CN101783777A - Digital modulation signal recognizing method - Google Patents

Abstract

The invention discloses a digital modulation signal recognizing method. The recognition process of the method comprises: inputting a characteristic parameter vector to be recognized into an SOFM part in a recognition network to extract effective characteristic parameters and determining the characteristic parameters adopted in each node in a decision tree from top down; constructing a recognition network structure by using RBF networks to replace all nodes in the decision tree; training each RBF network by using a known training set to minimize the recognition error, determining all connection weights and decision thresholds in a hidden layer and an output layer, and fixing the recognition network; inputting sampling values into the fixed recognition networks and calculating the values of the effective characteristic parameters according to the structure of the decision tree from top down; and inputting the values of the effective characteristic parameters into the RBF networks to calculate the output values and comparing the output values with the decision thresholds to determine a signal modulation mode. The method effectively extracts the characteristic parameters, lowers the complexity and inaccuracy of theoretical demonstration and improves recognition rate under multipath and low signal-to-noise ration conditions.

Description

Digital modulation signal recognizing method

Technical field

The invention belongs to communication technical field, be specifically related under a kind of multipath channel, the low signal-to-noise ratio condition recognition methods based on the various digital signal modulation modes of neural net.

Background technology

Along with development of Communication Technique, the system of modulation signal and modulation pattern become complicated more various, and signal environment is more and more intensive, and it is particularly important and urgent that the identification of signal modulation style seems.Aspect civilian, the signal authentication of blipology in radio spectrum management, disturb in the recognition technology and seem particularly important.Aspect military, no matter the especially communication countermeasures in the electronic countermeasures is to disturb enemy's communication or crack enemy's signal of communication, all must at first discern the modulation system of enemy's signal of communication.Modulation mode of communication signal identification also is the important technology in software radio and the reconstruct communication system.Especially, along with the extensive employing of orthogonal frequency division multiplex OFDM technology, the signal set of Modulation Signals Recognition technical research is increasing, and the complexity of identification is also more and more higher.

But it is at present less for the Study of recognition of ofdm signal, the recognition technology research of existing digital signal concentrates on ideal conditions or only has under the white Gaussian noise condition, simultaneously, existing recognition technology discrimination under the low signal-to-noise ratio condition is not high, can not satisfy the needs of practical application.Therefore, the blipology of the suitable many systems of a kind of energy of research seems very important under multipath, low signal-to-noise ratio condition.The information-distribution type that neural net had stores, extensive self-adaptive parallel is handled and the characteristics such as fault-tolerance of height, particularly its learning ability and fault-tolerance have distinctive feature to uncertain pattern recognition, are specially adapted to the signal identification under multipath, the low signal-to-noise ratio condition.

At present, based on existing two problems in the neural network method: the one, single neural net realizes the recognition system more complicated, required neuron number is many, and recognition performance is undesirable.The 2nd, in the recognition system characteristic parameter choose the experience that depends on the designer, also do not have a kind of blanket method of choosing and differentiating the validity feature value at present.At first problem, people have proposed the neural net based on decision tree, carry out the layering judgement by making up a plurality of graders, and this method has reduced the complexity of single grader.But, also be based on designer's experience for the design of decision tree.Feature Selection and decision tree design rely on designer's experience and have brought the low problem of discrimination, and under new communication environment, the structure of decision tree need rebuild.

Summary of the invention

The objective of the invention is to overcome the deficiency of above-mentioned prior art, a kind of digital modulation signal recognizing method is provided, improving the signal identification rate in the communication system, and need not to make up once more new decision tree and just can realize that the digital signal in the different communication environment discerns.

For achieving the above object, the present invention includes following process:

(1) feature parameter vector to be discriminated is input to the SOFM part of recognition network, and extracts effective characteristic parameters;

(2) according to the extraction result of SOFM network and the extraction order of validity feature, determine the characteristic parameter that adopts in each node of decision tree from top to bottom;

(3), constitute the network configuration of recognition system with each node in the RBF network replacement decision tree;

(4) utilize known training set, each RBF recognition network in the recognition system network is trained, make the identification error of training set reach minimum, each that determine each RBF recognition network hidden layer and output layer is connected weights and decision threshold, and fixes these recognition networks;

(5) signal to be identified that will receive is sampled, and sampled signal is input in the fixing recognition network, calculates the effective characteristic parameters value from top to bottom according to the structure of decision tree;

(6) the effective characteristic parameters value is input to its output valve of calculating in the RBF recognition network, compares, judge the modulation system of identification signal with this output valve and decision threshold.

The present invention utilizes the design of SOFM network extraction effective characteristic parameters and decision tree, has significantly reduced the complexity that theoretical proof brings and the inexactness of identification, and this method has good adaptive capacity to the change of communication identification system simultaneously.Utilize the RBF network to replace the node of decision tree to significantly reduce the high problem of complexity of classifying and causing with single RBF grader.Three effective characteristic parameters that propose among the present invention: based on once with the characteristic parameter V of twice wavelet transformation _w, instantaneous frequency characteristic parameter MAX_fre, signal transient amplitude box function characteristic parameter DBF2 can improve the discrimination of signal under multipath, the low signal-to-noise ratio condition well.Emulation shows that under multipath channel, low signal-to-noise ratio condition, method proposed by the invention not only can effectively be extracted characteristic parameter, and has improved the discrimination of system signal greatly.

Description of drawings

Fig. 1. be the recognition system structure chart that the present invention proposes based on neural net;

Fig. 2. be the structure chart of SOFM network used among the present invention;

Fig. 3. be that the present invention is through the decision tree structure chart after the extraction of effective characteristic parameters;

Fig. 4. be the structure chart of the RBF network that uses among the present invention;

Fig. 5. be among the present invention based on once with the leaching process schematic diagram of the characteristic parameter of twice wavelet transformation.

Embodiment

Referring to Fig. 1, the present invention includes following process:

Process 1 is input to SOFM part in the recognition network with feature parameter vector to be discriminated, carries out the extraction of effective characteristic parameters.

The structure of SOFM network is referring to Fig. 2, wherein x ₁, x ₂... x _nBe input neuron, n is the dimension of input feature value, i.e. neuron number, s ₁, s ₂... s _mBe the neuron of output layer, m is the number of output neuron, represents final clusters number, w _IjRepresent the weights that are connected of i input neuron and j output neuron,

Utilize the adjustment of Kohonen learning algorithm to connect weights.Then the output result of j output neuron is:

$s_j=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{ij}}*x_i---\left(1\right)$

The extraction of effective characteristic parameters comprises the steps:

(1.1) selected feature parameter vector is F=[F ₁, F ₂... F _N], F wherein ₁, F ₂... F _NRepresent different characteristic parameters, with the characteristic parameter F of all kinds of sampled values ₁Be input in the SOFM network, and calculate the output result of each output neuron by formula (1);

(1.2) statistical nature parameter F ₁SOFM network output result's class categories;

(1.3) number of the class categories that counts and the known prototype of training sample are compared, if statistics is consistent with known prototype, then keep this effective characteristic parameters, and with the effective characteristic parameters of selecting for use in the node of this characteristic parameter as decision tree, otherwise, with F ₁Be judged to invalid characteristic parameter;

(1.4) to characteristic parameter F ₂... F _NCarry out identical operations, finally selected 7 effective characteristic parameters are respectively: based on once with the characteristic parameter V of twice wavelet transformation _w, twice wavelet transformation characteristic parameter V ₂, signal box function characteristic parameter DBF1, signal transient amplitude box function characteristic parameter DBF2, instantaneous frequency characteristic parameter MAX_fre, based on the characteristic parameter σ of non-weak signal real part _IWith characteristic parameter F based on the FFT conversion _v

Process 2, according to the extraction result of SOFM network and the extraction order of validity feature, the characteristic parameter that adopts in definite from top to bottom each node of decision tree, the structure of decision tree is referring to Fig. 3.

Instantaneous frequency characteristic parameter MAX_fre differentiates digital signal from noise in the ground floor of decision tree, in the second layer based on once with the characteristic parameter V of twice wavelet transformation _wSingle carrier and multi-carrier signal are distinguished, in the 3rd layer of decision tree the box function characteristic parameter DBF2 of signal transient amplitude with DVB-T signal and 802.11a signaling zone separately, based on the characteristic parameter σ of non-weak signal real part _I2ASK, 4ASK, 2PSK are distinguished twice wavelet transformation characteristic parameter V in the 4th layer of decision tree from other single-carrier signal ₂The QAM signal is differentiated from 2FSK, 4FSK, 4PSK signal, based on the characteristic parameter F of FFT conversion _v2ASK, 2FSK signal are separated with the 4ASK signaling zone, the box function characteristic parameter DBF2 of signal transient amplitude separates 2FSK signal and 4FSK, 4PSK signaling zone in the decision tree layer 5, and the box function characteristic parameter DBF1 of signal differentiates 4PSK signal and 4FSK signal in the decision tree layer 6.

Process 3 replaces each node in the decision tree with the RBF network, constitutes the network configuration of recognition system, and the structure of RBF network is referring to Fig. 4, wherein, and X=[x ₁, x ₂X _L] be the input of network, Y=[y ₁, y ₂Y _M] be the output of network, $C=[c_1,c_2\·\·\·c_{N_1}]$ Be the central value of hidden layer, N ₁Node number for hidden layer. $F=[f_1,f_2\·\·\·f_{N_1}]$ Be RBF, then j node is output as in the ground floor:

$u_j=\exp (-\frac{(X-c_j){(X-c_j)}^{*}}{2{\mathrm{\σ}}_j^2}),j=\mathrm{1,2},\·\·\·N_1---\left(2\right)$

Wherein, σ _J ²Be the normalized parameter of j node, ^*The expression conjugate operation.

I output node equation of RBF network is:

y _i＝W _iU ^*，i＝1，2，…M????(3)

Wherein, $W_i=[w_{1i},w_{2i}\·\·\·w_{N_{1}i}]$ Be the weight vector of node, $U=[u_1,u_2\·\·\·U_{N_1}]$ Be the hidden layer output vector.

Process 4, utilize known training set, each RBF network in the recognition system network is trained, wherein, hidden layer carries out the K-nearest neighbor algorithm, and output layer is trained network with the LMS algorithm, makes the identification error of training set reach minimum, be connected weights and decision threshold with each that determine hidden layer and output layer, and fix this recognition network.

Process 5 is sampled the signal to be identified that receives, and sampled signal is input in the fixing recognition system network, calculates the effective characteristic parameters value from top to bottom according to the structure of decision tree, comprises following process:

(5.1) calculate instantaneous frequency characteristic parameter MAX_fre:

(5.1.1) according to formula

The instantaneous phase of calculating sampling value, wherein, Q (i) and I (i) are respectively the imaginary part and the real part of n signal sampling value;

(5.1.2) pass through formula The instantaneous frequency of calculating sampling value, and to instantaneous frequency

Maximizing obtains instantaneous frequency characteristic parameter MAX_fre.

(5.2) calculate based on once with twice wavelet transformation characteristic parameter V _w:

Based on once with the characteristic parameter V of twice wavelet transformation _wLeaching process referring to Fig. 5, specifically comprise following process:

(5.2.1) according to formula $W(\mathrm{\α},\mathrm{\τ})=\left|\frac{1}{\sqrt{\widehat{\mathrm{\α}}{f}_s}}\underset{k}{\mathrm{\Σ}}s\left(k\right){\mathrm{\Ψ}}^{*}\left(\frac{k-n}{\widehat{\mathrm{\α}}}\right)\right|$ Sampled value is carried out the Haar wavelet transformation one time, wherein, wherein,

Be the energy normalized factor, n is a shift factor,

Be scale factor, Ψ (t) is a wavelet mother function, and sampled value is carried out medium filtering, and calculates variance V ₁

(5.2.2) according to formula $W(\mathrm{\α},\mathrm{\τ})=\left|\frac{1}{\sqrt{\widehat{\mathrm{\α}}{f}_s}}\underset{k}{\mathrm{\Σ}}s\left(k\right){\mathrm{\Ψ}}^{*}\left(\frac{k-n}{\widehat{\mathrm{\α}}}\right)\right|$ Sampled value is carried out the Haar wavelet transformation again one time, sampled value is carried out medium filtering, and calculate variance V ₂

(5.2.3) by formula V _w=[V ₁V ₂] obtain based on once with twice wavelet transformation characteristic parameter V _w

(5.3) calculating is based on the characteristic parameter σ of non-weak signal real part _I:

(5.3.1) according to formula $m_a=1/N_s\underset{i=1}{\overset{N_s}{\mathrm{\Σ}}}{s}_i$ Obtain the mean value of sampled value, wherein, N _sBe the number of sampled value, s _iIt is sampled value;

(5.3.2) by formula a _i=s _i/ m _a-1 pair of sampled value is carried out zero normalized;

(5.3.3) pass through ${\mathrm{\σ}}_I=\sqrt{[1/c\underset{a_i>a_t}{\mathrm{\Σ}}{I}_t^2]-{[1/c\underset{a_i>a_t}{\mathrm{\Σ}}|I_i\left|\right]}^2}$ Calculate characteristic parameter σ based on non-weak signal real part _I, wherein, I _iBe the real part of i signal, a _tBe the decision threshold of non-weak signal, c is the number of non-weak signal.

(5.4) calculating is based on the characteristic parameter F of FFT conversion _v:

(5.4.1) the correlation matrix R of calculating sampling value _s(τ);

(5.4.2) to correlation matrix R _s(τ) ask the FFT conversion, and ask variance to obtain characteristic parameter F based on the FFT conversion to the result _v

(5.5) the box function characteristic parameter DBF1 of signal calculated:

(5.5.1) according to formula $d\left(\mathrm{\Δ}\right)=\underset{i=1}{\overset{N-1}{\mathrm{\Σ}}}|s_i-s_{i+1}|$ Calculate the absolute difference of neighbouring sample value, wherein,

Be the signals sampling value;

(5.5.2) pass through formula $d\left(2\mathrm{\Δ}\right)=\underset{i=1}{\overset{(N-1)/2}{\mathrm{\Σ}}}\left(\max \right\{a_{2i-1},a_{2i},a_{2i+1}\}-\min \{a_{2i-1},a_{2i},a_{2i+1}\}$ The comentropy of calculating sampling value, wherein,

Instantaneous amplitude for sampled value;

(5.5.3) by computing formula DBF1=1+log ₂(d (Δ)/d (2 Δ)) obtains the box function characteristic parameter DBF1 of signal, and wherein, d (Δ) is the absolute difference of neighbouring sample value instantaneous amplitude, and d (2 Δ) is the comentropy of sampled value

(5.6) the box function characteristic parameter DBF2 of signal calculated instantaneous amplitude:

(5.6.1) according to formula $b\left(\mathrm{\Δ}\right)=\underset{i=1}{\overset{N-1}{\mathrm{\Σ}}}|a_1-a_{i+1}|$ Calculate the absolute difference of neighbouring sample value instantaneous amplitude, wherein,

Instantaneous amplitude for sampled value;

(5.6.2) pass through formula $d\left(2\mathrm{\Δ}\right)=\underset{i=1}{\overset{(N-1)/2}{\mathrm{\Σ}}}\left(\max \right\{a_{2i-1},a_{2i},a_{2i+1}\}-\min \{a_{2i-1},a_{2i},a_{2i+1}\}$ The comentropy of calculating sampling value;

(5.6.3) by formula DBF2=1+log ₂(b (Δ)/d (2 Δ)) calculates the box function characteristic parameter DBF2 of signal, and wherein, b (Δ) is the absolute difference of neighbouring sample value instantaneous amplitude, and d (2 Δ) is the comentropy of sampled value.

Process 6 is input to the effective characteristic parameters value in the single RBF network, judges the modulation system of identification signal.

With reference to Fig. 3, deterministic process is as follows:

(6.1) for instantaneous frequency characteristic parameter MAX_fre, first its corresponding RBF network of training, and respectively the desired value of digital signal is made as 1, and the desired value of noise is made as 0, and threshold value is made as 0.5; Characteristic parameter with each sampled value to be identified carries out normalization again, and this normalized value is input to the RBF network as the input value of RBF network; At last, the output valve according to formula (3) calculating RBF network is judged to digital signal for output valve greater than the pairing sampled value of 0.5 normalized value, and the normalized value less than 0.5 is judged to noise;

(6.2) for based on once with the characteristic parameter V of twice wavelet transformation _w, first its corresponding RBF network of training, and respectively the desired value of single-carrier signal is made as 1, the desired value of multicarrier is made as 0, and threshold value is made as 0.5; Characteristic parameter with each sampled value to be identified carries out normalization again, and this normalized value is input to the RBF network as the input value of RBF network; At last, the output valve according to formula (3) calculating RBF network is judged to single-carrier signal for output valve greater than the pairing sampled value of 0.5 normalized value, and the normalized value less than 0.5 is judged to multi-carrier signal;

(6.3) for the box function characteristic parameter DBF2 of signal transient amplitude, its corresponding RBF network of training earlier, and respectively the desired value of DVB-T signal is made as 1, and the desired value of 802.11a signal is made as 0, and threshold value is made as 0.5; Characteristic parameter with each sampled value to be identified carries out normalization again, and this normalized value is as the input value of RBF network; At last, the output valve according to formula (3) calculating RBF network is judged to the DVB-T signal for output valve greater than the pairing sampled value of 0.5 normalized value, and the normalized value less than 0.5 is judged to 802.11a;

(6.4) for characteristic parameter σ based on non-weak signal real part _I, first its corresponding RBF network of training, and respectively the desired value of 2ASK, 4ASK, 2PSK signal is made as 1, the desired value of other single-carrier signal is made as 0, and threshold value is made as 0.5; Characteristic parameter with each sampled value to be identified carries out normalization again, and this normalized value is as the input value of RBF network; At last, the output valve according to formula (3) calculating RBF network is judged to 2ASK, 4ASK, 2PSK one class for output valve greater than the pairing sampled value of 0.5 normalized value, and the normalized value less than 0.5 is judged to another kind of.

(6.5) for twice wavelet transformation characteristic parameter V ₂, first its corresponding RBF network of training, the and respectively desired value of QAM signal is made as 1 is made as 0 with the desired value of 2FSK, 4FSK, 4PSK signal, and threshold value is made as 0.5; Characteristic parameter with each sampled value to be identified carries out normalization again, and this normalized value is as the input value of RBF network; At last, the output valve according to formula (3) calculating RBF network is judged to the QAM signal for output valve greater than the pairing sampled value of 0.5 normalized value, and the normalized value less than 0.5 is judged to 2FSK, 4FSK, 4PSK signal one class.

(6.6) for characteristic parameter F based on the FFT conversion _v, first its corresponding RBF network of training, the and respectively desired value of 4ASK signal is made as 1 is made as 0 with the desired value of 2ASK, 2FSK signal, and threshold value is made as 0.5; Characteristic parameter with each sampled value to be identified carries out normalization again, and this normalized value is as the input value of RBF network; At last, the output valve according to formula (3) calculating RBF network is judged to the 4ASK signal for output valve greater than the pairing sampled value of 0.5 normalized value, and the normalized value less than 0.5 is judged to 2ASK, 2FSK signal one class.

(6.7) for the box function characteristic parameter DBF2 of signal transient amplitude, its corresponding RBF network of training earlier, the and respectively desired value of 2FSK signal is made as 1 is made as 0 with the desired value of 4FSK, 4PSK signal, and threshold value is made as 0.5; Characteristic parameter with each sampled value to be identified carries out normalization again, and this normalized value is as the input value of RBF network; At last, the output valve according to formula (3) calculating RBF network is judged to the 2FSK signal for output valve greater than the pairing sampled value of 0.5 normalized value, and the normalized value less than 0.5 is judged to 4FSK, 4PSK signal one class.

(6.8) for signal box function characteristic parameter DBF1, first its corresponding RBF network of training, the and respectively desired value of 4FSK signal is made as 1 is made as 0 with the desired value of 4PSK signal, and threshold value is made as 0.5; Characteristic parameter with each sampled value to be identified carries out normalization again, and this normalized value is as the input value of RBF network; At last, the output valve according to formula (3) calculating RBF network is judged to the 4FSK signal for output valve greater than the pairing sampled value of 0.5 normalized value, and the normalized value less than 0.5 is judged to the 4PSK signal.

Effect of the present invention can further specify by following emulation:

1. simulated environment

Under multipath channel, low signal-to-noise ratio condition whole recognition system is carried out emulation, simulated environment is as shown in table 1.

Table 1 simulated environment

2. simulation result

Simulation result all is based on 100 Monte-carlo experiments, as table 2.

The discrimination of digital signal under the different signal to noise ratios of table 2

SNR(db)	?2ASK	4ASK	802.11a	DVB-T
					0	?98％	100％	100％	100％
4	?100％	100％	100％	100％
					8	?100％	100％	100％	100％
12	?100％	100％	100％	100％
					16	?100％	100％	100％	100％
20	?100％	100％	100％	100％

SNR(db)	?2ASK	4ASK	802.11a	DVB-T
					SNR(db)	?2FSK	4FSK	QAM	PSK
0	?95.5％	80％	98％	97％
					4	?97％	78％	99％	99％
8	?100％	78％	100％	90％
					12	?100％	100％	100％	82％
16	?100％	92％	100％	78％
					20	?100％	76％	100％	76％

As can be seen from Table 2, the method that the present invention proposes not only can well be discerned digital signal, has improved the discrimination under multipath, the low signal-to-noise ratio condition, and this recognition system greatly reduces the complexity of signal identification.

Claims (6)

1. digital modulation signal recognizing method comprises following process:

(1) feature parameter vector to be discriminated is input to the SOFM part of recognition network, and extracts effective characteristic parameters;

(2) according to the extraction result of SOFM network and the extraction order of validity feature, determine the characteristic parameter that adopts in each node of decision tree from top to bottom;

(3), constitute the network configuration of recognition system with each node in the RBF network replacement decision tree;

(4) utilize known training set, each RBF recognition network in the recognition system network is trained, make the identification error of training set reach minimum, each that determine each RBF recognition network hidden layer and output layer is connected weights and decision threshold, and fixes these recognition networks;

(5) signal to be identified that will receive is sampled, and sampled signal is input in the fixing recognition network, calculates the effective characteristic parameters value from top to bottom according to the structure of decision tree;

(6) the effective characteristic parameters value is input to its output valve of calculating in the RBF recognition network, compares, judge the modulation system of identification signal with this output valve and decision threshold.

2. digital modulation signal recognizing method according to claim 1, the wherein described extraction of carrying out effective characteristic parameters of step (1) is carried out according to the following procedure;

(2a) selected feature parameter vector is: F=[F ₁, F ₂... F _N], F wherein ₁, F ₂... F _NRepresent different characteristic parameters;

(2b) statistical nature parameter F ₁SOFM network output result's class categories;

(2c) number of the class categories that counts and the known prototype of training sample are compared, if statistics is consistent with known prototype, then keep this effective characteristic parameters, and with the effective characteristic parameters of selecting for use in the node of this characteristic parameter as decision tree, otherwise, F1 is judged to invalid characteristic parameter;

(2d) to characteristic parameter F ₂... F _NCarry out identical operations, finally selected 7 effective characteristic parameters are respectively: based on once with the characteristic parameter V of twice wavelet transformation _w, twice wavelet transformation characteristic parameter V ₂, signal box function characteristic parameter DBF1, signal transient amplitude box function characteristic parameter DBF2, instantaneous frequency characteristic parameter MAX_fre, based on the characteristic parameter σ of non-weak signal real part _IWith characteristic parameter F based on the FFT conversion _v

3. digital modulation signal recognizing method according to claim 1, wherein the described structure according to decision tree of step (5) is calculated the effective characteristic parameters value from top to bottom, carries out according to the following procedure:

(3a) calculate instantaneous frequency characteristic parameter MAX_fre;

(3b) calculate based on once with twice wavelet transformation characteristic parameter V _w

(3c) calculating is based on the characteristic parameter σ of non-weak signal real part _I

(3d) calculating is based on the characteristic parameter F of FFT conversion _v

(3e) the box function characteristic parameter DBF1 of signal calculated;

(3f) the box function characteristic parameter DBF2 of signal calculated instantaneous amplitude.

4. digital modulation signal recognizing method according to claim 3, wherein said calculating instantaneous frequency characteristic parameter MAX_fre, calculate according to the following procedure:

(4a) according to formula

The instantaneous phase of calculating sampling value, wherein, Q (i) and I (i) are respectively the imaginary part and the real part of n signal sampling value;

(4b) pass through formula

The instantaneous frequency of calculating sampling value, and to instantaneous frequency

Maximizing obtains instantaneous frequency characteristic parameter MAX_fre.

5. digital modulation signal recognizing method according to claim 3, wherein said calculating based on once with twice wavelet transformation characteristic parameter V _w, calculate according to the following procedure:

(5a) sampled value is carried out the Haar wavelet transformation one time, sampled value is carried out medium filtering, and calculate variance V ₁

(5b) sampled value is carried out the Haar wavelet transformation twice, sampled value is carried out medium filtering, and calculate variance V ₂

(5c) by formula V _w=[V ₁V ₂] obtain based on once with twice wavelet transformation characteristic parameter V _w

6. digital modulation signal recognizing method according to claim 3, the box function characteristic parameter DBF2 of wherein said signal calculated instantaneous amplitude, calculate as follows:

(6a) according to formula $b\left(\mathrm{\Δ}\right)=\underset{i=1}{\overset{N-1}{\mathrm{\Σ}}}|a_1-a_{i+1}|$ Calculate the absolute difference of neighbouring sample value instantaneous amplitude, wherein,

Instantaneous amplitude for sampled value;

(6b) pass through formula $d\left(2\mathrm{\Δ}\right)=\underset{i=1}{\overset{(N-1)/2}{\mathrm{\Σ}}}\left(\max \right\{a_{2i-1},a_{2i},a_{2i+1}\}-\min \{a_{2i-1},a_{2i},a_{2i+1}\}$ The comentropy of calculating sampling value;

(6c) by formula DBF2=1+log ₂(b (Δ)/d (2 Δ)) calculates the box function characteristic parameter DBF2 of signal, and wherein, b (Δ) is the absolute difference of neighbouring sample value instantaneous amplitude, and d (2 Δ) is the comentropy of sampled value.

Images