JPH0773194B2 - Adaptive filter - Google Patents

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an adaptive filter that obtains a desired signal output by filtering one-dimensionally or two-dimensionally distributed signals by spatial calculation.

2. Description of the Related Art When a digital signal is transmitted using a telephone line or a wireless transmission line, an adaptive filter is used as a method for reducing the effects of transmission distortion, intersymbol interference, and noise. Things are commonly done. (For example, an adaptive equalizer using decision-learning by Robert W. Lucky: Bernard Widlow.
d-Widrow) and others, Neural Nets for Adaptive Pattern Recognition (Neural Nets for
Adaptive pattern Recognition) Computer (Comput
er) Described in the March 1988 issue) Problems to be solved by the invention This method is extremely practical and effective because the characteristics of the filter are adaptively determined according to various expected line characteristics. It is one of the filters.

However, the filter configuration is fixed in advance,
Therefore, the delay amount is fixed. Therefore, there is a drawback that an excessive time delay is given to a signal with little distortion.

The present invention provides an optimal adaptive filter with reduced circuit complexity, elimination of unnecessary delay time.

Means for Solving the Problems According to the present invention, an input layer in which a filter is composed of a multilayer perceptron and an input signal is sampled along a space axis or a time axis to obtain n (n is a natural number) sampled signals And a plurality of switches for switching between outputs of a plurality of weight multiplying means and constant value output means for obtaining a product of each value of the n sampled signals obtained from the input layer and a weighting coefficient held inside. A plurality of weight switches, a hidden layer composed of a plurality of first node arithmetic units for adding outputs of the plurality of weight switches, and outputting a value obtained by nonlinearly converting the added value; A weighting circuit for weighting each of the outputs of the node arithmetic unit, an output layer of the second node arithmetic unit that adds the outputs of the weighting circuits, and makes the final output the value obtained by nonlinearly converting the added value, In advance All of the plurality of weight switches are connected to the output of the weight multiplication means, and the weight multiplication is performed so that the error of the final output with respect to the training sampling signal is smaller than a first threshold value given in advance. Second threshold value for outputting a control signal for increasing / decreasing the weighting coefficient of the means and the weighting circuit, and a variance of values output by the plurality of weighting switches with respect to the training sampling signal is predetermined. A learning controller that switches the weight switch to the output of the constant value output means with the average value of the values output by the weight switch as the output of the constant value output means only for the weight switch having a smaller value. It has

By these means, the filter according to the present invention keeps all the weight switches connected to the output of the weight multiplication means during the training period so that the error of the final output becomes smaller than the first threshold value given in advance. Up to the weighting coefficient of the weighting means and the weighting circuit, the variance of the value output by the weighting switch with respect to the sampling signal during the training period is smaller than a predetermined second threshold value. For only the weight switch, the average value of the values output by the weight switch can be given. As a result, the complexity of the circuit is reduced, unnecessary delay time is eliminated, and the optimum adaptive filter configuration is possible.

Example An example of the present invention will be described below with reference to the drawings.

FIG. 1 shows the basic configuration of one embodiment of the adaptive filter according to the present invention. The input time series signal 1 becomes the sampling signals 11 to 19 by the delay circuits 2 to 10, and becomes the input of the weighting coefficient switch 20. The delay times of the delay circuits 2 to 10 are values determined by the frequency band of the input time series signal 1, and for example, when the input time series signal 1 is a baseband signal and its maximum frequency is fc, The delay times τ _{2 to} τ 10 of the delay circuits 2 to ₁₀ are selected to be τ _k ≦ 1 / 2fc (where k = 1, 2, 3, ...) (1). Also, the delay times do not necessarily have to be equal.

The first weight outputs 21 to 25 from the weight coefficient switch 20 are
It becomes an input to the first node arithmetic units 26 to 30. The configuration of the weight coefficient switch 20 will be described later.

In the first node arithmetic units 26 to 30, the outputs 31 to 35 are transmitted to the next stage according to the predetermined non-linear transmission characteristic with respect to the first weight outputs 21 to 25 which are the inputs. To do. As for how to select this transfer characteristic,
Documents already known, for example, "IEEE SSP Magagene" published in April 1987, "Unintroduction Two Computer Computing With Neural Net" (Richard B. Lippmann, "An Introduction to Co
mputing with Neural Nets ", IEEE ASSP MAGAZINE APRIL
1987)], etc., and therefore omitted.

The transmitted outputs 31 to 35 are passed through the weighting circuits 36 to 40 and are output to the second
Weighted outputs 41 to 45 are added to the second node calculator 46. The input / output transfer characteristic of the second node arithmetic unit 46 is predetermined as in the case of the first node arithmetic units 26 to 30, but need not necessarily be the same. The transfer characteristic of the second node calculator 46 can be determined in the same manner as the transfer characteristics of the first node calculators 26 to 30 described above. The output of the second node calculator 46 is transmitted to the outside as a filter output 47 for the input time series signal 1.

Further, the learning controller 48 controls the weight coefficient switch 20 and the weight circuits 36 to 40 by a weight control signal 49 and a weight coefficient / switch control signal 50 in order to control the amount of weight and the input signal. Is output. Each of these signals has a weighting circuit 36.
˜40 and the control signal for the weight coefficient switch 20. These control signals control the increase and decrease of the respective weight amounts so that the final filter output 47 has a desired value according to the already known learning rules of the multilayer perceptron. At the same time, control of the changeover switch inside the weighting factor switch 20 described later and control of the threshold level are performed. Therefore, the learning controller 48 is realized by a combination of the already known signal processor and its program, and the details will be omitted.

FIG. 2 is a block connection diagram for explaining the internal structure of the weighting factor switch 20 shown in FIG. 1 in more detail. The sampled signal 11 is applied to the same number of weight switches 51 to 55 as the number of weight output signals (five in this example). The other sampled signal 12 is similarly applied to weight switches 56-60. The other sampling signals are added to the same weight switch group. In FIG. 2, only the weight switches for the sampled signals 11 and 12 are shown, but actually, all the sampled signals have the same configuration.

Also, the outputs of the weight switches are added one by one for each sampled signal, and the weighted outputs 21 to 25 are output.
Becomes Therefore, if the transfer function of the weight switch is g (x), the i-th weight output 0i is 0i = Σg _{ij (xj)} (2). However, g _ij is the j-th sampled signal x _j and i
Is the transfer function of the weight switch that connects to the th weight output o _i .

FIG. 3 is a block connection diagram for explaining the inside of the weight switch in FIG. To the input terminal 61, one of the sampling inputs of the weighting factor switch 20 in FIG. 1 is connected. The variable conductance 62 shows the voltage-current characteristic controlled by the weighting factor / switch control signal 50 from the learning controller 48 in FIG. The output current of the variable conductance 62 becomes an output 65 via the switch 63. Output 65 is connected to one of the weight outputs of weighting factor switch 20 in FIG. The switch 63 is a changeover switch, and can also be changed over to the variable current source 64 by the weighting factor / switch control signal 50. Variable current source 64 may also have its output current controlled by weighting factor and switch control signal 50. Therefore, the output 65 of the weight switch shown in FIG. 3 is a current signal represented by u = αx _i + k (3), where u is the value. Here, α is the value of the variable conductance 62, x _i is the signal level of the input terminal 61, and k is a constant. However, either α or k is zero, but which becomes zero is determined by the position of the switch 63. Therefore, each term on the right side of the equation for each weight output shown in the equation (2) is the output u of each weight switch shown in the equation (3).
And the left side is the current of the sum of u.

Next, how the filter of the present invention adaptively determines its characteristics by these configurations will be described.

First, prior to the actual transmission of the signal, the code sequence as the training signal is transmitted, and the input time series signal 1 of FIG.
Is added to the filter as. At this time, in the weighting factor switch 20, all the switches inside thereof are set to the side a as the switch 63 shown in FIG. Therefore, the output of each weight switch has a current value determined by k = 0 in the equation (3). At this time, the learning controller 48 holds the true value of the code string, which is the training signal, as information. Therefore, according to the difference between the filter output 47 and the true value, Weight and weight circuit 36 ~
And the transfer characteristic of 40 to control the filter output 47 so as to best approximate the true value. A specific method of controlling each weight may be based on a known adaptive filter training method, and will not be described here. Filter output 47 during this training period
And the error between the true value and the true value changes as shown in FIG.

That is, as shown in FIG. 4, the training is a first threshold value in which the error value is a predetermined small enough error.
Repeat until it reaches 70. When the learning controller 48 detects that the value of the error reaches the first threshold value 70, the learning controller 48 terminates the training, and from the value of each weight in FIG. Calculate what value to respond to the input. Then, for the weight switch whose variance of the value of the response is smaller than the predetermined second threshold value, the switch is switched by the weight coefficient / switch control signal 50.
At the same time, the variable current source 64 in FIG. 3 is controlled so as to output the average value of the response values. As a result, the path that contributes little to the transmission of the signal is
Separated from the filter device. If the input time-series signal 1 has a small code distortion, the path for the sampled signal with a large delay amount naturally satisfies the above condition, so that there is no influence on the output. That is, the filter output 47
However, the amount of time delay with respect to the input time series signal 1 is the smallest. In other words, according to this embodiment, the adaptive filter with the smallest delay time is realized.

Although these components use the concept of hardware for convenience of explanation, the same operation can be realized by a known signal conversion circuit, an electronic computer, and a control program for controlling the same.

As described above, according to the present invention, during the training period, all the weight switches are connected to the output of the weight multiplication means, and the error of the final output becomes smaller than the first threshold value given in advance. Up to a value smaller than a predetermined second threshold value by increasing or decreasing the weighting coefficient of the weight multiplying means and the weighting circuit, and the variance of the value output by the weighting switch with respect to the sampling signal during the training period. For only the weight switch, the average value of the values output by the weight switch can be given. That is, this reduces the complexity of the circuit, eliminates unnecessary delay time, and enables the optimum adaptive filter configuration.

[Brief description of drawings]

FIG. 1 is a block connection diagram of an adaptive filter according to an embodiment of the present invention, FIG. 2 is a block connection diagram showing the internal configuration of the weighting factor switch shown in FIG. 1 in more detail, and FIG. FIG. 4 is a block connection diagram showing the inside of the weight switch in more detail, and FIG. 4 is a waveform diagram showing the residual of the output of the adaptive filter during the training period. 2-10 ... Delay circuit, 20 ... Weighting factor switch, 26-30
...... First node calculator, 36 to 40 ...... Weight circuit, 46 ......
Second node calculator, 48 ... Learning controller.

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification code Internal reference number FI Technical display location G06F 17/10 H03H 17/02 G 8842-5J

Claims (1)

[Claims] 1. When filtering one-dimensionally or two-dimensionally distributed signals spatially or temporally, the signals are sampled along the spatial axis or the time axis, and n samples (n is a natural number) are sampled. An input layer for obtaining a digitized signal, outputs of a plurality of weight multiplying means for obtaining a product of each value of the n sampled signals obtained from the input layer and a weighting coefficient held therein, and a constant value output means A plurality of weight switches for switching between outputs of the plurality of weight switches, a hidden layer including a plurality of first node arithmetic units for adding outputs of the plurality of weight switches and outputting a value obtained by performing non-linear conversion of the added value, A weighting circuit for weighting each of the outputs of the plurality of first node computing units, and a second node computing unit for adding the outputs of the weighting circuits and nonlinearly converting the added value to be the final output An output layer consisting of All of the plurality of weight switches are connected in advance to the output of the weight multiplying means, and the error of the final output with respect to the sampling signal for training becomes smaller than the first threshold value given in advance. A control signal for increasing / decreasing the weighting coefficient of the weight multiplying means and the weighting circuit is output, and the variance of the values output by the plurality of weighting switches with respect to the sampling signal for training is predetermined. Learning only by switching the weight switch to the output of the constant value output means with respect to only the weight switch having a value smaller than the threshold value of, the average value of the values output by the weight switch as the output of the constant value output means. An adaptive filter consisting of a multilayer perceptron consisting of a controller.