CN101917369A - Method for identifying modulation mode of communication signal - Google Patents

Abstract

The invention discloses a method for identifying a modulation mode of a communication signal. The method comprises the following steps that: a network side receives a communication signal, constructs and stores characteristic parameters of the communication signal according to the characteristics of the received communication signal; the network side processes the constructed characteristics parameters of the communication in a GDA mode, and stores the processed characteristic parameters; and the network side judges the processed characteristic parameters in a mode of a decision tree judgment classifier of an AdaBoost, and identifies the modulation mode of the received communication signal. Due to the adoption of the method, the problem that the conventional identification method has low identification accuracy in a low signal-to-noise ratio is solved; and a higher correct identification rate is achieved.

Description

A kind of recognition methods of modulation mode of communication signal

Technical field

The present invention relates to the signal of communication process field, relate in particular to a kind of recognition methods of modulation mode of communication signal.

Background technology

The signal automatic Modulation Recognition all has using value very widely in the military and civilian field, bringing into play important effect in fields such as signal authentication, interference identification, spectrum managements.It also is one of key technology of software radio and cognitive radio.

The modulation type identification of signal is divided into feature extraction and classifier design two big steps.Based on time domain, frequency domain and various transform domain, it is good and need the feature of less priori to extract separability, is the important research direction of field of signal identification always.Feature extracting method commonly used has based on methods such as time-domain information, frequency domain information, Higher Order Cumulants, cycle specificity, wavelet transformations.Feature extraction computing based on time domain and wavelet transformation is simple, but is subjected to the channel influence of fading bigger, and characteristic performance is relatively poor under low signal-to-noise ratio; Though based on the feature extraction stable performance of frequency domain information, Higher Order Cumulants, cyclo-stationary, operand is bigger.

Feature selecting and processing are significant after the feature extraction.In the feature extraction of reality, signal of communication signal to noise ratio excursion is bigger, and it is relatively poor to extract characteristic mass under low signal-to-noise ratio.In order to guarantee the accuracy and the precision of signal identification, usually can extract the multidimensional signal feature.If use whole features to carry out grader identification, bring heavy calculating pressure will for the rear end grader; And may have nonindependence between the characteristic parameter that extracts, this tends to make recognition result to lack robustness.Characteristic processing method commonly used at present has genetic algorithm, and principal component analysis (Principal Component Analysis, PCA) etc.But the genetic algorithm calculation of complex, easily be absorbed in locally optimal solution.Though the feature selection approach based on PCA is farthest removed information redundancy, realize the compression of characteristic vector, can not guarantee that the separability of feature is constant.When feature separability variation, accuracy of identification is variation on the contrary.

In classifier design algorithm commonly used, based on the grader of decision tree judgement always to realize the extensive use on engineering of simple advantage.But judge because it only carries out feature with single threshold value, lack flexibility, and accuracy of identification is relatively poor under low signal-to-noise ratio.

Summary of the invention

Technical problem to be solved by this invention provides a kind of recognition methods of modulation mode of communication signal, has solved traditional recognition methods in the not high problem of low signal-to-noise ratio accuracy of identification.

In order to address the above problem, the invention provides a kind of recognition methods of modulation mode of communication signal, comprising:

The network side receiving communication signal is constructed the characteristic parameter of this signal of communication and is stored according to the characteristic of the signal of communication that receives;

Described network side is handled the characteristic parameter of posttectonic signal of communication by generalized discriminant analysis GDA mode, the characteristic parameter after the processing that obtains is stored;

Characteristic parameter after the processing that described network side will obtain is adjudicated the mode of grader and is differentiated the modulation system of the signal of communication that identification receives by the decision tree of AdaBoost.

Further, said method also can comprise, described network side is meant according to the characteristic parameter of the characteristic structure signal of communication of the signal of communication that receives: network side is according to the characteristic parameter of the information structuring signal of communication of amplitude, frequency spectrum, square spectrum and/or the phase place of the signal of communication that receives.

Further, said method can comprise that also the characteristic parameter of the signal of communication of described network side structure comprises:

Characteristic parameter R, it is defined as:

Wherein μ is the average of signal envelope square, σ ²Variance for signal envelope square;

Power spectrum compactness feature

It is defined as:

P wherein _Cn(i) be the amplitude of zero center normalized power spectrum, P (i) is the power spectrum of signal, P _Cn(i)=P _nAnd P (i)-1, _n(i)=P (i)/m _p, m _pAverage for the power spectrum amplitude;

Absolute nothing is inserted folded differential phase standard deviation d Δ _Ap, wherein the differential phase of signal is: Δ φ (i)=φ (i)-φ (i-1), φ are that the nothing of various types of signal is inserted folded phase place;

Effectively compose the peak number; Square spectrum single-frequency components detected parameters and biquadratic spectrum single-frequency components detected parameters.

Further, said method can comprise that also described network side is the step of effectively composing the peak number according to the characteristic parameter of the characteristic structure signal of communication of the signal of communication that receives, and may further comprise the steps:

Described network side is searched for its maximum and minimum by the power spectrum chart P of welch method signal calculated to spectrogram P,

If described network side judges P (i)＞P (i-1) ﹠amp; P (i)＞P (i+1) ﹠amp; P (i)＞P ₁The time, there is a maximum, the position I=[i of record maximum ₁, i ₂..., i _m] and corresponding maximum

The tentative number of the number M of statistics maximum for the spectrum peak;

If described network side judges P (j)＜P (j-1) ﹠amp; During P (j)＜P (j+1), there is a minimum, writes down minimizing position J=[j ₁, j ₂..., j _m] and corresponding minimum

Described network side is searched for minimum position corresponding between per two maximum, has i if judge ₁＜j ₁＜i ₂, then right Judge, if judge

M=M-1 then; Otherwise M remains unchanged;

Described network side travels through the minimum of all maximum correspondences, and it is judged, obtains effective spectrum peak number that final M is signal.

Further, said method can comprise that also described network side is square step of spectrum single-frequency components detected parameters according to the characteristic parameter of the characteristic structure signal of communication of the signal of communication that receives, and may further comprise the steps:

Square spectrum P of described network side signal calculated ₁, to a square spectrum P ₁Search for its maximum Y ₁And to write down peaked position be I ₁

Described network side is selected length L, calculates from P ₁(I ₁-L: I ₁-2) and P ₁(I ₁+ 2: I ₁+ L) average of the amplitude of all spectral lines is Y ₂Calculate both ratio Y ₁/ Y ₂, judgement then judge to have a single-frequency components, otherwise there is not single-frequency components in judgement if it surpasses a certain threshold values.

Further, said method can comprise that also described network side is handled the characteristic parameter of posttectonic signal of communication by the GDA mode, and the characteristic parameter after obtaining handling comprises:

Described network side is formed input sample data collection x at its posttectonic characteristic parameter of every class signal extraction;

Described network side select suitable kernel function k (x, y) and the kernel function parameter, the characteristic vector dimension d that setting will be extracted;

Described network side calculates nuclear matrix K according to following formula _bAnd K _w

Wherein, $K_b=\underset{i=1}{\overset{c}{\mathrm{\Σ}}}\frac{N_i}{N}({\mathrm{\μ}}_i-{\mathrm{\μ}}_0){({\mathrm{\μ}}_i-{\mathrm{\μ}}_0)}^T;$

N is the signal of communication feature samples number of getting; C is the classification number that sample can divide, and the number of every class sample is N _iξ _xBe sample vector based on nuclear,

Average for each sample inner product nuclear;

Described network side calculating K _bα=λ K _wThe characteristic vector of α, and carry out orthogonalization, obtain α;

Described network side is according to formula y=W ^Tφ (x)=[w ₁, w ₂..., w _d] ^Tφ (x)=[α ₁, α ₂..., α _d] ^Tξ _x, ξ wherein _x=(k (x ₁, x), k (x ₂, x) ..., k (x _N, x)) ^TBe sample vector, to the signal characteristic vector of input, through the characteristic parameter after obtaining handling after the GDA mode based on nuclear.

Further, said method also can comprise, described network side is with the characteristic parameter of posttectonic signal of communication, handle by the GDA mode, with the characteristic parameter after the processing that obtains, in the step that the mode of the decision tree judgement grader by AdaBoost is differentiated, also comprise the step of determining parameter σ and T, carry out the differentiation of mode of the decision tree judgement grader of the processing of GDA mode and AdaBoost by parameter σ and T, wherein σ is the undetermined parameter of control kernel function width, and T is the number of Weak Classifier.

Further, said method can comprise that also described network side is determined the step of parameter σ and T, comprising:

Described network side extracts the characteristic parameter after the Signal Processing from the signal of communication of one group of known signal modulation system type, the composition characteristic Vector Groups, and the characteristic vector group is divided into two parts at random, be respectively training sample and test sample book;

Described network side setting search width L, L is the integer greater than 0, make σ [σ-L, σ-L+1 ..., σ+L-1, σ+L] and the interior variation of scope, wherein σ is the undetermined parameter of control kernel function width; Be limited to D in the search of setup parameter T, D is the integer greater than 0, and T is the number of Weak Classifier;

The whenever selected parameter σ of described network side is from 1,2,, D travels through parameter T, selects for use training sample that the decision tree judgement grader of GDA and AdaBoost is trained respectively, and adopt test sample book to calculate correct recognition rata, the value of σ and T when writing down the highest accuracy of identification;

Described network side with σ [σ-L, σ-L+1 ..., σ+L-1, σ+L] and traversal in the scope, σ and T when selecting the highest accuracy of identification, if judge and satisfy required precision, then parameter is selected to finish; Otherwise, have the σ of high accuracy of identification ₁σ with inferior high accuracy of identification ₂Between, search for again with littler stepping, wherein σ ₁＜σ ₂, select to have the σ and the T of high accuracy of identification, after obtaining satisfied correct recognition rata, determine parameter σ and T.

Compared with prior art, use the present invention, by based on generalized discriminant analysis (GeneralizedDiscriminant Analysis, GDA) and AdaBoost (Adaptive Boosting, signal modulation style recognition methods AdaBoost) can need not under the condition of any priori, under low signal-to-noise ratio, finish AM/SSB/CW/MFSK/MPSK (M=2,4,8) identification of signal, and have higher correct recognition rata.

Description of drawings

Fig. 1 is the flow chart of the recognition methods of modulation mode of communication signal of the present invention;

Fig. 2 is that feature R is with the signal to noise ratio change curve;

Fig. 3 is that power spectrum compacts feature with the signal to noise ratio change curve;

Fig. 4 is that absolute nothing is inserted folded phase place standard deviation with the signal to noise ratio change curve;

Fig. 5 effectively composes the peak number with the signal to noise ratio change curve;

Fig. 6 is that a square spectrum single-frequency components detects ratio with the signal to noise ratio change curve;

Fig. 7 is that biquadratic spectrum single-frequency components detects ratio with the signal to noise ratio change curve.

Embodiment

The invention will be further described below in conjunction with the drawings and specific embodiments.

The present invention proposes a kind of method (Generalized DiscriminantAnalysis based on generalized discriminant analysis, GDA) carry out the processing of signal characteristic, not only compressed information dimension, and made feature reach same sex cluster, opposite sex separation to a certain extent, classification performance is better.The present invention utilizes the statistical property of signal, has proposed two new features: power spectrum compactness and absolute nothing are inserted folded differential phase standard deviation, have also redefined the detection method of single-frequency components.These parameters do not rely on priori, and are insensitive to the variation of signal to noise ratio, have good robustness, and algorithm is simple, and still have preferable performance under lower signal to noise ratio; Aspect classifier design, the present invention proposes a kind of decision tree judgement (AdaBoost-Decision Tree, AdaBoost-DT) classifier design based on AdaBoost.It forms a strong classifier by several Weak Classifiers, has higher nicety of grading than decision tree decision method commonly used.The present invention passes through based on generalized discriminant analysis (Generalized Discriminant Analysis, GDA) and AdaBoost (Adaptive Boosting, the method of common signal modulation type identification AdaBoost), can need not under the condition of any priori, finish AM/SSB/CW/MFSK/MPSK (M=2,4,8) identification of signal, and have higher correct recognition rata.Can be applicable to software radio and cognitive radio system, in shortwave or the ultra short wave communication receiving system, good basis is provided for the demodulating and decoding of rear end.

It should be noted that: method of the present invention will be carried out the collection of information and/or data by the information interaction of network side (as the base station) in specific implementation, and (can be that CPU etc. carries out control and treatment information and/or data by the controller in it, the present invention does not do any qualification to this), can also carry out the storage and the transmission of information and/or data by various memories (can be internal memory, hard disk or other memory devices) therebetween, the present invention does not do any qualification to this.

Main design of the present invention is: at first extract signal characteristic vector from the signal of communication sample (training sample) of known modulation type; Feature is handled, and the compression dimension extracts its feature that is easier to follow-up classification (adopting the GDA algorithm).Design is suitable for the grader (adopting based on the AdaBoost-DS algorithm) of signal characteristic then, determines its parameter, finishes the training stage of grader.In the performance test stage, select other one group of sample of signal, promptly test sample book is extracted its feature, finishes characteristic processing, detects recognition performance according to its correct recognition rata in the input category device.

Given signal of communication to be classified set among the present invention: AM, CW, SSB, 2FSK, 4FSK, 8FSK, BPSK, QPSK and 8PSK.

The characteristic parameter of the present invention's structure is based on the information such as amplitude, frequency spectrum, square spectrum and phase place of signal and sets up.

As shown in Figure 1, the recognition methods of modulation mode of communication signal of the present invention may further comprise the steps:

Step 10, network side receiving communication signal are according to the characteristic structural feature parameter of the signal of communication that receives and store;

Network side specifically is meant according to the characteristic parameter of the characteristic structure signal of communication of the signal of communication that receives: network side is according to the characteristic parameter of the information structuring signal of communication of amplitude, frequency spectrum, square spectrum and the phase place of the signal of communication that receives.

Wherein, the characteristic parameter of the signal of communication of network side structure, as follows:

1. characteristic parameter R, it is defined as:

$R=\frac{{\mathrm{\σ}}^2}{{\mathrm{\μ}}^2}---\left(1\right)$

μ is the average of signal envelope square in the formula, σ ²Variance for signal envelope square.Under certain signal to noise ratio condition, R can distinguish AM, SSB and (CW, MPSK, MFSK).Fig. 2 is the schematic diagram of feature R with the signal to noise ratio change curve.

2. power spectrum compactness feature

Definition as the formula (2):

${\mathrm{\μ}}_{42}^p=\frac{E\left\{P_{\mathrm{cn}}^4\left(i\right)\right\}}{E{\left[P_{\mathrm{cn}}^2\left(i\right)\right]}^2}---\left(2\right)$

P in the formula _Cn(i) be the amplitude of zero center normalized power spectrum, P (i) is the power spectrum of signal, and adopting improved period map method is that the welch method is calculated.P _Cn(i)=P _nAnd P (i)-1, _n(i)=P (i)/m _p, m _pAverage for the power spectrum amplitude.

Be mainly used to distinguish AM, CW, SSB and MFSK, mpsk signal.For the power spectrum of AM and CW signal owing to be to be main frequency spectrum with single carrier, so compactness is stronger, other signal then compactness a little less than.Fig. 3 is the power spectrum schematic diagram of feature with the signal to noise ratio change curve that compact.

3. absolute nothing is inserted folded differential phase standard deviation d Δ _ApThe differential phase of signal is:

Δφ(i)＝φ(i)-φ(i-1) (3)

Wherein φ is the slotting folded phase place of nothing of various types of signal.With the mpsk signal is example, and signal expression is:

Wherein, m=0,1 ..., M-1, A are signal amplitude, f _cBe the carrier frequency of signal, T _bBe symbol interval, g (t) generally adopts rectangular pulse or moulding pulse.

It does not have slotting folded phase place and is:

I=1 wherein, 2 ..., N, N are signal length.The d Δ _ApFor there not being the standard deviation of inserting folded phase delta phi.This feature can effectively be distinguished MFSK and other signal.Fig. 4 is the absolute schematic diagram that does not have slotting folded phase place standard deviation with the signal to noise ratio change curve.

4. effectively compose the peak number.Adopting improved period map method is that welch method calculating spectrogram can obtain more level and smooth power spectrum curve, is convenient to the statistics of peak value.Though the method for traditional search local maximum can well be estimated MFSK signal modulation number M, cause the erroneous judgement of other type signal spectrum peak number easily.The present invention has provided new peak value searching method, and its step is as follows:

Step 1401: network side adopts the power spectrum chart P of welch method signal calculated; Spectrogram P is searched for its maximum and minimum;

Step 1402: as P (i)＞P (i-1) ﹠amp; P (i)＞P (i+1) ﹠amp; P (i)＞P ₁The time, there is a maximum, write down the position I=[i of maximum ₁, i ₂..., i _m] and corresponding maximum

The tentative number of the number M of statistics maximum for the spectrum peak;

Step 1403: as P (j)＜P (j-1) ﹠amp; During P (j)＜P (j+1), there is a minimum, writes down minimizing position J=[j ₁, j ₂..., j _m] and corresponding minimum

Step 1404: search for minimum position corresponding between per two maximum, if there is i ₁＜j ₁＜i ₂, right

Judge, if

M=M-1 then; Otherwise M remains unchanged;

Step 1405: travel through the minimum of all maximum correspondences, it is judged, obtain the peak value number that final M is signal.

This searching method can be judged the spectrum peak number of various signals exactly.Fig. 5 effectively composes the schematic diagram of peak number with the signal to noise ratio change curve.

5. square spectrum single-frequency components detected parameters.For bpsk signal, square spectrum has very strong single-frequency components at 2 times of carrier frequency places; Other psk signal and SSB signal then do not have this feature.Whether we can exist single-frequency components to come it is classified by square spectrum of detection signal.In order more accurately single-frequency components to be detected, the present invention proposes a kind of new single-frequency detected parameters method, its step is as follows:

Step 1501: square spectrum P of network side signal calculated ₁, to spectrum P ₁Search for its maximum Y ₁And to write down peaked position be I ₁

Step 1502: select length L, calculate from P ₁(I ₁-L: I ₁-2) and P ₁(I ₁+ 2: I ₁+ L) average of the amplitude of all spectral lines is Y ₂

Step 1503: the ratio Y that calculates both ₁/ Y ₂, be existence one single-frequency components when it surpasses a certain threshold values, otherwise do not exist.

This method can be distinguished BPSK and other signal effectively, and robustness and separability are better.Fig. 6 is that a square spectrum single-frequency components detects than the schematic diagram with the signal to noise ratio change curve.

6. biquadratic is composed the single-frequency components detected parameters.Signal biquadratic spectrum is the quadruplicate frequency spectrum of signal.The biquadratic spectrum of signal is mainly used in the mpsk signal of distinguishing M=4 and M＞4.There is tangible single-frequency components in the QPSK signal afterwards through the biquadratic spectrum, and 8PSK does not exist, and can distinguish QPSK and 8PSK signal thus.For the detection method of single-frequency components, the detection method of using in the use characteristic 5 still.Fig. 7 is that biquadratic spectrum single-frequency components detects than the schematic diagram with the signal to noise ratio change curve.

From Fig. 2-Fig. 7 as can be seen, these parameters that the present invention constructed do not rely on priori, and are insensitive to the variation of signal to noise ratio, have good robustness, and still separability is good under lower signal to noise ratio, is suitable for practical application.

Step 20, network side are handled the characteristic parameter of posttectonic signal of communication by generalized discriminant analysis (GDA) mode, the characteristic parameter after the processing that obtains is stored;

Generalized discriminant analysis (GDA) basic idea is: at first will import data map to higher dimensional space by kernel function, make it at this space linear separability or linear separability as far as possible, calculate at higher dimensional space then and satisfy linear discriminant analysis (Linear Discriminnt Analysis, LDA) the best projection direction of criterion, thus processing finished to feature.

If training sample is x ₁, x ₂..., x _N, in the signal modulation classification, each sample x _iForm by the q dimensional feature, promptly

Represent the characteristic vector group (q=6) of such signal.If sample x can be divided into the c class, the number of every class sample is N _i

The definition Nonlinear Mapping is mapped to high-dimensional feature space with sample from the original input space: promptly:.Discrete matrix in the class of higher dimensional space then Discrete matrix between class

Overall discrete matrix

Be respectively:

$S_w^{\mathrm{\&phi;}}=\frac{1}{N}\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N_i}{\mathrm{\&Sigma;}}}(\mathrm{\&phi;}\left(x_j^i\right)-m_i^{\mathrm{\&phi;}}){(\mathrm{\&phi;}\left(x_j^i\right)-{\mathrm{<figure-callout\; id="0"\; label="m"\; filenames="HSA00000219585600021.png"\; state="\{\{state\}\}">m</figure-callout>}}_0^{\mathrm{\&phi;}})}^T---\left(4\right)$

$S_b^{\mathrm{\&phi;}}=\frac{1}{N}\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}\frac{N_i}{N}(m_i^{\mathrm{\&phi;}}-{\mathrm{<figure-callout\; id="0"\; label="m"\; filenames="HSA00000219585600021.png"\; state="\{\{state\}\}">m</figure-callout>}}_0^{\mathrm{\&phi;}}){(m_i^{\mathrm{\&phi;}}-{\mathrm{<figure-callout\; id="0"\; label="m"\; filenames="HSA00000219585600021.png"\; state="\{\{state\}\}">m</figure-callout>}}_0^{\mathrm{\&phi;}})}^T---\left(5\right)$

Wherein

Represent the average of i class sample at feature space F,

(i=1 ..., c; J=1 ..., N _i) represent j sample in the i class sample set in the F space.

In feature space F, linear discriminant analysis is expressed as the maximization target function:

$J_1\left(w\right)=\frac{w^T{S}_b^{\mathrm{\&phi;}}w}{w^T{S}_w^{\mathrm{\&phi;}}w}---\left(7\right)$

The target of GDA characteristic processing is to seek the projecting direction w that helps classifying most, makes classification separability criterion function J ₁(w) satisfy orthogonality in the time of for maximum, promptly

I, j=1 ..., d, wherein d is the dimension in the high-dimensional feature space projection.

Know by theory of reproducing kernel space, separate w arbitrarily in the F space ^φOpened by the sample in the F space, that is:

Therefore:

Wherein:

${\mathrm{\&mu;}}_i=(\frac{1}{N_i}\underset{k=1}{\overset{N_i}{\mathrm{\&Sigma;}}}(\mathrm{\&phi;}\left(x_1\right)\&CenterDot;\mathrm{\&phi;}\left(x_k^i\right)),\&CenterDot;\&CenterDot;\&CenterDot;,\frac{1}{N_i}\underset{k=1}{\overset{N_i}{\mathrm{\&Sigma;}}}(\mathrm{\&phi;}\left(x_N\right)\&CenterDot;\mathrm{\&phi;}\left(x_k^i\right)))---\left(10\right)$

With (9) formula substitution formula (7), then have:

${\mathrm{<figure-callout\; id="1"\; label="J"\; filenames="HSA00000219585600032.png"\; state="\{\{state\}\}">J</figure-callout>}}_1^{\&prime;}\left(\mathrm{\&alpha;}\right)=\frac{w^T{S}_b^{\mathrm{\&phi;}}w}{w^T{S}_w^{\mathrm{\&phi;}}w}=\frac{{\mathrm{\&alpha;}}^T{K}_b\mathrm{\&alpha;}}{{\mathrm{\&alpha;}}^T{K}_w\mathrm{\&alpha;}}---\left(11\right)$

Condition of orthogonal constraints becomes:

Wherein:

$K_b=\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}\frac{N_i}{N}({\mathrm{\&mu;}}_i-{\mathrm{\&mu;}}_0){({\mathrm{\&mu;}}_i-{\mathrm{\&mu;}}_0)}^T---\left(12\right)$

$K_w=\frac{1}{N}\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{N_i}{\mathrm{\&Sigma;}}}({\mathrm{\&xi;}}_{x_j^i}-{\mathrm{\&mu;}}_i){({\mathrm{\&xi;}}_{x_j^i}-{\mathrm{\&mu;}}_i)}^T---\left(13\right)$

$K_t=\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}({\mathrm{\&xi;}}_{x_j}-{\mathrm{\&mu;}}_i){({\mathrm{\&xi;}}_{x_j}-{\mathrm{\&mu;}}_i)}^T=K_b+K_w---\left(14\right)$

The characteristic value characteristic of correspondence vector of formula (11) is carried out orthogonalization can obtain conversion coefficient vector α, substitution (8) formula get final product best weight value w.

The GDA characteristic processing is output as:

y＝W ^Tφ(x)＝[w ₁，w ₂，…，w _d] ^Tφ(x)＝[α ₁，α ₂，…，α _d] ^Tξ _x (15)

ξ wherein _x=(k (x ₁, x), k (x ₂, x) ..., k (x _N, x)) ^TBe sample vector based on nuclear.

Described step 20 specifically may further comprise the steps:

Step 201, network side are formed input sample data collection x at its posttectonic characteristic parameter of every class signal extraction;

Step 202, select suitable kernel function k (x, y) and the kernel function parameter, the characteristic vector dimension d that setting will be extracted;

Step 203, by formula (12) and (13) calculating nuclear matrix K _bAnd K _w

Step 204, calculating K _bα=λ K _wThe characteristic vector of α, and carry out orthogonalization, obtain α;

Step 205, signal characteristic vector to be imported is arbitrarily extracted through the characteristic parameter after the GDA according to (15) formula.

In above-mentioned steps, need to the kernel function k of GDA (x, y) and parameter wherein be provided with, to reach optimum characteristic processing effect, need the combining classification device carry out parameter and select experiment.

Use the GDA algorithm that the feature of extracting is handled, not only can reduce intrinsic dimensionality, simplify the design of rear end grader; And owing to used dispersion maximum and the interior dispersion minimum criteria of class between kernel function and class, the feature after feasible the processing is compared with methods such as LDA, PCA commonly used, KPCA and is applicable to the signal classification more.

Characteristic parameter after step 30, network side will be handled, the decision tree judgement by AdaBoost (AdaBoost-Decision Tree, AdaBoost-DT) differentiate, the modulation system of the signal of communication that identification receives by the mode of grader.

Classifier design is the key link in the signal identification, and good grader both should have higher classification performance, should calculate simultaneously simply, is easy to Project Realization.Based on decision tree (Decision Tree, DT) though the grader of differentiating is a common classification device in the engineering, easy to use, but its accuracy of identification is lower, the present invention proposes a kind of decision tree (AdaBoost-Decision Tree based on AdaBoost, AdaBoost-DT) grader has improved nicety of grading.

The basic thought of AdaBoost is that a strong classifier can be obtained by several Weak Classifier combinations.Weak Classifier obtains having on the training set of different weights, and final strong classifier is the linear combination of these a series of Weak Classifiers.Its process is an iteration each time, and AdaBoost improves by its wrong those sample weights of dividing, and reduces the weight of the sample point of correct classification.Arthmetic statement is as follows:

Step1: the training set of establishing input is s={ (x ₁, y ₁), (x ₂, y ₂) ..., (x _N, y _N), y _i∈ 1 ,+1}.Wherein N is the number of training sample, y _iBe classification under the signal.If the number of Weak Classifier is T.The weight d of initialization training sample ₁(i)=and 1/N, i=1,2 ..., N.Circulation t=1,2 ..., T.

Step2: the training weight distribution is d _t=[d _t(1), d _t(2) ..., d _t(N)] weak learner obtains making weighted error function J _tMinimum Weak Classifier h _t(x).Wherein:

$J_t=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{d}_t\left(i\right)I(h_t\left(x_i\right)\&NotEqual;y_i)---\left(16\right)$

I(h _t(x _i)≠y _i)＝1，I(h _t(x _i)＝y _i)＝0。

Step3: calculate h _t(x) error ε _t=P _D(h _t(x _i) ≠ y _i), if ε _t=0 or ε _t＞0.5, make t=T and jump out circulation.

Step4: order

And renewal sample weights:

Z wherein _tBe to make

Normalization coefficient.

Step5: the last strong classifier that forms of output:

AdaBoost is based on two class problems and proposes, and need it be extended to multiclass by certain method, and multicategory classification method commonly used has: AdaBoost.M1, AdaBoost.M2 and AdaBoost.MH.Wherein AdaBoost.M1 is the most direct a kind of method, and Weak Classifier directly is divided into sample set the individual classification of k (k for can be sub-category number), and computing formula is identical with two class algorithms, and the present invention adopts the method for AdaBoost.M1 to experimentize.

In GDA algorithm of the present invention, need to kernel function k (x, y) and parameter be provided with.Common kernel function have polynomial kernel function, neural net kernel function and radially basic kernel function (Radial BasisFunction, RBF) etc.Wherein the expression formula of RBF kernel function is:

$k(x,x^{\&prime;})=\exp (-\frac{{\left|\right|x-x^{\&prime;}\left|\right|}^2}{{\mathrm{\&sigma;}}^2})---\left(17\right)$

Wherein σ is a undetermined parameter, and it has controlled the kernel function width.Because RBF characteristic of correspondence space is infinite dimensional, limited sample is linear separability certainly in this feature space, and it only has a undetermined parameter σ, so the most normal being used, the present invention also can select radially basic kernel function.

In the AdaBoost algorithm, the number T that Weak Classifier need be set is to reach best classifying quality.If the T value is less, then can not obtain optimum classification results because Weak Classifier is very few; If the T value is bigger, then can cause AdaBoost too to pay close attention to those noise spots, cause wrong the branch, the study phenomenon appearred.In the present invention, the characteristic processing of GDA is in order to carry out the AdaBoost algorithm classification better, and the Weak Classifier number T of AdaBoost algorithm also is subjected to the influence of GDA algorithm, and both interdepend.The present invention provides GDA and the AdaBoost combined parameters selection algorithm based on cross validation for this reason, and the step of parameter selection course is as follows:

Characteristic parameter after step 301, network side extract from the signal of communication of one group of known signal modulation type and handle, the composition characteristic Vector Groups, and the characteristic vector group is divided into two parts at random, be respectively training sample and test sample book;

Step 302, network side setting search width L (L is the integer greater than 0), make σ [σ-L, σ-L+1 ..., σ+L-1, σ+L] and the interior variation of scope; Be limited to D (D is the integer greater than 0) in the search of setup parameter T;

The whenever selected parameter σ of step 303, network side, from 1,2, D travels through parameter T, select for use training sample that the decision tree of GDA and AdaBoost judgement grader is trained respectively, and adopt test sample book to calculate the correct recognition rata of GDA-AdaBoost-DT algorithm, the value of σ and T when noting the highest accuracy of identification;

Step 304, network side with σ [σ-L, σ-L+1 ..., σ+L-1, σ+L] and traversal in the scope, σ and T when selecting the highest accuracy of identification.If satisfy required precision, then parameter is selected to finish.Otherwise, have the σ of high accuracy of identification ₁σ with inferior high accuracy of identification ₂Between, with littler stepping, serve as at interval at [σ for example with 0.1 ₁+ 0.1, σ ₁+ 0.2 ..., σ ₂] in the scope again search (suppose σ here ₁＜σ ₂), select to have the σ and the T of high accuracy of identification, so repeatedly,, determine parameter σ and T up to obtaining satisfied correct recognition rata.

Described network side is with the characteristic parameter of posttectonic signal of communication, handle by the GDA mode, with the characteristic parameter after the processing that obtains, in the step that the mode of the decision tree judgement grader by AdaBoost is differentiated, also comprise the step of determining parameter σ and T, carry out the differentiation of mode of the decision tree judgement grader of the processing of GDA mode and AdaBoost by parameter σ and T, wherein σ is the undetermined parameter of control kernel function width, and T is the number of Weak Classifier.It should be noted that said process is the parameter selection course of grader, is not the assorting process of grader.

The invention will be further described below by experiment; By under the different signal to noise ratios of simulation calculation based on the correct recognition rata of Adaboost decision Tree algorithms, and compare with traditional decision tree classification device.Computer Simulation select for use AM, CW, SSB totally 3 kinds of modulated-analog signals and 2FSK, 4FSK, 8FSK, BPSK, QPSK, 8PSK totally 6 kinds of digital signals experimentize.In the experiment, signal sampling frequency 25kHz, chip rate 200～1000Baud, signal length 8192 points.Wherein the frequency interval of 2FSK, 4FSK, 8FSK is respectively 2kHz, 1kHz and 500Hz.Under the 10dB condition, every class signal is independently produced 20 groups of sample datas and extract the characteristic vector group, select the RBF kernel function for use, adopt the system of selection of cross validation combined parameters to select kernel function parameter σ among the GDA and the Weak Classifier number T in the AdaBoost-DT algorithm.Obtain preceding 3 characteristic components by the GDA computing.Respectively DT grader and AdaBoost-DT grader being trained, in 0～20dB scope is to produce 100 groups of all kind of modulations sample of signal data at random at interval to discern then with 5dB, and its correct recognition rata as shown in Table 1 and Table 2.

Table 1GDA+DT algorithm signal correct recognition rata (%)

Table 2GDA+AdaBoost-DT algorithm signal correct recognition rata (%)

Table 1 and table 2 provided traditional decision Tree algorithms respectively and based on the accuracy of identification of the decision Tree algorithms of AdaBoost with the signal to noise ratio change list.Two kinds of employed features of grader are all carried out characteristic processing through the GDA algorithm.As can be seen, after the GDA algorithm, the feature separability is good, even adopt decision Tree algorithms correct recognition rata when 0dB also to reach 84%.Wherein adopt GDA+AdaBoost-DT algorithm correct recognition rata higher.

The present invention proposes a kind of recognition methods of the signal modulation system that combines based on generalized discriminant analysis (GDA) and AdaBoost decision Tree algorithms.Even this method also has good recognition performance under low signal-to-noise ratio.Compare with common method, in bigger signal to noise ratio scope, all have higher accuracy of identification; Proposed some effective features, these features are without any need for priori, and are insensitive to the variation of signal to noise ratio.Separability is strong, and algorithm is simple, is suitable for engineering and uses; The generalized discriminant analysis method is applied to the Modulation Recognition of Communication Signal field, has proposed a kind of modulation signal characteristic processing method based on generalized discriminant analysis.After process generalized discriminant analysis algorithm is handled feature, feature foreign peoples's separability is strengthened, similar aggregation increases, for follow-up high-precision signal classification is laid a good foundation; (Adaptive Boosting, AdaBoost) method is applied to signal of communication modulation type identification field, has proposed a kind of based on decision tree adaptive boosting classifier design method with the adaptive boosting in the integrated study.This method amount of calculation is moderate, but operational precision is apparently higher than traditional decision tree method of discrimination.

The above; only for the preferable embodiment of the present invention, but protection scope of the present invention is not limited thereto, and anyly is familiar with the people of this technology in technical scope disclosed in this invention; the variation that can expect easily or replacement all should be encompassed within protection scope of the present invention.Therefore, protection scope of the present invention should be as the criterion with the protection range of claim.

Claims (8)

1. the recognition methods of a modulation mode of communication signal is characterized in that, comprising:

The network side receiving communication signal is constructed the characteristic parameter of this signal of communication and is stored according to the characteristic of the signal of communication that receives;

Described network side is handled the characteristic parameter of posttectonic signal of communication by generalized discriminant analysis GDA mode, the characteristic parameter after the processing that obtains is stored;

Characteristic parameter after the processing that described network side will obtain is adjudicated the mode of grader and is differentiated the modulation system of the signal of communication that identification receives by the decision tree of AdaBoost.

2. the method for claim 1 is characterized in that,

Described network side is meant according to the characteristic parameter of the characteristic structure signal of communication of the signal of communication that receives: network side is according to the characteristic parameter of the information structuring signal of communication of amplitude, frequency spectrum, square spectrum and/or the phase place of the signal of communication that receives.

3. the method for claim 1 is characterized in that,

The characteristic parameter of the signal of communication of described network side structure comprises:

Characteristic parameter R, it is defined as:

Wherein μ is the average of signal envelope square, σ ²Variance for signal envelope square;

Power spectrum compactness feature

It is defined as:

P wherein _Cn(i) be the amplitude of zero center normalized power spectrum, P (i) is the power spectrum of signal, P _Cn(i)=P _nAnd P (i)-1, _n(i)=P (i)/m _p, m _pAverage for the power spectrum amplitude;

Absolute nothing is inserted folded differential phase standard deviation d Δ _Ap, wherein the differential phase of signal is: Δ φ (i)=φ (i)-φ (i-1), φ are that the nothing of various types of signal is inserted folded phase place;

Effectively compose the peak number; Square spectrum single-frequency components detected parameters and biquadratic spectrum single-frequency components detected parameters.

4. method as claimed in claim 3 is characterized in that,

Described network side is the step of effectively composing the peak number according to the characteristic parameter of the characteristic structure signal of communication of the signal of communication that receives, and may further comprise the steps:

Described network side is searched for its maximum and minimum by the power spectrum chart P of welch method signal calculated to spectrogram P,

If described network side judges P (i)＞P (i-1) ﹠amp; P (i)＞P (i+1) ﹠amp; P (i)＞P ₁The time, there is a maximum, the position I=[i of record maximum ₁, i ₂..., i _m] and corresponding maximum

The tentative number of the number M of statistics maximum for the spectrum peak;

If described network side judges P (j)＜P (j-1) ﹠amp; During P (j)＜P (j+1), there is a minimum, writes down minimizing position J=[j ₁, j ₂..., j _m] and corresponding minimum

Described network side is searched for minimum position corresponding between per two maximum, has i if judge ₁＜j ₁＜i ₂, then right Judge, if judge

M=M-1 then; Otherwise M remains unchanged;

Described network side travels through the minimum of all maximum correspondences, and it is judged, obtains effective spectrum peak number that final M is signal.

5. method as claimed in claim 3 is characterized in that,

Described network side is square step of spectrum single-frequency components detected parameters according to the characteristic parameter of the characteristic structure signal of communication of the signal of communication that receives, and may further comprise the steps:

Square spectrum P of described network side signal calculated ₁, to a square spectrum P ₁Search for its maximum Y ₁And to write down peaked position be I ₁

Described network side is selected length L, calculates from P ₁(I ₁-L: I ₁-2) and P ₁(I ₁+ 2: I ₁+ L) average of the amplitude of all spectral lines is Y ₂Calculate both ratio Y ₁/ Y ₂, judgement then judge to have a single-frequency components, otherwise there is not single-frequency components in judgement if it surpasses a certain threshold values.

6. the method for claim 1 is characterized in that,

Described network side is handled the characteristic parameter of posttectonic signal of communication by the GDA mode, the characteristic parameter after obtaining handling comprises:

Described network side is formed input sample data collection x at its posttectonic characteristic parameter of every class signal extraction;

Described network side select suitable kernel function k (x, y) and the kernel function parameter, the characteristic vector dimension d that setting will be extracted;

Described network side calculates nuclear matrix K according to following formula _bAnd K _w

Wherein, $K_b=\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}\frac{N_i}{N}({\mathrm{\&mu;}}_i-{\mathrm{\&mu;}}_0){({\mathrm{\&mu;}}_i-{\mathrm{\&mu;}}_0)}^T;$

N is the signal of communication feature samples number of getting; C is the classification number that sample can divide, and the number of every class sample is N _iξ _xBe sample vector based on nuclear,

Average for each sample inner product nuclear;

Described network side calculating K _bα=λ K _wThe characteristic vector of α, and carry out orthogonalization, obtain α;

Described network side is according to formula y=W ^Tφ (x)=[w ₁, w ₂..., w _d] ^Tφ (x)=[α ₁, α ₂..., α _d] ^Tξ _x, ξ wherein _x=(k (x ₁, x), k (x ₂, x) ..., k (x _N, x)) ^TBe sample vector, to the signal characteristic vector of input, through the characteristic parameter after obtaining handling after the GDA mode based on nuclear.

7. the method for claim 1 is characterized in that,

Described network side is with the characteristic parameter of posttectonic signal of communication, handle by the GDA mode, with the characteristic parameter after the processing that obtains, in the step that the mode of the decision tree judgement grader by AdaBoost is differentiated, also comprise the step of determining parameter σ and T, carry out the differentiation of mode of the decision tree judgement grader of the processing of GDA mode and AdaBoost by parameter σ and T, wherein σ is the undetermined parameter of control kernel function width, and T is the number of Weak Classifier.

8. method as claimed in claim 7 is characterized in that,

Described network side is determined the step of parameter σ and T, comprising:

Described network side extracts the characteristic parameter after the Signal Processing from the signal of communication of one group of known signal modulation system type, the composition characteristic Vector Groups, and the characteristic vector group is divided into two parts at random, be respectively training sample and test sample book;

Described network side setting search width L, L is the integer greater than 0, make σ [σ-L, σ-L+1 ..., σ+L-1, σ+L] and the interior variation of scope, wherein σ is the undetermined parameter of control kernel function width; Be limited to D in the search of setup parameter T, D is the integer greater than 0, and T is the number of Weak Classifier;

The whenever selected parameter σ of described network side is from 1,2,, D travels through parameter T, selects for use training sample that the decision tree judgement grader of GDA and AdaBoost is trained respectively, and adopt test sample book to calculate correct recognition rata, the value of σ and T when writing down the highest accuracy of identification;

Described network side with σ [σ-L, σ-L+1 ..., σ+L-1, σ+L] and traversal in the scope, σ and T when selecting the highest accuracy of identification, if judge and satisfy required precision, then parameter is selected to finish; Otherwise, have the σ of high accuracy of identification ₁σ with inferior high accuracy of identification ₂Between, search for again with littler stepping, wherein σ ₁＜σ ₂, select to have the σ and the T of high accuracy of identification, after obtaining satisfied correct recognition rata, determine parameter σ and T.

Images