CN105072062A - Signal source number estimation method based on resampling - Google Patents

Abstract

The invention discloses a signal source number estimation method based on resampling. The signal source number estimation method comprises the following steps: determining sampling times and an antenna array sub-set; sampling a received signal set without replacement to construct sub-sample sets; performing row-column transform on a covariance matrix constructed by the sub-sample sets to construct a unitary matrix; performing multiple Gerschgorin disk estimator processing on the unitary matrix to obtain a signal source number obtained by ith iteration; and when current iteration times are less than total iteration times, adding the signal source number obtained by the ith iteration into a set, wherein a final signal source number is an element with highest occurrence frequency in the set. According to the method, a conventional Gerschgorin signal source estimation algorithm is improved, so that influences caused by accidental errors are reduced. After adding of the thought of resampling, a finite sample set is utilized more fully; the occurrence probability of true values is increased; and probability of realizing consistency estimation is increased.

Description

A kind of number of source method of estimation based on resampling

Technical field

The present invention relates to number of source and estimate field, particularly relate to a kind of number of source method of estimation based on resampling.

Background technology

At present, forth generation mobile communication system starts scale commercialization in the whole world, universal, the Mobile solution of intelligent terminal flourish, linking number will be impelled to increase sharply and the rapid growth of wireless flow.Meanwhile, internet of things service also launches gradually in multiple industry, and wireless application presents the trend of diversified development, and its ubiquitousization feature manifests day by day.The extensive use of cloud computing and background service proposes higher transmission quality and system capacity requirements by 5G mobile communication system.

In order to improve spectrum efficiency further, large-scale antenna array technology will be introduced into the main flow of Radio Transmission Technology, when dual-mode antenna quantity is very large, approximately linear increases by MIMO (multi input and multi output) channel capacity, and target is to enable 4G to communicate to improve a magnitude in channel capacity, power efficiency and spectrum efficiency in these.In this process, the accurate location of antenna is one of key technology of wave beam forming.But the significant challenge restricting this technical application is high-dimensional Channel Modeling and estimation, and complexity controls.Meanwhile, wave beam forming needs to know channel condition information in advance, such as: angle of arrival information, amplitude information, if to channel model dysgnosia face, cannot select the fast algorithm of the process mass data matched with scene.

Linear array source signal number estimates it is the main contents that wireless channel parameter extracts, and is also the important process probing into wireless channel model characteristic.The incident source signal number of Obtaining Accurate is also applied to the basis that array signal model is channel parameter extraction algorithm, and be also the prerequisite setting up or upgrade channel model structure, whether pair array signal disposal and analysis algorithm performance accurately produces a very large impact by it.In complicated, unknown, changeable electromagnetic environment, utilize, extract and recover the basic principle that the useful information be contained in the general collection of Received signal strength sample is modern Array Signal Processing as much as possible.Channel model state information reflects mainly through the signal propagated, and how parameter to received signal effectively detects and estimates just to seem and is even more important.The accurate estimation of number of source meets the raising of orientation of information source, input and estimated accuracy, and therefore in numerous signal processing problems, its application prospect is very wide.

But there is two problems in traditional lid formula circle number of source algorithm for estimating:

On the one hand, receive in model at coloured noise multiple element antenna, interchannel additive noise cross correlation measure is large, and each array-element antenna is different to signal component responsiveness.Structure unitary transformation matrix needs the covariance matrix dimensionality reduction to sampled data, because the receive path cross correlation measure of array element each in aerial array is changed significantly, the minimum component of effect of signals cannot be predicted in advance, the vector be rejected if to the component that signal response is large time, array guiding A space and noise feature vector u spatial orthogonality will reduce, the Gerschgorin radii that unitary transformation matrix is corresponding also will become greatly, cause non-positive value in decision rule to delay to reach, bring estimated risk.Once array guide space and the mutually orthogonal hypothesis of spatial noise are broken, the principle of Gerschgorin radii also will be overturned.

On the other hand, the sampling complete or collected works that conventional estimated method is only formed for T snap once calculate, and snap collection is constructed to a covariance matrix.But sample general collection but can excavate more information, only once sample and there will be error unavoidably.

Summary of the invention

The invention provides a kind of number of source method of estimation based on resampling, present invention, avoiding accidental error, improve the extraction rate that number of source estimates accuracy and channel parameter, described below:

Based on a number of source method of estimation for resampling, described number of source method of estimation comprises the following steps:

Determine sampling number Z and oversampling ratio, sample set does not put back to sampling to received signal afterwards, obtains subsample collection;

The covariance matrix R formed by antithetical phrase sample set _iprocession conversion and dimensionality reduction reconstruct, construct unitary matrice U _m; Lid formula is carried out repeatedly to the M constructed a different unitary matrice and estimates process, obtain the number of source that i-th iteration obtains

When current iteration number of times i is less than iteration total degree Z, the number of source i-th iteration obtained adds in set; Final number of source is the element that in set, frequency of occurrence is maximum.

Wherein, describedly estimating process to carrying out repeatedly lid formula to the M constructed a different unitary matrice, obtaining the number of source that i-th iteration obtains step be specially:

By obtaining current unitary matrice U after the row-column transform of the m time and dimensionality reduction reconstruction processing _m, to U _mcarry out Gai Shiyuan and estimate process, obtain the GDE value of the m time this process is from m=1;

The GDE value obtained after the m time lid formula is estimated record, be MGDE and judge: $MGDE\left(k\right)=\frac{1}{M}\underset{k=1}{\overset{m}{\Σ}}{\hat{d}}_k;$

Wherein, m=1,2 ..., M; K=1,2 ..., m, when MGDE (k) reaches first non-positive value, the number of source that i-th iteration obtains otherwise make m=m+1, circulate next time.

The beneficial effect of technical scheme provided by the invention is: this method is improved traditional lid formula information source algorithm for estimating, reduces the impact that accidental error is brought.After adding the thought of resampling, finite sample collection obtains and utilizes more fully, improves actual value probability of occurrence, also improves the possibility obtaining Uniform estimates.The method is compared conventional Gerschgorin radii and is had higher accuracy of estimation and stability, is more conducive to being applied in actual array signal processing problems.

Accompanying drawing explanation

Fig. 1 is a kind of flow chart of the number of source method of estimation based on resampling;

Fig. 2 is uniform linear array Received signal strength schematic diagram;

Fig. 3 is the estimated performance comparison diagram of various information source number method under coloured noise;

Fig. 4 is the performance comparison figure of coloured noise ULA array method for resampling and conventional method;

Fig. 5 is with the estimated performance comparison diagram that fast umber of beats changes under coloured noise;

Fig. 6 is the estimated performance comparison diagram with differential seat angle change under coloured noise;

Fig. 7 is the performance comparison that resampling MDGE method of estimation changes with oversampling ratio;

Fig. 8 is that resampling iterations is to estimated performance effect diagram.

Embodiment

For making the object, technical solutions and advantages of the present invention clearly, below embodiment of the present invention is described further in detail.

Embodiment 1

The present invention proposes a kind of number of source method of estimation based on resampling, see Fig. 1, the method comprises the following steps:

101: determine sampling number, and oversampling ratio; Sample set does not put back to sampling to received signal, forms subsample collection;

102: the covariance matrix procession conversion consisted of antithetical phrase sample set and dimensionality reduction reconstruct, build unitary matrice; Lid formula is carried out repeatedly to the M constructed a different unitary matrice and estimates process, obtain the number of source that i-th iteration obtains

103: when current iteration number of times i is less than iteration total degree Z, the number of source i-th iteration obtained adds in set; Final number of source is the element that in set, frequency of occurrence is maximum.

104: traditional lid formula information source algorithm for estimating is improved, define new MGDE algorithm and be used for estimating number of source.

In sum, the embodiment of the present invention avoids accidental error by above-mentioned steps 101-step 104, improves the extraction rate that number of source estimates accuracy and channel parameter.

Embodiment 2

Be described in detail below in conjunction with the scheme in concrete computing formula, accompanying drawing 2 pairs of embodiments 1, wherein step 202,203,204 is outer systemic circulations, described below:

201: determine sampling number Z, oversampling ratio r, obtain M unit aerial array sub-set size T _r=r*L, wherein L is the fast umber of beats general collection size that the aerial array sampling of M unit is formed, and makes iterations initial value i=1;

202: sample set does not put back to sampling to received signal, formation size is T _rsubsample collection;

203: by building unitary matrice after the conversion of the covariance matrix procession of antithetical phrase sample set and dimensionality reduction, lid formula MGDE being carried out repeatedly to unitary matrice and estimates process, obtain the number of source that i-th iteration obtains;

1) the subsample collection of i-th iteration is calculated covariance matrix (wherein, X ⁽ⁱ⁾be the sample set of i-th iteration), and make unitary transformation number of times m=1, identity transformation battle array is E;

$E=\left[\begin{array}{cccccc}0& 1& 0& 0& ...& 0\\ 0& 0& 1& 0& ...& 0\\ 0& 0& 0& 1& ...& 0\\ ...& ...& ...& ...& ...& 1\\ 1& 0& 0& 0& ...& 0\end{array}\right]$

2) covariance matrix is converted ranks order: $R_{X^{\left(i\right)}}^{\left(m\right)}={\mathrm{ER}}_{X^{\left(i\right)}}^{(m-1)}{E}^{-1},m=1,2,\mathrm{...},M.$

Wherein, represent the covariance matrix after having carried out m row-column transform and the ranks removing most end reduce to M-1 dimension, build unitary matrice U with this ^(m).

3), after completing above-mentioned unitary transformation, carry out estimating of GDE (GerschgorinDiskEstimator, Gai Shiyuan estimating), can obtain GDE value according to GDE estimation criterion is

(GDE estimation criterion is:

$GDE\left(k\right)={\mathrm{\ρ}}_k-\frac{D\left(L\right)}{M-1}\underset{i=1}{\overset{M-1}{\Σ}}{\mathrm{\ρ}}_i$

Wherein, ρ _krepresent a kth lid formula radius of unitary matrice; The fast umber of beats size of L representative sampling; D (L) ∈ [0,1] is the nonincreasing function about L, is the regulatory factor relevant with the fast umber of beats of sampling.K is value between 1 ~ M-2, large from little change along with k, and when first negative value appears in GDE (k), number of source can be defined as if m=M, circulate end.)

Repeatedly lid formula estimates that the estimation criterion of MGDE is:

$MGDE\left(k\right)=\frac{1}{M}\underset{k=1}{\overset{m}{\Σ}}GDE\left(k\right)$

Wherein, GDE (k) be this M time circulation Inner eycle each time in through kth (k=1,2 ..., m; M=1,2 ..., M) and the value that obtains according to estimation criterion after secondary unitary transformation; When MGDE (k) reaches first non-positive value, record k value now, then the number of source that i-th iteration obtains is otherwise make m=m+1, start to circulate next time.

In fact, MGDE carries out GDE estimation respectively to the different matrixes obtained after the procession conversion in certain sequence of same matrix, and each GDE estimated result value being added up is averaged again, and the M of MGDE is exactly the multiple meaning of multiple.

204: judge whether current iteration number of times i is less than iteration total degree Z, if i<Z, the number of source that i-th iteration is obtained add set $Z_{\hat{k}}=\left\{{\hat{k}}^{\left(i\right)}\right\};$

Wherein, gather in element be the information source number obtained after iteration each time, and make i=i+1, enter step 202 and start to circulate next time; Otherwise circulation terminates information source number and finally can be defined as namely gather the element that middle frequency of occurrence is maximum.

M unit antenna see Fig. 2, Fig. 2 to be adjacent spacing be d forms uniform linear array, receives from { θ _l, l=1,2 ..., arrowband, the far field source signal of K} (K is number of source) direction incidence.Then Received signal strength can be expressed as: X (t)=AS (t)+N (t).

Wherein, X (t)=[x ₁(t), x ₂(t) ..., x _m(t)] ^tfor M unit antenna is at the Received signal strength vector of the t time snap;

Wherein, A=[a (θ ₁), a (θ ₂) ..., a (θ _k)] be array guiding matrix, a (θ _l) represent l signal respectively to the response of M unit antenna d is antenna spacing, θ _lbe the incident direction of l source signal, λ is wavelength, i=1,2 ..., M.

Wherein, S (t)=[s ₁(t), s ₂(t) ..., s _k(t)] ^tfor source signal vector, suppose source signal Gaussian distributed and separate.N (t)=[n ₁(t), n ₂(t) ..., n _m(t)] ^trepresent additive noise, can for obeying white Gaussian noise or the coloured noise with cross correlation.

Then, actual Received signal strength X (t) covariance matrix is:

R _x＝E[X(t)X ^H(t)]＝AR _SA ^H+R _n＝AR _sA ^H+σ _n ²∑。

In above formula, R _xit is the covariance matrix of Received signal strength X (t); X ^ht () is the associate matrix of X (t); A is array guiding matrix; R _s=[s (t) s ^h(t)]; R _n=[n (t) n ^h(t)]; S (t) is source signal; s ^ht associate matrix that () is s (t); N (t) is noise; n ^ht associate matrix that () is n (t); σ _n ²represent the variance of noise, ∑ is unit matrix.

In sum, the embodiment of the present invention avoids accidental error by above-mentioned steps 201-step 204, improves the extraction rate that number of source estimates accuracy and channel parameter.

Embodiment 3

Below in conjunction with concrete example, accompanying drawing, feasibility checking is carried out to the scheme in embodiment 1,2, described below:

Receiving terminal adopts 8 yuan of uniform linear arrays, antenna spacing is 0.5 λ, coloured noise adopts a zero-mean white noise to export this noise model by ARMA system (AutoRegressionMovingAverage, autoregressive moving average system), and coefficient correlation is set to 0.8.The index of measure algorithm performance, for detecting correct probability, is defined as in the Monte-Carlo test of 1000 times the ratio correctly estimated shared by number of times.

Fig. 3 be under inspection Colored Noise based on information theory with based on the theoretical algorithm performance difference of Gai Shi circle: be provided with four independent constant power narrow-band source signals incident, incident angle is respectively [-35,8,45,75], fast umber of beats of sampling is 100, and signal to noise ratio changes from-10dB to 5dB.MDL (MinimumDescriptionLength, the shortest description length criteria), PET (PredictedEigen-threshold, predicted characteristics threshold value), GDE (GerschgorinDiskEstimator, Gai Shiyuan estimates) and MGDE (MultipleGerschgorinDiskEstimator, repeatedly lid formula circle is estimated) algorithm estimated performance is as shown in the figure, when channel noise is coloured noise, MDL and PET criterion based on information theory is difficult to obtain correct estimated result, along with signal to noise ratio increases, the lifting of Detection accuracy also and not obvious.By contrast, the algorithm based on Gerschgorin radii can obtain good effect under coloured noise environment.Meanwhile, the estimated performance of single GDE when adopting the GDE method after multiple unitary transformation to improve low signal-to-noise ratio, and increase along with signal to noise ratio, performance also will tend towards stability.

Fig. 4 is for proving when traditional technique in measuring accuracy is greater than 50%, and this method can obtain higher detection perform from less sampling fast umber of beats.Be provided with four independent constant power narrow-band source signals incident, incident angle is respectively [10,20,30,40], and signal to noise ratio changes from-8dB to 6dB, and oversampling ratio is 0.6, and resampling iterations is 10.Fast umber of beats of sampling be 50 this method and the fast umber of beats single GDE conventional statistical methods that is respectively 50,500,1000 contrast as shown in Figure 4, conventional method performance can increase along with the fast umber of beats of sampling and improve, and still can not tend towards stability along with signal to noise ratio increases detection probability and can not reach probability 100%.When signal to noise ratio is lower, the conventional method that this method under identical fast umber of beats compares more sampling snaps reaches more high detection accuracy, and reaches 100% stable probability very soon along with signal to noise ratio increase.Meanwhile, when traditional technique in measuring correct probability is greater than 50%, the detection accuracy of this method of the fast umber of beats of identical sampling will significantly promote.

Fig. 5 compares the fast umber of beats of sampling proposes method of estimation impact on conventional method and this method.Be provided with four independent constant power narrow-band source signals incident, incident angle is respectively [10,20,30,40], and fast umber of beats is from 30 to 200 changes, and signal to noise ratio-3dB, oversampling ratio is 0.6, and resampling iterations is 10.The curve chart that the resampling MGDE method giving traditional GDE, single MGDE and this method as shown in the figure changes with fast umber of beats of sampling.Along with fast umber of beats increases, three kinds of algorithm detection perform have the trend of rising, but the fluctuation of traditional GDE uphill process is large.The GDE performance after multiple unitary transformation is adopted to bring obvious improvement when fast umber of beats of sampling is 50, this is because repeatedly unitary transformation inhibits the evaluated error brought because giving up signal response large component when building unitary matrice after being averaged, but even if fast umber of beats increase always still exist detection probability can not stablize level off to 100% problem.Compare first two method, this method improves the large phenomenon of performance boost process variation, and is make that correct detection probability is stable remains on 100% after 100 at fast umber of beats.

Fig. 6 compares the impact of information source differential seat angle on conventional method and this method.Be provided with two independent constant power narrow-band source signals incident, the first bundle signal is fixed as 0 degree, and another bundle signal from-180 to 180 changes, and signal to noise ratio-2dB, fast umber of beats is 50, and oversampling ratio is 0.6, and resampling iterations is 10.Fig. 5 has investigated the impact of incoming signal differential seat angle on estimated performance, can find out that from curve chart two kinds of methods significantly reduce for 0 and the incident information source estimated performance of positive and negative 180 degree of differential seat angles, this method does not have clear improvement yet, its main cause is that in signal model, information source has sine term to bay response, and 0 of this emulation hypothesis creates identical response with positive and negative 180 degree of signals.But the frequency of occurrences is very low in the actual measurement that requires in high-resolution of the identical situation of this incident angle sine value, this curve chart shows that the estimated performance for other angle values is also all very high, especially this method almost when full angle estimated performance be all better than single MGDE method, and maintain the stable detection probability being greater than 80% accuracy.

Fig. 7 investigates the selection of oversampling ratio to the impact of detection perform, and provides optional sampling ratio reference value corresponding to different fast umber of beats.Be provided with four independent constant power narrow-band source signals incident, incident angle is respectively [10,20,30,40], signal to noise ratio-3dB, and oversampling ratio is in 0.5 to 1 change, and resampling iterations is 10.For obtaining the optimal detection probability of this method, Fig. 7 has investigated the impact of oversampling ratio value on algorithm performance, has carried out emulation experiment during the fast umber of beats of three differences respectively.Along with the increase of the fast umber of beats of sampling, Received signal strength also contains more useful information, and detection perform can promote gradually, and this is a common recognition.But in each test, all there is the oversampling ratio that can make detection perform optimum, and the oversampling ratio value corresponding to optimal detection probability does not increase with the fast umber of beats of sampling and promotes, increase when subset comprises information, detection perform declines on the contrary.Increase with fast umber of beats, optional sampling ratio value successively decreases in [0.5,1] interval arrangement, and this illustrates for this method, and the selection of subset is not the bigger the better.

Fig. 8 analyzes iterations setting for derivation of equation result.Be provided with four independent constant power narrow-band source signals incident, incident angle is respectively [10,20,30,40], and sampling snap 100, oversampling ratio 0.6, signal to noise ratio is changed by-10dB to 7dB.Fig. 8 has compared resampling iterations and has increased to 20 processes from 1, estimated performance situation of change.When iterations is 1, the fast umber of beats of sample reduces estimated performance and declines.Along with iterations increases from 2, the estimation accuracy after resampling is higher than the multiple unitary transformation method of single.But estimated performance is not increase along with iterations and infinitely increase.Represent in Fig. 8, when iterations is 10 and 20, curve overlaps substantially, estimates the limit under namely reaching this situation.Therefore, in order to give full play to the advantage of resampling and not waste system resource, iterations needs the Choice setting a compromise with reference to complexity.

In addition, it should be noted that, although resampling process is increased to space complexity the lifting that cost exchanges estimated accuracy for, popularize with existing internal memory treatment technology and the extensive of cloud computing, be enough to the feasibility ensureing this parallel algorithm.In addition, in the process of repeatedly resampling and MGDE, each calculating is separately all separate, in resource, only needs shared drive.In last statistic processes, adding up maximum frequency estimated number is the final step of algorithm.Above-mentioned emulation experiment shows, in coloured noise, low snap, low signal-to-noise ratio situation, the method is compared conventional Gerschgorin radii and had higher accuracy of estimation and stability, is more conducive to being applied in actual array signal processing problems.

It will be appreciated by those skilled in the art that accompanying drawing is the schematic diagram of a preferred embodiment, the invention described above embodiment sequence number, just to describing, does not represent the quality of embodiment.

The foregoing is only preferred embodiment of the present invention, not in order to limit the present invention, within the spirit and principles in the present invention all, any amendment done, equivalent replacement, improvement etc., all should be included within protection scope of the present invention.

Claims (2)

1. based on a number of source method of estimation for resampling, it is characterized in that, described number of source method of estimation comprises the following steps:

Determine sampling number Z and oversampling ratio, sample set does not put back to sampling to received signal afterwards, obtains subsample collection;

The covariance matrix R formed by antithetical phrase sample set _iprocession conversion and dimensionality reduction reconstruct, construct unitary matrice U _m; Lid formula is carried out repeatedly to the M constructed a different unitary matrice and estimates process, obtain the number of source that i-th iteration obtains

When current iteration number of times i is less than number of times Z, the number of source i-th iteration obtained adds in set; Final number of source is the element that in set, frequency of occurrence is maximum.

2. a kind of number of source method of estimation based on resampling according to claim 1, is characterized in that, the described M to constructing a different unitary matrice is carried out repeatedly lid formula and estimated process, obtains the number of source that i-th iteration obtains step be specially:

By obtaining current unitary matrice U after the row-column transform of the m time and dimensionality reduction reconstruction processing _m, to U _mcarry out Gai Shiyuan and estimate process, obtain the GDE value of the m time this process is from m=1;

The GDE value obtained after the m time lid formula is estimated record, be MGDE and judge: $MGDE\left(k\right)=\frac{1}{M}\underset{k=1}{\overset{m}{\&Sigma;}}{\hat{d}}_k;$

Wherein, m=1,2 ..., M; K=1,2 ..., m, when MGDE (k) reaches first non-positive value, the number of source that i-th iteration obtains otherwise make m=m+1, circulate next time.