JPH02291712A - Method and device for eliminating noise - Google Patents

Abstract

PURPOSE:To shorten the converging time by controlling a step size based on a fact that the mean electric power of the remaining noises is reduced together with the convergence of a coefficient and at the same time subtracting sequentially the signal component electric power which disturbs the detection of the remaining noises after obtaining the estimated value of the signal component electric power. CONSTITUTION:An averaging circuit 8 calculates the estimated value of the signal component electric power, and a subtracter 9 subtracts the estimated value from the output signal of a multiplier 7 to obtain the output signal of an averaging circuit 10 via a polarity detector 12. The input of the circuit 10 is almost equal to +1 right after an adaptive filter 3 is started and then set at + or -1 in the same probability with progress of adaptation of the filter 3. Thus the positive value which increases gradually and set at zero together with convergence of the filter 3 is obtained as the output of the circuit 10. This positive value is multiplied by a constant and then by an original step size (a). As a result, an effective step size is set large at first and then set equal to the size (a) after convergence of the filter 3. Thus the converging time is shortened.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a method and apparatus for removing noise that enters a main input terminal and interferes with a signal by using an adaptive filter. Such a device is generally called a noise canceller, and is used to cancel background noise in automobiles, during communications between aircraft crews, divers, and the ground.

(Prior Art) A noise canceler is known as a well-known technology for removing noise that enters the main input terminal and interferes with the signal using an adaptive filter (Proceedings of IEE).
N.G.S.O. FIEEE) Volume 63, No. 12, 1975, 1
See pages 692-1716; hereinafter, [Reference 1]).

noise. The canceller uses an adaptive filter that has a transfer function that approximates the impulse response of the path that noise takes from the noise source to the main input terminal. ), it operates to suppress noise that mixes into the main input terminal and interferes with the signal. At this time, each tap coefficient of the adaptive filter is successively corrected by correlating the difference signal obtained by subtracting the noise replica from the mixed signal containing noise and signal with the reference noise obtained at the reference input terminal. . The LMS algorithm (LMS ALGORITHM) is a typical example of the convergence algorithm of the noise canceler that corrects the coefficients of such an adaptive filter.
) (Reference 1) and the Learning Identification Method (LEARNING IDENTIFI)
CATION METHOD; LIMX IEEE Transactions on Automatic Control
AUTOMATIC CONTROL) Volume 12, No. 3, 19
See 1967, 282-287'<-ji; Hereinafter, [Reference 2J) is known.

Figure 3 shows conventional noise. FIG. 2 is a block diagram showing a configuration example of a canceller. A mixed signal of a signal and noise detected at the main input terminal 1 is supplied to a subtracter 4 .

On the other hand, the reference noise detected at the reference input terminal 2 is supplied to the adaptive filter 3. The noise replica generated by the adaptive filter 3 is
The subtracter 4 subtracts the noise component from the mixed signal to eliminate the noise component, and the signal is supplied to the output terminal 6. The output of the subtracter 4 is simultaneously supplied to the multiplier 5, multiplied by 20, and used to update the coefficients of the adaptive filter 3.

Here α is a constant and is called the step size. Now, let us assume that the signal is sk (where k is an index indicating time), the reference noise is nk, the noise to be erased is vk, and the additional noise received by sk is σk.
The signal uk supplied to is expressed by the following equation.

uk=sk+vk+ak. ．． ．． ．． ．． ．． ．． ．． ．． ．． ．． ．． ．． ．．
．． ．． ．． ．． ．． (1) The purpose of the noise canceller is to generate a replica 9k of the noise component vk in equation (1) and cancel the noise. In FIG. 3, by using a closed loop circuit consisting of an adaptive filter 3, a subtracter 4, and a multiplier 5 to adaptively generate a noise replica 9k, the output signal of the subtracter 4 is a difference expressed by the following equation. A signal dk can be obtained. However, in general σk
is considered to be sufficiently small compared to sl(, so this is ignored.Then, the output dk of the subtractor 4 can be expressed as follows.

dk=sk+vk-%. ．． ．． ．． ．． ．． ．． ．． ．． ．． ．． ．． ．． ．．
．． ．． ．． ．． ．． (2) In equation (2), (vk-Ok
) is called residual noise.

Assuming the LMS algorithm, the m-th coefficient cm,k of the adaptive filter 3 is updated according to the following equation.

Cm,k" Cm,k-1 + 20"k"m,k-
1-・-・- (3) Equation (3
) is expressed in matrix form, ck=ck1+2a・dk−n
k1. ．． ．． ．． ．． ．． ．． ．． ．． ．． ．． ．． (4) becomes. Here, ck and nk are respectively given by the following equations.

Cl("[CO+CI,"'+CN-1] +.+.
．． +. +++. ．． ．． ++. (5) nk”[”O+nlv
”'+nN-1] ........
(6) On the other hand, in Document 2 (LIM algorithm), coefficients are updated according to equation (7) instead of equation (4).

ck=ck-1+(2p/Nσn2)dk"k-1-・
- (7) p is the step size for LIM,
σ. is the average power input to the adaptive filter 3.

σ. is used to make the value of the up size μ inversely proportional to the average power and to achieve stable convergence. σ1
There are many ways to calculate, for example, the formula (
8).

The step size in equations (4) and (7) defines the speed of convergence of the adaptive filter and the residual noise level after convergence. In the case of LMS, the larger α, the faster the convergence, but the larger the residual noise level. On the other hand, in order to achieve a sufficiently small residual noise level, it is necessary to adopt a commensurately small α, which leads to a decrease in the convergence speed. LIM steps. The same applies to the size P.

In order to meet the conflicting demands on convergence speed and residual noise step size, the ■S algorithm is proposed. (IEEEE Transactions on Acoustakes Speech and Sequential
FROSSE SYNC (IEEEETRANSACTIoNSON
ACOUSTICS, SPEECH AND
SIGNAL PROCESSING) Volume 34, No. 2,
1986, pp. 309-316; Hereinafter, [Reference 3J) VS algorithm uses a step size matrix A instead of the step size α of the LMS algorithm in equation (4), and the size of each component of A is ? Adaptive
It is controlled by the progress of convergence of the filter. By using individual step sizes given by the step size matrix A instead of a common step size for each coefficient, the optimal step size corresponding to the variation in the size of the autocorrelation matrix components can be determined for each coefficient. It can be used as a coefficient to improve convergence speed. The actual coefficient update is based on the following equation.

ck=ck-1+2A-dk・nk-1. ．． ．． ．． ．． ．．
．． ．． ．． ．． ．． ．． (9) A=[aij] aij=O for i+j ≠Ofori=j am,m is the slope component of the corresponding m-th tap■,k
The polarity of sgn[. , kl, and the magnitude is controlled by the polarity change pattern. However, ■, k is m, k
"2゜αm,m"k'nm,k-1-・・・・−・(1
0). In an ideal case where dk = vk- Ok holds true, the polarity of m and k directly represents the progress of convergence, but dk is generally influenced by sk, as expressed by dk = sk + vk- %. So, to alleviate this? gn [■, kl is m. If it changes continuously, αe, m is 1/2, and if it is equal to m1 times in a row, α. , m is multiplied by 2. That is, a succession of the same polarity,
Alternatively, a feature of the VS algorithm is that by detecting a series of changes in polarity, sk+vk-Okjik-9k is equivalently established. However, α
. , m has a limited range of variation, and the maximum value αmax=1
lλ and αmin are defined by residual noise after convergence. Here, λ is the maximum eigenvalue of the autocorrelation matrix. Whether this method works without problems depends largely on the relationship between Sk and vk-%. Said. , k is determined by the signal-to-noise ratio (Signal-to-Noise ratio) of sk.
e Ratio (SNR) and spectrum. When the SNR is good, skl > Ivk-%l almost always holds, which seriously affects the polarity detection. Considering that the SNR is the ratio of the mathematically expected values of the instantaneous power of the signal and the noise, even if the SNR is bad, the sk
The more high-frequency components there are, the more instantaneously ISkl
> Ivk- %l becomes more likely. Stated another way, the more peaks and dips Sk has, the more the SNR is bad enough that some peaks
There is a high possibility that kl will be much larger than Ivk-9k.

(Problems to be Solved by the Invention) In order to minimize these effects, the above m. and m must be made large and αmin small, which reduces the superiority of the vS algorithm. Furthermore, as can be seen from equation (7), the ■S algorithm must maintain the same number of step sizes as coefficients in memory, and as the number of taps increases, more memory is required, and the hardware becomes a burden. An object of the present invention is to provide a method and apparatus for noise removal using an adaptive filter, which has a short convergence time and a small hardware scale.

(Means for Solving the Problems) The noise removal method of the present invention uses an adaptive filter based on a reference signal obtained at a reference input terminal that receives only noise as input, in order to remove noise mixed into the main input terminal. A noise replica is generated from the mixed signal obtained at the main input terminal in which the received signal and noise are mixed. In a noise canceller that operates to reduce the difference signal obtained by subtracting the replica, the difference signal power is obtained by squaring the difference signal, and the polarity is determined by subtracting the averaged difference signal power from the difference signal power. The above-mentioned adaptive. It is characterized by adaptively changing the coefficient correction amount of the filter.

Further, the noise removal device of the present invention includes an adaptive filter that receives reference noise from a reference input terminal and generates a noise replica when removing noise that enters the main input terminal and mixes with the signal; a first subtractor that subtracts a noise replica from a mixed signal from a main input terminal consisting of a signal and noise; a first multiplier that squares the output of the first subtractor; and a first multiplier that squares the output of the first subtractor. The first to average the output of
and an averaging circuit of the first multiplier.
a second subtracter that subtracts the output of the averaging circuit, a polarity detector that detects the polarity of the output of the second subtractor, and a second averaging circuit that averages the output of the polarity detector. , a second multiplier that multiplies the output of the second averaging circuit by a constant;
at least a third multiplier for multiplying the output of the multiplier by the output of the first subtracter, and updating the coefficients of the adaptive filter using the output of the third multiplier. It is characterized by having been configured.

(Operation) The noise removal method and device using an adaptive filter of the present invention controls the step size by using the fact that the average power of residual noise becomes smaller as the coefficients converge, and at this time, the residual noise detection method and apparatus control the step size. The convergence time can be shortened by calculating the estimated value of the signal component power that causes interference and subtracting it sequentially.

(Example) Next, the present invention will be explained in detail with reference to the drawings.

FIG. 1 is a block diagram showing one embodiment of the present invention.

In this figure, functional blocks given the same reference numbers as in FIG. 3 have the same functions as in FIG. 3.

The difference between FIG. 1 and FIG. 3 is that the step size supplied to the multiplier 5 changes according to the magnitude of the residual noise. Along with this, multipliers 7 and 11 for controlling the step size, averaging circuits 8 and 10, a subtracter 9, and a polarity detector 12 are added. Although the signal required for step size control is the residual noise vk-%, it is not possible to directly detect only the residual noise. The signal sk becomes a disturbance to the residual noise vk-Ok. The averaging circuit 8 and the subtracter 9 have the effect of removing the influence of the signal sk.

Since the output of the multiplier 7 is the square of dk, it is given by the following equation.

dk2=(sk+vk-%)2. ．． ．． ．． ．． ．． ．． ．． ．． ．． ．．
．． ．． ．． ．． (11) The averaging circuit 8 takes a moving average with a long time constant for the input signal. Since a moving average with a long time constant can be approximately regarded as an estimated value of a mathematical expected value, it is expressed as E[·]. In general, Sk and nk, and therefore sk and vk, are independent of each other, so taking into account the properties of mathematical expectation values, the output of the averaging circuit 8 is: E[(sk+vk-9k)'']=E[sk21+E[(
vk-%)2]. (12). Considering that the second term in equation (12) decays rapidly, equation (13) is obtained.

E[(Sk+Vk-%)2]-E[sk21. ．． ．．
．． ．． ．． ．． ．． ．． ．． (13) Therefore, the output of the subtracter 9 is (sk+vk-9k)2-E[sk2]. After the polarity detector 12 detects the polarity of the output signal of the subtractor 9,
An averaging circuit 10 takes a moving average with a short time constant for the human input signal. If we express a moving average with a short time constant by d, we get
The output of the averaging circuit 10 is given by equation (14).

In equation (14), immediately after the adaptive filter starts operating, (vk-%)2 is sufficiently large and 1sk2-E
[sk2]14 (vk-Ok)2 holds true, so the formula (
15) can be approximated as follows.

msgn{2sk・(vk-vk)+(vk-?k)2
} ． ．． - (15) Equation (
15) The polarity of the right side is 2sk. (vk-4k). The probability of 2sk・(vk%)>0 is 1/2, 128
Since the probability that kl < Ivk- +kl is greater than 0, the input to the averaging circuit 10 will be +1 with a probability of 1/2 or more. On the other hand, after the adaptive filter converges, Isk2
1E[sk2]I, 21sk-(vk-4k)l, (v
Since k-Ok)2 becomes an independent random process with zero mean, it becomes. As can be seen from the above description, the averaging circuit 8 provides an estimated value of sk. Subsequently, the estimated value is subtracted from the output signal of the multiplier 7 by the subtracter 9, and the right side of equation (14) is obtained as the output signal of the averaging circuit 10 via the polarity detector 12. The input to the averaging circuit 10 is almost +1 immediately after the adaptive filter starts operating, and becomes ±1 with the same probability as the adaptive filter progresses. Therefore, the output of the averaging circuit 10 is a positive value that gradually increases and approaches zero as the adaptive filter converges. Multiply this value by a constant to get the original step size α
By multiplying .

However, when the output of the multiplier 11 becomes smaller than 1, it is forcibly set to 1. The amplitude of the signal obtained as the output of the subtractor 9 is greatly influenced by the signal strength supplied to the input terminal 1, but this influence can be reduced by the polarity detector 12, and the parameter can be adjusted over a wide range of signal strengths. The settings become easier.

Figure 2 shows an example of the averaging circuit, with a leakage coefficient of 13 (0<
13<1) is known as a first-order leakage integration circuit. The signal supplied to the human input terminal 21 is multiplied by p in the multiplier 22 and supplied to the adder 23. The output signal of adder 23 reaches output terminal 26 and is also supplied to delay element 25 . The signal delayed by one clock in the delay element 25 is multiplied by 1-13 in the multiplier 24 and is supplied to the adder 23.

In total, the signal supplied to the human input terminal 21 is transmitted to the delay element 2
5 is delayed by one clock and then repeatedly added by the adder 23, so that it is integrated. At this time, the multiplier 24 causes "leakage 2". When the average value of the signal supplied to the human input terminal 21 is approximately constant, the multiplier 2
As can be seen from the values 3 and 24, the output signal obtained at the output terminal 26 gradually increases and then saturates. By appropriately selecting the constant p, the average value of the human input signal can be approximated by this saturation value. 1- if p is small
The pH becomes 1, and the signal at the output terminal 26 is fed back almost unchanged to the adder 23, forming a moving averaging circuit with a long time constant. On the other hand, if p is large, the signal fed back from the output terminal 26 to the adder 23 will attenuate rapidly, and the contribution of the current signal fed from the input terminal 21 through the multiplier 22 will become large, so that the shift The time constant of the averaging circuit becomes shorter. Therefore,
The averaging circuit shown in FIG. 2 can be used as both averaging circuit 8 and 10 by appropriately setting the leakage coefficient p. When used as the averaging circuit 8, the input signal 21 is the signal from the multiplier 7, and the output signal 26
corresponds to the signal to the subtracter 9, and to the averaging circuit 10, the input signal 21 corresponds to the signal from the subtracter 9, and the output signal 26
corresponds to the signal to the multiplier 11. In FIG. 2, a cyclic averaging circuit is shown as an example, but a circuit having a transversal configuration can be similarly used.

As already explained, the difference between LIM and LMS is that the step size μ divided by the average power On′ input to the adaptive filter 3 is used instead of α. The method of making the size variable can be applied as is.

(Effects of the Invention) As described above in detail, according to the present invention, the step size is controlled by using the fact that the average power of the residual noise becomes smaller as the coefficients converge, and at this time, the average power of the residual noise is The convergence time can be shortened by calculating the estimated value of the signal component power that becomes an obstacle to detecting the power and subtracting it successively. Furthermore, since the circuit that adaptively controls the step size consists of only a squaring multiplier, a 2-averaging circuit, a subtracter, and a polarity detector, a noise removal method using an adaptive filter with small hardware scale and equipment can be provided.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing one embodiment of the present invention, FIG. 2 is a block diagram showing one embodiment of an averaging circuit, and FIG. 3 is a block diagram showing a conventional example. In FIG. 1, 1 is the main input terminal, 2 is the reference input terminal,
3 is an adaptive filter, 4 and 9 are subtractors, 5,
7 and 11 are multipliers, 6 is an output terminal, and 8 and 10 are averaging circuits, respectively.

Claims (1)

[Claims] 1) In order to remove noise mixed into the main input terminal, a noise replica is generated by an adaptive filter based on a reference signal obtained at a reference input terminal that receives only noise as input, and a noise replica is generated from the main input terminal. In a noise canceller that operates to reduce a difference signal obtained by subtracting the noise replica from a mixed signal obtained at an input terminal in which a received signal and noise are mixed, the difference signal power is obtained by squaring the difference signal. ,
The polarity is determined by subtracting the averaged difference signal power from the difference signal power, and the coefficient modification amount of the adaptive filter is adaptively changed in accordance with the value obtained by further averaging. noise removal method. 2) An adaptive filter that receives reference noise from a reference input terminal and generates a noise replica when removing noise that enters the main input terminal and mixes with the signal; a first subtracter that subtracts the mixed signal from the main input terminal, a first multiplier that squares the output of the first subtracter, and a first multiplier that averages the output of the first multiplier. an averaging circuit, a second subtracter that subtracts the output of the first averaging circuit from the output of the first multiplier, and a polarity detector that detects the polarity of the output of the second subtracter. a second averaging circuit that averages the output of the polarity detector; a second multiplier that multiplies the output of the second averaging circuit by a constant; a third multiplier that multiplies the output of the first subtracter, and uses the output of the third multiplier to
A noise removal device characterized by updating filter coefficients.