CN101510858B - Channel long-range forecast method based on slope correction - Google Patents

Abstract

The invention discloses a channel long-term prediction method based on gradient correction, which comprises the following steps of: (1) conducting short-term prediction to a channel and recording the channel gain during the short-term prediction of the channel; (2) utilizing strong association of the channel sending frames in coherence time and utilizing short-term prediction algorithm to conduct rough long-term prediction to the channel; and (3) correcting the prediction results obtained from step (2) in accordance with the gradient correction method. By correcting the long-term prediction of the channel, the method leads the long-term prediction value to conform to the real channel better without obvious hysteresis phenomenon. The method can significantly improve the precision of direction application of short-term prediction to the long-term prediction and can be used for the long-term prediction of the channel of a wideband wireless communication system with a sub-packet transmission mode (such as OFDM, SC-FDE, etc.). When the correction is not carried out, long-term prediction still has a correct trend but can not conform to the real channel well and has obvious hysteresis phenomenon.

Description

A kind of channel long-range forecast method based on the slope correction

Technical field

The present invention relates to wideband digital communications method, belong to the broadband wireless communication technique field.

Background technology

Along with the development of Internet and multimedia service, can provide the requirement of broadband high-speed data transport service more and more higher to mobile radio system.Because available frequency resource is very limited,, improves message transmission rate and can only rely on development to have the more new technology of spectral efficient simultaneously because wireless communication system generally is subjected to strict Power Limitation.OFDM (Orthogonal Frequency division Multiplexing, hereinafter to be referred as OFDM) and single carrier frequency domain equalization (SingleCarrier Frequency Domain Equalization is hereinafter to be referred as SC-FDE) because the advantage of its high data rate, high spectrum utilization is subjected to people day by day more and more pays close attention to.

OFDM is a kind of multi-carrier modulation technology, and its main thought is to use the mode of parallel data and frequency multiplexing to alleviate multipath and disturbs the intersymbol interference that causes, thereby avoids using the equalizer of high complexity, and has reached the higher availability of frequency spectrum simultaneously.OFDM is a kind of piecemeal transmission technology, and the piecemeal transmission is meant information waiting for transmission is divided into the identical data block of length, adds corresponding Cyclic Prefix before the rf modulations before each data block of time domain.

In ofdm system, the sub-carrier number of being divided when the length of a time-domain symbol and frequency domain equalization is N, and (cyclic prefix, CP) length is L to Cyclic Prefix.The frequency domain response of time varying channel is the vectorial H in N * 1 _{N, k}=[H _{N, 0}..., H _{N, N-1}] ^T(n represents to send the frame label of n frame, k=0, and 1 ..., N-1 represents k subcarrier).Channel impulse response is h _{N, l}, (l=0,1 ..., L-1), supposed that here the length of Cyclic Prefix equals the length of channel impulse response.The mapping symbols that the sends vectorial b in N * 1 _{N, k}=[b _{N, 0}, b _{N, 1}..., b _{N, N-1}] ^T(n represents to send the frame label of n frame, k=0, and 1 ..., N-1 represents k subcarrier) and expression, its frequency domain form vectorial B in N * 1 _{N, k}=[B _{N, 0}..., B _{N, N-1}] ^T(n represents to send the frame label of n frame, k=0, and 1 ..., N-1 represents k subcarrier) and expression, receiving symbol is expressed as N * 1 matrix X before frequency domain equalization _{N, k}=[X _{N, 0}..., X _{N, N-1}] ^T(n represents to send the frame label of n frame, k=0, and 1 ..., N-1 represents k subcarrier).Z _{N, k}=[Z _{N, 0}..., Z _{N, N-1}] ^T(n represents to send the frame label of n frame, k=0, and 1 ..., N-1 represents k subcarrier) and the expression additive white Gaussian noise (additive white Gaussian noise, AWGN).Because the effect of Cyclic Prefix, the relation of input and output symbol can be expressed as:

X _n，k＝H _n，kB _n，k+Z _n，k，k＝0，1，…，N-1

The same with OFDM, SC-FDE also is a kind of important piecemeal transmission technology.The implementation procedure of SC-FDE is as follows: the data of transmission are after sign map, add CP again, then successively through the D/A conversion, intermediate frequency, send after the rf modulations, receiving terminal is finished radio frequency successively, intermediate frequency demodulation, after the A/D conversion, remove the CP in the received signal earlier, then it is fast fourier transform (Fast Fourier Transform, hereinafter to be referred as the FFT conversion), the channel condition information that obtains according to channel estimating carries out equilibrium to received signal again, and balanced data is inverse fast fourier transform (Inverse Fast Fourier Transform, hereinafter to be referred as IFFT), at last the data behind the IFFT are carried out symbol detection, obtain final output signal.

In mobile communication, because reflection, diffraction and scattering exist all the time, unavoidably there is multipath transmisstion, because factor affecting such as intensity, time delay, signal bandwidths, composite signal intensity and phase place can change, the decline that causes thus is called multipath fading.Abominable wireless propagation environment causes the transmission signals distortion.In order to reduce the error rate of system, the signal that the common channel information equilibrium that utilizes estimation to obtain receives.But for Quick-Change channel, the channel condition information that utilizes traditional channel estimation methods based on decision-feedback to obtain is out-of-date information, can not reflect current channel conditions.This just requires to utilize channel estimating to obtain following channel state information comparatively accurately.Channel estimating is meant current according to channel and following channel condition information (CSI) of the prediction of historical data in the past, the current and past data information of channel wherein, or to be called the observed value source mainly be the channel response value that channel estimating obtains.No matter be that the 3G system or following 4G system have all adopted technology for self-adaptively transmitting in large quantities in order to support data service at a high speed, such as adaptive power control (Power Control), Adaptive Modulation, mix automatic request retransmission, adaptive coding technology etc.These technology for self-adaptively transmitting all need to obtain timely and accurately the variation tendency of channel: the variation tendency of state information on the variation of channel overall gain, each subchannel, thus carry out Adaptive Transmission.

Channel estimating divides two kinds of long-term forecast and short-term forecasts.Short-term forecast refers to the channel gain of accurately predicting interval one frame or several frames, is used for equilibrium, improves systematic function.Fig. 1 is the mechanism block diagram of fallout predictor in system.Long-term forecast refers to the situation of change of the channel gain of predicted channel after longer a period of time of interval, and the adaptive algorithm that can be multiple channel provides foundation.

The meaning of channel long-range forecast promptly is: no matter change fast still slow, the channel condition information that utilizes channel estimation methods to obtain all can't obtain following channel condition information, channel estimating can go out CSI in the future according to channel state prediction before, carries out self-adaptive processing.Channel long-range forecast is with respect to short-term forecast, and precision of prediction is generally relatively poor, and some application are had a negative impact.Existing short-term forecast method also can be expanded and carry out channel long-range forecast, can have problems. such as: channel condition information is not accurate enough, and trend changes hysteresis or the like.The present invention is based on existing channel short-term accurately predicting, a kind of method of long-term channel accurately predicting is proposed: based on the short-term forecast algorithm, predicted value is carried out the slope correction, to revising by result to short-term forecast, revise the long-term forecast result to reach, can significantly improve the precision of long-term forecast.

Existing channel short-term forecast algorithm has following several:

1.MMSE algorithm:

Suppose with sample rate f _sChannel response is carried out discretization, so, for l (l=0,1 ..., L-1) footpath channel is based on M past sample h _l(n), h _l(n-1) ..., h _l(n-M+1), the following response of prediction h _l(n+p) model is as follows:

${\hat{h}}_l(n+p)=\underset{j=0}{\overset{M-1}{\mathrm{\Σ}}}{c}_{l,j}{h}_l(n-j)---\left(1\right)$

M is the forecast model exponent number in the formula, coefficient c _{L, j}(l=0,1 ..., L-1) be $E\left[{\left|e_l\left(p\right)\right|}^2\right]=E{|h_l\left(p\right)-{\hat{h}}_l\left(p\right)|}^2]$ Minimum MMSE

The optimum coefficient of correspondence as a result.Promptly work as: $E\left[{\left|e_l\left(p\right)\right|}^2\right]=1-\underset{j=0}{\overset{M-1}{\mathrm{\Σ}}}{c}_{l,j}{h}_l(n-j)$ The time, optimum coefficient is:

c _l＝R _l ^-1r _l (2)

C wherein _l=(c _{L, 1}, c _{L, 2}..., c _{L, M}) ^T, R _l, for (autocorrelation matrix of M * M), its each component is R _{L, ij}=E[h _l(n-i) h ' _l(n-j)].r _lBe (M * 1) auto-correlation vector, its each component is r _Ij=E[h _l(n) h ' _l(n-j)].If p=1 then claim that fallout predictor is the single step fallout predictor.Formerly do not know to estimate R from the sample of observing under the situation of maximum doppler frequency or scattering wave number _{L, ij}Notice that the sample sampling rate must meet Nyquist rate, promptly be at least the twice of maximum doppler frequency.The sampling rate of selecting is several times as much as Nyquist rate, is significantly less than data rate, predicts the complexity that reduces prediction greatly with the channel response of this speed, also obtains the estimated performance better than data rate situation.The channel response of higher rate can be realized by interpolation method based on predicted value.

2.Volterra algorithm:

This is the nonlinear algorithm that a kind of secondary filter is realized.Wherein import data (or observation data) for x (n)=[x (n), x (n-1) ..., x (n-N+1)] ^T, according to a p rank Nonlinear Volterra state extended operation H _p, with observation data expand to u (n)=[u (n), u (n-1) ..., u (n-M+1)] ^T, and M 〉=N, the Nonlinear Volterra operation is defined as:

H ₀＝h ₀

$H_1\left[x\left(n\right)\right]=\underset{i}{\mathrm{\Σ}}{h}_{i}x(n-i)---\left(3\right)$

$H_2\left[x\left(n\right)\right]=\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}{h}_{i,j}x(n-i)x(n-j)$

The system of it has been generally acknowledged that is a second nonlinear, for such (N ₁, N ₂) second-order system, linear adaptive filter is output as:

$y\left(n\right)={\mathrm{<figure-callout\; id="0"\; label="h"\; filenames="H2009100197396E00022.png"\; state="\{\{state\}\}">h</figure-callout>}}_0+\underset{i=0}{\overset{N_1-1}{\mathrm{\&Sigma;}}}{h}_{i}x(n-i)+\underset{i=0}{\overset{N_2-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{N_2-1}{\mathrm{\&Sigma;}}}{h}_{i,j}x(n-i)x(n-j)---\left(4\right)$

Can see that total coefficient number is:

M＝1+N ₁+N ₂(N ₂+1)/2 (5)

If the input of definition linear adaptive filter

u(n)＝[1，x(n)，x(n-1)，…，x(n-N ₁+1)，x ²(n)，x(n)x(n-1)，…，x ²(n-N ₂+1)] ^T

(6)

And coefficient vector is written as:

$H\left(n\right)={[h_0,{\mathrm{<figure-callout\; id="1"\; label="h"\; filenames="H2009100197396E00022.png,H2009100197396E00041.png"\; state="\{\{state\}\}">h</figure-callout>}}_1,...,h_{N_1-1},...,{\mathrm{<figure-callout\; id="0"\; label="h"\; filenames="H2009100197396E00022.png"\; state="\{\{state\}\}">\; <figure-callout\; id="0"\; label="h"\; filenames="H2009100197396E00022.png"\; state="\{\{state\}\}">h</figure-callout>\; </figure-callout>}}_{\mathrm{0,0}},{\mathrm{<figure-callout\; id="0"\; label="h"\; filenames="H2009100197396E00022.png"\; state="\{\{state\}\}">\; <figure-callout\; id="1"\; label="h"\; filenames="H2009100197396E00022.png,H2009100197396E00041.png"\; state="\{\{state\}\}">h</figure-callout>\; </figure-callout>}}_{\mathrm{0,1}},...,h_{N_2-1,N_2-1}]}^T---\left(7\right)$

Then the output of linear adaptive filter can be expressed as:

y(n)＝H ^T(n)u(n) (8)

So just can utilize the various adaptive algorithms of linear filtering to ask coefficient H (n).

3. subspace channel prediction arithmetic:

The Subspace Decomposition algorithm is used in the channel estimating, also is a relatively more popular direction in the current prediction algorithm research, and this respect ESPRIT and MUSIC algorithm have typicalness.For each tap of flat fading channel or broadband system, its frequency domain channel function can be thought the stack of a large amount of scattering components.ESPRIT class channel prediction method integral body in the algorithm of subspace is divided into two steps:

At first, be the stack of harmonic component with Channel Modeling, according to the rotational invariance of signal subspace, structural matrix bundle or title matrix are right, and the generalized eigenvalue of pencil of matrix just provides harmonic frequency so.

Then, utilize least square method to calculate the range value of each frequency component.So just can obtain the predicted value of channel status.

4. adaptive filter algorithm:

This class algorithm belongs to the MMSE algorithm, adopts gradient algorithm.Complexity than the MMSE algorithm of front is low.Introduce the RLS algorithm below, and algorithm is placed on specifically introduction in the SC-FDE system.

5.RLS algorithm:

If the sub-carrier number of being divided when the length of a time-domain symbol of branch block system and frequency domain equalization is N, cyclic prefix CP length is L.The frequency domain response of time varying channel is the vectorial H in N * 1 _{N, k}=[H _{N, 0}..., H _{N, N-1}] ^T(n represents to send the frame label of n frame, k=0, and 1 ..., N-1 represents k subcarrier).Channel impulse response is h _{N, l}, (l=0,1 ..., L-1) (n represents to send the frame label of n frame, l=0, and 1 ..., L-1 represents the 1st footpath channel), supposed that here the length of Cyclic Prefix equals the length of channel impulse response.The mapping symbols that the sends vectorial b in N * 1 _{N, k}=[b _{N, 0}, b _{N, 1}..., b _{N, N-1}] ^T(n represents to send the frame label of n frame, k=0, and 1 ..., N-1 represents k subcarrier) and expression, its frequency domain form vectorial B in N * 1 _{N, k}=[B _{N, 0}..., B _{N, N-1}] ^T(n represents to send the frame label of n frame, k=0, and 1 ..., N-1 represents k subcarrier) and expression, receiving symbol is expressed as N * 1 matrix X before frequency domain equalization _{N, k}=[X _{N, 0}..., X _{N, N-1}] ^T(n represents to send the frame label of n frame, k=0, and 1 ..., N-1 represents k subcarrier).Z _{N, k}=[Z _{N, 0}..., Z _{N, N-1}] ^T(n represents to send the frame label of n frame, k=0, and 1 ..., N-1 represents k subcarrier) and expression additive white Gaussian noise AWGN.Because the effect of Cyclic Prefix, the relation of input and output symbol can be expressed as:

X _n，k＝H _n，kB _n，k+Z _n，k，k＝0，1，…，N-1 (9)

The cost function of RLS algorithm is:

$\mathrm{\&xi;}\left(n\right)=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}^{n-i}{\left|e\left(i\right)\right|}^2+\mathrm{\&delta;}{\mathrm{\&lambda;}}^n{\left|\right|w\left(n\right)\left|\right|}^2---\left(10\right)$

Wherein, n represents current time, just n the moment that transmission signals arrives.I, (i=1,2 ..., n) the expression n and the moment in the past are corresponding one by one with the time sequencing of n symbol.λ, (0＜λ≤1) is forgetting factor.E (i)=d (i)-y (i) expression Expected Response and the error between the i actual output of filter constantly.The memory span of supposing filter is M ₁, then i input vector constantly be defined as u (i)=[u (i), u (i-1) ..., u (i-M ₁+ 1)] ^TN tap weights vector constantly is defined as $w\left(n\right)={[w_0\left(n\right),w_1\left(1\right),...,w_{M_1-1}\left(n\right)]}^T.$ δ, (0＜δ≤1) is called stability factor or regularization parameter.The process of optimizing is exactly to seek the process of the tap weights vector that makes the cost function minimum.Utilize optimum tap weights vector i output constantly can be expressed as y (i)=w ^H(n) u (i).P step RLS Forecasting Methodology is operated in towards decision pattern, prediction is carried out after system's frequency domain equalization is finished, its basic structure in system is as shown in Figure 1: channel predictor is made up of L parallel RLS fallout predictor, and L path of L fallout predictor and channel impulse response is corresponding.For l fallout predictor, its input vector is $y_{n,l}={[y_{n,l},y_{n-1,l},...,y_{n-{\mathrm{<figure-callout\; id="1"\; label="M"\; filenames="H2009100197396E00022.png,H2009100197396E00041.png"\; state="\{\{state\}\}">M</figure-callout>}}_1+1,l}]}^T.$ The tap weights vector is $w_l\left(n\right)={[{\mathrm{<figure-callout\; id="0"\; label="w"\; filenames="H2009100197396E00022.png"\; state="\{\{state\}\}">w</figure-callout>}}_{0,l}\left(n\right),{\mathrm{<figure-callout\; id="1"\; label="w"\; filenames="H2009100197396E00022.png,H2009100197396E00041.png"\; state="\{\{state\}\}">w</figure-callout>}}_{1,l}\left(n\right),...,w_{M_1-1,l}\left(n\right)]}^T.$ L fallout predictor is output as Among Fig. 1 Be the frequency domain form that reconstruct sends data, reconstruct data is to obtain by the data that ruled out being carried out again the planisphere mapping, if system has adopted error correction coding, the judgement data also will be reconstructed through coding.In the part of back, convenient for the narration problem, do not consider to adjudicate the error that causes, promptly think the reconstruct symbol Identical with the frequency domain form that sends mapping symbols.

The step of RLS prediction algorithm is:

Algorithm initialization:

w _l(n)＝[1，0，0…0]，n＝0，1，…，M ₁-1，

$k_{0,l}=\frac{y_{0,l}}{{\left|\right|y_{0,l}\left|\right|}^2+\mathrm{\&delta;}},$

$P_{0,l}={\mathrm{\&delta;}}^{-1}(I-\frac{y_{0,l}{y^H}_{0,l}}{{\left|\right|y_{0,l}\left|\right|}^2+\mathrm{\&delta;}}),$

Wherein ${\left|\right|{\mathrm{<figure-callout\; id="0"\; label="y"\; filenames="H2009100197396E00022.png"\; state="\{\{state\}\}">y</figure-callout>}}_{0,l}\left|\right|}^2=\underset{i=0}{\overset{M-1}{\mathrm{\&Sigma;}}}{\left|y_{0-i,l}\right|}^2,$ N=0 represents training frames, the moment y before 0 _{N, l}=0.

To each moment n=1,2 ... ... the circulation below calculating:

Cyclic variable l=1 is set ..., L

Step 1, the prior estimate error of calculating filter:

e _n，l＝y _n，l-w _l ^H(n-1)y _n-p，l，n≥p

Step 2 is upgraded the tap weights vector, and calculates fallout predictor output:

w _l(n)＝w _l(n-1)+k _n-p，le ^* _n，l，n≥p

${\hat{h}}_{n+p,l}={w_l}^H\left[n\right]y_{n,l}$

Step 3, upgrade gain vector:

$k_{n,l}=\frac{P_{n-1,l}{y}_{n,l}}{\mathrm{\&lambda;}+{y^H}_{n,l}{P}_{n-1,l}{y}_{n,l}},n\&GreaterEqual;1$

Step 4, upgrade inverse correlation matrix:

P _n，l＝λ ^-1(I-k _n，ly ^H _n，l)P _n-1，l，n≥1

Loop ends.

For short-term forecast, this algorithm is based on the strong correlation that adjacent a few frames have, and this moment, short-term channel predicting interval M was 1, promptly predicted the next frame channel condition information.Channel has strong correlation in how many frames, can describe by the coherence time of channel, and be the inverse of Doppler frequency shift coherence time, T _Coh=1/f _m=c/vf _cFor example: under the COST207 environment, carrier frequency is 2G, and sample rate is 10MHz, the SC-FDE symbol lengths, and promptly sub-carrier number is 256, CP length is 64.When Doppler was 100Hz, be 0.01s coherence time.The frame number that transmits in coherence time is: N _ZP=10M/ (100* (256+64))=312.5.And the like, when Doppler was 500Hz, the frame number that transmits in coherence time was: N _ZN=10M/ (300* (256+64))=62.15.

Hence one can see that, under different Doppler frequency shifts, coherence time difference, the frame number with strong correlation is also different.Still can simply carry out long-term forecast with RLS short-term forecast method, long-term forecast also is based on the correlation between the same intervals frame.No longer be 1 with the interval M that the short-term forecast algorithm carries out rough channel long-range forecast this moment, but a predefined natural number, M=10 for example, but the predicting interval can not surpass the frame number of permission in correlation time.Because the correlation of the used frame of this simple long-term forecast is strong not as the correlation of the used frame of short-term forecast, so the time prediction can not be used for the precision height of short-term forecast as it, but its trend of predicting is correct.

Summary of the invention

The present invention is directed to the low problem of precision that the existing channel long-range forecast method exists, a kind of channel long-range forecast method based on the slope correction is provided, can significantly improve short-term forecast and be directly used in the precision of long-term forecast, can be used for the channel long-range forecast of the system of broadband wireless communication of piecemeal transmission means (as OFDM, SC-FDE etc.).

Performing step of the present invention is as follows:

(1) at first channel is carried out short-term forecast, and the channel gain of record short-term forecast at this moment;

(2) utilize the strong correlation of the channel of transmit frame in coherence time, utilize the short-term forecast algorithm to carry out rough channel long-range forecast;

(3) predicting the outcome in the step (2) revised according to the slope revised law.

The detailed implementation method of above-mentioned each step is as follows:

(1) at first channel is carried out short-term forecast;

If the sub-carrier number of being divided when the length of a time-domain symbol of branch block system and frequency domain equalization is N, cyclic prefix CP length is L, and the frequency domain response of time varying channel is the vectorial H in N * 1 _{N, k}=[H _{N, 0}..., H _{N, N-1}] ^T(n represents to send the frame label of n frame, k=0, and 1 ..., N-1 represents k subcarrier), channel impulse response is h _{N, l}(n represents to send the frame label of n frame, l=0, and 1 ..., L-1 represents the 1st propagation path), suppose that here the length of Cyclic Prefix equals the length of channel impulse response, the mapping symbols of the transmission vectorial b in N * 1 _{N, k}=[b _{N, 0}, b _{N, 1}..., b _{N, N-1}] ^T(n represents to send the frame label of n frame, k=0, and 1 ..., N-1 represents k subcarrier) and expression, its frequency domain form vectorial B in N * 1 _{N, k}=[B _{N, 0}..., B _{N, N-1}] ^T(n represents to send the frame label of n frame, k=0, and 1 ..., N-1 represents k subcarrier) and expression, receiving symbol is expressed as N * 1 matrix X before frequency domain equalization _{N, k}=[X _{N, 0}..., X _{N, N-1}] ^T(n represents to send the frame label of n frame, k=0, and 1 ..., N-1 represents k subcarrier).Z _{N, k}=[Z _{N, 0}..., Z _{N, N-1}] ^T(n represents to send the frame label of n frame, k=0, and 1 ..., N-1 represents k subcarrier) and expression additive white Gaussian noise AWGN, because the effect of Cyclic Prefix, the relation table of input and output symbol is shown:

X _n，k＝H _n，kB _n，k+Z _n，k，k＝0，1，…，N-1

Data are by transmission system, and front short-term forecast algorithm is by the agency of, is example with the RLS prediction algorithm now:

Algorithm initialization:

N tap weights vector constantly is defined as: w _l(n)=[1,0,0 ... 0], n=0,1 ..., M ₁-1, M ₁Memory span for filter

The gain vector initialization: ${\mathrm{<figure-callout\; id="0"\; label="k"\; filenames="H2009100197396E00022.png"\; state="\{\{state\}\}">k</figure-callout>}}_{0,l}=\frac{{\mathrm{<figure-callout\; id="0"\; label="y"\; filenames="H2009100197396E00022.png"\; state="\{\{state\}\}">y</figure-callout>}}_{0,l}}{{\left|\right|{\mathrm{<figure-callout\; id="0"\; label="y"\; filenames="H2009100197396E00022.png"\; state="\{\{state\}\}">y</figure-callout>}}_{0,l}\left|\right|}^2+\mathrm{\&delta;}},$

The inverse correlation matrix initialization: ${\mathrm{<figure-callout\; id="0"\; label="P"\; filenames="H2009100197396E00022.png"\; state="\{\{state\}\}">P</figure-callout>}}_{0,l}={\mathrm{\&delta;}}^{-1}(I-\frac{{\mathrm{<figure-callout\; id="0"\; label="y"\; filenames="H2009100197396E00022.png"\; state="\{\{state\}\}">y</figure-callout>}}_{0,l}{y^{\mathrm{<figure-callout\; id="0"\; label="H"\; filenames="H2009100197396E00022.png"\; state="\{\{state\}\}">H</figure-callout>}}}_{0,l}}{{\left|\right|{\mathrm{<figure-callout\; id="0"\; label="y"\; filenames="H2009100197396E00022.png"\; state="\{\{state\}\}">y</figure-callout>}}_{0,l}\left|\right|}^2+\mathrm{\&delta;}}),$

Wherein ${\left|\right|{\mathrm{<figure-callout\; id="0"\; label="y"\; filenames="H2009100197396E00022.png"\; state="\{\{state\}\}">y</figure-callout>}}_{0,l}\left|\right|}^2=\underset{i=0}{\overset{M-1}{\mathrm{\&Sigma;}}}{\left|y_{0-i,l}\right|}^2,$ N=0 represents training frames, the moment y before 0 _{N, l}=0.

To each moment n=1,2 ... ... the circulation below calculating:

Cyclic variable l=1 is set ..., L

Step 1, the prior estimate error of calculating filter:

e _n，l＝y _n，l-w _l ^H(n-1)y _n-p，l，n≥p

Step 2 is upgraded the tap weights vector, and calculates fallout predictor output:

w _l(n)＝w _l(n-1)+k _n-p，le ^* _n，l，n≥p

${\hat{h}}_{n+p,l}={w_l}^H\left[n\right]y_{n,l}$

Step 3, upgrade gain vector:

$k_{n,l}=\frac{P_{n-1,l}{y}_{n,l}}{\mathrm{\&lambda;}+{y^H}_{n,l}{P}_{n-1,l}{y}_{n,l}},n\&GreaterEqual;1$

Step 4, upgrade inverse correlation matrix:

P _n，l＝λ ^-1(I-k _n，ly ^H _n，l)P _n-1，l，n≥1

Loop ends;

Work as y _{N, l}Be frequency domain information H _nThe time, obtain short-term forecast channel gain H _N+1(n represents to send the frame label of n frame), and to the prediction data that obtains is carried out record, this moment, short-term channel predicting interval M was 1, promptly predicted the next frame channel condition information.For the frequency domain revised law, the channel frequency domain value H=[H that the record prediction obtains ₁, H ₂...]; For the time domain revised law, need pass through the channel gain H of prediction and write down real part respectively after the FFT conversion become time domain again $h_l^r=\mathrm{Re}\left(\mathrm{IFFT}\left(H\right)\right)$ And imaginary part $h_l^i=\mathrm{Im}\left(\mathrm{IFFT}\left(H\right)\right),(l=\mathrm{0,1},...,L-1)$ These data will be as the foundation of long-term forecast algorithm of the present invention.

(2) utilize the strong correlation of the channel of transmit frame in coherence time, carry out rough channel long-range forecast with the short-term forecast algorithm.No longer be 1 with the interval M that the short-term forecast algorithm carries out rough channel long-range forecast this moment, but one according to the frame number that transmits in coherence time, predefined natural number, M=10 for example, but the predicting interval can not surpass the frame number of permission in coherence time.RLS prediction algorithm for example: (following each parameter front is all described)

Algorithm initialization:

w _l(n)=[1,0,0 ... 0], n=0,1 ..., M ₁-1, M ₁Memory span for filter

${\mathrm{<figure-callout\; id="0"\; label="k"\; filenames="H2009100197396E00022.png"\; state="\{\{state\}\}">k</figure-callout>}}_{0,l}=\frac{{\mathrm{<figure-callout\; id="0"\; label="Y"\; filenames="H2009100197396E00022.png"\; state="\{\{state\}\}">Y</figure-callout>}}_{0,l}}{{\left|\right|{\mathrm{<figure-callout\; id="0"\; label="Y"\; filenames="H2009100197396E00022.png"\; state="\{\{state\}\}">Y</figure-callout>}}_{0,l}\left|\right|}^2+\mathrm{\&delta;}},$

${\mathrm{<figure-callout\; id="0"\; label="P"\; filenames="H2009100197396E00022.png"\; state="\{\{state\}\}">P</figure-callout>}}_{0,l}={\mathrm{\&delta;}}^{-1}(I-\frac{{\mathrm{<figure-callout\; id="0"\; label="Y"\; filenames="H2009100197396E00022.png"\; state="\{\{state\}\}">Y</figure-callout>}}_{0,\mathrm{<figure-callout\; id="0"\; label="l\; Y\; H"\; filenames="H2009100197396E00022.png"\; state="\{\{state\}\}">l</figure-callout>}}{Y^H}_{}{{}^{}}_{0,l}}{{\left|\right|{\mathrm{<figure-callout\; id="0"\; label="Y"\; filenames="H2009100197396E00022.png"\; state="\{\{state\}\}">Y</figure-callout>}}_{0,l}\left|\right|}^2+\mathrm{\&delta;}}),$

Wherein $Y_n=\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}\left(\sqrt{{h_{\mathrm{rn},k}}^2+{h_{\mathrm{in},k}}^2}\right)$ Be the overall gain of channel, n=0 represents training frames, the moment y before 0 _{N, l}=0.

To each moment n=1,1+M ... ... pre-set integer M is at interval, the circulation below calculating:

Cyclic variable l=1 is set ..., L

Step 1, the prior estimate error of calculating filter:

e _n，l＝Y _n，l-w _l ^H(n-1)Y _n-M，l，n≥M

Step 2 is upgraded the tap weights vector, and calculates fallout predictor output:

w _l(n)＝w _l(n-1)+k _n-M，le ^* _n，l，n≥M

H _n+M，l＝w _l ^H[n]Y _n，l

Step 3, upgrade gain vector:

$k_{n,l}=\frac{P_{n-1,l}{Y}_{n,l}}{\mathrm{\&lambda;}+{Y^H}_{n,l}{P}_{n-1,l}{Y}_{n,l}},n\&GreaterEqual;1$

Step 4, upgrade inverse correlation matrix:

P _n，l＝λ ^-1(I-k _n，lY ^H _n，l)P _n-1，l，n≥1

Loop ends;

Through M ₁After the inferior circulation, obtain the M+1 frame, 2M+1 frame, 3M+1 frame ... M ₁The channel condition information of these several frames of M+1 frame, and the at interval channel condition information of M frame, M simultaneously after can predicting successively ₁The difference of the frame number of M+1 frame and M+1 frame will allow to send within the frame number in coherence time.

As seen from Figure 7, this moment, the information of prediction was very inaccurate, but trend is predicted out.

(3) predicting the outcome in the step (2) revised according to the slope revised law

The slope revised law has two kinds of methods: frequency domain correction and time domain correction

Frequency domain correction: the two frame channel overall gain A that utilize the interval M frame that rough long-term forecast obtains _n, A1 _n, overall gain is each subchannel gains mould square sum in every frame channel:

$A_n=\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}\left(\sqrt{{H_{n,k}^r}^2+{H_{n,k}^i}^2}\right)$

${A1}_n=\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}\left(\sqrt{{H_{n+M,k}^r}^2+{H_{n+M,k}^i}^2}\right)$

Wherein

N is the frame label of n frame, and k is the subchannel label, utilizes the intercept formula to come slope calculations then.The channel overall gain long-term forecast algorithm that provides as Fig. 2: A _n, A1 _nBe respectively the data that rough long-term forecast obtains, wherein A1 _nBe the channel status overall gain behind the following M frame.With A _nFor initial point is set up coordinate system, A _n, A1 _nThe difference of abscissa be long-term forecast M at interval, A _n, A1 _nThe difference of ordinate be the difference L1 of the coarse value of long-term forecast _n, Ci Shi slope K 1 then _n=L1 _n/ M., this slope just can regard the variation tendency of channel as, utilize the channel overall gain B of this frame short-term channel prediction then _nSlope information K1 with this moment _n, come the long-term forecast information B1 of counting period M frame by the intercept formula _n=K1 _n* M+B _nAt this moment, we can predict the state information H=[B1 that obtains whole channel _n, B1 _N+M, B1 _N+2M...].Because short-term forecast and real channel meet better, so also can meet the real channel overall gain preferably through revised long-term forecast information.As seen from Figure 2, B1 _nCompare A1 _nMore near real channel.

The time domain correction: the frequency domain correction is just revised on overall gain, and the time domain correction is the real part h in the every footpath of time domain _rWith imaginary part h _iCarry out the slope correction respectively, short-term forecast and long-term rough channel estimating information H are converted to time domain by the IFFT conversion, write down the real part of every footpath channel condition information And imaginary part

(n is the frame label of n frame), (l=0,1 ..., information L-1).The channel overall gain real part long-term forecast algorithm that provides as Fig. 3: with certain footpath real part h _{N, l} ^r(n is the frame label of n frame), long-term forecast is spaced apart M, $a1=h_{n,l}^r,$ $b1=h_{n+M,l}^r,$ The real part of the time domain channel information of this interval, footpath M frame that obtains for rough long-term forecast.With a1 is that initial point is set up coordinate system, a1, and the difference of the abscissa of b1 is long-term forecast M at interval.A1, the difference of the ordinate of b1 is the difference L2 of the coarse value of long-term forecast.Ci Shi slope K 2=L2/M then.This slope just still can be regarded the variation tendency of channel as.Utilize the real part information a2 and the slope information K2 of this moment of this footpath time-domain information of short-term forecast then, come the long-term forecast information b2=K2 * M+a2 of counting period M frame by the intercept formula.This moment should footpath real part information $h_{n,l}^r=b2.$ After predicting respectively with the method, every footpath real part imaginary part can obtain the time-domain information of this moment $h_n=[h_{n,1}^r+j{h}_{n,1}^i,h_{n,2}^r+j{h}_{n,2}^i...h_{n,L}^r+j{h}_{n,L}^i\text{].}$ As seen from Figure 3, through the slope revised long-term forecast value short-term forecast result that can coincide very accurately.This moment the long-term forecast information that obtains in time domain relatively accurately, so when transforming to frequency domain, its channel condition information also can well meet real channel.So just do the precision of prediction that has significantly improved long-term channel.The time domain correction has more credibility than frequency domain correction.As shown in Figure 4, the state information that can see its single frames channel also can well meet the real channel state information.

Fig. 5 has provided the long-term forecast that do not add correction and the comparison of short-term forecast and real channel information, and Fig. 6 has provided the comparison of carrying out revised long-term forecast and short-term forecast and real channel information.

Through revised channel long-range forecast still is the real information that the channel condition information of single frames can both reflect current channel more accurately in the channel overall gain not only.In addition, time domain prediction is real part h _rWith imaginary part h _iPrediction respectively is so comprise original phase information in last the predicting the outcome.After the correction of time domain slope, channel overall gain not only, every frame channel condition information precision all significantly improves, and phase information is also relatively accurately.

The present invention is by revising channel long-range forecast, make the long-term forecast value can meet real channel preferably, there is not tangible hysteresis, can significantly improve short-term forecast and be directly used in the precision of long-term forecast, can be used for the channel long-range forecast of the system of broadband wireless communication of piecemeal transmission means (as OFDM, SC-FDE etc.).And when not revising, the trend of long-term forecast is still correct, but can not extraordinaryly meet real channel, and tangible hysteresis is arranged.

Description of drawings

Fig. 1 is the structured flowchart of fallout predictor in system.

Fig. 2 is a channel overall gain long-term forecast algorithm schematic diagram.

Fig. 3 is a channel overall gain real part long-term forecast algorithm schematic diagram.

Fig. 4 is the state information of the single frames channel that obtains of real channel state information and long-term forecast algorithm.

Fig. 5 does not add the long-term forecast of correction and the comparison diagram of short-term forecast and real channel information.

Fig. 6 is the comparison diagram that carries out revised long-term forecast and short-term forecast and real channel information.

Fig. 7 is real channel overall gain and short-term forecast and the revised channel overall gain of long-term forecast time domain figure.

Embodiment

Embodiment

This embodiment simulation parameter:

Simulated environment: Matlab 7.0

Channel model: COST 207

Simulated environment: self adaptation single carrier frequency domain equalization (SC-FDE)

Subchannel sum: N=256

CP length: 64

Sign map: 4QAM

Bandwidth: 10M

Doppler frequency shift: 200Hz

Recursive least-squares RLS algorithm predicts forgetting factor: 0.99

RLS prediction normalized parameter is: 0.001

Short-term forecast step-length: 1

Memory span: 6 frame SC-FDE symbols

To little energy footpath zero setting thresholding: 1E-6

Long-term forecast is spaced apart: 10 frame SC-FDE symbols

The operation frame number is: 10 Doppler's cycles

Fig. 7 is the comparison of real channel overall gain and short-term forecast and the revised channel overall gain of long-term forecast time domain with this understanding.Can see the superiority of the inventive method among Fig. 7.Clearly as can be seen, through time domain revised long-term forecast obtain the channel overall gain than when revising more near real channel.Fig. 4 is under this condition, each subchannel gains and the short-term forecast of real channel, the comparison of each subchannel gains after the correction of long-term forecast time domain.As can be seen, each subchannel gains of this frame that the revised long-term forecast of process obtains almost overlaps with the real channel gain, almost reaches the effect of short-term accurately predicting.

Claims (1)

1. channel long-range forecast method based on the slope correction, implementation step is as follows:

(1) at first channel is carried out short-term forecast, and the channel gain of record short-term forecast at this moment;

(2) utilize the strong correlation of the channel of transmit frame in coherence time, utilize the short-term forecast algorithm to carry out rough channel long-range forecast;

(3) predicting the outcome in the step (2) revised according to the slope revised law; The slope modification method has frequency domain correction and two kinds of methods of time domain correction:

Frequency domain correction: the two frame channel overall gain A that utilize the interval M frame that rough long-term forecast obtains _n, A1 _n, overall gain is each subchannel gains mould square sum in every frame channel:

Wherein

N is the frame label of n frame, and k is the subchannel label, utilizes the intercept formula to come slope calculations, H then _nThe expression frequency domain information;

A _n, A1 _nBe respectively the data that rough long-term forecast obtains, wherein A1 _nFor the channel status overall gain behind the following M frame, with A _nFor initial point is set up coordinate system, A _n, A1 _nThe difference of abscissa be long-term forecast M at interval, A _n, A1 _nThe difference of ordinate be the difference L1 of the coarse value of long-term forecast _n, Ci Shi slope K 1 then _n=L1 _n/ M, the channel overall gain B that utilizes this frame short-term channel to predict then _nSlope information K1 with this moment _n, come the long-term forecast information B1 of counting period M frame by the intercept formula _n=K1 _n* M+B _n

The time domain correction: the frequency domain correction is just revised on overall gain, and the time domain correction is the real part h in the every footpath of time domain _rWith imaginary part h _iCarry out the slope correction respectively, short-term forecast and long-term rough channel estimating information H are converted to time domain by the IFFT conversion, write down the real part of every footpath channel condition information

And imaginary part Information, n is the frame label of n frame, l=0,1 ..., L-1; With certain footpath real part

N is the frame label of n frame, and long-term forecast is spaced apart M,

The real part of the time domain channel information of this interval, footpath M frame that obtains for rough long-term forecast, with a1 is that initial point is set up coordinate system, a1, the difference of the abscissa of b1 is long-term forecast M at interval, the difference of the ordinate of a1, b1 is the difference L2 of the coarse value of long-term forecast, Ci Shi slope K 2=L2/M then, utilize the real part information a2 and the slope information K2 of this moment of this footpath time-domain information of short-term forecast then, come the long-term forecast information b2=K2 * M+a2 of counting period M frame by the intercept formula, this moment should footpath real part information After predicting respectively with the method, every footpath real part imaginary part promptly obtains the time-domain information of this moment

Images