CN113518050B

CN113518050B - Modulation identification method, system, readable storage medium and device

Info

Publication number: CN113518050B
Application number: CN202110700954.3A
Authority: CN
Inventors: 赵军辉; 秦子杰; 马小婷
Original assignee: East China Jiaotong University
Current assignee: East China Jiaotong University
Priority date: 2021-06-24
Filing date: 2021-06-24
Publication date: 2021-12-24
Anticipated expiration: 2041-06-24
Also published as: CN113518050A

Abstract

The invention provides a modulation identification method, a modulation identification system, a readable storage medium and a device, wherein the method comprises the following steps: acquiring a received signal data set; performing cluster analysis on the received signals according to characteristic parameters carried by the received signals so as to divide a received signal data set into a plurality of clusters; and when all the received signals in the cluster contain more than two modulation modes, adopting a convolutional neural network-based supervised learning method to perform supervised identification on the received signals in the cluster. According to the invention, through unsupervised learning cluster analysis, one part of modulation modes with obvious distinguishing characteristics can be directly identified, and the other part of modulation modes which are easy to be confused are classified into the same cluster due to extremely high similarity, and further supervised identification is carried out by adopting convolutional neural network-based supervised learning aiming at the clusters containing two or more modulation types, so that the problem of classification nonuniformity existing in the existing modulation identification method can be solved.

Description

Modulation identification method, system, readable storage medium and device

Technical Field

The present invention relates to the field of communications technologies, and in particular, to a modulation identification method, a modulation identification system, a readable storage medium, and a device.

Background

Automatic Modulation Classification (AMC) refers to a technique for determining a Modulation scheme of an unknown signal by analyzing electromagnetic characteristics, spectral characteristics, statistical characteristics, and the like of a transmitted signal. In the field of military communication, modulation parameters are estimated by identifying the modulation mode of an intercepted signal, the signal is demodulated, analyzed and processed, radio interference or electromagnetic interference is applied to the signal more pertinently, and the wireless communication of the opposite side is blocked or even paralyzed. In the field of civil communication, the modulation identification technology provides important prior information for the realization of emerging technologies such as Cognitive Radio (CR), Software Defined Radio (SDR) and the like, thereby effectively improving the user capacity and Cognitive skills of the system. Therefore, the automatic modulation recognition technology has become a key technology for the development of wireless communication efficiency and intelligence.

At present, many research results are obtained for the signal modulation identification of a single simple channel model, but with the diversification of modulation modes and the influence of complex and variable signal transmission environments, the signal modulation identification method of the single simple channel model is not suitable for the complex environments of various different channels. For this reason, new modulation recognition algorithms need to be developed to meet the current demand.

In recent years, a modulation recognition algorithm based on machine learning has been widely applied to the solution of the modulation recognition problem, and exhibits superior recognition performance under a variety of different channel environment conditions. However, there is a serious drawback in the method of modulation recognition using supervised learning, i.e. classification non-uniformity, i.e. difficulty in distinguishing different classes is different. For example, in the conventional image classification problem, "person" and "car" are easy to distinguish because the image shapes are different greatly, but "dog" and "sheep" have higher visual similarity and are difficult to distinguish. The same problem exists with modulation identification.

As the number of modulation schemes increases, the similarity between modulation schemes also varies greatly. However, at present, only one neural network is often used to distinguish all modulation modes in a data set, and optimization of a network model is realized by reducing an average value of a loss function, so that a high average recognition accuracy is achieved, so that one part of modulation modes can be well classified by using the model, and the other part of modulation modes are easily confused with each other, thereby causing classification nonuniformity. Therefore, how to solve the classification nonuniformity is the key point for improving the modulation identification precision.

Disclosure of Invention

Based on this, the present invention provides a modulation recognition method, a modulation recognition system, a readable storage medium and a device, so as to solve the technical problem of the existing modulation recognition classification nonuniformity.

According to the embodiment of the invention, the modulation identification method comprises the following steps:

acquiring a data set of a received signal, wherein the received signal carries characteristic parameters for representing a modulation mode;

performing cluster analysis on the received signal data according to the characteristic parameters carried by the received signals, dividing the modulation signals with the discrimination degree greater than a preset value, and aggregating the modulation signals with the discrimination degree less than the preset value into a cluster;

judging whether the obtained cluster contains two or more modulation signals;

if so, performing supervised identification on the received signals in the cluster by adopting a supervised learning method based on a Convolutional Neural Network (CNN) to identify a modulation mode of the received signals in the mixed cluster;

and if not, taking the modulation mode corresponding to the cluster as the identification result of each received signal in the cluster.

In addition, a modulation identification method according to the above embodiment of the present invention may further have the following additional technical features:

further, performing cluster analysis on the received signal data according to the characteristic parameters carried by the received signals, and dividing the modulation signals with the discrimination degree greater than a preset value, wherein the step of aggregating the modulation signals with the discrimination degree less than the preset value into a cluster comprises:

dividing the received signal data set into a plurality of completely disjoint received signal subsets by adopting a hyperplane division method according to the characteristic parameters carried by the received signals;

and according to the characteristic parameters carried by the received signals, performing cluster analysis of a modulation mode on each received signal subset respectively so as to divide the received signal subsets into a plurality of clusters.

Further, after the step of performing a cluster analysis of a modulation scheme on each received signal subset according to the characteristic parameters carried by the received signals, so as to divide the received signal subsets into a plurality of clusters, the method further includes:

judging whether the sum of mean square errors obtained by clustering is smaller than a threshold value;

and if not, performing clustering analysis again on the clustering center of each cluster obtained by clustering before until the sum of mean square errors obtained by clustering is smaller than the threshold value.

Further, the total number of clusters is 6.

Further, the cluster analysis is fuzzy clustering.

Further, after the step of acquiring a received signal data set, the method further comprises:

and adopting an unsupervised sparse autoencoder to perform dimensionality reduction on the received signals in the received signal data set.

Further, the step of performing supervised identification on the received signals in the cluster by using a convolutional neural network-based supervised learning method includes:

and taking the original received data of the received signals and the data subjected to the dimensionality reduction as input data, and carrying out supervised identification on the received signals in the cluster by adopting a supervised learning method based on a convolutional neural network.

A modulation identification system according to an embodiment of the present invention includes:

the data acquisition module is used for acquiring a data set of a received signal, wherein the received signal carries characteristic parameters for representing a modulation mode;

the cluster analysis module is used for carrying out cluster analysis on the received signal data according to the characteristic parameters carried by the received signals, dividing the modulation signals with the discrimination degree larger than a preset value, and aggregating the modulation signals with the discrimination degree smaller than the preset value into a cluster;

the result judging module is used for judging whether the obtained cluster contains two or more modulation signals;

the modulation identification module is used for carrying out supervised identification on the received signals in the clusters by adopting a supervised learning method based on a convolutional neural network when the obtained clusters contain two or more modulation signals so as to identify the modulation mode of the received signals in the mixed clusters; and when judging that all the received signals in the cluster only belong to the same modulation mode, taking the modulation mode corresponding to the cluster as the identification result of each received signal in the cluster.

The invention also proposes a computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, implements the modulation identification method described above.

The invention also provides a modulation identification device, which comprises a memory, a processor and a computer program stored on the memory and capable of running on the processor, wherein the processor executes the program to realize the modulation identification method.

The invention has the beneficial effects that: firstly, through unsupervised learning cluster analysis, a part of modulation modes with obvious distinguishing characteristics can be directly identified, and the other part of modulation modes which are easy to be confused can be classified into the same cluster due to extremely high similarity. Aiming at clusters containing two or more modulation types, further supervised recognition is carried out by adopting CNN supervised learning, so that the problem of classification nonuniformity of multiple modulation method recognition by adopting the same network structure in the prior art can be solved, the CNN can carry out special training and optimization aiming at the modulation modes which are easy to be confused, and the recognition rate of a single modulation mode is ensured to be optimal while the integral average recognition rate is higher. Meanwhile, by adopting a method of clustering and then classifying, the repeated training overhead of the neural network for easily classified data can be reduced, so that the neural network is more focused on solving difficult problems, and the training efficiency is improved.

Drawings

Fig. 1 is a flowchart of a modulation identification method in a first embodiment of the present invention;

FIG. 2 is a flow chart of a modulation identification method according to a second embodiment of the present invention;

FIG. 3 is a block diagram of a semi-supervised identification algorithm provided by an embodiment of the present invention;

FIG. 4 is a graph illustrating the impact of different cluster numbers and fuzzy coefficients on the FCM performance according to an embodiment of the present invention;

FIG. 5 is a diagram illustrating the result of HD-FCM clustering according to an embodiment of the present invention;

fig. 6 is a diagram of recognition performance of CNN in supervised classification according to an embodiment of the present invention;

FIG. 7 is a comparison diagram of a semi-supervised identification algorithm and a supervised identification method according to an embodiment of the present invention;

fig. 8 is a block diagram showing a modulation recognition system in a third embodiment of the present invention;

the following detailed description will further illustrate the invention in conjunction with the above-described figures.

Detailed Description

To facilitate an understanding of the invention, the invention will now be described more fully with reference to the accompanying drawings. Several embodiments of the invention are presented in the drawings. This invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

It will be understood that when an element is referred to as being "secured to" another element, it can be directly on the other element or intervening elements may also be present. When an element is referred to as being "connected" to another element, it can be directly connected to the other element or intervening elements may also be present. The terms "vertical," "horizontal," "left," "right," and the like as used herein are for illustrative purposes only.

Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. The terminology used in the description of the invention herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. As used herein, the term "and/or" includes any and all combinations of one or more of the associated listed items.

At present, all modulation modes in a data set are often distinguished by only one neural network, optimization of a network model is realized by reducing the average value of a loss function, and high average identification accuracy is achieved, so that one part of modulation modes can be well classified by using the model, and the other part of modulation modes are easily confused with each other, and thus classification unevenness is caused. Therefore, the invention aims to provide a modulation identification method, a system, a readable storage medium and equipment, so as to provide a modulation identification method based on a semi-supervised hierarchical structure by combining an unsupervised clustering method and a convolutional neural network. The general process is as follows: first, an unsupervised Sparse Auto Encoder (SAE) is used to perform dimensionality reduction on the received signal. Then, aiming at the characteristic of huge data volume of a modulation identification task, a Hyperplane division-Fuzzy C-means clustering method (HD-FCM) is provided, one part of modulation modes with obvious distinguishing characteristics can be directly identified through the HD-FCM, and the other part of modulation modes which are easy to be confused with each other are classified into the same cluster due to extremely high similarity. And finally, performing further supervision and identification by adopting CNN supervision and learning aiming at the clusters containing two or more modulation types. The method has the advantages that the problem of classification nonuniformity of a plurality of modulation method identification by adopting the same network structure in the prior art is solved, so that the CNN can carry out special training and optimization aiming at the confusable modulation method, and the identification rate of a single modulation method is ensured to be optimal while the integral average identification rate is higher.

The embodiments will be described in detail with reference to the following specific examples.

Example one

Referring to fig. 1, a modulation identification method according to a first embodiment of the present invention is shown, where the modulation identification method can be implemented by software and/or hardware, and the method includes steps S01 to S05.

Step S01, a data set of a received signal is obtained, where the received signal carries a characteristic parameter for characterizing a modulation mode.

Wherein the received signal data set comprises N received signals, the N received signals comprise M modulation modes,M ≤ N. In addition, the characteristic parameter carried by the received signal may be, but is not limited to, an electromagnetic characteristic, a spectral characteristic, a statistical characteristic, and the like.

Step S02, performing cluster analysis on the received signal data according to the characteristic parameters carried by the received signals, and dividing the modulated signals with the discrimination degree greater than a preset value, and aggregating the modulated signals with the discrimination degree less than the preset value into a cluster.

The discrimination degree is greater than the preset value, which means that the discrimination degree is higher, namely, the modulation signals can be well discriminated; on the contrary, the discrimination smaller than the preset value means that the discrimination is smaller and belongs to the same or similar modulation signals. In specific implementation, the received signal data may be subjected to cluster analysis according to the principles of "minimizing inter-class similarity and maximizing intra-class similarity", so as to divide the modulated signals with higher discrimination, and the same or similar modulated signals are gathered into a cluster.

It should be noted that, because different characteristic parameters correspond to different signal modulation modes, the received signals may be subjected to cluster analysis of the modulation modes according to the characteristic parameters carried by the received signals, that is, each received signal is clustered according to the modulation mode predicted by the characteristic parameter, so as to cluster the received signals of the same or similar modulation modes in the same cluster.

Specifically, in the clustering process, a proper objective function can be established, and the first coarse classification of the modulation signals is realized by continuously adjusting the membership degree of the samples and the central point of the cluster. Clustering enables samples in the same cluster to have extremely high similarity, and samples in different clusters have high difference. If one cluster only contains one modulation mode, the modulation mode can be better distinguished from other modulation methods only through unsupervised learning; if a cluster is composed of two or more modulation signals, the modulation modes contained in the cluster have extremely high similarity, which is the difficulty of supervised classification.

For example, the QAM16 and the QAM64 both adopt amplitude-phase modulation methods, and are different only in that the orders of the two modulation methods are different, and constellation poles of the two modulation methods are similar under a lower signal-to-noise ratio, and thus the two modulation methods are divided into the same cluster during cluster analysis; in contrast, QAM16 differs from FSK using frequency modulation in modulation scheme and modulation order, and can be distinguished more easily, and thus will be divided into different clusters during cluster analysis.

In step S03, it is determined whether the obtained cluster includes two or more modulation signals.

Wherein, when it is determined that the obtained cluster includes two or more modulation signals, step S04 is executed; if it is determined that the obtained cluster does not include two or more modulation signals, i.e., all received signals in the cluster belong to the same modulation scheme, step S05 is executed.

Step S04, performing supervised identification on the received signal in the cluster by using a supervised learning method based on a convolutional neural network, so as to identify a modulation mode of the received signal in the mixed cluster.

In step S05, the modulation scheme corresponding to the cluster is used as the identification result of each received signal in the cluster.

For example, when the modulation modes of the received signals in the cluster are FSK, the modulation mode corresponding to the cluster is FSK, and the system determines that the modulation mode of the received signals in the cluster is FSK.

In summary, in the modulation identification method in the above embodiments of the present invention, through unsupervised learning cluster analysis, some modulation modes with obvious distinguishing features can be directly identified, and other modulation modes that are easily confused with each other are classified into the same cluster due to the extremely high similarity. Aiming at clusters containing two or more modulation types, further supervised recognition is carried out by adopting CNN supervised learning, so that the problem of classification nonuniformity of multiple modulation method recognition by adopting the same network structure in the prior art can be solved, the CNN can carry out special training and optimization aiming at the modulation modes which are easy to be confused, and the recognition rate of a single modulation mode is ensured to be optimal while the integral average recognition rate is higher. Meanwhile, by adopting a method of clustering and then classifying, the repeated training overhead of the neural network for easily classified data can be reduced, so that the neural network is more focused on solving difficult problems, and the training efficiency is improved.

Example two

Referring to fig. 2, a modulation identification method according to a second embodiment of the present invention is shown, and the modulation identification method can be implemented by software and/or hardware, and includes steps S11 to S17.

Step S11, a data set of a received signal is obtained, where the received signal carries a characteristic parameter for characterizing a modulation mode.

Step S12, perform dimensionality reduction on the received signal in the received signal data set by using an unsupervised sparse autoencoder.

The sparse autoencoder is an artificial neural network for acquiring deep feature representation of input data through unsupervised learning. The constraint condition of the self-coding network is that the input is a data sample, the input which is continuously close to the network is output, the neural network is continuously trained according to the principle, and after the loss function of the network is converged, the neuron value of the middle hidden layer is the deep feature of the sample data, so that the deep feature is extracted, and the dimension reduction of the input data is realized.

Step S13, according to the characteristic parameters carried by the received signals, the received signal data set is divided into a plurality of completely disjoint received signal subsets by using a hyperplane division method.

Assume that the received signal data set is represented asX={x ₁ ,x ₂ ,…,x _NTherein ofNUsually on the order of millions, representing the number of samples involved, the firstkA signal samplex _k ={ x ₁ ,x ₂ ,…,x _n}. The principle of hyperplane division is to use the data setXLocal information ofXIs divided intoPA subset of.

In a one-dimensional dataset, the way to partition disjoint subsets is often to sort the data samples incrementally, and partition the data by selecting a threshold. This method is also applicable to high dimensional data. Assume for dimension size ofnReceived signal ofx _kIf it is divided intoPClass data set, then a threshold existso ₁ ,o ₂ ,…,o _p-1And is ando ₁≤ o ₂≤…≤ o _p-1. Therefore, it is necessary tox _kMapping to one-dimensional data pointsy _k：

（1）

Whereinx _kj ,j=1,2,…,nRepresenting received signalsx _kTo (1) ajThe value of the characteristic is used as the characteristic value,φrepresenting a functional mapping such that the data pointsy _kThe partitioning into disjoint subsets may be based on a threshold, e.g., a weighted sum or a weighted sum of squares. For the sake of simplicity, hereφIt is selected as a weighted sum of the sum,y _kcan be expressed as:

（2）

thus, by function mapping and threshold selection, a set of numbers ofP -1 ofn-a 1-dimensional hyperplane forNOf dimension ofnOf received signal samplesPA subset of classes. At the same time, it can be based on the threshold valueoThe same number of data points are assigned to each sub data set. For example, in a three-dimensional coordinate system, a planez=-x–y+50 may divide a sample set containing 100 data points into two types of subsets. For modulation identification tasks, the dimensions of the samples of the received signalnThe number of the data samples is far larger than 3, and a large number of the data samples can be divided into subsets which are not intersected with each other and are equal in number by adopting a hyperplane segmentation method.

Step S14, according to the characteristic parameters carried by the received signals, FCM fuzzy clustering analysis of the modulation mode is performed on each received signal subset, and the modulated signals with the discrimination degree greater than a preset value are divided, and the modulated signals with the discrimination degree less than the preset value are aggregated into a cluster, so as to divide the received signal subsets into a plurality of clusters.

Compared with a hard clustering method, fuzzy clustering can consider the fuzzy relation among samples, a fuzzy matrix is constructed according to the attribute of a research object, and the clustering relation is determined through the membership degree. FCM clustering can be expressed as:

（3）

wherein the content of the first and second substances,UandVrespectively representing a membership matrix and a clustering center matrix,μ _ijis shown asjA sample relative to the firstiThe degree of membership of the individual cluster centers,x _jis shown asjThe number of the sample points is one,v _iis shown asiThe center of each cluster is determined by the center of each cluster,mindicating the blur coefficient, which is often 2.2. Solved by Lagrange

（4）

（5）

In some optional embodiments of this embodiment, after the step of performing FCM fuzzy cluster analysis of a modulation scheme on each received signal subset according to the characteristic parameters carried by the received signals, so as to divide the received signal subsets into a plurality of clusters, the method further includes:

after the step of performing a cluster analysis of a modulation scheme on each received signal subset according to the characteristic parameters carried by the received signals, so as to divide the received signal subsets into a plurality of clusters, the method further includes:

and if not, performing clustering analysis again on the clustering center of each cluster obtained by clustering before until the sum of mean square errors obtained by clustering is smaller than the threshold value. Therefore, the HD-FCM clustering process can be divided into three steps:

step 1: partitioning sub-dataSet, dividing the whole data set into super-planePSub-data sets of completely disjoint classes, each having a number of samplesN/P。

Step 2: performing FCM fuzzy clustering on the sub-data sets, and dividing the obtained samples in each sub-data set into sub-data sets by adopting an FCM fuzzy clustering methoddCluster, co-obtainedPdAnd (4) clustering centers.

And step 3: FCM clustering is carried out on the known clustering centers, the clustering center of each cluster obtained before is clustered into a c cluster again through FCM, and c is the preset threshold value to obtain a further clustering centerV[i]=[v _ij]_{c n×}Receiving signal samplesx _kAlso based on previous membership matrices

Dividing the data into different clusters to obtain the final clustering result of the whole data setX[1], X[2],…, X[c]。

In addition, the final clustering number in the FCM clustering algorithmcAnd coefficient of blurmAre determined in advance. Therefore, firstly, partial samples are extracted from the whole data set for fuzzy clustering, different cluster numbers and fuzzy coefficients are studied, and accordingly better parameter selection is determined. We extract 1100 signal samples for fuzzy clustering, select the sum of squares of errorsJ _FCMAs a measure, the results are shown in FIG. 4.

As can be seen from FIG. 4, the three curves correspond to the blur coefficients respectivelymThe three values are commonly used, and it can be seen thatm=2.2, the sum of squared errors is minimal at different cluster numbers, and therefore,m=2.2 is a better choice. Second, with cluster numbercThe division of the signal samples is more and more fine, the aggregation degree of the samples in each cluster is gradually increased, and the sum of squares of errorsJ _FCMAnd also gradually decreases. When clustering cluster number c<At the time of 6, the reaction kettle is,cthe increase in (b) will greatly increase the degree of aggregation of the sample,J _FCMamplitude of descentDegree is also very large when c>At 6 hours, increase againcThe resulting aggregate return is rapidly diminishing,J _FCMthe magnitude of the decrease is reduced therewith, and finally, the magnitude of the decrease is increased withcGradually becomes gentle. Therefore, the most suitable cluster number for the sample set is 6.

In step S15, it is determined whether the obtained cluster includes two or more modulation signals.

Wherein, when it is determined that the obtained cluster includes two or more modulation signals, step S16 is executed; when it is determined that the obtained cluster does not include two or more modulation signals, that is, all the received signals in the cluster belong to the same modulation scheme, step S17 is executed.

Step S16, using the original received data of the received signal and the data after the dimensionality reduction as input data, and performing supervised identification on the received signal in the cluster by using a supervised learning method based on a convolutional neural network, so as to identify a modulation mode of the received signal in the mixed cluster.

In step S17, the modulation scheme corresponding to the cluster is used as the identification result of each received signal in the cluster.

Specifically, in combination with the HD-FCM algorithm, the structure of the semi-supervised identification algorithm proposed in this embodiment is shown in fig. 3, and can be divided into two parts:

(1) dimensionality reduction and clustering based on unsupervised learning: firstly, the unsupervised SAE is adopted to carry out dimensionality reduction processing on an original receiving signal, deep layer feature extraction is carried out, and subsequent clustering and classification are facilitated. And then, taking the sample characteristics obtained through SAE as input information of the HD-FCM algorithm, and clustering the received signal samples. In the clustering process, a proper target function is established, and the first coarse classification of the modulation signals is realized by continuously adjusting the membership degree of the samples and the central point of the cluster. Clustering enables samples in the same cluster to have extremely high similarity, and samples in different clusters have high difference. If one cluster only contains one modulation mode, the modulation mode can be better distinguished from other modulation methods only through unsupervised learning; if a cluster is composed of two or more modulation signals, it indicates that the modulation modes contained in the cluster have extremely high similarity, which is the difficult point of supervised classification.

(2) Classification recognition based on supervised learning: and performing further supervised classification identification by using the CNN aiming at the clusters containing various modulation signals. Note that the input data of CNNX _CNNFrom the original received signalX _SAEAnd output characteristics from SAE

The output of the CNN is the predictive tag of the received signal. And finally, combining the recognition results of unsupervised clustering and supervised classification to realize the classification recognition function of the whole model.

The embodiment also has the following beneficial effects on the basis of the first embodiment:

(1) the SAE is adopted to process the original data sample, so that the received signal can be reduced to a proper dimension, the important characteristics in the sample can be learned, and the characteristic preprocessing is carried out for the subsequent supervision and classification.

(2) By adopting the unsupervised clustering method, under the condition of not being restricted by the supervision information, the data rules learned can lead the sample data to be gathered into the same cluster due to the similarity, and the differences are mutually distinguished.

(3) Aiming at the clustering process with huge data volume, a hyperplane division method is provided, and a complete data set is divided into a plurality of completely disjoint subsets, so that the subsequent FCM clustering method can accurately focus on a local area of a data space, and the clustering efficiency is greatly improved.

(4) The shallow feature and the deep feature are combined to serve as the input feature of supervised learning, on one hand, original sample information is completely reserved, the feature globalization is realized, on the other hand, a data structure after SAE dimension reduction is effectively utilized, and the feature refinement is realized. The method and the device have the advantages that the recognition accuracy of the whole model is improved better.

For example, the following steps are carried out: modulation identification based on semi-supervised architecture proposed for evaluationThe radio receive signal data set RadioML 2016.10a is used as the receive signal data set. The data set comprises 11 modulation methods commonly used for wireless communication, wherein the modulation methods comprise 8 digital modulation modes, namely BPSK, QPSK, 8PSK, QAM16, QAM64, GFSK, CPFSK and PAM4, and three analog modulation modes, namely WBFM, AM-SSB and AM-DSB. The received signal is at a sampling rate of 1kHzNCollection in 128 samples. In order to simulate a real communication scene, the data set respectively comprises a central frequency offset, a sampling rate offset, additive white gaussian noise, multipath and a random process of fading. The signal-to-noise ratio is-18 dB to 20dB, and each 2dB step size comprises 20 signal-to-noise ratios. The number of samples contained in the entire dataset is 220000. The data set was partitioned in the experiment with 110000 samples for training and 110000 samples for testing.

Firstly, in the unsupervised dimension reduction stage, the signal samples in the data set are sampled because the input layer of the sparse self-encoder SAE is one-dimensionalxReal part ofR(x) And imaginary partI(x) The splicing is performed, which can be expressed as:

x _SAE =( R(x), I(x)) （10）

since the sample length in the dataset is 128, then

And is the input data to the SAE network. In this experiment, the SAE network used contains two hidden layers, with 128 and 16 neurons, respectively. Thus, output after SAE dimension reduction

Wherein

。

Then, according to HD-FCM algorithm pair

Fuzzy clustering is carried out, and in the clustering process, a hyperplane drawing method is adopted firstly

Dividing the data into 100 sub data sets, wherein each sub data set comprises 1100 samples, then performing FCM fuzzy clustering on each sub data set, dividing each sub data set into 6 clusters, and setting the number of clusters finally obtained by the whole data set to be 6.

Before supervised classification is performed, the sample dimensions of the input CNN need to be designed. The shallow feature and the deep feature are combined together to be used as the input sample of the CNN, and the shallow feature refers to the signal sample after real and virtual separation

The deep layer features refer to SAE dimension reduction output information

The combination mode is as follows:

（11）

thus, the resulting CNN samples

The corresponding labely ⁱAnd adopting one-hot coding.

In the supervised classification process, the adopted CNN comprises two convolutional layers and two fully-connected layers, wherein the convolutional layer in the first layer comprises 64 convolutional kernels with the size of 2 multiplied by 4 and the step length is 1, the convolutional layer in the second layer comprises 16 convolutional kernels with the size of 1 multiplied by 4 and the step length is 1, the fully-connected layer in the first layer comprises 64 neurons, and the fully-connected layer in the second layer comprises 16 neurons. ReLU nonlinear function is selected as the activation function for all the convolutional layers, 11 neurons are included in the output layer, and Softmax is selected as the activation function. And updating the network parameters by adopting an Adam optimization algorithm. It is composed ofMiddle, training batch size 128, initial learning rateη=0.001，β ₁=0.9，β ₂=0.999, the number of training sessions is 100.

From the training result, when the number of the cluster clusters is 6, the AM-SSB and CPFSK modulation signals are clustered and then are divided into two clusters, and the two clusters are obviously demarcated from other clusters, which shows that the AM-SSB and the CPFSK modulation signals have higher difference compared with other modulation modes, and can be better distinguished from other modulation modes through unsupervised clustering.

Secondly, 8PSK and QPSK can be divided from other modulation schemes, but they are difficult to distinguish from each other and are grouped into the same cluster. The reason is that the two modulation modes have extremely high similarity. The two modulation modes are both phase shift keying, the bits of the digital information are coded on the phase of the carrier wave, the difference is only that the modulation orders of the two are different, 8PSK has 8 state symbols, each symbol can code 3-bit data, QPSK utilizes 4 different phases in the carrier wave to represent the input data information, and one symbol represents 2-bit information. Therefore, only two modulation signals, namely 8PSK and QPSK, can be divided from 11 modulation signals by unsupervised clustering, but the two modulation signals cannot be distinguished from each other, so that the supervised classification is needed for more accurate identification. Meanwhile, QAM16 and QAM64 also have the same phenomenon, both of which are amplitude-phase joint modulation, and realize transmission of information bits by using the amplitude and phase of the carrier. Therefore, the two have extremely high similarity, and are difficult to distinguish by clustering, and further classification processing is required. In addition, BPSK and PAM4 are grouped into one class, and AM-DSB, GFSK and WBFM are grouped into one class. This requires supervised learning to further identify the data samples in the cluster.

Fig. 5 shows a specific result of the received signal after FCM fuzzy clustering, and it can be seen that two modulation signals, AM-SSB and CPFSK, can be accurately distinguished from other signal samples after unsupervised clustering, the remaining 9 modulation signals are divided into 4 groups which are easily confused with each other, namely 8PSK and QPSK, BPSK and QAM4, GFSK, WABFM and AM-DSB, and QAM16 and QAM64, and further identification and distinction are performed by adopting a CNN algorithm with supervised training, respectively. The 4 CNNs shown in fig. 5 are all initialized in the same structure, and the CNNs are adjusted toward different optimization directions for different classification tasks in repeated training, so that the trained CNN network can adapt to specific classification tasks.

Fig. 6 shows the performance of CNN including different identification tasks during the supervised classification phase, CNN1 for mutual discrimination between 8PSK and QPSK, CNN2 for accurate identification of BPSK and PAM4, CNN3 for identification tasks of AM-DSB, WBFM and QM-DSB, and CNN4 for accurate discrimination between QAM16 and QAM 64. It can be seen from the results that the features of each signal sample are more prominent with the improvement of the signal-to-noise ratio, so that the recognition accuracy of the CNN is greatly improved, which proves that the CNN structure adopted in this chapter can be better adapted to the recognition classification of 4 groups of modulation signals. Under the condition of high signal to noise ratio, all modulation signals which cannot be effectively distinguished by an unsupervised clustering method can be accurately identified by a supervised classification method.

In the subsequent simulation experiments, the proposed semi-supervised recognition algorithm is compared with the existing commonly used supervised learning based recognition methods (GoogleNet, cafnenet and AlexNet) to more fully recognize the signal recognition performance of the proposed method. Among the compared supervision algorithms, GoogleNet, cafenenet and AlexNet are all task processes that migrate the classical CNN network in the field of computer vision into the field of communications, and are constantly being improved and applied. In addition, Linear Support Vector Machines (LSVM) and Deep Belief Networks (DBN) are also widely used in the modulation recognition problem.

Fig. 7 shows the performance of the proposed semi-supervised identification algorithm compared with several representative supervised identification methods at different signal-to-noise ratios. It can be seen from the experimental results that as the signal-to-noise ratio increases, the correlation of the received noise signal vector decreases, the recognition accuracy of the proposed semi-supervised recognition algorithm is also continuously improved, and finally, the recognition accuracy is about 99.72% at about 0 dB. Meanwhile, the performance of the semi-supervised identification algorithm is better than that of the compared supervised learning algorithm in the whole signal-to-noise ratio range. The method fully proves that the provided semi-supervised identification algorithm can effectively solve the problem of classification nonuniformity of the supervised algorithm, breaks through suboptimal identification performance and brings remarkable gain for a signal identification task.

In specific implementation, the semi-supervised identification algorithm proposed by the present application is compared with the confusion matrix of other supervised learning algorithms (such as DBN algorithm, AlexNet algorithm, LSVM algorithm, etc.) under the condition that SNR =10 dB. The abscissa in the confusion matrix represents the actual label of the sample, and the ordinate represents the label predicted by the classification algorithm, so that the higher the value of the diagonal in the confusion matrix, the better the recognition result performance. Experiments show that most AM-SSB signals and CPFSK signals can be correctly classified in a confusion matrix of a supervised learning algorithm, the highest recognition accuracy can reach 99%, but other types of modulation signals are seriously confused with each other. For example, in the DBN algorithm, a partial 8PSK signal is mistakenly split into QPSK signals, a small amount of AM-DSB signals are mistakenly split into WBFM signals, and QAM16 and QAM64 signals are mistaken for each other with a very high probability of being present. Due to the existence of the problem of uneven classification, the supervised learning algorithm is optimized only aiming at a simple recognition task in the training process, and the bottleneck of suboptimal performance is involved. In the confusion matrix of the semi-supervised identification algorithm provided by the application, when the signal-to-noise ratio is 10dB, basically all signal samples can be accurately identified, and the semi-supervised identification algorithm can be suitable for AM-SSB and CPFSK which are easy to distinguish, or 8PSK and QPSK, QAM16, QAM64 and other modulation signals which are easy to confuse, and each modulation signal can achieve higher identification performance.

EXAMPLE III

Another aspect of the present invention further provides a modulation identification system, please refer to fig. 8, which shows a modulation identification system according to a third embodiment of the present invention, the system includes:

a data obtaining module 11, configured to obtain a data set of a received signal, where the received signal carries a characteristic parameter used for characterizing a modulation mode;

a cluster analysis module 12, configured to perform cluster analysis on the received signal data according to the characteristic parameters carried by the received signals, and divide the modulation signals with a degree of distinction greater than a preset value, where the modulation signals with a degree of distinction less than the preset value are aggregated into a cluster;

a result judgment module 13, configured to judge whether the obtained cluster includes two or more modulation signals;

the modulation identification module 14 is configured to perform supervised identification on the received signals in the clusters by using a supervised learning method based on a convolutional neural network when it is determined that the obtained clusters include two or more modulation signals, so as to identify a modulation mode of the received signals in the mixed clusters; and when judging that all the received signals in the cluster only belong to the same modulation mode, taking the modulation mode corresponding to the cluster as the identification result of each received signal in the cluster.

Further, in some optional embodiments of the present invention, the cluster analysis module 12 includes:

and the signal data set dividing unit is used for dividing the received signal data set into a plurality of completely disjoint received signal subsets by adopting a hyperplane division method according to the characteristic parameters carried by the received signals.

And the cluster analysis unit is used for respectively carrying out cluster analysis of a modulation mode on each received signal subset according to the characteristic parameters carried by the received signals so as to divide the received signal subsets into a plurality of clusters.

Further, in some optional embodiments of the present invention, the cluster analysis module 12 further includes:

a threshold judgment unit for judging whether the sum of mean square errors obtained by clustering is less than a threshold;

and the cluster analysis unit is also used for carrying out cluster analysis again on the cluster center of each cluster obtained by clustering until the sum of mean square errors obtained by clustering is less than the threshold value when the sum of mean square errors obtained by clustering is judged to be not less than the threshold value.

Wherein the total number of clusters is 6. The clustering analysis is fuzzy clustering.

Further, in some optional embodiments of the present invention, the modulation identification system further comprises:

and the signal dimension reduction module is used for performing dimension reduction processing on the received signals in the received signal data set by adopting an unsupervised sparse self-encoder.

Further, in some optional embodiments of the present invention, the modulation identification module 14 is further configured to use raw received data and dimension-reduced data of the received signal as input data, and perform supervised identification on the received signal in the cluster by using a supervised learning method based on a convolutional neural network.

In summary, in the modulation identification system in the above embodiment of the present invention, through unsupervised learning cluster analysis, some modulation modes with obvious distinguishing features can be directly identified, and other modulation modes that are easily confused with each other are classified into the same cluster due to the extremely high similarity. Aiming at clusters containing two or more modulation types, further supervised recognition is carried out by adopting CNN supervised learning, so that the problem of classification nonuniformity of multiple modulation method recognition by adopting the same network structure in the prior art can be solved, the CNN can carry out special training and optimization aiming at the modulation modes which are easy to be confused, and the recognition rate of a single modulation mode is ensured to be optimal while the integral average recognition rate is higher. Meanwhile, by adopting a method of clustering and then classifying, the repeated training overhead of the neural network for easily classified data can be reduced, so that the neural network is more focused on solving difficult problems, and the training efficiency is improved.

The invention also proposes a computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, implements a modulation recognition method as described above.

The invention also proposes a modulation recognition device comprising a processor, a memory and a computer program stored on the memory and executable on the processor, which when executing the computer program implements a modulation recognition method as described above.

The modulation identification device may be a computer, a signal receiver, a radio reaction instrument, or the like. The processor may be, in some embodiments, a Central Processing Unit (CPU), controller, microcontroller, microprocessor or other data Processing chip for executing program code stored in memory or Processing data, such as executing access restriction programs.

Wherein the memory includes at least one type of readable storage medium including flash memory, hard disks, multimedia cards, card-type memory (e.g., SD or DX memory, etc.), magnetic memory, magnetic disks, optical disks, etc. The memory may in some embodiments be an internal storage unit of the modulation identification device, for example a hard disk of the modulation identification device. The memory may also be an external storage device of the modulation recognition apparatus in other embodiments, such as a plug-in hard disk provided on the modulation recognition apparatus, a Smart Media Card (SMC), a Secure Digital (SD) Card, a Flash memory Card (Flash Card), and the like. Further, the memory may also include both an internal storage unit of the modulation identification apparatus and an external storage device. The memory may be used not only to store application software installed in the modulation recognition apparatus and various kinds of data, but also to temporarily store data that has been output or will be output.

In summary, in the modulation identification device in the above embodiment of the present invention, through unsupervised learning cluster analysis, some modulation modes with obvious distinguishing features can be directly identified, and other modulation modes that are easily confused with each other are classified into the same cluster due to the extremely high similarity. Aiming at clusters containing two or more modulation types, further supervised recognition is carried out by adopting CNN supervised learning, so that the problem of classification nonuniformity of multiple modulation method recognition by adopting the same network structure in the prior art can be solved, the CNN can carry out special training and optimization aiming at the modulation modes which are easy to be confused, and the recognition rate of a single modulation mode is ensured to be optimal while the integral average recognition rate is higher. Meanwhile, by adopting a method of clustering and then classifying, the repeated training overhead of the neural network for easily classified data can be reduced, so that the neural network is more focused on solving difficult problems, and the training efficiency is improved.

Those of skill in the art will understand that the logic and/or steps represented in the flowcharts or otherwise described herein, e.g., an ordered listing of executable instructions that can be viewed as implementing logical functions, can be embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus, or device, such as a computer-based system, processor-containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions. For the purposes of this description, a "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device.

More specific examples (a non-exhaustive list) of the computer-readable medium would include the following: an electrical connection (electronic device) having one or more wires, a portable computer diskette (magnetic device), a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium could even be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, via for instance optical scanning of the paper or other medium, then compiled, interpreted or otherwise processed in a suitable manner if necessary, and then stored in a computer memory.

It should be understood that portions of the present invention may be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, the various steps or methods may be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any one or combination of the following techniques, which are known in the art, may be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application specific integrated circuit having an appropriate combinational logic gate circuit, a Programmable Gate Array (PGA), a Field Programmable Gate Array (FPGA), or the like.

In the description herein, references to the description of the term "one embodiment," "some embodiments," "an example," "a specific example," or "some examples," etc., mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above do not necessarily refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

The above-mentioned embodiments only express several embodiments of the present invention, and the description thereof is more specific and detailed, but not construed as limiting the scope of the present invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the inventive concept, which falls within the scope of the present invention. Therefore, the protection scope of the present patent shall be subject to the appended claims.

Claims

1. A modulation identification method, the method comprising:

judging whether the obtained cluster contains two or more modulation signals;

if so, carrying out supervised identification on the received signals in the clusters by adopting a supervised learning method based on a convolutional neural network so as to identify the modulation mode of the received signals in the mixed clusters;

2. The modulation identification method according to claim 1, wherein the step of performing cluster analysis on the received signal data according to the characteristic parameters carried by the received signals to divide the modulated signals with a degree of discrimination greater than a predetermined value, and the modulated signals with a degree of discrimination less than the predetermined value are aggregated into a cluster of characteristic parameter signal data sets comprises:

3. The modulation identification method according to claim 2, wherein after the step of performing cluster analysis of the modulation scheme on each received signal subset according to the characteristic parameters carried by the received signals to divide the received signal subsets into a plurality of clusters, the method further comprises:

4. The modulation identification method according to claim 3, wherein the total number of the clusters is 6.

5. The modulation identification method according to any one of claims 1-4, wherein the cluster analysis is fuzzy clustering.

6. The modulation identification method according to any one of claims 1-4, further comprising, after the step of acquiring a received signal data set:

7. The modulation identification method according to claim 6, wherein the step of performing supervised identification on the received signals in the cluster by using a supervised learning method based on a convolutional neural network comprises:

8. A modulation identification system, the system comprising:

9. A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out a modulation recognition method according to any one of claims 1 to 7.

10. A modulation recognition apparatus comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the modulation recognition method according to any one of claims 1 to 7 when executing the program.