Fast modulation recognition method for multilayer perceptron based on multimodally-distributed test data fusion

Abstract

The present invention discloses a fast modulation recognition method for a multilayer perceptron (MLP) based on multimodally-distributed test data fusion. The method sequentially includes: preprocessing a received signal, obtaining a signal feature sequence, generating a matrix of decision statistics data o*hj−, generating an MLP an input feature by fusing the decision statistics data, recognizing a modulation mode by using the MLP, and matching an output with a corresponding classification label. The present invention has a low algorithm complexity as compared with a classical likelihood algorithm, and at the same time improves the recognition precision of a single distribution test algorithm.

Description

This application is the National Stage Application of PCT/CN2020/110119, filed on Aug. 20, 2020, which claims priority to Chinese Patent Application No. 201911077041.X, filed on Nov. 6, 2019, which is incorporated by reference for all purposes as if fully set forth herein.

FIELD OF THE INVENTION

The present invention relates to the field of communication, and specifically, to A fast modulation recognition method for a multilayer perceptron (MLP) based on multimodally-distributed test data fusion.

DESCRIPTION OF THE RELATED ART

Modulation recognition, also referred to as modulation classification, aims to obtain modulation mode information of unknown signals by analyzing the characteristics of the unknown signals. Modulation recognition has been widely applied to the military field, the civilian field, among other fields. In-depth researches have been carried out on a modulation recognition algorithm in a single-input single-output (SISO) system over 30 years. However, the application of the modulation recognition algorithm in a multiple-input multiple-output (MIMO) system is still in an early stage. The first field to which a modulation recognition algorithm is applied is the field of modulation recognition in the MIMO system. In a large-scale MIMO system, as a quantity of antennas increases, the complexity of most existing algorithms correspondingly grows exponentially. Especially, in a large-scale MIMO system, a modulation recognition algorithm that satisfies an actual application requirement is still a great challenge.

There are mainly two types of existing modulation recognition algorithms, that is, a likelihood-based algorithm and a feature-based algorithm. In the likelihood-based algorithm, optimal recognition precision can be obtained in a Gaussian white noise channel by using a statistical model of additive noise. However, the premise of the likelihood-based algorithm is that there are accurate noise power and distribution. During the use of an eigenvalue-based recognition mechanism, although this type of statistical features can adequately complete a modulation recognition task and obtain relatively desirable recognition precision, the mechanism lacks the capability of handling additive noise, and the anti-interference capability of the algorithm is improved by using information such as a signal-to-noise ratio (SNR).

In addition, existing modulation recognition algorithms have one common problem: The calculation costs of signal processing of the algorithms are very high, and the complexity of calculation grows exponentially as a quantity of antennas and a quantity of orders of a modulation mode increase. During the consideration of a high-order modulation mode and a large-scale MIMO system, a common likelihood-based modulation recognition method for a single antenna system has extremely high calculation complexity.

SUMMARY OF THE INVENTION

A technical problem to be resolved by the present invention is to provide a fast modulation recognition method for an MLP based on multimodally-distributed test data fusion, so that algorithm complexity can be reduced and recognition precision can be improved.

To resolve the foregoing technical problem, the present invention provides a fast modulation recognition method for an MLP based on multimodally-distributed test data fusion, including the following steps:

step 1) preprocessing a received signal, a used normalization formula being:

$r_k=\frac{\left(r_k\right)-\stackrel{\_}{\left(r\right)}}{\sigma \left(\left(r\right)\right)}+j\frac{\left(r_k\right)-\stackrel{\_}{\left(r\right)}}{\sigma \left(\left(r\right)\right)},$ and

obtaining a signal feature sequence {z_k}_k=1^N from the received signal {r_k}_k=1^N;

step 2) obtaining decision statistics data o*_hj− by using four distribution test algorithms: a Kolmogorov-Smirnov (KS) test, a Cramer-Von Mises (CVM) test, an Anderson-Darling (AD) test, and a variance (Var) test, o*_hj− being defined as:

$t_{\mathrm{mod}}^{\mathrm{KS}}=\underset{1\le n\le N}{\max }❘F_1\left(z_n\right)-F_0\left(z_n❘M\right)❘,t_{\mathrm{mod}}^{\mathrm{CVM}}={\int }_{-\infty }^{\infty }{\left[F_1\left(z_n\right)-F_0\left(z_n❘M\right)\right]}^2{\mathrm{dF}}_0\left(z_n\right),t_{\mathrm{mod}}^{\mathrm{AD}}={\int }_{-\infty }^{\infty }\frac{{\left[F_1\left(z_n\right)-F_0\left(z_n❘M\right)\right]}^2}{F_0\left(z_n❘M\right)\left(1-F_0\left(z_n❘M\right)\right)}{\mathrm{dF}}_0\left(z_n❘M\right),t_{\mathrm{mod}}^{\mathrm{Var}}=\frac{1}{N}\sum _{i=1}^N{\left(d_i-\mu \right)}^2,d_i=F_1\left(z_n\right)-F_0\left(z_n❘M\right),\mu =\frac{1}{N}\sum _{i=1}^N{d}_i,$

where F₁(z_n) is an empirical cumulative distribution of the received signal, F₀(z_n|M) is a theoretical cumulative distribution of a candidate modulation mode, and M is the candidate modulation mode;

step 3) obtaining a matrix of o*_hj−:

$\left(\begin{array}{cccc}{t}_{\mathrm{mod}1}^{\mathrm{KS}}& t_{\mathrm{mod}2}^{\mathrm{KS}}& \cdots & t_{\mathrm{mod}n}^{\mathrm{KS}}\\ t_{\mathrm{mod}1}^{\mathrm{CVM}}& t_{\mathrm{mod}2}^{\mathrm{CVM}}& \cdots & t_{\mathrm{mod}n}^{\mathrm{CVM}}\\ t_{\mathrm{mod}1}^{\mathrm{AD}}& t_{\mathrm{mod}2}^{\mathrm{AD}}& \cdots & t_{\mathrm{mod}n}^{\mathrm{AD}}\\ t_{\mathrm{mod}1}^{\mathrm{Var}}& t_{\mathrm{mod}2}^{\mathrm{Var}}& \cdots & t_{\mathrm{mod}n}^{\mathrm{Var}}\end{array}\right);$

step 4) generating an input feature of an MLP classifier by using the matrix, the input feature being defined as:t*_mod=t_{mod i}^KS+t_{mod i}^CVM+t_{mod i}^AD+t_{mod i}^Var, i=1,2, . . . ,n where

an input feature of each modulation mode is inputted into the MLP classifier; and step 5) obtaining the input features of different modulation modes after the fusion of decision statistics data, recognizing the modulation modes by the MLP classifier with the input features, and matching an output with a corresponding classification label:label={0,1,2, . . . ,n−1}

Further, {z_k}_k=1^N is a real part, an imaginary part, an amplitude or a phase taken from a received complex signal.

Further, the step of generating the matrix of the decision statistics data o*_hj− is as follows: first, calculating the empirical cumulative distribution F₁(z_n) of the received signal, next, calculating the theoretical cumulative distribution F₀(z_n|M) of the candidate modulation mode, and finally, calculating the decision statistics data o*_hj− in different modulation modes based on the four distribution test algorithms.

Further, data fusion is performed on the matrix of o*_hj− in a column addition manner.

Further, the MLP classifier uses a forward propagation technology and a backward propagation technology to train a model and update weights.

Further, recognition precision is calculated after the training ends, and the recognition precision

$P_{\mathrm{acc}}=\frac{N_c}{N_{\mathrm{total}}}\times 100\%$ is calculated for a classification result of the MLP classifier, where N_c is a quantity of accurately classified samples, and N_total is a total quantity of test signal samples.

Beneficial effects of the present invention are as follows:

1. In a MIMO system, compared with a maximum likelihood (ML) modulation recognition algorithm, the proposed modulation recognition algorithm may obtain lower algorithm complexity.

2. For modulation modes such as BPSK, 8PSK, 4QAM, and 16QAM, recognition precision not less than 90% is implemented in the case of an SNR greater than 10 dB, and a recognition time is only related to a quantity of signal samples.

3. A modulation recognition algorithm proposed for a scenario of a fading channel has very high robustness.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a flowchart of a DTE (Distribution Test Ensemble) algorithm according to the present invention.

FIG. 2 is a diagram of experimental simulation results in a Gaussian noise channel of a DTE algorithm according to the present invention.

FIG. 3 is a diagram of experimental simulation results in a phase shift channel of a DTE algorithm according to the present invention.

FIG. 4 is a diagram of experimental simulation results in a frequency shift channel of a DTE algorithm according to the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

The present invention is further described below with reference to the accompanying drawings and specific embodiments, to enable a person skilled in the art to better understand and implement the present invention. However, the embodiments are not used to limit the present invention.

Referring to FIG. 1, an embodiment of a fast modulation recognition method for a MIMO communication system based on multimodally-distributed test data fusion of the present invention is provided. The method sequentially includes steps of preprocessing a received signal, obtaining a signal feature sequence, generating a matrix of decision statistics data o*_hj−, generating an input feature of an MLP classifier and an output result of the MLP classifier, where the MLP classifier requires learning, and calculating recognition precision.

In the present invention, an experimental simulation stage includes four modulation signals M={BPSK,8-PSK,4-QAM,16-QAM}. Amplitude components of the received signal are used. The length (N) of a signal sample is 128. A quantity (N_T) of transmit antennas is 2. A quantity (N_R) of receive antennas is 4.

The invention has good recognition performance in different channel conditions. In a Gaussian noise channel, when an SNR is greater than 16 dB, the method in the present invention may reach a recognition accuracy of 90%. In a fading channel, the method in the present invention has robustness in both a frequency shift case and a phase shift case, and the recognition rate is nearly not affected.

Specifically, in the step of preprocessing the received signal, the received signal is first preprocessed before the recognition of a modulation mode. In a preprocessing manner, the distribution of real parts and imaginary parts of imaginary signals are normalized. A normalization formula is:

$r_k=\frac{\left(r_k\right)-\stackrel{\_}{\left(r\right)}}{\sigma \left(\left(r\right)\right)}+j\frac{\left(r_k\right)-\stackrel{\_}{\left(r\right)}}{\sigma \left(\left(r\right)\right)}.$

In the step of obtaining the signal feature sequence, because the real part component, imaginary part component, amplitude, and phase of a signal all carry information of the signal to different degrees, the signal feature sequence {z_k}_k=1^N of a distribution test can be extracted from the real part component, imaginary part component, amplitude, and phase. In the present invention, an amplitude component z_k=|r_k|=√{square root over ({r_k})²+({r_k})²)} of the signal is selected in an experimental simulation stage.

In the step of generating the matrix of the decision statistics data o*_hj− is as follows: An empirical cumulative distribution F₁(z_n) of the received signal is first calculated. Next, a theoretical cumulative distribution F₀(z_n|M) of a candidate modulation mode is calculated. The decision statistics data o*_hj− in different modulation modes is then calculated based on distribution test algorithms: a KS test, a CVM test, an AD test, and a Var test, to construct a matrix of the decision statistics data:

$\left(\begin{array}{cccc}{t}_{\mathrm{mod}1}^{\mathrm{KS}}& t_{\mathrm{mod}2}^{\mathrm{KS}}& \cdots & t_{\mathrm{mod}n}^{\mathrm{KS}}\\ t_{\mathrm{mod}1}^{\mathrm{CVM}}& t_{\mathrm{mod}2}^{\mathrm{CVM}}& \cdots & t_{\mathrm{mod}n}^{\mathrm{CVM}}\\ t_{\mathrm{mod}1}^{\mathrm{AD}}& t_{\mathrm{mod}2}^{\mathrm{AD}}& \cdots & t_{\mathrm{mod}n}^{\mathrm{AD}}\\ t_{\mathrm{mod}1}^{\mathrm{Var}}& t_{\mathrm{mod}2}^{\mathrm{Var}}& \cdots & t_{\mathrm{mod}n}^{\mathrm{Var}}\end{array}\right).$

In the step of generating the input feature of the MLP classifier, for a 4×n matrix of the decision statistics data obtained in the foregoing step, in the present invention, a column addition manner is used for data fusion,

to obtain n t*_mod=t_{mod i}^KS+t_{mod i}^CVM+t_{mod i}^AD+t_mod^Var, i=1, 2, . . . , n.

An input feature of each modulation mode is inputted into the MLP classifier. Finally, the modulation mode is recognized by using the MLP classifier, and an output is matched with a corresponding classification label:label={0,1,2, . . . ,n−1}

Before formal use, the MLP classifier requires learning. A forward propagation technology and backward propagation technology are used to train a model and update weights. In the present invention, a quasi-Newton method is used, an activation function is set to an inverse trigonometric function (tanh), the quantity of layers of a neural network is 3, the quantity of hidden neurons is 8, there are both 1000 training samples and 1000 test samples, and the maximum quantity of iterations is 300 epochs.

In the step of calculating the recognition precision, the recognition precision

$P_{\mathrm{acc}}=\frac{N_c}{N_{\mathrm{total}}}\times 100\%$ is calculated for a classification result of the MLP classifier, where N_c is a quantity of accurately classified samples, and N_total is a total quantity of test signal samples.

In the method in the present invention, it is fully considered that in the case of a high-order modulation mode and a large-scale MIMO system, a common likelihood-based modulation recognition method for a single antenna system has very high calculation complexity. To resolve the problem of calculation complexity, the method in the present invention uses a distribution test-based algorithm to resolve a modulation recognition problem in a MIMO environment.

FIG. 2 shows the comparison of recognition performance of different recognition algorithms in different SNRs. The algorithm in the present invention may be referred to as DTE for short, and is better than a classical distribution test algorithm in the entire SNR. Particularly, at a high SNR, when SNR=16 dB, the recognition rate is higher than that of an AD test algorithm by 2%.

FIG. 3 and FIG. 4 show the robustness of the algorithm in the present invention in different fading channel conditions. It is very clear that an ML algorithm is highly susceptible to the impact of a frequency shift and a phase shift, the shift degree is excessively large, and the recognition performance of the ML algorithm is greatly reduced.

As shown in Table 1:

Classifier	Addition	Multiplication	Logarithm	Exponent
ML	$6{\mathrm{NM}}^{N_T}{N}_R·{\sum }_{m=1}^{M^{N_T}}{I}_m$	$5{\mathrm{NM}}^{N_T}{N}_R·{\sum }_{m=1}^{M^{N_T}}{I}_m$	NMNTNR	${\mathrm{NM}}^{N_T}{N}_R·{\sum }_{m=1}^{M^{N_T}}{I}_m$
K-S test	MNR(log2 N + 2N)	0	0	0
C-v-M	MNR(log2 N + 3N)	NMNR	0	0
test
A-D test	MNR(log2 N + 3N)	2NMNR	0	0
Var test	MNR(log2 N + N)	0	0	0
DTE	MNR(4log2 N + 9N)	3NMNR	0	0
MLP	${\sum }_{k=2}^{{}^{{}_l}}{n}_{k-1}{n}_k$	${\sum }_{k=2}^{{}^{{}_l}}{n}_{k-1}{n}_k$	0	0

The comparison of calculation complexity of different algorithms is given. A classical ML algorithm includes exponential operation, and as a result the complexity of the algorithm is greatly increased. However, the algorithm in the present invention only involves addition and multiplication and has clear advantages.

The foregoing embodiments are merely preferred embodiments used to fully describe the present invention, and the protection scope of the present invention is not limited thereto. Equivalent replacements or variations made by a person skilled in the art to the present invention all fall within the protection scope of the present invention. The protection scope of the present invention is as defined in the claims.

Claims

1. A fast modulation recognition method for a multilayer perceptron based on multimodally-distributed test data fusion, comprising the following steps: step 1) preprocessing a received signal, a used normalization formula being:

2. The fast modulation recognition method for a multilayer perceptron based on multimodally-distributed test data fusion according to claim 1, wherein {zk}k=1N is a real part, an imaginary part, an amplitude or a phase taken from a received complex signal.

3. The fast modulation recognition method for a multilayer perceptron based on multimodally-distributed test data fusion according to claim 1, wherein the step of generating the matrix of the decision statistics data t*mod is as follows: first, calculating the empirical cumulative distribution F1(zn) of the received signal, next, calculating the theoretical cumulative distribution F0(zn|M) of the candidate modulation mode, and finally, calculating the decision statistics data t*mod in different modulation modes based on the four distribution test algorithms.

4. The fast modulation recognition method for a multilayer perceptron based on multimodally-distributed test data fusion according to claim 1, wherein data fusion is performed on the matrix of t*mod in a column addition manner.

5. The fast modulation recognition method for a multilayer perceptron based on multimodally-distributed test data fusion according to claim 1, wherein the MLP classifier uses a forward propagation technology and a backward propagation technology to train a model and update weights.

6. The fast modulation recognition method for a multilayer perceptron based on multimodally-distributed test data fusion according to claim 5, wherein recognition precision is calculated after the training ends, and the recognition precision