A modulation recognition method based on residual generation adversarial network

Abstract

The embodiment of the present disclosure relates to a modulation identification method based on a residual generation adversarial network. The modulation identification method based on the residual generation adversarial network comprises: generating a high-quality data expansion dataset by a convolutional neural network as a generator network, a residual network composed of a new residual unit as a discriminator network and performing gradient descent training; and output results of the discriminator are fed back to the generator and the discriminator and gradient descent training is performed respectively. The generator network of the embodiment of the present disclosure first learns feature information of different distribution domain data, generates high-quality data to expand the dataset, and then a residual network composed of a new residual unit is used as a discrimination network, which effectively improves the feature extraction capability of the network and better distinguishes similar signals; the identification accuracy is significantly improved under the condition of a small sample, the model convergence speed is accelerated, and the method complexity is obviously reduced.

Description

TECHNICAL FIELD

[0001] The embodiment of the present disclosure relates to the technical field of modulation signal identification, in particular to a modulation identification method based on residual generation adversarial network. BACKGROUND

[0002] Automatic modulation classification (AMC) of communication signals is one of the key technologies of cognitive radio systems. In the condition of non-cooperative communication system, whether the receiving end can quickly and accurately identify the modulation style of the received signal under the condition that the signal modulation information is unknown is the key prerequisite for subsequent signal demodulation. In complex electromagnetic environment, AMC technology plays an important role in battlefield reconnaissance, signal parameter estimation and spectrum monitoring.

[0003] Modulation identification technology has developed so far, whether it is a traditional method based on artificial design features or a deep learning-based algorithm has made fruitful research results. Traditional modulation identification methods rely on artificial design features, have fewer identifiable signal types, and have weak network generalization ability. While deep learning-based modulation identification algorithms often need to use a large amount of labeled data to train the network by gradient descent, and have high algorithm complexity.

[0004] Therefore, it is necessary to improve one or more problems existing in the related technical solutions.

[0005] It should be noted that this part aims to provide background or context for the technical solutions of the present disclosure stated in the claims. The description herein is not admitted to be prior art merely because it is included in this part. SUMMARY

[0006] The purpose of the embodiment of the present disclosure is to provide a modulation identification method based on residual generation adversarial network, and to at least overcome one or more problems caused by the limitations and defects of the related art.

[0007] According to the embodiment of the present disclosure, a modulation identification method based on residual generation adversarial network is provided, comprising:

[0008] A convolutional neural network is used as a generator network to generate high-quality data augmentation dataset, and a residual network composed of new residual units is used as a discriminator network and performs gradient descent training;

[0009] The output result of the discriminator is fed back to the generator and the discriminator and gradient descent training is performed respectively.

[0010] In an embodiment of the present disclosure, the generator input is noise, and the output is generated data; the discriminator input is unlabeled real data, labeled real data and the generated data, and the output is a similarity between the generated data and the unlabeled real data, the labeled real data determined by a true-false classifier, and a class attribute of the generated data and the unlabeled real data, the labeled real data determined by a class classifier.

[0011] In an embodiment of the present disclosure, the gradient descent training adopts an alternating iterative method to perform gradient descent training on one of the generator and the discriminator while fixing the network model and the weight of the other.

[0012] In an embodiment of the present disclosure, the new residual unit takes Leakyrelu as a hidden layer activation function, and the formula of the Leakyrelu is:

[0013] (1)

[0014] wherein a is a constant between 0 and 1, represents input data;

[0015] The structural expression of the new residual unit is:

[0016] (2)

[0017] wherein, represents direct mapping, represents residual part, represents activation function, represents convolution operation.

[0018] If there are preset new residual units, the structural expression of the new residual unit becomes:

[0019] (3)

[0020] According to the chain rule of derivatives used in the gradient descent training, the loss function The gradient of may be expressed as:

[0021] (4)

[0022] wherein, is input data.

[0023] In an embodiment of the present disclosure, during the entire gradient descent training process, will not be -1 all the time, that is, there will be no gradient disappearance problem during the gradient descent training process, express The gradient of a layer is directly transferred to any layer shallower than it. layer.

[0024] In one embodiment of this disclosure, the objective function is modified to consist of two parts: supervised learning and unsupervised learning, based on the input-output structure of the network and the loss function. The supervised learning uses labeled data. The unsupervised learning process utilizes the generated data generated by the generator and calculates the multivariate cross-entropy loss function. or The discriminator is calculated using the binary cross-entropy loss function, where 0 represents fake data and 1 represents real data. The output layer activation functions of the discriminator are the Sigmoid function and the Softmax function, respectively.

[0025] In one embodiment of this disclosure, the Sigmoid function outputs a number between 0 and 1, where a number closer to 1 indicates a more accurate judgment result, and a number closer to 0 indicates a more accurate judgment result, which is closer to false data. The Softmax function outputs a K-dimensional vector. The probability is converted to:

[0026] (5)

[0027] in, The network will Predicted as the first The probability of a vector is given by its dimension, and the vector with the highest probability is the label attribute of the data.

[0028] In one embodiment of this disclosure, the complexity of gradient descent training is simplified. Based on the authenticity of the data, the (K+1)th dimension vector is determined to be any one of the categories among the first K vectors. Then, the loss function is expressed as follows:

[0029] (6)

[0030] in, The loss function representing supervised learning, The loss function represents unsupervised learning. The differentiable function representing the D mapping, It is about seeking expectations. This indicates that the tag data was determined to be the first. dimensional vector ( ), This indicates that unlabeled data is classified as a vector of dimension K+1. This indicates that the generated data is determined to be real data;

[0031] The objective function is to minimize input data subject to distributed, the objective function is:

[0032] (7)

[0033] wherein represents that the label data is judged as a correct class, represents that the generated data is judged as real data.

[0034] In an embodiment of the present disclosure, the new residual unit is provided with a preset layer asymmetric convolution structure, the generator network is provided with a preset layer convolution structure, and the discriminator network is provided with a preset number of the new residual unit.

[0035] In an embodiment of the present disclosure, the new residual unit, the generator network and the discriminator network are respectively removed from a batch normalization layer and a pooling layer, a zero padding layer is added, and an asymmetric decreasing convolution kernel with a stride of 1 is adopted.

[0036] The technical scheme provided by the embodiment of the present disclosure can include the following beneficial effects:

[0037] In the embodiment of the present disclosure, by the above method, the generator network first learns the feature information of data in different distribution domains, generates high-quality generated data through noise to expand the data set, and the residual network composed of the new residual unit is used as the discriminator network, which effectively improves the feature extraction capability of the network and better distinguishes similar signals; the recognition accuracy is significantly improved under the condition of small samples, the model convergence speed is accelerated, and the method complexity is obviously reduced. BRIEF DESCRIPTION OF DRAWINGS

[0038] The accompanying drawings, which are incorporated into and form part of the specification, illustrate embodiments consistent with the present disclosure and, together with the specification, serve to explain the principles of the present disclosure. Obviously, the drawings in the following description are only some embodiments of the present disclosure, and other drawings can be obtained from these drawings without creative labor for those skilled in the art.

[0039] Figure 1 A flow chart of a residual generative adversarial network recognition method in an exemplary embodiment of the present disclosure is shown;

[0040] Figure 2 A schematic diagram of a residual generative adversarial network structure in an exemplary embodiment of the present disclosure is shown;

[0041] Figure 3 A schematic diagram of a new residual unit network structure in an exemplary embodiment of the present disclosure is shown;

[0042] Figure 4FIG. 1 shows a schematic diagram of the structure parameters of the new residual unit in the exemplary embodiments of the present disclosure;

[0043] Figure 5 FIG. 2 shows a schematic diagram of the structure parameters of the generator network and the discriminator network in the exemplary embodiments of the present disclosure;

[0044] Figure 6 FIG. 3 shows a comparison chart of the results of different network layers in the exemplary embodiments of the present disclosure;

[0045] Figure 7 FIG. 4 shows a comparison chart of the results of different gradient optimization algorithms in the exemplary embodiments of the present disclosure;

[0046] Figure 8 FIG. 5 shows a chart of the results of Leakyrelu with different parameters in the exemplary embodiments of the present disclosure;

[0047] Figure 9 FIG. 6 shows a comparison chart of the results of different activation functions in the exemplary embodiments of the present disclosure;

[0048] FIG. 10(a) shows a curve chart of the discriminator loss function in the exemplary embodiments of the present disclosure;

[0049] FIG. 10(b) shows a curve chart of the generator loss function in the exemplary embodiments of the present disclosure;

[0050] FIG. 11(a) shows a chart of the confusion matrix (SNR=0dB) in the exemplary embodiments of the present disclosure;

[0051] FIG. 11(b) shows a chart of the confusion matrix (SNR=10dB) in the exemplary embodiments of the present disclosure;

[0052] FIG. 11(c) shows a chart of the confusion matrix (SNR=18dB) in the exemplary embodiments of the present disclosure;

[0053] Figure 12 FIG. 12 shows a comparison chart of the classification accuracy in the exemplary embodiments of the present disclosure. DETAILED DESCRIPTION

[0054] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations can be implemented in any

[0055] In addition, the accompanying drawings are only schematic and are non-limiting illustrative of the present disclosure. Wherever possible, the same reference indicators will be used throughout the drawings, and similar reference indicators will be used for similar elements, and the description will not be repeated. Some of the blocks in the drawings can be functional building blocks, and can not necessarily correspond to physical or logical independent units.

[0056] In the present example embodiment, a modulation recognition method based on a residual generation adversarial network is first provided. Referring to FIG. 1, the modulation recognition method based on the residual generation adversarial network can include steps S101-S102. Figure 1

[0057] Step S101: generating a high-quality data expansion dataset by a convolutional neural network as a generator network, and performing gradient descent training by a residual network composed of a new residual unit as a discriminator network;

[0058] Step S102: the output result of the discriminator is fed back to the generator and the discriminator to perform gradient descent training.

[0059] Through the above modulation recognition method based on the residual generation adversarial network, the generator network first learns the feature information of different distribution domain data, generates high-quality generated data by noise to expand the dataset, and then the residual network composed of a new residual unit is used as a discriminator network, which effectively improves the feature extraction capability of the network and better distinguishes similar signals. The recognition accuracy is significantly improved under the condition of small sample, the model convergence speed is accelerated, and the method complexity is obviously reduced.

[0060] In the following, reference will be made to Figures 1 to 5 The above modulation recognition method based on the residual generation adversarial network in the present example embodiment will be described in more detail.

[0061] Step S101: generating a high-quality data expansion dataset by a convolutional neural network as a generator (Generator, G) network, and performing gradient descent training by a residual network composed of a new residual unit as a discriminator (Discriminator, D) network;

[0062] ​Specifically, since the traditional modulation recognition method relies on manual design of features, the number of recognizable signals is small, and the network generalization ability is weak, while the modulation recognition algorithm based on deep learning often needs to use a large amount of labeled data to train the network, and the algorithm complexity is high; Because the convolutional neural network can extract a wide range of features from information, the convolutional neural network is used as a generator network first to generate a large amount of high-quality generated data to expand the data set, so as to obtain a large amount of data to train the network, and the residual network has been proved to have good effect in the extraction of subtle features of signals and the improvement of network gradient descent training, so the residual network composed of new residual units is used as the discriminator network.

[0063] In addition, the new residual unit takes Leakyrelu as the hidden layer activation function and removes the pooling layer, which on the one hand better preserves the initial information amount, and on the other hand can also calculate the gradient for the part of the input less than zero in the process of executing back propagation, better realizing the optimal update of network parameters and bias terms, and the formula of the Leakyrelu is:

[0064] (1)

[0065] Wherein, a is a constant between (0, 1), represents the input data;

[0066] The structural expression of the new residual unit is:

[0067] (2)

[0068] Wherein, represents direct mapping, represents the residual part, represents the activation function, represents the convolution operation;

[0069] The preset new residual units, then the structural expression of the new residual unit becomes:

[0070] (3)

[0071] According to the chain rule of derivative used in gradient descent training, the loss function The gradient of can be expressed as:

[0072] (4)

[0073] From the above formula, two attributes of the residual network of the embodiment can be obtained:

[0074] 1. In the whole gradient descent training process Will not always be -1, that is, there is no gradient disappears under the gradient descent training process;

[0075] 2. Indicates The gradient of a layer can be directly passed to any one of its shallower Layer.

[0076] Combined with the input and output structure of the network and the loss function, the objective function is modified into two parts of supervised learning and unsupervised learning, the supervised learning uses label data And multivariate cross-entropy loss function for calculation, unsupervised learning uses the data Or Generated by the generator (0 represents false data, 1 represents real data) and binary cross-entropy loss function for calculation, the output layer activation function of the discriminator uses Sigmoid function and Softmax function respectively.

[0077] The Sigmoid function outputs a number between 0 and 1, the closer to 1 represents the closer to the real data of the discrimination result, the closer to 0 represents the closer to the false data of the discrimination result, and the Softmax function outputs a K-dimensional vector , converted into probability:

[0078] (5)

[0079] Among them, Represents the probability that the network predicts the first Dimensional vector, the dimension vector with the maximum probability is the label attribute of the data.

[0080] In order to simplify the complexity of the algorithm, the output vector is expanded to a K+1-dimensional vector, the K+1-dimensional vector is used to judge the gradient descent training condition of the unlabeled data, since only the true and false nature of the data needs to be judged, the K+1-dimensional vector can be judged as any one of the first K vectors. Category, the loss function expression is as follows:

[0081] (6)

[0082] Among them Represents the loss function of supervised learning, Represents the loss function of unsupervised learning, Represents the differentiable function of D mapping, Is the expectation, Indicates that the label data is judged as the first Dimensional vector ( ), ​represents that the unlabelled data is determined as the K+1 dimensional vector, represents that the generated production data is determined as real data;

[0083] The objective function is to minimize , input data subject to distribution, the objective function is:

[0084] (7)

[0085] wherein represents that the label data is determined as the correct class, represents that the generated data is determined as real data.

[0086] Step S102: the output result of the discriminator is fed back to the generator and the discriminator, and gradient descent training is performed respectively.

[0087] Specifically, the generator and the discriminator adopt an alternating iterative gradient descent training method. After the discriminator outputs the result, the generator and the discriminator are trained by gradient descent respectively to obtain more real data. The generator and the discriminator are alternately executed in the gradient descent training stage. The convergence result of the final network is that the generated data generated by the generator has the highest similarity with the real data, and the discriminator can accurately distinguish the modulation style of the signal.

[0088] In one embodiment, as shown in Figure 2 , the input of the generator is noise, and the output is generated data. The input of the discriminator is unlabelled real data, labelled real data and the generated data. The output is the similarity of the generated data to the unlabelled real data and the labelled real data determined by the true-false classifier, and the class attribute of the generated data to the unlabelled real data and the labelled real data determined by the class classifier.

[0089] Specifically, in order to achieve the purpose of generating generated data by the network and realizing signal classification, the output layer of the network is modified into two parts: judging the true-false attribute and the class attribute of the data. The input of the generator is noise, and the output is generated data. The input of the discriminator is unlabelled data, labelled data and generated data. The output is two parts: the similarity of the generated data to the unlabelled real data and the labelled real data determined by the true-false classifier, and the class attribute of the generated data to the unlabelled real data and the labelled real data determined by the class classifier.

[0090] In one embodiment, the gradient descent training adopts an alternating iterative method for gradient descent training. One of the generator and the discriminator is trained by gradient descent, and the network model and weight of the other one are fixed.

[0091] Specifically, the gradient descent training adopts an alternating iterative gradient descent training method. When the generator and the discriminator are trained by gradient descent, gradient descent is performed on only one network each time. For example, the discriminator is first trained by gradient descent once, then the network model and the weight are fixed, and then the generator is trained by gradient descent again, so as to realize the purpose of adversarial gradient descent training and mutual improvement between the two.

[0092] In one embodiment, as shown in Figure 3 , Figure 4 , Figure 5 The new residual unit sets a preset layer asymmetric convolution structure, the generator network sets a preset layer convolution structure, and the discriminator network sets a preset number of the new residual unit.

[0093] Specifically, Figure 3 is a network structure diagram of the new residual unit. Considering that the present application is under small sample conditions, in terms of network layers, it is assumed that the new residual unit sets a layer of asymmetric convolution structure, the generator network sets b layers of convolution structure, and the discriminator network sets c new residual units. For example, as shown in Figure 4 , the new residual unit sets a = 3 layers of asymmetric convolution structure, as shown in Figure 5 , the generator network selects b = 3 layers of convolution structure, and the discriminator network selects c = 4 new residual units. Here, a, b, and c are not specifically limited in value.

[0094] In one embodiment, the new residual unit, the generator network, and the discriminator network each remove a batch normalization layer and a pooling layer, add a zero padding layer, and adopt an asymmetric decreasing small convolution kernel with a stride of 1.

[0095] Specifically, removing the batch normalization layer and the pooling layer can maximize the preservation of the original information of the signal; adding the zero padding layer can increase the receptive field range, and the asymmetric convolution with a stride of 1 helps to extract edge feature information; the decreasing small convolution kernel can extract different levels of features of the signal, significantly reducing network parameters and lowering algorithm complexity.

[0096] Through the above modulation recognition method based on the residual generative adversarial network (Res-GAN), the generator network first learns the feature information of different distribution domain data, generates high-quality generated data through noise to expand the data set, and then the residual network (Resnet) composed of the new residual unit is used as the discriminant network, which effectively improves the feature extraction capability of the network and better distinguishes similar signals. Under small sample conditions, the recognition accuracy is significantly improved, the model convergence speed is accelerated, and the method complexity is obviously reduced.

[0097] The present embodiment will be further described below in combination with a specific simulation example.

[0098] The network model gradient descent training of the embodiment uses the Keras environment of Python, is configured as an Nvidia GTX 1650 GPU, gradient descent training is performed with the help of the Tensorflow back end, the modulation signal data set adopts the DeepSig public modulation recognition data set RML 2016.10b, and table 1 is the data set information adopted in the embodiment.

[0099] Table 1 Modulation signal data set

[0100]

[0101] Table 2 is the specific steps of the Res-GAN method for one cycle.

[0102] Table 2 Res-GAN method specific steps

[0103]

[0104] Experiment one: network initial parameter setting

[0105] According to the 3-channel principle, the data dimension is increased from (2, 128) to (2, 128, 1), because the data set signal has been normalized, so the noise subject to (0, 1) Gaussian distribution is directly used to fit the data, and the gradient descent training number of rounds is set to 90 epochs.

[0106] The embodiment compares the influence of network layer number, different gradient optimization algorithms and hidden layer activation functions on network performance through a large number of experiments. In terms of network layer number, n represents the number of convolution layers of the new residual unit, and N represents the number of new residual units of the D network. The network performance under different values of n and N is compared, and the experimental results are as shown in Figure 6 .

[0107] From Figure 6 , it can be concluded that the new residual unit is a 3-layer convolution structure, and the D network selects 4 new residual units to maximize the optimization of the network model.

[0108] The result comparison chart of different gradient optimization algorithms is as shown in 7.

[0109] Compared with the other three algorithms, the Adam gradient optimization algorithm has faster convergence speed under low signal-to-noise ratio and stronger stability under high signal-to-noise ratio, and the algorithm has good adaptability. The specific reason is that the Adam optimization algorithm uses the first moment and the second moment to correct the error of the weighted gradient. At the same time, because the number of samples used in network testing is small, the curve will present a non-monotonic fluctuation trend.

[0110] When the default setting of Adam learning rate is changed, the network cannot be trained by gradient descent, so the Adam learning rate is set to the default value 0.0002. In combination with the Leakyrelu expression, the network performance under different values of a and different activation functions is compared respectively, and the experimental results are shown in Figs. Figure 8 and Figure 9 .

[0111] In combination with Figure 8 , Figure 9 results, the network performance is best when the parameter value a in Leakyrelu is 0.25, and compared with the other two activation functions, it converges faster and the recognition accuracy is improved by 2%-4%. In addition, under the premise of Adam optimization algorithm, the network performance under different batch-size, Dropout and Adam parameter settings is compared through a large number of experiments, and part of the experimental results are as follows.

[0112] Table 3 Comparison of network performance under different parameters

[0113]

[0114] Based on the above experimental results, the residual generation adversarial network (Res-GAN) algorithm selects the Adam gradient optimization algorithm (learning rate is set to 0.0002, the first exponential decay rate is set to 0.5), the activation function selects Leakyrelu (a=0.25), Dropout is set to 0.05, and Batch-size is set to 64. By comparing the experimental results, it is found that different initialization parameters have certain influence on the model performance (the accuracy fluctuation range is 3%-5%). Therefore, it is necessary to select the initialization parameter set of the network model through experiments.

[0115] Experiment two: modulation signal classification and identification

[0116] According to the network parameter initial value setting of experiment one and the specific algorithm steps of table 2, 660, 1800, 3600, 7200, 15000, 30000 and 60000 labeled data (m value) of each signal under the signal-to-noise ratio of (-4dB~18dB) are randomly selected, L is 86400, and q is 500. The selected data is divided into gradient descent training set, validation set and test set according to the ratio of 8:1:1. The optimal result of the selected model is shown in Fig. 10, and the confusion matrix result after network test is shown in Fig. 11.

[0117] The Res-GAN method gradient descent training stage first gradient descent trains the D network. As can be seen from FIG. 10(a), the loss function curve overall maintains a steady downward trend, and the verification set has reached a convergence state after 60 epochs. After the D gradient descent training is completed, the parameters are fixed, and then the G network is gradient descent trained. As can be seen from the loss function curve of FIG. 10(b), the fitting degree of the verification set and the gradient descent training set is high, and the network gradient descent training does not have overfitting problem.

[0118] As can be seen from the confusion matrix diagram after the network test, the Res-GAN algorithm can achieve an identification accuracy of 91% when SNR=0dB, and an identification accuracy of 98% when SNR=18dB. Except for WBFM and AM-DSB, the signals can achieve accurate identification. Because the amplitude graphs of the two signals in the original data set are relatively high in similarity, identification confusion may occur.

[0119] Experiment three: comparative experiment

[0120] In order to fully illustrate that the new residual unit designed in the embodiment can better extract the multi-dimensional information of the signal and better distinguish the modulation signal, the Res-GAN algorithm is compared with the CLDNN model, the AUCNN model, the Resnet-WSMF algorithm and the ACGAN model, and the optimization results of the five networks are selected, Figure 12 and Table 4 are the comparison between the network classification accuracy and the network parameters.

[0121] Table 4 Comparison of different network parameters

[0122]

[0123] The running time represents 1 epoch time, from Figure 12 and Table 4 results can be obtained. Under the small sample condition set in the embodiment, the network parameters of ACGAN are the least, and the identification accuracy is the lowest; the network parameters of CLDNN are more, and the identification performance is general; the network performance of AUCNN and Resnet-WSMF is stable, but the gradient descent training time is longer. Compared with the above four networks, the Res-GAN algorithm achieves the highest identification accuracy, and the network parameters and time complexity are only slightly higher than ACGAN.

[0124] In order to verify the performance of the method in the embodiment under the small sample condition, different numbers of labeled samples are selected for comparative experiment, and the results are recorded in Table 5.

[0125] Table 5 Influence of different labeled sample quantities on network performance

[0126]

[0127] The results of Table 5 can be obtained that, when the label sample amount is about 660, the recognition accuracy of other networks is low except the method of the embodiment, and the performance of the comparison network is improved to different degrees with the increase of the label sample amount. In combination with the results of Table 4 and Table 5, the method of the embodiment verifies the practicability under the small sample condition on the one hand, and has certain advantages in reducing the algorithm complexity on the other hand. The minimum label sample amount of network convergence is about 660.

[0128] In view of the problems of low signal recognition accuracy and high algorithm complexity under the small sample condition, the embodiment proposes a modulation recognition algorithm based on Res-GAN, uses Resnet composed of CNN and a new residual unit as a basic network by means of the GAN model, combines semi-supervised learning with supervised learning by defining a new objective function, fully utilizes the unlabeled data and effectively realizes the expansion of the sample quantity. The simulation experiment result proves that the new residual unit can not only enrich the complexity of feature extraction, but also better realize the optimal update of network parameters; compared with other algorithms widely used in recent years, the method of the embodiment has the characteristics of high recognition accuracy, simple implementation process, good stability and low complexity. The recognition accuracy of 10 kinds of modulation signals under the small sample condition can reach 98% when SNR=18dB, which verifies the effectiveness of the method of the embodiment. In the next step, the confused signals will be preprocessed, and auxiliary information will be added to enable the network to better distinguish similar signals.

[0129] It should be understood that the orientations or positional relationships indicated by the terms "center", "longitudinal", "transverse", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise" and the like in the above description are based on the orientations or positional relationships shown in the drawings, and are only for the convenience of describing the embodiments of the present disclosure and simplifying the description, and therefore cannot be understood as indicating or implying that the devices or elements indicated must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the embodiments of the present disclosure.

[0130] In addition, the terms "first" and "second" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first" and "second" can explicitly or implicitly include one or more of the features. In the description of the embodiments of the present disclosure, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0131] In the embodiments of the present disclosure, unless specifically defined and limited otherwise, the terms "mount", "connect", "connection", "fixed", and the like should be understood broadly, for example, can be fixed connection, can also be detachable connection, or integrated; can be mechanical connection, can also be electrical connection; can be direct connection, can also be indirect connection through an intermediate medium, can be internal communication of two elements or interaction relationship between two elements. For those skilled in the art, the specific meanings of the above terms in the present disclosure can be understood according to the specific circumstances.

[0132] In the embodiments of the present disclosure, unless specifically defined and limited otherwise, the first feature "on" or "under" the second feature can include that the first and second features are in direct contact, or that the first and second features are not in direct contact but are in contact through another feature between them. Moreover, the first feature "on", "above" and "over" the second feature includes that the first feature is directly above and obliquely above the second feature, or only indicates that the horizontal height of the first feature is higher than that of the second feature. The first feature "under", "below" and "under" the second feature includes that the first feature is directly below and obliquely below the second feature, or only indicates that the horizontal height of the first feature is less than that of the second feature.

[0133] In the description of the specification, the description of the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present disclosure. In the specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in the specification.

[0134] Other embodiments of the present disclosure will be apparent to those skilled in the art upon consideration of the specification and practice of the applications disclosed. The present application is intended to cover any variations, uses, or adaptations of the present disclosure following the general principles thereof and including modifications and equivalents of the present disclosure that are obvious to those skilled in the art. The specification and examples are to be regarded as illustrative only, and the true scope and spirit of the present disclosure are indicated by the appended claims.

Claims

1. A modulation recognition method based on residual generative adversarial networks, characterized in that, include: High-quality data augmentation datasets are generated by using a convolutional neural network as a generator network, and a residual network composed of new residual units is used as a discriminator network and gradient descent training is performed. The new residual unit uses Leakyrelu as the hidden layer activation function, and the formula for Leakyrelu is: （1） Where a is a constant between (0,1), Represents input data; The structural expression of the new residual unit is: （2） in, Represents direct mapping, Represents the residual portion. Represents the activation function. Represents the convolution operation; Preset If the new residual unit is given, then the structural expression of the new residual unit becomes: （3） Based on the chain rule of derivatives used in gradient descent training, the loss function... about The gradient is expressed as: （4） in, Input data; The output of the discriminator is fed back to the generator and the discriminator, and gradient descent training is performed on them respectively. The new residual unit is configured with a preset layer of asymmetric convolution structure, the generator network is configured with a preset layer of convolution structure, and the discriminator network is configured with a preset number of the new residual units; the final convergence result of the network is that the generated data generated by the generator has the highest similarity to the real data, and the discriminator can distinguish the modulation pattern of the signal.

2. The modulation recognition method based on residual generative adversarial networks according to claim 1, characterized in that, The generator takes noise as input and generates data as output. The discriminator takes unlabeled real data, labeled real data, and the generated data as input and outputs a true / false classifier that determines the similarity between the generated data and the unlabeled real data and the labeled real data, and a category classifier that determines the category attribute between the generated data and the unlabeled real data and the labeled real data.

3. The modulation recognition method based on residual generative adversarial networks according to claim 1, characterized in that, The gradient descent training employs an alternating iterative method, where gradient descent trains one of the generators and the discriminator while fixing the network model and weights of the other.

4. The modulation recognition method based on residual generative adversarial networks according to claim 1, characterized in that, Throughout the gradient descent training process It will not remain at -1, meaning that the gradient vanishing problem will not occur during the gradient descent training process. express The gradient of a layer is directly transferred to any layer shallower than it. layer.

5. The modulation recognition method based on residual generative adversarial networks according to claim 4, characterized in that, Combining the network's input-output structure and the loss function, the objective function is modified to consist of two parts: supervised learning and unsupervised learning. The supervised learning uses labeled data. The unsupervised learning process utilizes the generated data generated by the generator and calculates the multivariate cross-entropy loss function. or The discriminator is calculated using the binary cross-entropy loss function, where 0 represents fake data and 1 represents real data. The output layer activation functions of the discriminator are the Sigmoid function and the Softmax function, respectively.

6. The modulation recognition method based on residual generative adversarial networks according to claim 5, characterized in that, The Sigmoid function outputs a number between 0 and 1. The closer the number is to 1, the closer the result is to the real data; the closer it is to 0, the closer the result is to fake data. The Softmax function outputs a K-dimensional vector. The probability is converted to: （5） in, The network will Predicted as the first The probability of a dimensional vector, and the maximum probability. That one-dimensional vector is the label attribute of the data.

7. The modulation recognition method based on residual generative adversarial networks according to claim 1, characterized in that, The new residual unit, the generator network, and the discriminator network have their batch normalization layer and pooling layer removed, zero-filling layers added, and asymmetric decreasing convolutional kernels with a stride of 1 used.

Images