You are on page 1of 8

484 IEEE TRANSACTIONS ON ACOUSTICS,

SPEECH,
ANDSIGNAL
PROCESSING, VOL. ASSP-25, NO. 6, DECEMBER 1977

We assume that the complex signal s(m,n ) is zero outside of REFERENCES


the square B. P. Bogert, M. J. R. Healy, and J . W. Tukey, “The quefrency
alanysis of time series for echoes,” in Proc. Symp. Time Series
Analysis, M. Rosenblatt,Ed. New York: Wiley, 1963, pp.
209-243.
A. V. Oppenheim, R.W. Schafer, and T. G. Stockham, Jr., “Non-
linear filtering of multiplied and convolved signals,” Proc. IEEE,
V O ~ .56, pp. 1264-1291, Aug. 1968.
A. V. Oppenheim and R. W. Schafer, Digital Signal Processing.
and that M is significantly less than N , say M is 10 percent of Englewood Cliffs, NJ: Prentice-Hall, 1975, pp. 480-531.
D. E. Dudgeon,“Twodimensional recursivefiltering,”Sc.D.
N . Using the algorithm described in Section I11 and using thesis, Dep. Elec. Eng., M.I.T., May 1974.
Tribolet’s algorithm [7] to do the phase unwrapping, we find M. P. Ekstromand J. W. Woods,“Two-dimensionalspectral
that we need to do 4N + M t 2 N-point FFT’s plus N 2 t N factorizationwithapplications in recursivedigitalfiltering,”
IEEE Trans. Acoust., Speech, Signal Processing, vol.ASSP-24,
complex-logarithm evaluations. The number of real multiplies pp. 115-128, Apr. 1976.
inan N-pointFFT is roughly proportional to 2N log2N, D. E. Dudgeon, “The existence of cepstra for twodimensional
so that the number of real multiplies in the DFT method of rational polynomials,” IEEE Trans. Acoust.,Speech, Signal
Processing, vol. ASSP-23, pp. 242-243, Apr. 1975.
computing the cepstrum is proportional to 8 N 2 log2N J. M. Tribolet, “A new phaseunwrappingalgorithm,” ZEEE
(roughly) with complex-logarithm evaluations going up as Trans. Acoust.,Speech, Signal Processing, vol.ASSP-25,pp.
N2 +N. 170-177, API. 1977.
N. Brenner, “Three Fortran programs that perform the Cooley-
Using the recursive method and taking advantageof the TukeyFouriertransform,”LincolnLab.Tech.Rep.1967-2,
knowledge that s(m,n ) is zero outside of the M X M square, July 28, 1967.
the recursive method needs M 2 - 1 complex and 2 real multi- P. Pistor,“Stabilitycriterionforrecursivefilters,” IBM. Res.
Develop., vol. 18, pp. 59-71, Jan. 1974.
plies to compute a cepstral point. We desire N 2 such points, S. Winograd, “On computingthediscreteFouriertransform,”
so the number of real multiplies for the recursive method goes Proc. Nat. Acad. Sci. US., vol. 73, pp. 1005-1006, Apr. 1976.
up as N 2 ( 4 M 2 - 2). (Because of edge effects, this estimate is M. P. Ekstromand R. E. Twogood, “A stabilitytestfor 2-D
recursivedigitalfiltersusing thecomplex cepstrum,” in Proc.
somewhat pessimistic, but it indicates the order of the de- 1977 ZEEE Int. Con$ Acoustics, Speech, and Signal Processing,
pendence of the numberof multiplies on M and N . ) pp. 535-538.

. , . ,

Adaptive Noise Canceling Applied to


Sinusoidal Interferences

Abstract-This paper investigates a new method for eliminating sin- I. INTRODUCTION


usoidal or other periodic interference corrupting a signal. This task is
typicallyaccomplishedbyexplicitlymeasuring
interferenceandimplementinganotchfiiter
the frequency of the
at that frequency. The
method proposed herein uses an adaptive filter to eliminate the inter-
T HE ELIMINATION of a sinusoidal interference corrupt-
ing a signal is typically accomplished with a fured notch
filter tuned to the frequency of the interference. A very nar-
ference. The procedure is called adaptive noise canceling and is appli- row notch is usually desired in order to filter out the inter-
cable when an auxiliary reference input available
is containing the inter- ference without distorting the signal.However,if the in-
ferencealone. The referenceinput is filtered insuchawaythatit
closely matches the interfering sinusoid, and is then subtracted from
terference is not precisely known, and if the notch isvery
the primary input leaving the signal alone. narrow, the center of the notch may not fall exactly over the
The results of thisresearch show that when a sum of sinusoids is ap- interference. Also there are many applications where the
plied to an adaptive filter, the filter converges to a dynamic solution in interfering sinusoid drifts slowly in frequency. A fured notch
which the weights of the filter are timevarying. This time-varying solu- cannot work here at all unless it is designed wide enough to
tion implements a tunable notch filter, with a notch located at each of
the referencefrequencies.Whenused in noise-cancelingapplications,
cover the range of the drift, with the consequent distortion of
this adaptive notch filterprovides a simple alternative to other methods the signal. In situations such as these it is often necessary to
of tracking and eliminating sinusoidal interferences. measure in some way the frequency of the interference, and
then implement anotch filter atthatfrequency. However,
estimating the frequency of several sinusoids embedded among
Manuscript received June 24, 1976;revisedMay 2, 1977. signal and other broad-band noises can require a great deal of
The author is with the Department ofElectricalEngineering,Uni-
versity of Houston, Houston, TX 77004. calculation. An alternative simpler method proposed here can
GLOVER: NOISE CANCELING 48 5

The adaptation algorithm most often used to set the weights


ofthe filter is the least-mean-squares (LMS) algorithm [3]
given by the equations
T
Yk = WkXk

l l
RELATED
NOISE €k = d k - Y k

jb I 1
I

D
REFERENCE
I

REFERENCE
NOISE
F FILTER
Y

ERROR
wk+l = wk t (YfkXk.
LMSis an iterative gradient-descent algorithm that usesan
estimate of the gradient on the mean-square error surface to
(1)

INPUT
seek the optimum weight vector at the minimum mean-square
Fig. 1. Adaptive noise canceling (ANC) system.
errorpoint. The term f k x k represents the estimate ofthe
beused when a reference forthe interference is available, negative gradient, and the adaptation constant 01 determines
thereby making explicit measurement of its frequency un- the step size taken at each iteration along that estimated nega-
necessary. This reference is adaptively filtered to match the tive-gradient direction. The true negative gradient is given by
interfering sinusoids as closelyas possible, allowing them to the expected value of ekxk. If cr is chosen properly, such that
then be subtracted out. This application of adaptive filtering small steps are taken, adaptation noise due to error in the gra-
is called adaptive noise canceling (ANC) and is fully described dient estimate isaveraged out. Widrow has shown [3] that
for broad-band interferences in [ 11 and [2] . when adapting with LMS on stationary stochastic processes,
In the broad-band case the solution for the adaptive filter is the expected value of the weight vector converges to an op-
a constant set of filter weights. Any motion in the weights timal solution w*.
after convergence to this solution is considered to be simply With sinusoidal inputs, however, this solution is not relevant,
noise in the adaptive process. The results presented herein and a new approach to the analysis is needed. The suggested
show that when the reference is sinusoidal, significant time- approach is to select a different set of inputs and outputs as
varying components in the weights give rise to a tunable notch shown in Fig. 2. Our goal is to show that when xk is a cosine
filter centered atthe frequency of each reference sinusoid. xk = C COS (w,kT t 6)
This adaptive notch filter can be a very useful methodfor
the dashed box can be approximated, if certain conditions are
automatically tracking and eliminating sinusoidal interferences.
met, by a linear, time-invariant (LTI) filter G(z) from input f k
11. THE ADAPTIVE
NOTCH FILTER to output Y k . With this model, the sinusoidal input xk, the N
weights of the filter, and the weight-update equation of the
A. Adaptive Noise Canceling
LMS algorithm are all lumped together in the transfer func-
An adaptive filter is used in ANCas shown in Fig. 1. The tion G(z).
primary consists of the signal plus noise, s t no. The reference With this goal in mind let us compute the ztransform of the
input is the related noise n l . The reference n is filtered to output Y(z).Fig. 3 is a block diagram representation of (1).
match no and thensubtracted from the primary. The error For generality, consider theith element ofa general input
signal to the adaptation algorithm is therefore the output of x-vector, Xik, with arbitrary phase angle 0
the ANC system. For more background on adaptive filtering
and the use of ANC applied to broad-band interferences, the Xik = c COS (w,kT f 6i)
reader is referred to [ 11 .

B. ANC Appliedto Sinusoidal Interference


Following the path through the ith weight in Fig. 3, we obtain
The analysis of ANC applied to sinusoidal interferences re-
quires an approach different from that used in [ I ] and [2] . It
is necessary to look into the adaptive fiter and the particular
adaptation algorithm being used.
The adaptive filter used to perform the noise canceling is a =-C [,io. E(ze-jwrT) ,-jei E(ZeiWyT
transversal filter. The filter input x(t) is sampled at a rate 1/T 2 13. (3)
to give the discrete “reference input” x(”) =‘Xk.l This se- E(ze-iwyT) is E(z) rotated counterclockwise around the unit
quence is then applied to an N-stage tapped delay line (TDL). circle through an angle w, T . E(zejw rT> represents a clockwise
The values at the N taps of the TDL at time k constitute the rotation.
elements ofthe reference N-vector xk. The outputs of the The response between 01 and Wi(z)is that of a delayed digital
taps are weighted and summed to give the adaptive filter out- integrator representing the equation W i , k + 1 = W j k -t QEkXik,
T
put Yk = WkXk. and is combined in the inner dashed box as U(z) = l/(z - 1).
The ith weight is then
‘Throughout this paper a subscript of k will represent the time index.
Any other subscript will represent an element of avector,matrix,
series, etc.
486 IEEE TRANSACTIONS ONACOUSTICS,SPEECH, AND SIGNAL PROCESSING, VOL. ASSP-25, NO. 6 , DECEMBER 1977

I---------- 1 .
I
I x, - ADAPTIVE
I .
I (SINUSOID)
F ILTER
I
I

I 'I
I
I
I
I I
I _ _ - _ - -I
(b) I
Fig. 2. Newmodel for the
adaptivesystem.
(a)
Block diagram. Fig. 3. Block diagram of the LMS adaptive algorithm.
(b) Transfer-function diagram.

The ith contribution to the output at Y ( z )is Equation (5) can now be simplified as
yi(z) = 2 {wjk Xik 1

where TV represents the undesired time-varying terms of (5b).


The exponential coefficient in (6) has unity amplitude and
has been included in the TV terms. From inspection of these
equations, it is clear that a fair measure of the relative strengths
of the TV and TI components in the output Y(z), as a func-
tion of theinput frequency and adaptive filter parameters,
would be the ratio [P(O, T , N ) IN. ] When this ratio is small
compared to unity, we would hope that the TV components
would be insignificant compared to the TI components. In
fact, when PIN = 0, we are left with only the TI components.
P/N is plotted in Fig. 4 as a function of f,T(wr = 27rf,) for
several values of N . It is clear that as long as f r T is not too
The terms in (5a) represent the time-invariant (TI) part of the close to 0 or 0.5, there are several points where PIN = 0 or
response from E(z) to Y ( z ) , since only frequencies of E(z) ap- PIN 0. In general, there are two ways to reduce P(w,T, N)/N
pear at the output. The time-varying terms in (5b) introduce for a given reference w,. First, the sampling rate 1/T can be
at Y(z) unwanted, frequency-shifted components ofE(z). adjusted so that the product f r T is well placed between 0 and
Since our goal is to have only G(z), let us look at the ex- 0.5 at a low point on the PIN curve. Second, the number of
ponential summations multiplying the time-varying terms. weights N in the adaptive filter can be increased to obtain a
Recall that €Ji is an arbitrary phase shift for the ith element of better PIN curve. If the proper choice of parameters is made,
the x-vector. For the particular case where the xjk are ob- the transfer function between E(z) and Y(z) is approximated
tained from the taps of a TDL, we have by an LTI filter.

Oi=0 - w r T [ i - 1 1 . C, The Time-InvariantApproximation


Substituting for €Ji,the summations are easily found to be If we assume that one of these methods isused so that
PIN zs 0, then the TV terms in (8) are insignificant, and we
obtain the time-invariant transfer function
i= 1

where the function /3 is defined as


GLOVER: NOISE CANCELING 481

- --
1-
N = 8 WEIGHTS
3 /1

- 1-

1- N.18 WEIGHTS
# _ _ -
A / - .
v - - - - - - " -P
.5

-1-

1
N=32 WEIGHTS

-1

N.64 WEIGHTS

1 c

Fig. 5. Experimentalsweepfrequencyplot of H(z). (a) 01 = 0.05,


Fig. 4. Plots of p/N function for various N. C=l,N=2. (b)a=O.O02,C=l,N=2.

-- E [ l t This is the transfer function for a 2nd-order digital notch


4 Ze-jWrT - 1 zejwrT - 1

NaC2
1~ filter at the frequency 0,. The zeros of H(z) are at the poles
of G(z), z = e t i W r Tprecisely
, on the unit circle. For slow
adaptation rates (narrow-bandwidth notch, aswill be seen
(Z COS wrT - 1)
- 2 shortly) such that NaC2/4 << 1, the pole locations are ap-
(9) proximated by
Z' - 22 COS w,T t 1 *

G(z) has poles on the unit circle at z = e tiwr.T, and a zero at


z = l/cos wrT. G(z) can be viewed as a pair of iritegrators that
have been rotated to fwrT. Actually, it is the input that is
shifted in frequency by an amount fw,due to the first multi- The zeros lie on the unit circle at frequencies + w r ,with the
plication by the reference sinusoid, integrated at dc, and then poles a distance approximately NaC2 /4 behind them radially
shifted back again by the second multiplication. One can see toward the center of ,the circle. Near the frequency w = w,,
the similarity with techniques for obtaining resonant filters by H(z) can be approximated by the nearby pole and zero
using two low-pass filters and heterodyning with sine and
cosine at the resonant frequency [4].

D. Notch-Filter Response
Now that we have justified the model in Fig. 2, the transfer The 3-dB bandwidth.(BW) is then obtained by finding the two
function of interest in ANC is that from d k to e k , which will points on the unit circle which are 4 times as far from the
be denoted H(z). (See the Appendix for a discussion ofJ(z), pole as they are from the zero. The result is found geomet-
the corresponding transfer function from d k to yk.) H(z) is rically as approximately
now easily found to be
BW= -
NQ!c2rad/s.
2T
Under the assumption of approximate linearity, anex-
perimental sweep frequency response of the actual H(z)
was obtained for a reference frequency of f r T = 0.2, and the
result is shown inFig.S(a). Note that the value o f f = 0 is
greater thanunity.Infact, using the final-value theorem
- Z' - 22 COS w r T + 1 for z transforms on (lo), the value at f = 0 is found to be
z 2 - 2 (1 -
NaC2
q ) z c o s w r T t (1 - y) * (10)
1/ [ 1 - (NaC2/4)] . A similar response, but with much smaller
adaptation rate Q! and thus much smaller BW, is shown in Fig.
488 IEEE TRANSACTIONS ON ACOUSTICS,
SPEECH, AND SIGNAL PROCESSING, VOL. ASSP-25, NO. 6, DECEMBER 1917

therefore consists of the sum and difference frequencies be-


tween a d and a,,filtered by (aC/2) u(z). If w, and (.+ are
close, since (aC/2) U(z) is a very low-pass transfer function,
the difference frequency component in the weightswillbe
much stronger than the sum frequency component.
While each weight is principally sinusoidal at the difference
frequency, the individual weights differ according to the factor
e f ioi . Since fora TDL Bi = 6 - a,[i - 11 T , neighboring
weights differ by the phaseangle a,T. Therefore, plotting
the “instantaneous impulse response,” the values of the Wik
as a function of i at a single time instant k = k o , would yield
64 a sinusoid at frequency a,.
wts
The combination of these two effects, the movement of each
weight sinusoidally at ( a d - a,),and the instantaneous re-
sponse at a,,gives the series of impulse responses shown in
Fig. 6 . The instantaneous sinusoidal pattern of frequency w,
“moves” to the left at the difference frequency ( a d - a,)(or
to the right if a,> a d ) .
In summary, when wd f w,, the weights indeed do have a
zero constant solution. They also have a nonzero time-varying
Fig. 6. Series of instantaneous impulse responses in a 64-weight TDL. solution which should not be considered as simply noise in the
There are five iterations between plots. (f, > fd). adaptation process. The time-varying weights “modulate” the
reference frequency a,and heterodyne it into theprimary fre-
5(b). For this smallervalue of a, the value at f = 0 is also
quency a d ,thereby creating the notch filter effect.
much closer to unity.

E. Conclusion
Iv.
EXTENSIONTO A SUM OF SINUSOIDS AS
THE REFERENCE
Thus far we haveseen that when the reference is a pure
The analysis for a single-reference sinusoid is easily extended
sinusoid, and /3/N X 0, a linear, time-invariant notch-filter re-
to a sum of reference sinusoids, resulting in a multiple-notch
sponse is obtained from the primary input D(z) to the ANC
filter. Suppose the reference is a sum of M sinusoids
output E@). Three important characteristics of this adaptive
notch filter are as follows. M
1) It is tunable, since the null point of the notch is deter- xk = cmCOS (amkT + 6,)
m=1
mined by, and thus will move with, the reference frequency.
2) The notch can be made very sharp at precisely the fre- then
quency of reference-input sinusoid. An example of the im-
M
portanceof this will be seen in an adaptive noise-canceling Xik = c, COS (am[k - i t 11 T + 6,)
experiment described later. m= 1
3) The effects of the notch are clearly not due to the usual
M
constant-weight converged solution. If the weights were con-
= Cm COS (a,k T + Sim) (12)
stant, they could only pass the input reference frequency to
m=1
the y output, and could not cause any cancellation of nearby
frequencies. where
111. SOLUTIONOF THE WEIGHTS OF THE FILTER eim.=em - am[i- 11 T.
Now that wehaveseen the input-output behavior of the Proceeding as in (3)-(8), we obtain for the multiple-sinusoid
adaptive system, let us lookatthe time-varying solution of case
weights which brings about this behavior. We have actually NOr M

-
already derived the z transform for the weights in (4). In
order to study the implications of this solution, still under the
assumption that p/N 0, consider that dk , and therefore E k ,
is sinusoidal at frequency a d . The pole-zero plotfor E(z)
~ ( z=)-~ ( z )
4
M

m=1 n=1
[U(ze-’”m T > t U(zei”m T>l

M
m=1

2 T , N ) [TV]
would indicate poles at z = e ‘iw T . Since E(ze-iwr T , repre-
n#m
sents a counterclockwise rotation of E(z) through anangle
a , T , and E(zeiwrT) a clockwise rotation, the pole-zero plot
of lE(ze-iarT) ei9i + E(zei”,T) would show poles at
. -
+(a, t ad)T and at &(a, - ad) T . This rotated spectrum is
then filtered through (aC/2) U(z) to give Wi(z). Each weight where TV again
~ ” renresents undesired time-varying
. . -components.
GLOVER: NOISE CANCELING 489

1-
ah/,=
2T

IHk)I /
c,=.707
\
C*=l.O
6 4= 9 6 4=
T

4 fT
0 .l S6 .20 .3 .4 .5
FG. 7. Experimentalsweepfrequencyplot of sum oftwo sinusoids
as reference. Q! = 0.05/16,N = 32.

C I-------- - 1

ACC EL.€ROME1 3- PRIMARY I


I
I
I MOTORdl AT C

\ I MOTORtt2 AT C
64 W t s
- I (BANDPASS OUTPUT)
I
L - - - - - - l
MULTIPLE NOTCH AND
BANDPASS F I L T E R S
Fig. 8. Separating vibrations from two variable-speed motors.

Herewehave both sum and difference frequencies of the generalcase of sinusoids into the adaptive filter of an ANC
reference sinusoids to contend with in order to make the p/N system. The previous single-frequency results can be viewed
ratios small. Large sum frequencies may be handled as before, as a special'case of the multiple-frequency TDL results.
by decreasing T or increasing N so that p [ ( o m t wJ2) T,
N] / N = 0. However, the difference frequencies may already v. APPLICATIONS
OF THE ADAPTIVE
NOTCH FILTER
be so small that they are far to the left on the PIN curve. De-
In actual practice, some special configurations of the adap-
creasing T would push them' even further to the left,increasing
tive notch filter are sometimes used. An example of a 2-weight
p [ ( w m - w,/2) T, N ] IN. Therefore, when the reference sinus-
notch filter was introduced in [l] . If there are only two com-
oids are close together, often the only way to eliminate all of
ponents to the x k vector, and the delay between them :is
the unwanted TV termsjs to increase N , the number of weights
equivalent to a 90" phase shift atthe reference frequency,
in the adaptive filter. That is, a long filter is required to give
then they represent a sine and cosine pair, and p/N is easily
good resolution between adjacent frequencies.
shown to be exactly zero. Further, if instead of a TDL there
If we again assume that each of the 6's in (13) is such that
aie N = 2~ components of x k consisting of M sine-cosine
PIN = 0, the transfer function from E(z) to Y(z) is
pairs at different frequencies, then again all p/N ratios are ex-
actly zero. This configuration is preferred when separate
references can be obtained for each of several interfering
-r m=1 sinusoids.
The H(z) transfer function from D(z) to E(z) is, therefore, An example of the use of a single long adaptive filter to ob-
1
tain multiple-notch filters is shown in Fig. 8. At pointsA and
B two variable-speed motors operating at different but .very
close frequencies generate odd harmonics (for purposes of this
-t m=1
simulation) through the chassis on which they are mounted.
Due .to multipath and other reasons;,each harmonic received
H(z) represents a series of notches, one notch at each reference at point C on. the chassis has an unknown and unpredictable
frequency am.The 3-dB BW at each wm is amplitude and phase. The goalis to separate the harmonics
NaC& coming from each motor so that the vibrations at point C due
BWm = -rad/s. to each motor can be determined. Fig. 8 is the ANC config-
2T
uration for eliminating motor 2 so that motor 1 can be ob-
An experimental sweep frequency response for a 32-weight served. As shown in the Appendix motor 2 can be observed
TDL is shown in Fig. 7. As indicated on the plot, one band- at the same time at thebandpass filter output at Y k .
width is twice the other because the amplitude ratio of the A simple method of generating the odd reference harmonics
two reference sinusoids is fito 1. is used. The fundamental frequency of motor 2 is measured
In summary, these results comprise a solution for the most directly off the motor and is clipped to give a square wave at
490 IEEE TRANSACTIONS
ONACOUSTICS,
SPEECH,
ANDSIGNAL
PROCESSING,
VOL.
ASSP-25, NO. 6 , DECEMBER 1977

(b) (dl
Fig. 9. (a) Spectrum of square-wave reference. (b) Spectrum of primary: both motors and their harmonics. (c) Spectrum
of notch-filter output: motor 1 and its harmonics. (d) Spectrum of bandpass-filter output: motor 2 and its harmonics.

the same frequency. A more efficient method of obtaining Although this exampleissomewhat contrived, it isgiven to
higher harmonics with equal amplitudes could be used in ac- demonstrate the type of situation in which the adaptive filter
tual practice. can be used to eliminate multiple-sinusoid interference. There
The fundamental frequency of motor 2 is 0.078. A digital art: many similar practical applications in which the adaptive
amplitude spectrum of the square-wave reference derived from notch filter has proved to be an extremely useful tool.
it is shown in Fig. 9(a). ,Note the aliasing of the higher har-
monics. The sumof thevibration hakmonicsfrom both VI. CONCLUSIONS
motors is received at point C and is used as the primary irlput The results presented here have provided a new approach to
to the noise canceler. The amplitude of this primary is shown the analysis of the adaptive filter and theadaptation algo-
in Fig. 9(b). rithm. When theinputtothe adaptive filter is sinusoidal,
A 64-weight adaptive filter was used for this simulation. The there are significant time-varying components in the filter
operation of the noise canceler was to place a notch filter at weights that havepreviouslybeenignored as noise in the
each of the frequencies seeri in the spectrum of the reference adaptation process. In ANC applications the effect is that of
in Fig. 9(a). By adapting very slowly, the flotches were made a tunable adaptive notch filter exactly at the frequency of the
very narrow so that tlie nearby harmonics from motor 1 were sinusoidal interference. As long as a reference is available that
undisturbed. The amplitude spectrum of the resulting noise- includesevery sinusoidal interference,the ANC systemwill
canceller output at ek is showninFig. s ( ~ ) . Only thehar- automatically create a notch over each sinusoid and follow it
monics from motor 1 remain, and they have the same relative if it drifts in frequency. This adaptive notch filter provides a
amplitude as they do in Fig. 9(b). Since for each notch filter simple and effective alternative to other methods of tracking
there is a correspondingbandpass fiter, the harmonicsof and eliminating sinusoidal interference.
motor 2 are simultaneouslyavailable at the Y k output. The One further point should be made about the relationship of
result is shown in Fig. 9(d), the results obtained here to previous ANC results. In an ANC
Thus we have seen that this combination of multiple-notch application, when the noise reference is sinusoidal, the noise to
fdters and bandpassfilters obtained by using multiple-reference be canceled is generally at exactly the same frequency as the
sinusoids into a long TDL allows us to separate the vibration reference. The solution desired is therefore probably the con-
harmonics caused by the two motors. Each of the notch and stant-weight solution previously assumed. This constant solu-
bandpass filters is precisely atthe, required harmonicfre- tion corresponds to an infinitely narrow notch fiter, the limit-
quency, even if one or both of the motors change frequency. ing case as a: + 0. However, in practice a: is not set to zero in
GLOVER: NOISE CANCELING 49 1

order to maintain adaptivity of thefilter, in order to have NaC2


w,.T
finite settling time in case the interference components change
in amplitude and/or frequency. A consequence of this is that
- ( ~

L
N;C2)
(Z COS - 1)
( N;C2) * (17)
signal components falling within the finite notch will be dis- z2-2 1-- z COS w,T+ 1- -
torted. Therefore, it is important to keep a as small as pos-
siblewhenlarge sinusoidal components are present in the This is a tunable digital bandpass filter centered at frequency
reference. f,.. Its poles are the same as those of H(z). Since the points of
It should also be pointed out that in many practical cases the unit gain are exactly at ff,., where the zeros of H(z) were,
reference interference is a mixture of both sinusoidal and non- then the 3-dB BW is found to be the sameas (1 1) for the
periodic components. For slow adaptation rates one canap- notch filter.
proximate that the resulting ANC solution is a combination of Multiple bandpass filters, corresponding to themultiple-
steady-state solution described in [I] and the notch-filter solu- notch filters, are obtained when a long TDL has more than one
tion described here. However, the exact nature of this inter- sinusoidal reference. Using (14) we obtain the following series
action is still being studied. of bandpass filters.

APPENDIX
BANDPASSFILTER
THE ADAPTIVE
In each instance where a notch filter is obtained as the trans-
fer function from dk to Ek, there is a corresponding bandpass
filter from d k to Y k , which will be denoted J(z). Ifwe con- The 3-dB BW at each reference frequency is the same as (16).
sider first the case of single-reference frequency into an REFERENCES
N-weight TDL, the results are easily found using the expres-
[ l ] B. Widrow et al., “Adaptive noise cancelling: Principles and appli-
sion for G(z) found earlier in (9). Substituting to find J(z) cations,” Proc. IEEE, vol. 63, pp. 1692-1716, Dec. 1975.
[ 2 ] J. Kaunitz, “Adaptive filtering of broadband signalsas applied to
J(z) = -
Y(Z> noise cancelling,” Stanford Electron. Labs., Stanford Univ., Rep.
SEL-72-038, Aug. 1972.
Hz) [3] B. Widrow,“Adaptivefilters I: Fundamentals,”StanfordElec-
tron. Labs., Stanford Univ., Rep. SEL-66-126, Dec. 1966.
-- G(z) [ 4 ] J. M. Wozencraft
and I. M. Jacobs, Principles of Communications
1+ G ( z ) Engineering.
Wiley, York: New 1965, p. 496.

You might also like