JPH0350909A - Prediction device - Google Patents

Abstract

PURPOSE:To attain prediction with high accuracy when there is any periodicity in an input signal by using a correlation coefficient so as to find out a waveform similar to a waveform at a current sample time from a signal string resulting from oversampling an input signal string and obtaining a prediction value through the use of a past signal. CONSTITUTION:Let a current sampling time be (j), a signal X(j) is inputted from a terminal 1 and a storage circuit 2 stores L sets of past signals X. Moreover, the signal X(j) is inputted to an oversampling circuit 3, a zero insertion circuit 31 in the circuit 3 inserts a zero at an interval of one sampling signal string to output a signal X2(j) whose sampling frequency Fs is twice that of the signal X(j). Then the frequency spectrum from Fs/4 to Fs/2 of the signal X2(j) is equal to the reflected frequency spectrum of the signal X(j). Then an LPF 32 outputs a signal XL whose frequency component is nearly equal to that of the signal X(j) and whose sampling frequency only differs. Then the storage circuit 4 stores by 2K preceding over-sampled signal. Then a correlation coefficient calculation section 5 applies the correlation evaluation between the L sets of past signals and the signal XL.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a technology for predicting digital signals such as voice, music, and images.

[Conventional technology]

There is an adaptive predictor as a conventional predictor that predicts the input signal at the current sample time from the input signal up to the current sample time. For a detailed overview of this method, refer to the theory of estimation and adaptive signal processing in ``Theory of Digital Signal Processing'' published by Corona Co., Ltd. Hereinafter, as an example of an adaptive predictor, a prediction principle based on a probability approximation method using a gradient, which is a type of linear prediction method, will be briefly described.

FIG. 3 is a block diagram of a conventional adaptive predictor based on the probability approximation method. 11 is an input terminal, 12 is a storage circuit, 13 is a subtracter, 14 is a prediction coefficient update section, 15 is a prediction value calculation section, 1
6 is a delay device.

The predictor uses the signals before visual sampling time j, that is, X(
It predicts the input signal X N+1) at the main time of the flash using the signal from X (j) to X (j), and if the predicted value is set as XP N+1), then
At(j)·X(j−N+1)·(1). Here, A, N) are prediction coefficients at time j,
If the prediction error signal E(j) at time j is set as the following equation, E (J)=X (j)-XP (j)...
-(2) Change each coefficient so as to minimize the power E''(j) of the prediction error signal. The power EZ(j) of the prediction error signal
) is E”(j) − (X (j) − is =-2E(j) (j i) ・(4) Therefore, in order to minimize the prediction error signal power E''(j), each prediction coefficient A = (j) is as shown below. change it like this.

A; (N+1) =A+ N)+g-E (j) ・X(j-i)...
(5) g is a correction coefficient, which is a constant determined according to the characteristics of the input signal, and usually uses a positive value sufficiently smaller than 1.

The signal flow at sample time j will be explained below. The storage circuit 12 stores the signal X input from the input terminal 11 in the number of prediction coefficients. The subtracter 13 receives the input signal
The prediction error E (j) for the human input signal at the current sample time is calculated by subtracting the prediction signal XP (j) from (j). The prediction coefficient update unit 14 calculates the prediction error E(j) according to 2(5).
and the input signal X in the memory circuit 12 (+1 to j-) to X
Using the signal of (j), predict coefficient Az(j) to A(j
) and update the new prediction coefficient A+ (j + 1)
Find Am(N+1) from The predicted value calculation unit 15 calculates A+=(N+1)− from the predicted coefficient Az(N+1) updated according to equation (5) and the input signal X(
+1) to j-, and from X (j), the predicted value XP N+1) of the input signal at the main time of the spark is determined. The delay unit 16 stores the predicted value XP (N+1) used when the subtracter 13 calculates the prediction error signal at the actual time j+1.

[Problem to be solved by the invention]

Conventional adaptive predictors have the disadvantage that high-order prediction coefficients are required for input signals with complex frequency characteristics, which increases the time required for the prediction coefficients to converge.

The purpose of the present invention is to eliminate such conventional drawbacks, and to eliminate the need for convergence time to achieve high prediction accuracy even for signals with complex frequency characteristics as long as the input signal has periodicity. The object of the present invention is to provide a predictor that makes highly accurate predictions.

[Means to solve the problem]

The predictor of the present invention includes: a first storage circuit that stores an input signal; an oversampling circuit that multiplies the sampling frequency of the input signal by an integer; and a second storage circuit that stores the output signal of the oversampling circuit. and a correlation coefficient calculation unit that calculates a correlation coefficient and an average amplitude ratio between the signal in the first storage circuit and the signal in the second storage circuit, and a correlation coefficient calculation unit that calculates a correlation coefficient and an average amplitude ratio between the signal in the first storage circuit and the signal in the second storage circuit, and a correlation coefficient calculation unit that calculates a correlation coefficient and an average amplitude ratio between the signals in the first storage circuit and the second storage circuit, and a maximum value detection circuit for selecting a time at the time; and calculating the product of the average amplitude ratio at the time, the sign of the correlation coefficient at the time, and a signal one sample time after the time in the second storage circuit. It is characterized by consisting of a multiplier.

[Operation] Past input signals are oversampled and saved, and a portion whose waveform is similar to the signal near the current sample time is selected.

If the input signal has periodicity, it is predicted that the signal part at the current time of the flash will be similar to the signal sequence of the past similar part, so the average amplitude ratio is applied to the next signal value of the past similar part. The multiplied value can be used as a predicted value. By using this predictor, accurate prediction can be made regardless of the frequency characteristics of the input signal.

〔Example〕

FIG. 1 is an embodiment of a predictor according to the present invention.

1 is an input terminal, 2 is a storage circuit, 3 is an oversampling circuit, 4 is a storage circuit, 5 is a correlation coefficient calculating section, 6 is a maximum value detection circuit, 7 is a multiplier, and 8 is an output terminal.

Letting the current sample time be j, the signal X (j) is input to input terminal 1.
Input from The memory circuit 2 stores the past signals X. That is, the signal X(j-L) is discarded and Xl-L
+1) to X(j) are stored in the memory circuit 2.

Further, the input signal XN) is input to the oversampling circuit 3. The oversampling circuit 3 multiplies the sampling frequency of the input signal by an integral number, and in this embodiment, an example in which the sampling frequency is doubled is shown. The oversampling circuit 3 includes a zero insertion circuit 31 and a low-pass filter 32. The zero insertion circuit 31 inserts a zero value every other time between the sampling signal strings, and converts the sampling frequency F into the signal X (j)
It outputs a signal X2 (j) which is twice the value. Signal X2
The frequency spectrum of (,r) from zero to F,/4 is equal to the frequency spectrum of signal X(j). signal
The frequency spectrum from Fs/4 to F2/2 for 2<j) is equal to the folded frequency spectrum of the signal X(j). The low-pass filter 32 filters the signal X2 (
The frequency components from F,/4 to F2/2 of signal X(j) are attenuated, and a signal XL whose frequency components are approximately equal to those of signal X(j) and differs only in sampling frequency is output. For details on the design method of this low-pass filter 32, please refer to the above-mentioned "Digital Filter Theory in Theory of Digital Signal Processing J" published by Corona Co., Ltd.
Letting S be the delay time that occurs in the input and output of ) is output. The storage circuit 4 stores the oversampled signal XL in the past two times, that is, XL(j −s− +
(1/2)) to XLN-s). Since oversampling is performed, the sampling time of signal XL is 0.
．． It will be 5 units. The correlation coefficient calculation unit 5 evaluates the correlation between the oversampled signal XL and the past signals from the current sample time j stored in the storage circuit 2, and calculates the average amplitude ratio. In this embodiment, as an evaluation of similarity, a correlation coefficient P (t,) is calculated as shown in the following equation.

Here, XI and x2 are respectively X(j
− to +1) to X (j) signal, and XL(mu−
+1) to XL (t).

It is.

The correlation coefficient P (t) indicates the degree of similarity of the components of the signal from which the DC component is removed. Therefore, even if a DC component is superimposed on the interval for which the correlation coefficient is calculated, the degree of similarity can be calculated without being influenced by the DC component. The value takes a value between -1 and 1, and the closer the absolute value is to 1, the more similar the two waveforms are. When P(t,) is positive, the two waveforms are in phase, and when P(t,) is negative, it indicates that they are in opposite phases.

The average amplitude ratio AC(t) is the average value of the amplitude ratios of the past signals from the current sample time j stored in the storage circuit 2 and the oversampled signal XL,
The square root of the power ratio is used as shown in the following equation.

・ ・ ・(8) Since the number of signals XL stored in the memory circuit 4 is 2 times, the correlation coefficient calculation unit 5 calculates P (j−s−+(L/2
)) to P(j-s) (2 to -L+1) correlation coefficients P and average amplitude ratio AC are calculated and output to the maximum value detection circuit 6. The maximum value detection circuit 6 is connected to the above-mentioned (2 to -L+
1) Find the time M at which the absolute value of P is maximum from among the correlation coefficients P, and output the time M, the correlation coefficient P (M), and the average amplitude ratio AC (M). The relationship between signal X and signal XL is as shown in FIG. 2(a) shows the signal X in the memory circuit 2, and FIG. 2(b) shows the signal XL in the memory circuit 4. A past signal sequence XL that has been oversampled so that the sample ratio value series is most similar to the sample ratio value series of the signal from +1 at sample time j- to the current sample time j.
Section C of XL(M) is selected from (M-(L/2)+(1/2)).

When the human signal has periodicity, the signal X(j
+1) is predicted to be similar to the signal XL(M+1) at the past sample time M+1, so the value obtained by multiplying the signal XL(M+1) by the average amplitude ratio AC(M) and the sign of the correlation coefficient P(M) is the predicted value signal XP (j+1) of the signal X (j+1)
) can be used as Therefore, the multiplier 7 calculates the product of the signal XL(M+1) stored in the storage circuit 4, the average amplitude ratio AC(M), and the sign of the correlation coefficient P(M),
Predicted value XP(j+1
) from the output terminal 8.

[Effects of the Invention] According to the present invention, a portion similar to the waveform near the current sample time is found from a signal sequence obtained by oversampling an input signal sequence using a correlation coefficient, and a predicted value is obtained using a past signal. When the input signal has periodicity, accurate prediction can be performed without requiring a convergence time until the prediction accuracy becomes highly accurate, regardless of the frequency characteristics of the input signal.

Furthermore, since oversampling is performed on the input signal to be stored, the influence of the sampling frequency of the input signal on prediction accuracy can be reduced.

[Brief explanation of drawings]

FIG. 1 is a diagram showing an embodiment of the present invention, FIG. 2 is a diagram showing a signal X stored in a storage circuit, and a signal XL stored in a storage circuit, and FIG. 3 is a diagram showing a conventional application prediction. FIG. 1... Input terminal 2... Memory circuit 3... Oversampling circuit 31...
Zero insertion circuit 32...Low pass filter 4...Storage circuit 5...Correlation coefficient calculation section 6...Maximum value detection circuit 7...Multiplier 8 ・Output terminal

Claims (1)

[Claims] (1) a first storage circuit that stores an input signal; an oversampling circuit that multiplies the sampling frequency of the input signal by an integer; a second storage circuit that stores the output signal of the oversampling circuit; a correlation coefficient calculation unit that calculates a correlation coefficient and an average amplitude ratio between a signal in the first storage circuit and a signal in the second storage circuit; a maximum value detection circuit to select; and a multiplier that calculates the product of the average amplitude ratio at the time, the sign of the correlation coefficient at the time, and a signal one sample time after the time in the second storage circuit. A predictor comprising: