JPH03117206A - Adaptive filter - Google Patents

Abstract

PURPOSE:To obtain an adaptive filter whose tap coefficient correction calculation is executed in a processing time nearly equal to that for product sum calculation by applying pipeline processing to the tap coefficient correction calculation. CONSTITUTION:Switches 106, 107 are thrown to the position of R and a switch 108 is thrown to the position of CV respectively and respective data are read from a data memory 101 and a coefficient memory 102. The data are subject to prescribed product sum calculation by a multiplier 104 and an accumulator 105. Then a switch 112 is thrown to the position of W to bring an auxiliary coefficient memory 103 to the write mode and a coefficient data from the coefficient memory 102 is written in the memory 103 via a delay circuit 109 in a pipeline mode. Then a constant alpha is multiplied with a residual signal e(k) generated by an external subtractor at a multiplier 110, the result is inputted to a multiplier 104 via the switch 108 and the data from the memory 101 is inputted to generate a correction coefficient. Then the switch 112 is thrown to the position of R, the said correction coefficient, and a coefficient data from the memory 103 are corrected by addition, and the obtained coefficient is written in the memory 102 through a delay circuit 113.

Description

[Detailed Description of the Invention] [Object of the Invention] (Industrial Application Field) The present invention is an echo canceller that is widely used in I5DN subscriber line transmission systems, acoustic echo prevention in remote conferences, echo perception prevention in satellite lines, etc. It relates to adaptive filters, which are the main part of equalizers and equalizers.

(Prior Art) Adaptive filters are usually used in main parts of echo cancellers, equalizers, and the like.

The adaptive filter 201 in the echo cow Jansera is the third
As shown in the figure, the characteristic H(ω) (where ω is the angular frequency) of the unknown system 202 is estimated to obtain its approximate characteristic R(ω). Here, the accuracy of the approximation characteristic R(ω) obtained by the adaptive filter 201 is high, and H(ω)−H
(ω), the output of the adaptive filter 201? (ω) and the output Y(ω) of the unknown system 202 are almost equal, and the power of the error signal E(ω), which is the output of the subtracter 203, becomes very small. Therefore, the output Y(ω) of the unknown system 202
If the signal is an echo, the adaptive filter 201 can be used to cancel the echo.

On the other hand, similarly to the echo canceller described above, the adaptive filter 301 in the equalizer estimates the inverse characteristic R(ω) of the unknown system 302 having the characteristic H-'(ω), and estimates the inverse characteristic R(ω) of the unknown system 302 that is A transmission signal X(ω), which is an unknown input signal, is reproduced from the signal Y(ω).

By the way, the above-mentioned adaptive filter is often realized in the time domain, and the estimation operation is generally performed by sequential processing such as the steepest descent method or the learning identification method.

In this case, for example, in the steepest descent method, the adaptive filter can be realized by a product-sum operation, a correction coefficient generation operation, and a tap coefficient correction operation as follows.

1) Product-sum operation y (k) − H” (k)
)3) Tap coefficient correction calculation H(k+1)-H(k)+Δ(k) Here, H"(k) and X"(k) are the characteristic H(ω) of the adaptive filter and the adaptive filter input signal, respectively. A time-domain sequence of samples at time X(ω) of the characteristic H” (k) = (h + (k), h 2 (k)
, , h N (k))X ” (k) = (x
(k-N+1), , x (k-N+1)). Further, α is a coefficient that determines estimation accuracy, and y (k) is a reference signal of the adaptive filter.

These calculations can be realized by an adaptive transversal filter, as shown in FIG. 5, for example.

In this figure, 401 is the characteristic H'' of the adaptive filter.
402 is a data memory that stores the input signal sequence X'' (k) of the adaptive filter.

The product-sum operation in this adaptive filter is performed on the first and second
The switches 403 and 404 are connected to the R side, and the third switch 405 is connected to the Cv side.
01.01, 402, and from these data, the first multiplier 406 and the accumulator 407
This is done by calculating the following formula using .

9 (k)-Σ h I(k) x (k-1+
1) 1st ■ In addition, the correction coefficient generation operation is performed by multiplying the residual signal e (k) generated by an external subtracter by a constant α in a second multiplier 408, and then adding it to the signal connected to the CR side. third switch 40
5 to the first multiplier 406, and further inputs the data in the data memory 402 to the first multiplier 406 via the second switch 404 connected to the R side. This is done by performing calculations on formulas.

d h + (k) − a e (k) x (
k-1+1) (however, 1m1-N) The tap coefficient correction calculation is performed using the i-th tap coefficient b+ (k) read from the coefficient memory 401 via the first switch 403 connected to the R side, and the above-mentioned This is performed by adding the correction coefficient dhi (k) obtained by the correction coefficient generation operation in an adder 409.

h+ (k+1)-h+ (k)+dht (k)
(However, 1m1-N) By the way, a general time-domain adaptive filter can be realized by executing the above-mentioned calculation for each sample.
When trying to realize an adaptive FIR filter with a very high order of up to 00, or an adaptive FIR filter with about 32 taps at a very short sampling interval of 12.5 μsec, the time required for each of the above calculations is is a major obstacle.

Therefore, pipeline processing is usually used in which main operations such as memory access, multiplication, and cumulative addition (accumulator) are executed one after another at the same time using different hardware.

However, since some memory accesses cannot be inherently pipelined, this results in a significant loss of overall processing efficiency.

FIG. 6 shows how the coefficient memory and data memory are accessed.

As shown in the figure, in the product-sum operation, the coefficients from the coefficient memory and the data from the data memory are read out within one memory cycle, and the product-sum operation is simultaneously performed within one memory cycle by applying pipeline processing. By doing this, it is possible to perform a one-tap sum of products operation in one memory cycle.

On the other hand, in the tap coefficient correction operation, the tap coefficients read from the coefficient memory must be corrected and then stored in the original coefficient memory again, so pipeline processing cannot be applied and access to the coefficient memory is limited to one time. Two times per tap, or two memory cycles, are required. therefore,
Correcting the tap coefficients requires about twice as much processing time as the sum-of-products calculation.

(Problem to be solved by the invention) In this way, in conventional adaptive filters, the time required for the tap coefficient correction calculation is approximately twice the time spent on the product-sum calculation, which is a major problem in system implementation. It was a hindrance.

The present invention aims to solve these problems by implementing pipeline processing of tap coefficient correction calculations, and provides an adaptive filter that can realize the tap coefficient correction calculations in a processing time comparable to that of product-sum calculations. intended to provide.

[Structure of the Invention] (Means for Solving the Problems) The adaptive filter of the present invention achieves the above-mentioned objects, and also has the following features:
An adaptive filter that estimates characteristics of an unknown system by sequential processing from an input signal of the unknown system and an output signal of the unknown system,
a first storage means for storing input signals of an unknown system; a second storage means for storing coefficients of an adaptive filter which are estimated values of characteristics of the unknown system; and a third storage means for storing coefficients of an adaptive filter auxiliary. a storage means, a product-sum calculation means for performing a product-sum calculation from the data read from the first and second storage means, respectively; a transfer means for transferring to a third storage means;
correction coefficient generation means for generating information for sequentially changing the coefficients of the adaptive filter from the unknown system output signal and the product-sum calculation results; and coefficients read from the third storage means using the generated information. and coefficient correction means for correcting the correction result in the second storage means.

(Function) In the adaptive filter of the present invention, the conventional tap coefficient memory (
In addition to the second storage means), a third storage means for additionally storing the tap coefficients is newly prepared, and when executing the product-sum operation,
The coefficients read from the second storage means are transferred to the third storage means by the transfer means in a pipeline manner. In the tap coefficient correction process, the coefficients are read out from the third storage means, the corrections are made to the coefficients, and then the correction coefficients are written into the second storage means, which was previously impossible. Pipeline processing can be applied to modify tap coefficients.

As a result, access to the second storage means, which conventionally required two memory cycles per tap, can be reduced to one memory cycle, making it possible to perform tap coefficient correction processing in the same processing time as a sum-of-products operation. This can increase the ease of implementation of the adaptive filter.

(Example) Embodiments of the present invention will be described below based on the drawings.

FIG. 1 is a block diagram for explaining the configuration of an adaptive filter according to an embodiment of the present invention.

In the figure, 101 is a data memory that stores an input signal sequence to the adaptive filter, which is an unknown system input signal;
Reference numeral 2 denotes a coefficient memory that stores adaptive filters and evening tap coefficients that are estimated values of characteristics of an unknown system, and 103 represents an auxiliary coefficient memory that auxiliarily stores coefficients of the adaptive filter. Also,
104 and 105 are unknown system input signals read out from the data memory 101 via the first switch 106 and the second and third switches 107 and 108 from the coefficient memory 102;
These are a first multiplier and an accumulator that constitute a product-sum calculation circuit that performs a predetermined product-sum calculation from coefficients read out via the first multiplier and accumulator. Further, reference numeral 109 indicates an auxiliary coefficient memory 1 from the coefficient memory 102 as the product-sum operation is executed in the product-sum operation circuit.
This is a delay circuit for transferring coefficient data to 03 by pipeline processing. A second multiplier 110 multiplies the residual signal (unknown system output signal) generated by the external subtracter by a constant α for determining estimation accuracy. And this second
The output of the multiplier 110 is sent to the data memory 101 at the first multiplier 104 via the third switch 108.
The multiplication is combined with the unknown input signal read from the input signal, and the resulting value becomes information for sequentially changing the characteristics of the adaptive filter. Therefore, the first multiplier 104 and the second multiplier 110 constitute a correction coefficient generation circuit. Furthermore, 111 is an adder that adds the output of the above-mentioned correction coefficient generation circuit to the coefficient data read out from the auxiliary coefficient memory 103 via the fourth switch 112. The addition result of the adder 111 is transferred to the coefficient memory 102 via the delay circuit 113 and the second switch 107 in a pipeline manner and written therein. This constitutes a coefficient modification circuit for modifying the coefficient data in the coefficient memory 102.

Next, the operation of the adaptive filter configured as above will be explained.

First, connect the first and second switches 106 and 107 to the R side, and the third switch 108 to the CV side,
Each data is read from data memory 101 and coefficient memory 102. These data are input to a product-sum calculation circuit consisting of a first multiplier 104 and an accumulator 105, where a predetermined product-sum calculation is performed.

Here, the fourth switch 112 is connected to the W side, and the auxiliary coefficient memory 103 is in the write mode.

As a result, the coefficient memory 102
The coefficient data read from the auxiliary coefficient memory 103 is transferred to and written in the auxiliary coefficient memory 103 in a pipeline manner via the delay circuit 109.

After this, in the correction coefficient generation circuit, the residual signal e (k) generated by the external subtracter is multiplied by a constant α in the second multiplier 110, and then the third switch 108 is applied. A correction coefficient is generated by inputting the data read from the data memory 101 to the first multiplier 104 via the first multiplier 104 .

Next, the fourth switch 112 is connected to the R side, the coefficient memory 103 is set to read mode, and the correction coefficient generated in the correction coefficient generation circuit and the coefficient data read from the auxiliary coefficient memory 103 are transferred to the adder 111. perform correction calculations on the coefficient data.

Then, the coefficient data obtained as a result of this modification is transferred to the delay circuit 1.
13, the coefficients are transferred to the coefficient memory 102 in a pipeline manner and written therein.

Thus, in this embodiment, as shown in FIG. 2, the coefficient data correction process conventionally required two memory cycles per tap, but by completely pipelining the tap coefficient correction process, In general, only one memory cycle is required, which increases the ease of implementing the adaptive filter.

[Effects of the Invention] As explained above, according to the adaptive filter of the present invention, the tap coefficient correction calculation can be performed in a pipeline process, and the tap coefficient correction calculation can be realized in the same processing time as the product-sum calculation. This makes it possible to increase the ease of implementing the adaptive filter.

[Brief explanation of drawings]

FIG. 1 is a block diagram for explaining the configuration of an adaptive filter according to an embodiment of the present invention, FIG. 2 is a diagram showing the effect of the adaptive filter of FIG. 1, and FIGS. 3 and 4 are conventional filters. FIG. 5 is a block diagram showing the configuration of an adaptive transversal filter, and FIG. 6 is a diagram showing memory access in the adaptive filter of FIG. 101...Data memory, 102...Coefficient memory,
103... Auxiliary coefficient memory, 104... First multiplier, 105... Accumulator, 106.107.1
08.112...Switch, 109.113...Delay circuit, 110...Second multiplier, 111...Adder.

Claims (1)

[Scope of Claims] An adaptive filter that estimates the characteristics of the unknown system through sequential processing from an input signal of the unknown system and an output signal of the unknown system, comprising: a first storage means for storing the input signal of the unknown system; and a second storage means for storing adaptive filter coefficients that are estimated values of the characteristics of the unknown system; a third storage means for auxiliary storage of the adaptive filter coefficients; and the first and second storage means. A sum-of-product calculation means performs a sum-of-product calculation from data read from the respective storage means, and when the sum-of-product calculation means is executed, the coefficient read from the second storage means is transferred to the third storage means. transfer means; correction coefficient generation means for generating information for sequentially changing the coefficients of the adaptive filter from the unknown system output signal and the product-sum operation result; an adaptive filter comprising: a coefficient modifying means for modifying the coefficients read from the storage means, and storing the modification result in the second storage means.