JPH04267618A - Adaptive filter - Google Patents

Abstract

PURPOSE:To obtain an optimum convergent value by comparing mean square errors MSE of output signals of two IIR filters having a different initial tap coefficient of output signals of an unknown dynamic system. CONSTITUTION:Initial values of each tap coefficient of IIR filters 1, 2 are different from each other. Let respective MSEs be MSE(1), MSE(2). A difference is caused in the result of calculation in a comparator 4. For the first M samples, for example, the relation of MSE(1)>MSE(2) is discriminated and a re- initializing section 6 based on the result of comparison initializes the IIR filter 1 whose MSE is larger at random. Then the MSE(1) is gradually decreased and the relation of MSE(1)<MSE(2) is in existence at a succeeding 2M sampling. Then the IIR filter 2 is again initialized this time. The IIR filter whose MSE is smaller is selected for each M sample and a residual difference signal with an output signal of the unknown dynamic system 3 is outputted. Through the constitution above, the MSE is reduced and converged and the unknown system 3 obtains a sole optimum MSE.

Description

[Detailed description of the invention]

0001

[Industrial Application Field] The present invention relates to an adaptive filter.
In particular, IIR (Infinite
e Impulse Response) The present invention relates to an adaptive filter for performing adaptive processing by estimating the transfer characteristic using the filter.

0002

[Prior Art and Problems] Adaptive filters are well known as devices used in echo cancellers, data equalizers, etc. to estimate the unknown temporally changing transfer characteristics of a system, and as shown in FIG. 6(a), The mean squared error between the input signal and the signal obtained by passing through the unknown dynamic system 10 (Mean Squared-Error=
The unknown dynamic system 10 is estimated by changing the tap coefficients of the MSE (hereinafter abbreviated as MSE), and the residual component of the output signal is reduced.

[0003] Such adaptive filters are usually IIR filters or FIR (Finite Impu
lse Response) filter is used, but in this case, since the IIR filter has an infinite impulse response, it has the advantage that estimation can be performed with a small number of taps even if the unknown dynamical system has a long impulse response. be.

On the other hand, in order to obtain the above-mentioned MSE, computation must be performed in a cyclic (feedback) manner, and the computation function is more complex than that of an FIR filter, and the estimation operation for an unknown dynamic system is unstable. , there is a drawback that the MSE does not necessarily converge to the optimal value, as shown in FIG. 2(b).

[0005] Accordingly, an object of the present invention is to realize an adaptive filter that takes advantage of the IIR filter's ability to perform estimation with a small number of tap coefficients and that can obtain an optimal convergence value.

[0006]

[Means for Solving the Problems] A conventional adaptive filter as shown in FIG. 6 uses only a single IIR filter, and therefore, the instability of the IIR filter directly becomes an output signal.

Therefore, the adaptive filter according to the present invention focuses on using two IIR filters and selecting the one with better MSE.As shown in principle in FIG. two II with coefficients
R filters 1 and 2, and a comparator 4 that calculates each mean square error from each residual signal of the output signal of each filter 1 and 2 and the output signal of the unknown dynamic system 3 and compares the two for each sample. , a switch 5 for selecting the residual signal with the smaller mean square error as the output signal according to the comparison result of the comparator 4;
and a re-initialization unit 6 which re-initializes the filter with the larger mean square error based on the comparison result of the comparator 4 using random tap coefficients at predetermined sample intervals M.

Further, in the present invention, the predetermined sample M is
The number of taps is set to be greater than the number of taps of IR filters 1 and 2.

[0009]

[Operation] In the adaptive filter according to the present invention shown in FIG.
A comparator 4 calculates each mean square error MSE between the output signals of the IR filters 1 and 2 and the output signal of the unknown dynamic system 3 for each sample, and also compares the mean square errors MSE from each IIR filter 1 and 2. do. .

After receiving the comparison result from the comparator 4, the switch 5 selects the residual signal (the output signal of the IIR filter and the unknown dynamic system 3) with the smaller mean square error MSE.
(residual signal with the output signal of ) is selected as the output signal for each sample.

On the other hand, the reinitialization unit 6, which also receives the comparison result from the comparator 4, selects the IIR with the larger mean square error.
The filter is reinitialized at predetermined sample intervals M (M>1). In this case, the reinitialization section 6 reinitializes the IIR filter using random values. By setting the predetermined sample M to be greater than or equal to the number of taps of each IIR filter 1 and 2, it is possible to prevent each IIR filter from being reinitialized in a transient state.

[0012] To explain the above effects using the graph of FIG.
Since the initial values of the tap coefficients of both filters 1 and 2 are set to be different, for example, the MSE based on the output signal of IIR filter 1 is MSE(1), and the MSE based on the output of IIR filter 2 is MSE. (2) Then,
A difference occurs in the result of the calculation in the comparator 4, and for the first M samples shown in the figure, for example, MSE(1) >MS
E(2), so in response to this comparison result, the reinitialization unit 6 randomly reinitializes the IIR filter 1 with the larger MSE.

As a result, as shown in the figure, MSE(1) may become large once, but then it gradually decreases, so that at the next 2M samples, on the contrary, the MSE(1) becomes large.
Since E(1) <MSE(2), the reinitialization unit 6 reinitializes the IIR filter 2 based on this comparison result.

In this way, the IIR filter with the smaller MSE is selected for each M sample, and the unknown dynamic system 3
By using the residual signal with the output signal from the adaptive filter as the output signal, the possibility of instability of the entire adaptive filter can be greatly reduced, and the MSE is converged toward a smaller value as shown in Fig. 2. Optimal estimation of the unknown dynamic system 3 can be performed.

[0015]

[Embodiment] FIG. 3 shows an embodiment of the adaptive filter according to the present invention. In this embodiment, the comparator 4 shown in FIG. M
Two arithmetic units 41 and 42 that perform SE calculations, a subtracter 43 that subtracts the output signals of these two arithmetic units, and a selection signal to switch 5 and reinitialization unit 6 by the output signal of this subtracter 43. The re-initialization unit 6 includes a random tap coefficient memory 61 connected to the determination unit 44 and a changeover switch 62 controlled by the output signal of this memory 61. has been done.

Next, to explain the operation of this embodiment, first, a residual signal between the output signal of the IIR filter 1 and the output signal of the unknown dynamic system 3 is output from the subtracter 11, and the output of the IIR filter 2 is output from the subtracter 11. A residual signal between the signal and the output signal of the unknown dynamic system 3 is output from the subtracter 21 and sent to the calculation units 41 and 42 of the comparator 4, respectively.

In the calculation units 41 and 42, the multiplication unit 1
In steps 2 and 22, the above residual signals at sample n are multiplied together to calculate e21(n) and e22(n), and this is added to the attenuation of 1-β (β is 0.9<β<1.0). e21(n) (1
-β), e22(n) (1-β) is obtained. and,
Furthermore, from the previous sample, the delay units 16 and 26 and the multiplier 15
, 25, βE[e21(n-1) ], βE[
e22(n-1) ] (E[ ] indicates the average value) is obtained, and this and the above e21(n) (1-β), e
22(n) (1-β) in the adders 14 and 24, MSE(1) = E[e21(n) ] = (1-β
)e21(n) +βE[e21(n-1)] M
SE(2) =E[e22(n)]=(1-β)e2
2(n)+βE[e22(n-1)] is obtained.

These equations are as follows: E[e2(n)]=(1/n)Σe2(i)(i=
0~n) = (1/n)Σ
e2(i) (i=0~n-1)+(1/n)e2(n)
=(n-1/n)(1/
n-1) Σe2(i) (i=0~n-1)+(1/n
)e2(n) = (1-β
)e2(n)+βE[e2(n-1)] However, β=
(n-1/n).

In this manner, MSE(1) and MSE(2) are calculated from each IIR filter 1 and 2 and subtracted by the subtracter 43. Depending on whether the result of this subtraction is positive or negative, the determining section 44 generates an on/off signal.

That is, when MSE(1) >MSE(2) and the output of the subtracter 43 is positive, the output signal from the determination unit 44 causes the switch 5 to switch the subtracter 21 of the IIR filter 2 with the smaller MSE. Toggled to select residual signal as adaptive filter output (position shown)
At the same time, the switch 61 of the re-initialization unit 6 is switched to the IIR filter 1 with a large MSE (the position shown in the figure). Then, at this time, the memory 61 receives the output signal of the determination section 44 and when a time corresponding to M samples arrives, gives the stored random tap coefficients to the IIR filter 1 to reinitialize it.

The operation timings of the switch 5 and the switch 62 are shown in FIG. 4, and as can be seen from this, the switch 5 operates for each sample, but the switch 62 operates at the same time as shown in FIG. The operation is performed every M samples. However, the number of samples M is selected to be M>>N (N is the number of taps of the IIR filter) in order to avoid transient operating conditions of the IIR filter.

[0022]

As described above, according to the adaptive filter according to the present invention, two IIR filters are used to obtain the respective mean square errors, and the smaller IIR filter is used to calculate the adaptive filter output for each sample. As soon as the IIR filter is generated, the larger IIR filter is re-initialized with a random tap coefficient every predetermined M samples and the same operation is repeated.
Since the IR filter is not left as it is but constantly regenerated, stable operation of the adaptive filter as a whole can be realized.

To explain this with reference to FIG. 5, each local minimum value shown in the figure indicates the mean square error when the IIR filter with the smaller mean square error is selected within a predetermined sample. This corresponds to the adaptive operation in the example, and in the present invention, by further repeating the above operation, the initial tap coefficient (corresponding to the positions on both sides in the figure) is changed, and finally the value shown in the figure is changed. Thus, at a certain tap coefficient, the only non-local optimal mean squared error is obtained.

As a result, in the case of an echo canceller, for example, the amount of cancellation (ERLE) was only 15 dB in the conventional example, but it has been improved to 40 dB in the present invention.

Furthermore, when considering the number of multiplications in the present invention, as can be seen from FIG.
In total, filters 1 and 2 perform 6 multiplications for each sample, and since each IIR filter itself requires 2N multiplications, a total of 4N+6 multiplications are performed.

On the other hand, in the case of the conventional example in which only one IIR filter is used, it is necessary to check whether the poles of the filter exist within a single circle.
2 multiplications are required, for a total of 2N+N2 multiplications.

Therefore, if the number of taps N>4, 4N+
6>2N+N2, and the amount of calculations is also reduced.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the principle configuration of an adaptive filter according to the present invention.

FIG. 2 is a graph diagram for explaining the action of the adaptive filter according to the present invention.

FIG. 3 is a block diagram showing an embodiment of an adaptive filter according to the present invention.

FIG. 4 is an operation time chart of two switches used in the adaptive filter according to the present invention.

FIG. 5 is a diagram for explaining the control transition of the mean square error (MSE) by the adaptive filter according to the present invention.

FIG. 6 is a diagram for explaining a conventional example.

[Explanation of symbols]

1, 2 IIR filter 3 Unknown dynamic system 4 Comparator 5 Switch 6 Re-initialization section In the figures, the same reference numerals indicate the same or corresponding parts.

Claims (2)

[Claims] Claim 1: Two IIR filters (1, 2) having mutually different initial tap coefficients, and each filter (
a comparator (4) that calculates each mean square error from each residual signal between the output signal of 1 and 2) and the output signal of the unknown dynamic system (3) and compares the two for each sample; (4) A switch (5) which selects the residual signal with a smaller mean square error as the output signal based on the comparison result of the comparator (4), and a predetermined filter with a larger mean square error based on the comparison result of the comparator (4). A reinitialization unit (6) that performs reinitialization using random tap coefficients at sample (M) intervals.
An adaptive filter comprising: 2. The adaptive filter according to claim 1, wherein the predetermined samples (M) are greater than or equal to the number of taps of each IIR filter (1, 2).