path modeling is conducted in an off-line

T
method before the experiments.

IP
The first experiment is performed to verify the approximated analysis of noise reduction of

CR
narrowband ANC system with FM effect in section 3 and measure the parameter  f . The

verified experiment is conducted at three frequencies, specifically, 100Hz, 200Hz, and 300Hz.

US
As shown in Fig.12 (a)(b)(c), the noise reduction curve is fitted with the experiment data by

adjusting  f carefully in the standard of the least mean-square error. The noise reduction curve
AN

given by theory agrees well with the experiment data at all three frequencies except some tiny
M

errors, and these three noise reduction parameters 100Hz , 200Hz , 300Hz are adopted in all of the

following experiments. The theoretical noise reduction curve has a maximum value of over
ED

60dB at three frequencies while this value sharply decays to 10-13dB when there is a 1Hz FM

in system.
PT

In the second experiment, a comparison is given between narrowband ANC system with a
CE

LCANF algorithm and the proposed system in three-tone control with fixed frequencies. The

ANF form is chosen as the direct structure. Fig. 13 (a) shows the behavior of the narrowband
AC

ANC system with a LCANF algorithm, where experimental three-tone noise reduction is

compared with theoretical values to verify the analysis in section 3. To be more specific, the

sound pressure level (SPL) of primary tone with frequency 100Hz is 102.8dB at the error point

while this value is 101.3dB in the tone with frequency 200Hz and 100.5dB with frequency

300Hz. The ANF parameters are set as   0.99 and   0.997 , and the noise reduction

parameters  f have been obtained in the first experiment. The frequency estimation sequence is

25
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

set as 100Hz, 200Hz, 300Hz in the LCANF algorithm. As shown in Fig.13 (a), the narrowband

ANC system with a LCANF algorithm works unevenly in each tone. Although the frequency

estimation results is unbiased at 300Hz which leads to good noise reduction effect, the biases at

100Hz and 200Hz cause poor control performance with only around 20dB reduction which

agrees well with the theoretical prediction. By contrast, Fig.13 (b) illustrates satisfactory

behavior of proposed system in all three tones. The noise reduction is over 54dB in each tone

T
which suggests that the error of frequency estimation is smaller than 0.005%.

IP
The third experiment is carried out to verify the analysis of time delay effect. In this

CR
experiment, a tone changes around the frequency 200Hz with a speed raging from 0.5Hz/s to

3Hz/s. The pole zero contraction factor  is still set as 0.99 while the noise reduction

US
parameter 200Hz is provided in the first experiment. As shown in Fig.14, the experimental noise

reduction results match well with theoretical prediction. Specifically, the control performance
AN
rapidly decreases if the time delay is short. In addition, the forgetting factor  plays a key role in

the time delay effect, where the noise reduction is positively correlated with term (1   ) .
M

The last experiment is a comparison study of multiple tones ANC with time-varying
ED

frequencies. The experiment condition is similar to the second experiment, only differs in the

time-varying frequencies. Fig. 15 illustrates the three-tone signal with control and without,
PT

where the frequency-domain intervals of these tones are around 100Hz and all three tones are

time-varying signals. The changes of frequencies are continuous with slow rates between -2Hz/s
CE

to 2Hz/s only except an abrupt jump from 190Hz to 210Hz at about 8s. The frequency

estimation sequence is set from the lower frequencies to the higher in the LCANF algorithm. By
AC

comparing the time-frequency diagrams of Fig.15 (b) and (c), it is clear that the proposed

narrowband ANC system works better than the system with a LCANF algorithm on noise

control. Suffering from the frequency interference, the results of the frequency estimation at the

lower frequencies (around 100Hz and 200Hz) in the system with a LCANF algorithm are

always biased. So, the noise reduction is only about 17-20dB in some cases in the narrowband

26
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

ANC system with a LCANF algorithm. In contrast, in the proposed system with a PANF

algorithm, this value is improved to around 30dB or over as the frequency estimation can be

conducted more accuratly. In addition, the unsteady state caused by the abrupt jump has little

influence on both two systems as the sound pressure has been fast reduced to satisfactory level

(specifically, within 0.5s), which means good robustness. Further, as shown in Fig.16 (a)(b), the

instantaneous behavior of both systems at 18s are given to compare the theoretical and

T
experimental system performance. To be more specific, there are three peaks in the uncontrolled

IP
signal with the frequencies around 103Hz, 202Hz , 300Hz, and the SPL is 103.0dB, 101.5dB

CR
and 101.3dB. The theoretical values of noise reduction are 18.1dB, 17.7dB and 27.9dB at three

peaks in narrowband ANC system with a LCANF algorithm while these values are 30.3dB,

US
35.0dB and 27.9dB in the proposed system. The experimental results meet well with the

prediction with only tiny error at some frequencies. Compared with the narrowband ANC
AN
system with a LCANF algorithm, the frequency estimation deviation in the proposed system has

been decreased to the level governed by time delay effect as the bias caused by unwanted
M

frequency has been sharply reduced.

.
ED

6. Conclusion
PT

In this paper, a new narrowband ANC system is presented with a PANF frequency

estimation algorithm. On the account of the independent tracking, the proposed PANF
CE

algorithm shows accurate estimation, which improves the noise reduction system performance.

As shown in performance analysis, the frequency estimation deviation can be reduced to the
AC

level dominated by the time delay effect in the proposed PANF algorithm. Further, compared

with TCANF algorithm, the computational burden in the proposed PANF algorithm stays

around the level of LCANF algorithm which means the delay deviation  can be limited as a

tiny value (e.g. 0.1Hz or smaller) by adjusting sampling frequency. Simulations and experiment

results are given with discussion to show the system effect and verify the performance analysis.

27
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

In simulations, the analysis of bias and delay effect has been proven. Next, the experimental

results show that the proposed narrowband ANC system has fast convergence as well as low

error in both steady and unsteady state.

In general, based on the PANF algorithm, the proposed narrowband ANC system reduces the

bias caused by unwanted frequencies while the delay effect can also be decreased by increasing

the sampling frequency. As a result, it seems that the ill-effect caused by FM can be solved,

T
provided the secondary-path modeling is sufficiently precise. However, in some applications,

IP
secondary path might be a time-varying one which means that the secondary-path assumption in

CR
section 3 will lose efficacy. In the future, a secondary-path online modeling technique will be

presented to resolve this problem.

US
AN
Appendix A

In system with low SNR, the noise term in (47) can never be neglected. In order to analyze
M

the influence of noise, we follow the assumption of (58). By putting (51)(52)(53) into (47)

directly, removing higher order terms, after some straightforward derivation, we have
ED

(1   )4 2 sin 2 (1 )[(1   ) 2  4 cos 2 (1 )](1   ) 2

P P
 {C [cos(i )  cos(1 )]  4 2 [cos(l )  cos(1 )]2 }
PT

2
si
i2 l  i ,l 1
P (A1)
[(1   ) 2  4 cos 2 (1 )](1   ) 2  cos(1 )(1   )2 2  4 2 [cos(i )  cos(1 )]2
CE

i2
P
Cs21 k (1   )4 2 sin 2 (1 ) 4 2 [cos(i )  cos(1 )]2  0
AC

i2

As a result, the expression of k can be given as

[(1   ) 2  4 cos 2 (1 )](1   ) 2 P Csi2 2(1   )3 cos(1 ) 2
bias   k 
4 2
 
i  2 Cs1[cos(i )  cos(1 )]
2

(1   ) Cs21
 1 P
Csi2 (1   )3 cos(1 ) 2
 (1   ) 2 sin 2 (1 )  { } 
i2 Cs21[cos(i )  cos(1 )] Cs21

(A2)

28
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

From (A2), it is clear that the estimation results of direct algorithm are still biased in the

condition of poor SNR even though the unwanted frequencies have been removed by a TCANF

or PANF algorithm. Fortunately, by considering that the noise term has a higher order of

(1   ) , we can adjust (1   ) to ensure that the noise term is much smaller than any

component of the superposition. Then the frequency-estimation error can be made arbitrarily

small.

T
Acknowledgements

IP
This work is supported by National Natural Science Foundation of China(Grant

CR
No.11474306) and Youth Innovation Promotion Association, Chinese Academy of Sciences.

References

US
[1] S. Elliott, Signal Processing for Active Control, Academic Press, New York, 2000.
[2] S. Haykin, Adaptive Filter Theory, Prentice-Hall, Englewood Cliffs, NJ, 2002.
[3] F. R. Yang, M. Wu, J. Yang, and Z. Kuang, A fast exact filtering approach to a family of affine
AN
projection-type algorithm, Signal Process. 101, (August) (2014) pp. 1-10
[4] H. L. Sun, F. Y. An, M. Wu and J. Yang, Experiments on performances of active-passive hybrid
mufflers, Proc. Int. Congr Sound Vib. (ICSV22), Florence, Italy, (July) (2015) pp. 1-7
[5]Kai Chen, Roshun Paurobally, Jie Pan and Xiaojun Qiu, Improving active control of fan noise with
M

automatics spectral reshaping for reference signal, Applied Acoustics 87 (2015) pp142-152.
[6] L. S. Xiao, M. Wu and J. Yang, A new efficient filtered-x affine projection sign algorithm for active
control of impulsive noise, Signal Process. 120 (March) (2016) pp. 456-461
[7] S. M. Kuo, M. Tahernezhadi and W. Hao, Convergence analysis of narrow-band active noise control
ED

system, IEEE Trans. Circuits. Syst. II, Analog Digit. Signal Process. 46 (February) (1999) pp. 220-223.
[8] L. Luo, J.W. Sun, B.Y. Huang, and D. D. Quoc, Efficient combination of feedforward and feedback
structures for nonlinear narrowband active noise control, Signal Process. 128 (November) (2016) pp. 494-
503.
PT

[9] H. J. Jeon, T. G. Chang and S. M. Kuo, Analysis of frequency mismatch in narrowband active noise
control, IEEE Trans. Acoust, Speech, Signal Process. ASSP-18 (August) (2010) pp. 1632-1642.
[10] S. M. Kuo, and S. P. Nallabolu, Analysis and correction of frequency error in electronic mufflers
using narrowband active noise control, in Proc. IEEE Int. Conf. on Control Applicat. (October) (2007)
CE

pp.1353-1358.
[11] S. Kim and Y. Park, On-line fundamental frequency tracking method for harmonic signal and
application to ANC, J. Sound Vib. 241 (April) (2001) pp. 681-691.
[12] B. Mazeaud and M. A. Galland, A multi-channel feedback algorithm for the development of active
AC

liners to reduce noise in flow duct applications, Mech. Syst. Signal Process. 21 (February) (2007) pp.
2880-2899.
[13] H. J. Jeon, T. G. Chang, S. Yu, and S. M. Kuo, A narrowband active noise control system with
frequency corrector, IEEE Trans. Audio, Speech, Lang. Process. 19 (May) (2011) pp. 990-1002.
[14] T. W. Wang, W. S. Gan, and S. M. Kuo, New feedback active noise control system with improved
performance, in Proc. IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP) (May) (2014) pp.
6662-6666
[15] B. La Scala, R. Bitmead, and B. G. Quinn, An extended Kalman filter frequency tracker for high-
noise environments, IEEE Trans. Signal Processing, 44 (1996) pp. 431–434.
[16] B. La Scala and R. Bitmead, Design of an extended Kalman filter frequency tracker, IEEE Trans.
Signal Processing, 44 (1996) pp. 739–742.

29
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

[17] S. Bittanti and S. M. Savaresi, On the parametrization and design of an extended Kalman filter
frequency tracker, IEEE Trans. Autom. Control, 45 (Sepember) (2000) pp. 1718-1724.
[18] S. Corbetta, A. Dardanelli, I. Boniolo, S. M. Savaresi and S. Bittanti, Frequency estimation of
narrow band signals in Gaussian noise via Unscented Kalman Filter, in IEEE Conf. on Decision and
Control, (2010) pp. 2869-2874.
[19] J. I. Makhoul and L. K. Cosell, Adaptive lattice analysis of speech, IEEE Trans. Circuits. Syst.
CAS-28 (June) (1981) pp. 494-499.
[20] D. V. Bhaskar Rao and S. Y. Kung, Adaptive notch filtering for the retrieval of sinusoids in noise,
IEEE Trans. Acoust., Speech, Signal Process. 32 (August) (1984) pp. 791-802.
[21] A. Nehorai, A minimal parameter adaptive notch filter with constrained poles and zeros, IEEE Trans.
Acoust., Speech, Signal Process. 33 (February) (1985) pp. 983-996.
[22] N. I. Cho, C. H. Choi and S. U. Lee, Adaptive line enhancement by using an IIR lattice notch filter,

T
IEEE Trans. Acoust., Speech, Signal Process. 37 (April) (1989)
[23] T. Kwan, and K. Martin, Adaptive detection and enhancement of multiple sinusoids using a cascade

IP
IIR filter, IEEE Trans. Circuits Syst. 36 (July) (1989) pp. 937-947.
[24] J. F. Chicharo, and T. S. Ng, Gradient-based adaptive IIR notch filtering for frequency estimation,
IEEE Trans. Acoust., Speech, Signal Process. 38 (May) (1990) pp. 769-777.

CR
[25] P. A. Regalia, An improved lattice-based adaptive IIR notch filter, IEEE Trans. Signal Process. 39
(September) (1991) pp. 2124-2128.
[26] N. I. Cho and S. U. Lee, On the adaptive lattice notch filter for the detection of sinusoids, IEEE
Trans. Circuits. Syst. II, Analog Digit. Signal Process. 40 (July) (1993) pp. 405-416

US
[27] J. F. Chicharo, and H. H. Wang, Power-System Harmonic Signal Estimation And Retrieval for
Active Power Filter Applications, IEEE Trans. Power Electron. 9 (November) (1994) pp. 580-586.
[28] S. Bittanti, M. Campi, and S. M. Savaresi, Unbiased estimation of a sinusoid noise via adapted notch
filters, Automatica, 33 (1997) pp. 209-215.
AN
[29] R. Zhu, F. R. Yang and J. Yang, A RLS-based lattice form complex adaptive notch filter, IEEE
Signal Process. Lett. 23 (February) (2016) pp. 217-221.
[30] L. Ljung and T. Soderstrom, Theory and Practice of Recursive Identification. MA: MIT, Cambridge,
1983.
M
ED
PT
CE
AC

30
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

Figure Captions

x ( n )
Hˆ Si bi
FxLMS
Hˆ Si
xai (n)
y1(n)
d ( n)
xai (n) 
wai (n) yi (n) yi(n)  e( n )
 H Si  
xbi (n) wbi (n) y(n)
Hˆ Si yN (n)
yˆi(n)

T
̂1 yˆ1(n)  yˆ N (n)
yˆ (n)

IP
Signal ˆ i Cascaded

dˆ (n)
Generator ANF

CR
ˆ N

Fig. 1 Block diagram of the narrowband ANC system with a cascaded ANF algorithm

US
AN
M

u ( n) xh1 (n) xhi (n)

1 1 1 xhN (n)
D1 (q 1 ) N1 (q 1 ) Di (q 1 ) N i (q 1 ) DN (q 1 )
ED

Frequency Frequency Frequency

Estimation Estimation Estimation

̂1 ˆ i ANFi ˆ N
PT

Fig. 2 Block diagram of the LCANF algorithm

CE
AC

31
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

u (n) N (q 1 ) N 2 (q 1 ) N 3 (q 1 ) 1 ̂4
1 Frequency
D1 (q 1 ) D2 (q 1 ) D3 (q 1 ) D4 (q 1 ) Estimation

N 4 (q 1 ) 1 ̂3
Frequency
D4 (q 1 ) D3 (q 1 ) Estimation

N 4 (q 1 ) N 3 (q 1 ) 1 ̂2
Frequency
D4 (q 1 ) D3 (q 1 ) D2 (q 1 ) Estimation

N 4 (q 1 ) N 3 (q 1 ) N 2 (q 1 ) 1 Frequency
̂1
D4 (q 1 ) D3 (q 1 ) D2 (q 1 ) D1 (q 1 ) Estimation

T
IP
Fig. 3 Block diagram of the TCANF algorithm, N=4

xbi (n)

CR
Hˆ Si
FxLMS
Hˆ Si
xai (n)
y1(n) d ( n)
xai (n)

xbi (n)
wai (n)

wbi (n)

yi (n)

Hˆ Si
H Si
yi(n)

yN (n)


y(n)

 US e( n )
AN
yˆi(n)
̂1 yˆ1(n) yˆ N (n)

Signal ˆ i Parallel
M

Generator ANF

ˆ N
ED

Fig. 4 Block diagram of the proposed Narrowband ANC system with a PANF algorithm
PT

ei 1 (n) eN 1 (n)
e( n ) N1 (q 1 ) e1 (n) N i (q 1 )
D1 (q 1 ) Di (q 1 )
CE

yˆ1(n)
1

Frequency

u1 (n) D1 (q 1 ) Estimation
̂1
AC

yˆi(n)
1
 Frequency

ui (n) Di (q 1 ) Estimation
ˆ i

yˆ N (n) 1

Frequency
DN (q 1 ) Estimation
ˆ N
u N ( n)

Fig. 5 Block diagram of the proposed PANF algorithm

32
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

T
IP
Fig.6 Bias of ANF algorithm,   0.999 ,SNR=10dB.(a)Direct form   0.99 (b) Lattice form

CR
  0.985 (c) Lattice form   0.99 (d) Direct form   0.997

US
AN
M
ED

Fig.7 Noise influence on the bias,   0.99 ,   0.999 (a) Direct form (b) Lattice form
PT
CE
AC

33
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

Fig.8 Delay deviation of ANF based on Burg algorithm,   0.99 . (a)   0.99 ,(b)

  0.995 , (c)   0.997 , (d)   0.998

T
IP
CR
US
AN
(a)
M
ED
PT
CE

(b)
AC

Fig.9 Frequency estimation of four tones with additive Gaussian white noise of SNR=20dB. (a)

Overall behavior (b) Detailed behavior

34
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

T
IP
(a)

CR
US
AN
M
ED

(b)

Fig.10 Frequency estimation in challenging conditions (a) Two close tones, additional white
PT

noise with SNR=20dB (b) Two tones with very different amplitudes, additional white noise

with SNR=30dB
CE
AC

Fig.11 Narrowband ANC experiment system

35
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

(a)

T
IP
CR
US
AN
(b)
M
ED
PT

(c)
CE

Fig.12 Performance of narrowband ANC system with Frequency mismatch. (a) 100Hz

(b)200Hz (c) 300Hz

AC

36
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

T
IP
(a)

CR
US
AN
M
ED

(b)

Fig.13 Tracking and control of multiple tones (a) Narrowband ANC system with a LCANF
PT

algorithm (b) The proposed system with a PANF algorithm

CE
AC

Fig.14 Time delay effect

37
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

T
IP
CR
(a)

US
AN
M
ED

(b)
PT
CE
AC

(c)

38
 ACCEPTED MANUSCRIPT
Put your running title here: Elsevier General template for review

Fig. 15 Tracking and control of multiple tones with time-varying frequencies. (a)Without

control (b) Narrowband ANC system with a LCANF algorithm (c) The proposed narrowband

ANC system with a PANF algorithm

T
IP
CR
(a)
US
AN
M
ED
PT
CE

(b)

Fig. 16 Noise control performance at 18s. (a) Narrowband ANC system with a LCANF
AC

algorithm (b) The proposed narrowband ANC system with a PANF algorithm

39