Multi-terminal interference generation system and method of multi-discriminator generative adversarial network

Abstract

The invention relates to a multi-terminal interference generation system and method of a multi-discriminator generative adversarial network, and belongs to the technical field of communication. The system comprises an interference noise sample data set module, a sample data type classifier, a plurality of discriminators, a generator and at least a classifier. According to the method, interference noise data is generated through a generator, training and classification are carried out through a discriminator and a classifier, and the quality of the generated interference noise data is improved by calculating gradient parameters of the generator and the discriminator and updating the parameters. In a multi-category data type system, each category of data corresponds to one discriminator, the discrimination efficiency of the discriminators is improved, and in multi-terminal interference noise data generation, the performance of the adversarial neural network generated by using multiple discriminators is remarkably improved compared with the performance of the adversarial neural network generated by using a single discriminator.

Description

Technical Field

[0001] The invention belongs to the field of communication technology and relates to a multi-terminal interference generation system and method for a multi-discriminator generation adversarial network. Background Art

[0002] The first generation of mobile communication systems has developed to 5G and 6G today. The explosive growth of communication systems has led to a sharp increase in the number of communication terminals. On the one hand, in order to increase the number of terminals online at the same time and increase the total capacity of the system, various multiplexing technologies have begun to appear, such as frequency division multiplexing, time division multiplexing, wavelength division multiplexing and spectrum sharing. On the other hand, in response to the rapid development of industry applications, the test and verification of new equipment and new functions, large traffic and large connections of multiple users have increased the complexity and cost of the test plan. As a result, a multi-terminal simulator has emerged, and its key is to accurately model the terminal channel correlation and interference power distribution. Traditional interference modeling essentially uses a large amount of data to fit distribution parameters, but due to the influence of nonlinear distortion of the power amplifier and IQ imbalance of the interfering user, interference spectrum leakage and its mirror spectrum are introduced on the basis of the original user spectrum, resulting in the complexity of interference distribution. Traditional methods are obviously not competent.

[0003] In 2014, Goodfellow proposed the Generative Adversarial Network (GAN) network, which synthesizes data based on similar statistical distribution characteristics of a given data set. It uses adversarial training and implicit modeling methods to avoid analytical modeling of the high-dimensional distribution of input data. At present, there are dozens of GAN network structures and their main derivative structures, among which CGAN introduces category indicator vectors for multi-class sample generation, and WGAN introduces Wasserstein (abbreviated as: W distance) to improve distribution distance. In order to achieve W distance fitting for multi-class sample generation, the discriminator and classifier in CGAN are separated, and the discriminator is used to fit the W distance, and the classifier discriminates the sample category.

[0004] Although the above method can generate multi-category samples, there are still the following problems: a single discriminator fitting the W distance of multiple samples will make it impossible for the discriminator to take into account the W distance of each type of data, and the final generated data will be different from the actual distance; training a multi-discriminator network requires known training data categories and cannot be used to collect interference signals of unknown categories. Summary of the invention

[0005] In view of this, the purpose of the present invention is to provide a multi-terminal interference generation system and method for a multi-discriminator generative adversarial network. The purpose is to mark the training data set, then use a multi-discriminator network to separately calculate the distribution distance of different distribution categories, and obtain the fusion value of the loss function of multiple discriminator networks, and then use the fused loss to update the discriminator network D.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A multi-terminal interference generation system of a multi-discriminator generative adversarial network comprises: at least one interference noise sample data set module, at least one sample data type classifier, several discriminators, at least one generator and at least one classifier, wherein:

[0008] The interference noise sample data set module is used to generate multi-terminal interference noise sample data, and add interference noise classification codes and noise classification label values ​​to the multi-terminal interference noise data. Each type of terminal interference noise data is set with a unique corresponding interference noise classification code and noise classification label value;

[0009] The sample data type classification module inputs the data into the discriminator of the corresponding label value for training according to the noise classification label value in the interference noise sample data set;

[0010] The discriminator module includes several discriminators, each of which corresponds to a noise type specified by a noise classification label value;

[0011] The generator module is trained according to different noise classification label values ​​and different discriminators in the interference noise sample data set, and the trained generator is used to generate terminal interference noise data;

[0012] The small batch interference sample noise data in the training set is input into the generator to obtain the terminal interference noise data, and the terminal interference noise data is input into the discriminator corresponding to the noise classification label value of the small batch interference sample noise data to calculate the generator gradient parameter; the terminal interference noise data and the noise classification label value in the small batch interference sample noise data are input into the corresponding discriminator to calculate the discriminator gradient parameter; the parameters of the generator and the corresponding discriminator are updated respectively according to the calculated generator gradient parameter and discriminator gradient parameter;

[0013] The classifier module is used to correctly identify the noise classification label value corresponding to the interference noise data, and to determine the quality of the interference noise data generated by the generator in the neural network.

[0014] Furthermore, the generator module includes an input layer in, a classification layer labels, several deconvolution layers and an output layer, wherein the deconvolution layer is used to map the input data features from the low-dimensional space back to the high-dimensional space, thereby facilitating the reconstruction of the input signal details and converting the random noise into a high-quality output signal; the input layer is used to input a random number vector; the classification layer is used to input the classification code of the terminal interference noise data; and the output layer is completed by the Tanh function and is used to output the generated interference noise data.

[0015] Furthermore, the discriminator module includes an input layer imageinput, a classification layer labels, several convolution layers and an output layer, wherein the convolution layer is used to use convolution operations to extract the characteristics of the terminal interference noise data and effectively capture the spatial structure and local correlation of the terminal interference noise data; the input layer is used to input the terminal interference noise data, and the classification layer is used to input the terminal interference noise classification code; the output layer is composed of a softMax function, and the output layer outputs the probability value of each noise type determined by the discriminator.

[0016] Furthermore, the classifier module includes an input layer imageinput, several convolutional layers and an output layer; the input layer is used to input terminal interference noise data, and the labels layer is used to input terminal interference noise classification codes, that is, classification codes of terminal interference noise data; the output layer is composed of a softMax function, which is the probability value of the noise classifier to determine each noise type.

[0017] The present invention also proposes a multi-terminal interference generation method for a multi-discriminator generative adversarial network, which comprises the following steps:

[0018] S1. Construct an interference noise sample data set according to the multi-terminal interference noise data generation requirements, wherein the interference noise sample data set includes at least terminal interference noise data, interference noise classification code, and noise classification label value;

[0019] S2. Classify the interference noise sample data set according to the noise classification label value to obtain a number of small batches of interference sample noise data with the same noise classification label value, and all small batches of interference sample noise data constitute a training set;

[0020] S3, inputting the small batch interference sample noise data in the training set into the generator to obtain terminal interference noise data, and inputting the terminal interference noise data into the discriminator corresponding to the noise classification label value of the small batch interference sample noise data to calculate the generator gradient parameter;

[0021] S4, inputting the terminal interference noise data and the noise classification label value in the small batch interference sample noise data into the corresponding discriminator, and calculating the discriminator gradient parameter;

[0022] S5. Update the parameters of the generator and the corresponding discriminator respectively according to the calculated generator gradient parameters and discriminator gradient parameters;

[0023] S6. Use the generator with updated parameters to generate a set of terminal interference noise data, and classify the terminal interference noise data through the classifier. If the classification accuracy reaches the threshold, the training is completed, and the current generator is saved and used for interference data generation; if the classification accuracy does not reach the threshold, return to step S3 to repeat the training until the threshold is reached.

[0024] Further, in step S3, the generator gradient parameters are calculated as follows:

[0025] YGeneratedLoss=-crossentropy(YGenerated,T,'WeightsFormat','UC')

[0026] lossG = YGeneratedLoss

[0027] gradientsG=dlgradient(lossG,netG.Learnables,'RetainData',true,'EnableHigherDerivatives',true)

[0028] Among them, YGeneratedLoss represents the cross entropy value of the generated terminal interference noise data through the discriminator, YGenerated represents the terminal interference noise data, T is the noise classification label value corresponding to YGenerated; 'WeightsFormat' represents the option of specifying the dimension of the weight parameter, 'UC' indicates that the elements in the weight vector correspond to the channel dimension of the predicted value, and crossentropy() is the cross entropy calculation function; lossG represents the loss value of the terminal interference noise data generated by the generator obtained by the discriminator;

[0029] gradientsG is the generator gradient parameter, netG.Learnables is the learnable parameter of the generator network, 'RetainData', true means to use the intermediate data retained for calculating the gradient for calculation, 'EnableHigherDerivatives', true means to enable higher-order derivative calculations, dlgradient() neural network gradient calculation function.

[0030] Further, in step S4, the discriminator gradient parameter is calculated as:

[0031] YRealLoss=-crossentropy(YReal,T,'WeightsFormat','UC')

[0032] lossD=YRealLoss+(1-YGeneratedLoss)

[0033] gradientsD=dlgradient(lossD,netD.Learnables,'RetainData',true,'EnableHigherDerivatives',true)

[0034] Among them, YRealLoss represents the cross entropy value of the real terminal interference noise data passing through the discriminator, YReal is the terminal interference noise data in the small batch interference noise sample data; T is the noise classification label value corresponding to YReal; lossD represents the loss value of the discriminator, netD.Learnables represents the learnable parameters of the discriminator network, crossentropy() cross entropy calculation function; dlgradient() neural network gradient calculation function.

[0035] Further, in step S5, the parameters of the generator and the discriminator are updated as follows:

[0036] [netG, trailingAvgG, trailingAvgSqG] = adamupdate(netG, gradientsG, trailingAvgG, trailingAvgSqG, iteration, learnRate, gradientDecayFactor, squaredGradientDecayFactor).

[0037] Among them, netG represents the update generator network, trailingAvgG represents the moving average of the parameter gradient, trailingAvgSqG represents the moving average of the square gradient of the parameter, iteration is the number of iterations, learnRate represents the global learning rate, gradientDecayFactor is the gradient decay factor, squaredGradientDecayFactor is the square gradient decay factor, and adamupdate() represents the parameter update calculation function for deep learning using the Adam algorithm.

[0038] The beneficial effects of the present invention are:

[0039] First: In generative adversarial neural networks, generator training has always been a difficult problem. The essence of generative adversarial neural networks is to require that the data feature distribution generated by the generator has the same features as the training data. In generative adversarial neural networks, the role of the discriminator is to distinguish to the greatest extent whether the data comes from the training data or the generator. In a multi-category data type system, each category of data corresponds to a discriminator, which is conducive to improving the discriminator's identification efficiency. In the generation of multi-terminal interference noise data, the performance of the generative adversarial neural network using multiple discriminators is 15% higher than that of the single discriminator generative adversarial neural network.

[0040] Second: Multi-discriminator generative adversarial neural networks are suitable for multi-terminal interference noise generation. Multi-terminal interference noise training data sets can be classified using interference terminal data, so directly using the number of interference terminals and discriminators for one-to-one correspondence is beneficial to improving the training effect of multi-discriminator generative adversarial neural networks. In actual deployment and use, since only the number of discriminators is increased and the generator does not need to be modified, there is no increase in actual deployment complexity.

[0041] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0043] Figure 1 A schematic diagram of the overall process of a multi-terminal interference generation method for a multi-discriminator generative adversarial network;

[0044] Figure 2 A simplified schematic diagram of a multi-terminal interference generation system for a multi-discriminator generative adversarial network;

[0045] Figure 3 Schematic diagram of the structure of a generative adversarial neural network with multiple discriminators and a single discriminator under the embodiment, wherein: Figure 3 a is the four discriminators, Figure 3 b is the discriminator;

[0046] Figure 4 Schematic diagram of the training data generation process for multi-discriminator generative adversarial neural networks;

[0047] Figure 5 Schematic diagram of the training process of a multi-discriminator generative adversarial neural network;

[0048] Figure 6 This is a diagram of the training progress of the classifier;

[0049] Figure 7 It is a schematic diagram of the performance analysis results of the classifier;

[0050] Figure 8 Schematic diagram of the training progress of a generative adversarial neural network for a single discriminator;

[0051] Fig. 9 Schematic diagram of the training progress of a multi-discriminator generative adversarial neural network.

[0052] Fig.10 Generate variance and mean distribution maps of interference noise data for a single discriminator generative adversarial neural network;

[0053] Fig.11 Generate variance and mean distribution maps of interference noise data for multi-discriminator generative adversarial neural networks;

[0054] Fig.12 Generate time and frequency domain plots of interference noise data for a single discriminator generative adversarial neural network;

[0055] Fig.13 Generate time and frequency domain plots of interference noise data for multi-discriminator generative adversarial neural networks;

[0056] Fig.14 Schematic diagram of the accuracy of the interference noise data generated by the Generative Adversarial Neural Network for the single classifier;

[0057] Fig.15 Schematic diagram of the results of the accuracy of generating interference noise data for a multi-discriminator generative adversarial neural network. DETAILED DESCRIPTION

[0058] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0059] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0060] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0061] See also Figure 1 to Figure 15 , which is a multi-terminal interference generation system and method for a multi-discriminator generative adversarial network.

[0062] Example

[0063] The present invention provides a multi-terminal interference generation method for a multi-discriminator generative adversarial network. Specifically, Figure 1 As shown, it includes the following steps:

[0064] S1. According to the multi-terminal interference noise data generation requirements, an interference noise sample data set TrainDataSet (hereinafter referred to as S) is constructed, wherein the interference noise sample data set S at least includes terminal interference noise data dsTrainData (hereinafter referred to as ds), interference noise classification code dcTrainData (hereinafter referred to as dc), and noise classification label value dTTrainData (hereinafter referred to as dt);

[0065] S2. According to the noise classification label value dt, the interference noise sample data set S is classified to obtain a number of small batch interference sample noise data TrainMiniBatchData (hereinafter referred to as B) with the same noise classification label value LabelHotCode (hereinafter referred to as lc). All small batch interference sample noise data B constitute the training set TrainMiniBatchDataSet (hereinafter referred to as T);

[0066] S3, input the small batch interference sample noise data B in the training set T into the generator NetG (denoted as G) to obtain the terminal interference noise data YGenerated (denoted as YG), and input the terminal interference noise data YG into the discriminator NetD (denoted as D) corresponding to the noise classification label value lc of the small batch interference sample noise data B to calculate the generator gradient parameter gradientsG (denoted as g_G);

[0067] S4, input the terminal interference noise data NoiseSubCarrier (denoted as nc) and the noise classification label value lc in the small batch interference sample noise data B into the corresponding discriminator D, and calculate the discriminator gradient parameter grandientsD (denoted as g_D);

[0068] S5. Update the parameters of the generator G and the corresponding discriminator D respectively according to the calculated generator gradient parameter g_G and discriminator gradient parameter g_D;

[0069] S6. Use the generator G with updated parameters to generate a set of terminal interference noise data, and classify the terminal interference noise data through the classifier. If the classification accuracy reaches the threshold, the training is completed, and the current generator G is saved and used for interference data generation; if the classification accuracy does not reach the threshold, return to step S3 to repeat the training until the threshold is reached.

[0070] The present invention also provides a multi-terminal interference generation system for a multi-discriminator generative adversarial network, such as Figure 2 As shown, it includes: at least one interference noise sample data set module, at least one sample data type classifier, several discriminators, at least a generator and at least one classifier, wherein:

[0071] The interference noise sample data set module completes the generation of multi-terminal interference noise sample data, and adds interference noise classification codes and noise classification label values ​​to multi-terminal interference noise data. Each type of terminal interference noise data has a unique corresponding interference noise classification code and noise classification label value. The array formed by the terminal interference noise data NoiseSubCarrier is recorded as dsTrainData; the array synthesized by the interference noise classification code Clabel array is dcTrainData; the array formed by the noise classification label value LabelHotCode is recorded as dTTrainData. The data dcTrainData, dsTrainData, and dTTrainData constitute a complete interference noise sample data set, recorded as TrainDataSet.

[0072] The sample data type classification module trains different discriminators NetD in the multi-discriminator generative adversarial neural network according to the noise classification label value LabelHotCode in the interference noise sample data set TrainDataSet.

[0073] The discriminator module consists of multiple discriminators NetD, each discriminator NetD corresponds to a noise type specified by a noise classification label value, and the number of discriminators NetD and noise classification label value LabelHotCode is the same.

[0074] The generator module is trained according to different noise classification label values ​​LabelHotCode in the interference noise sample dataset TrainDataSet and different discriminators NetD. The trained generator NetG is used to generate terminal interference noise data.

[0075] The classifier module uses the data in the interference noise sample data set TrainDataSet to train the classifier neural network NetClassifier to correctly identify the noise classification label value corresponding to the interference noise data. It is used to determine the quality of the interference noise data generated by the generator NetG in the neural network.

[0076] The generator NetG consists of an input layer in layer, a classification layer labels layer, a deconvolution layer 1, a deconvolution layer 2, a deconvolution layer 3, a deconvolution layer and an output layer. Among them, the size of the deconvolution layer is 4x1, and the number of channels is 64, 32, 8, and 8 respectively; several deconvolution layers are used for the deconvolution layer to map the input data features from the low-dimensional space back to the high-dimensional space, so as to facilitate the reconstruction of the input signal details and convert the random noise into a high-quality output signal; among them, the input layer in is used to input a random number vector (Z), corresponding to a set of random bit sequences with a length of 100, and the classification layer labels layer is used to input the classification code (Clabel) of the terminal interference noise data. The output layer is completed by the Tanh function, which is used to output the interference noise data (X) generated by the generator NetG. Here, the random number vector consists of "1" or "0" and its length is 100 bits. The generated interference noise data (X) is a complex value with a length of 128. Input different classification codes (Clabel) of terminal interference noise data to obtain different types of terminal interference noise data. Since the input random number vector (Z) is different, the generated terminal interference noise data is also different. The structure of the generator NetG deep neural network is shown in Table 1:

[0077] Table 1

[0078]

[0079]

[0080] The discriminator NetD consists of an input layer imageinput layer, a classification layer labels layer, convolution layer 1, convolution layer 2, convolution layer 3, convolution layer and output layer. The size of the convolution layer is 4x1, and the number of channels is 16, 32, 64, and 128 respectively. The convolution layer is used to extract the features of the terminal interference noise data (X) using convolution operations, effectively capturing the spatial structure and local correlation of the terminal interference noise data; the input layer imageinput is used to input the terminal interference noise data (X), corresponding to the terminal interference noise data (NoiseSubCarrier), and the labels layer is used to input the terminal interference noise classification code (Clabel). The output layer is composed of a softMax function, which is the probability value of the discriminator (netD) to determine each noise type. According to the input terminal interference noise classification code (Clabel), the corresponding terminal interference noise label value (LabelHotCode) can be searched, which is used to calculate the loss value of the discriminator. The deep neural network structure of the discriminator NetD is shown in Table 2:

[0081] Table 2

[0082]

[0083]

[0084] The noise classifier NetClassifier consists of an input layer imageinput, convolution layer 1, convolution layer 2, convolution layer 3, convolution layer and output layer; the input layer imageinput is used to input terminal interference noise data (X), terminal interference noise data (NoiseSubCarrier), and the labels layer is used to input the terminal interference noise classification code (C), that is, the classification code of the terminal interference noise data (Clabel). The output layer is composed of a softMax function, which is the probability value of the noise classifier NetClassifier for determining each type of noise. The structure of the classifier and the discriminator is the same, but the parameters are different. After training, the parameters of the two are also different. The deep neural network structure of the classifier NetClassifier is shown in Table 3:

[0085] Table 3

[0086]

[0087]

[0088] In this embodiment, the generated adversarial neural network structure is as follows Figure 3 As shown. Figure 3 a is a multi-discriminator generated adversarial neural network, which is a network structure specifically implemented by the present invention. Figure 3In a, there are four discriminators, namely NetD1, NetD2, NetD8 and NetD16, which correspond to the types of interference noise data formed by 1 interference terminal, 2 interference terminals, 8 interference terminals and 16 interference terminals in the multi-terminal interference noise sample data set. In this embodiment, the neural network structures of NetD1, NetD2, NetD8 and NetD16 are the same. Figure 3 In the single discriminator generative adversarial neural network of b, there is only one discriminator NetD. The following describes this embodiment in several processes.

[0089] In this embodiment, the interference noise sample data set (TrainDataSet) consists of terminal interference noise data generated by 1 interference terminal, 2 interference terminals, 8 interference terminals and 16 interference terminals. Each number of terminal interference noise data samples is 2048, and the length of each interference noise data (frequency domain, complex representation) is 128, where:

[0090] The interference noise data of a terminal is expressed as: interference noise data (dsTrainSuData), interference noise classification code (dcTrainSuData), noise classification label value (dTTrainSuData);

[0091] The interference noise data of the two terminals are expressed as: interference noise data (dsTrainMu2Data), interference noise classification code (dcTrain Mu2Data), and noise classification label value (dTTrain Mu2Data);

[0092] The interference noise data of 8 terminals are expressed as: interference noise data (dsTrainMu8Data), interference noise classification code (dcTrain Mu8Data), noise classification label value (dTTrain Mu8Data);

[0093] The interference noise data of 16 terminals are expressed as: interference noise data (dsTrainMu16Data), interference noise classification code (dcTrain Mu16Data), and noise classification label value (dTTrain Mu16Data).

[0094] In the interference noise sample data set (TrainDataSet), dsTrainXData records multi-terminal interference noise data (NoiseSubCarrier), which comes from multi-terminal interference noise data generated by Matlab simulation (the data can also be obtained by actual testing, and in this embodiment, Matlab model simulation is used to generate the data), representing the interference to the frequency domain subcarrier, with a length of 128.

[0095] dcTrain XData records the classification code (Clabel) of multi-terminal interference noise data, randomly generates a random bit stream of length 100, and generates 1 classification code stream for each noise type, which remains unchanged during the training and use of adversarial generative neural networks.

[0096] In this embodiment, it is recorded as: single-terminal suClabel, 2-terminal mu2Clabel, 8-terminal mu8Clabel, 16-terminal mu16Clabel.

[0097] dTTrain XData records the label value (LabelHotCode) of multi-terminal interference noise data, and defines a label value for each noise type. In this embodiment, there are 4 types of interference noise, so the label value (LabelHotCode) is 4. The length is 4 bits, and the single terminal label value a_suLabelHotCode is [0 0 0 1], the 2 terminal label value a_mu2LabelHotCode is [0 0 10], the 8 terminal label value a_mu8LabelHotCode is [0 1 0 0], and the 16 terminal label value a_mu16LabelHotCode is [10 0 0]. X in (dsTrainXData, dcTrain XData, dTTrainXData) represents Su, Mu2, Mu8, and Mu16.

[0098] In the process of training the multi-discriminator generative adversarial neural network, in order to improve the training effect, small batch data are used for training. In this embodiment, Figure 4 The schematic diagram of the training data generation process of the multi-discriminator generative adversarial neural network shown in the figure shows that in the interference noise sample data set (TrainDataSet), the single-terminal, 2-terminal, 8-terminal and 16-terminal interference noise data are randomly sorted, so the small batch training data taken out from it may include interference noise data with different terminal label values, which does not meet the training requirements of the multi-discriminator generative adversarial neural network. Therefore, it is necessary to classify the data of the interference noise sample data set (TrainDataSet) and divide it into training set data of small batch size so that the terminal label value (LabelHotCode) in each small batch of data is the same. Specifically, it includes the following steps:

[0099] Step 1: Interference noise sample data set (TrainDataSet), including single terminal, 2 terminals, 8 terminals and 16 terminals interference noise data. In this embodiment, the number of interference noise samples for each terminal is 2048. Each type of terminal interference noise type has interference noise data (dsTrainXData), interference noise classification code (dcTrainXData), and noise classification label value (dTTrainXData);

[0100] Step 2: Classify according to the label value (LabelHotCode) in the noise classification label value (dTTrainXData) in the interference noise sample data set (TrainDataSet). In this embodiment, it is divided into 4 types of training data TrainXDataSet, where X represents single terminal Su, 2 terminals Mu2, 8 terminals Mu8 and 16 terminals Mu16. The full names are TrainSuDataSet, TrainMu2DataSet, TrainMu8DataSet, and TrainMu16DataSet respectively.

[0101] Step 3: According to the training requirements of the multi-discriminator generative adversarial neural network, that is, the amount of training data each time is miniBatchSize, TrainXDataSet is grouped to generate TrainXMiniBatchDataSet, where X represents single terminal Su, 2 terminals Mu2, 8 terminals Mu8 and 16 terminals Mu16. In this embodiment, miniBatchSize=4 is selected for training, so the group size of each label value classification is 2048 / 16=128.

[0102] Step 4: Merge the four categories of TrainXMiniBatchDataSet together to form a complete small batch interference noise sample dataset (TrainMiniBatchDataSet).

[0103] In this embodiment, the multi-discriminator generative adversarial neural network has four discriminators, namely NetD1, NetD2, NetD8 and NetD16.

[0104] In the use of training data, different discriminators are selected according to the label values ​​in the small batch interference noise sample data set. The training data of the single terminal label value a_suLabelHotCode[0 0 0 1] corresponds to the NetD1 discriminator, the training data of the 2-terminal label value a_mu2LabelHotCode[0 0 1 0] corresponds to the NetD2 discriminator, the training data of the 8-terminal label value a_mu8LabelHotCode[0 1 0 0] corresponds to the NetD8 discriminator, and the training data of the 16-terminal label value a_mu16LabelHotCode[1 0 0 0] corresponds to the NetD16 discriminator. During the training process, in order to achieve the training effect, the discriminator is randomly selected, so random sorting is required in the small batch interference noise sample data set (TrainMiniBatchDataSet). The specific operation process is as follows Figure 5 As shown, it includes the following specific steps:

[0105] Step 1: The small batch interference noise sample data set (TrainMiniBatchDataSet) is a grouping of the interference noise sample data set (TrainDataSet) according to the size of miniBatchSize. Since there is no randomness in the grouped data, the selection of the discriminator during the training process loses randomness, resulting in poor training results. Therefore, before training, the sample data of the small batch interference noise sample data set (TrainMiniBatchDataSet) is first randomly sorted in the minBatchSize size grouping unit. In this embodiment, miniBatchSize is 4.

[0106] Step 2: Take out minBatchSize terminal interference noise sample data from the small batch interference noise sample data set (TrainMiniBatchDataSet) in sequence, that is, a TrainMiniBatchData (the label values ​​of these sample data are the same) data.

[0107] Step 3: Generate a set of terminal interference data YGenerated using the generator (NetG) according to the classification label value (LabelHotCode) in TrainMiniBatchData. Then send YGenerated to the corresponding discriminator, and use cross entropy to calculate the cross loss value YGeneratedLoss of the generated interference noise data. The specific calculation method is:

[0108] YGeneratedLoss=-crossentropy(YGenerated,T,'WeightsFormat','UC');

[0109] Among them, T is the noise classification label value (LabelHotCode) corresponding to YGenerated.

[0110] Step 4: Find the corresponding discriminator according to the classification label value (LabelHotCode) in TrainMiniBatchData, send TrainMiniBatchData to the corresponding discriminator, and use cross entropy to calculate the cross loss value YRealLoss of the sample interference noise data. The specific calculation formula is:

[0111] YRealLoss=-crossentropy(YReal,T,'WeightsFormat','UC');

[0112] Among them: YReal is the terminal interference noise data (NoiseSubCarrier) in the small batch interference noise sample data (TrainMiniBatchData); T is the noise classification label value (LabelHotCode) corresponding to YReal; crossentropy() cross entropy calculation function;

[0113] Step 5: Use the calculated cross loss values ​​YRealLoss and YGeneratedLoss to calculate the loss value and gradient parameters of the generator (NetG) and the loss value and gradient parameters of the corresponding discriminator, and then use the gradient parameters to update the corresponding generator (NetG) and discriminator (NetD) parameters. The specific calculation formula is:

[0114] lossD=YRealLoss+(1-YGeneratedLoss);

[0115] gradientsD=dlgradient(lossD,netD.Learnables,'RetainData',true,'EnableHigherDerivatives',true);

[0116] [netD, trailingAvgNetD, trailingAvgSqNetD] = adamupdate(netD, gradientsD, trailingAvgNetD, trailingAvgSqNetD, netDiteration, learnRate, gradientDecayFactor, squaredGradientDecayFactor);

[0117] lossG = YGeneratedLoss;

[0118] gradientsG=dlgradient(lossG,netG.Learnables,'RetainData',true,'EnableHigherDerivatives',true);

[0119] [netG, trailingAvgG, trailingAvgSqG] = adamupdate(netG, gradientsG, trailingAvgG, trailingAvgSqG, iteration, learnRate, gradientDecayFactor, squaredGradientDecayFactor);

[0120] Among them: dlgradient() is the Matlab gradient calculation function; adamupdate() function is the neural network parameter update function in Matlab.

[0121] Step 6: Generate a set of terminal interference noise data using the generator (NetG) after updating the parameters. In this embodiment, 128 terminal interference noise data are generated, and then the classifier (NetGlassifier) ​​is used to obtain the probability of correct classification. If the correct probability meets the requirement, the training of the adversarial generative neural network ends, otherwise return to step 2 to continue training. After the small batch interference noise sample data set (TrainMiniBatchDataSet) is used up, return to the first step to continue.

[0122] Each training of a small batch of interference noise sample data set (TrainMiniBatchDataSet) is completed once, which is called an Epoch. In this embodiment, the maximum Epoch value is set to 128.

[0123] This embodiment also performs a performance analysis on the finally obtained generative adversarial neural network. In the present invention, the classifier (NetClassifier) ​​also adopts a deep neural network structure, and uses the interference noise sample data set (TrainDataSet) in this embodiment to prepare for training. After the training, the classification accuracy of the classifier (NetClassifier) ​​can reach more than 99%, which can be used to evaluate the performance of the generative adversarial neural network.

[0124] (1) Classifier (NetClassifier) ​​module training progress chart and performance evaluation

[0125] Figure 6This is a progress chart of classifier (NetClassifier) ​​training. The Accuracy chart indicates the prediction accuracy of the classifier. Red indicates prediction using validation data, and blue indicates prediction using training data. The FScore chart indicates a performance indicator that takes into account both precision and recall. The Loss chart indicates the inconsistency parameter between the model prediction value and the actual value. From the Accuracy chart, the Fscore chart, and the Loss chart, we can conclude that the model has converged, and the accuracy of the classifier (NetClassifier) ​​for the test set and validation set has reached more than 99%.

[0126] Figure 7 Further analysis of the classifier (NetClassifier) ​​performance shows that the classifier has a good classification effect on 16-terminal and 8-terminal interference noise, with an accuracy rate of 100%. The classification of single-terminal and 2-terminal is slightly worse, with a single accuracy rate of 99.2% and 98.6%. The overall average correct classification probability reaches 99%.

[0127] according to Figure 6 and Figure 7 It shows that the trained classifier (NetClassifier) ​​can be used to evaluate the quality of the interference noise data generated by the generator (NetG).

[0128] (2) Generate Adversarial Deep Neural Network Module Training Progress Chart

[0129] In this embodiment, if Figure 8 and Fig. 9 As shown, a comparison of the training progress of single-discriminator and multi-discriminator generative adversarial neural networks is provided.

[0130] exist Figure 8 A single discriminator generates an adversarial neural network, and the classifier (NetClassifier) ​​checks that the generator (NetG) has an accuracy rate of no more than 70% at best, and is usually only around 62%.

[0131] In the single discriminator generative adversarial neural network, miniBatchSize is selected as 64. After multiple parameter verifications, the best effect is achieved when miniBatchSize is 64.

[0132] exist Fig. 9 In the case of multi-discriminator generative adversarial neural networks, the classifier (NetClassifier) ​​checks the generator (NetG) with an accuracy of no more than 85% at best, and usually more than 72%. In the case of a single discriminator generative adversarial neural network, a miniBatchSize of 4 is selected. After multiple parameter verifications, the effect is best when the miniBatchSize is 4.

[0133] from Figure 8 and Fig. 9 Compared with the generator (NetG), the multi-discriminator neural network has a performance improvement of about 15% over the single-discriminator neural network.

[0134] (3) Comparison of mean and variance distribution of interference noise data generated by the generator (NetG)

[0135] In generative adversarial neural networks, in order to evaluate the quality of data generated by the generator (NetG), a comparative explanation of the sdt / mean distribution is often used.

[0136] Fig.10 The interference noise data generated for the single discriminator generative adversarial neural network is analyzed and compared with the sample noise data distribution. In terms of the 1-terminal, 2-terminal, 8-terminal and 16-terminal interference noise distributions, the mean and variance of the sample interference noise data and the interference noise data generated by the generator are well consistent.

[0137] Fig.11 The interference noise data generated for the multi-discriminator generative adversarial neural network is analyzed and compared with the sample noise data distribution. In terms of the 1-terminal, 2-terminal, 8-terminal and 16-terminal interference noise distributions, the mean and variance of the sample interference noise data and the interference noise data generated by the generator are well consistent.

[0138] from Fig.10 and Fig.11 Compared with the generator (NetG), the interference noise data generated by the multi-discriminator and single-discriminator GANs are consistent with the sample data distribution. The interference noise data generated by the single-discriminator GANs are slightly better than those generated by the single-discriminator GANs.

[0139] (4) Comparison of the time domain and frequency domain of the interference noise data generated by the generator (NetG)

[0140] In this embodiment, a single discriminator and a multi-discriminator generative adversarial neural network are used to generate interference noise data of a single terminal, 2 terminals, 8 terminals and 16 terminals, and a comparison is made from the frequency domain and the time domain.

[0141] Fig.12 In the method, a single discriminator generates an adversarial neural network to generate interference noise data. The more interference terminals there are, the more impact the subcarriers sent will have, both in the time domain and the frequency domain.

[0142] from Fig.12 and Fig.13Compared with the generator (NetG), the interference noise data generated by the multi-discriminator and single-discriminator generative adversarial neural networks are consistent with the sample data distribution, but it is impossible to accurately determine the quality of the interference noise data generated by the multi-discriminator and single-discriminator generative adversarial neural networks.

[0143] (5) Quality analysis of interference noise data generated by the generator (NetG)

[0144] In the generative adversarial deep neural network, the above-mentioned mean and variance distribution diagrams of the generated interference noise data, as well as the time domain and frequency domain diagrams are used for comparison, which can only perform preliminary comparative analysis. In this embodiment, a classifier (NetClassifier) ​​is used for comparison, which can more accurately illustrate the quality of the generated interference noise data.

[0145] exist Fig.14 In the experiment, the single discriminator generative adversarial neural network randomly generated 128 interference noise data, and then the classifier was used for classification and recognition, with an accuracy rate of 65.6%.

[0146] exist Fig.15 In the experiment, the multi-classifier generative adversarial neural network randomly generates 128 interference noise data, and then the classifier is used for classification and recognition, with an accuracy rate of 80%.

[0147] from Fig.14 and Fig.15 Compared with the generator (NetG), the interference noise data generated by the multi-discriminator and single-discriminator adversarial neural network are classified and judged by the classifier (NetClassifier). The multi-discriminator is about 15% better than the single-discriminator adversarial neural network in generating interference noise data.

[0148] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A multi-terminal interference generation system based on a multi-discriminator generative adversarial network, Features: It includes: at least one interference noise sample data set module, at least one sample data type classifier, several discriminators, at least a generator and at least one classifier, wherein, The interference noise sample data set module is used to generate multi-terminal interference noise sample data, and add interference noise classification codes and noise classification label values ​​to the multi-terminal interference noise data. Each type of terminal interference noise data is set with a unique corresponding interference noise classification code and noise classification label value; The sample data type classification module inputs the data into the discriminator of the corresponding label value for training according to the noise classification label value in the interference noise sample data set; The discriminator module includes several discriminators, each of which corresponds to a noise type specified by a noise classification label value; The generator module is trained according to different noise classification label values ​​and different discriminators in the interference noise sample data set, and the trained generator is used to generate terminal interference noise data; The small batch interference sample noise data in the training set is input into the generator to obtain the terminal interference noise data, and the terminal interference noise data is input into the discriminator corresponding to the noise classification label value of the small batch interference sample noise data to calculate the generator gradient parameter; the terminal interference noise data and the noise classification label value in the small batch interference sample noise data are input into the corresponding discriminator to calculate the discriminator gradient parameter; the parameters of the generator and the corresponding discriminator are updated respectively according to the calculated generator gradient parameter and discriminator gradient parameter; The classifier module is used to correctly identify the noise classification label value corresponding to the interference noise data, and to determine the quality of the interference noise data generated by the generator in the neural network.

2. According to the multi-discriminator generative adversarial network multi-terminal interference generation system of claim 1, Features: The generator module includes an input layer in, a classification layer labels, several deconvolution layers and an output layer, wherein the deconvolution layer is used to map the input data features from the low-dimensional space back to the high-dimensional space; the input layer is used to input the random number vector; the classification layer is used to input the classification code of the terminal interference noise data; the output layer is completed by the Tanh function and is used to output the generated interference noise data.

3. According to the multi-discriminator generative adversarial network multi-terminal interference generation system of claim 1, Features: The discriminator module includes an input layer imageinput, a classification layer labels, several convolutional layers and an output layer, wherein the convolutional layer is used to extract the characteristics of the terminal interference noise data through convolution operations; the input layer is used to input the terminal interference noise data, and the classification layer is used to input the terminal interference noise classification code; the output layer is composed of a softMax function, which outputs the probability value of each noise type determined by the discriminator.

4. According to the multi-discriminator generative adversarial network multi-terminal interference generation system of claim 1, Features: The classifier module consists of an input layer imageinput, several convolutional layers and an output layer; the input layer is used to input terminal interference noise data, and the labels layer is used to input terminal interference noise classification codes, that is, the classification codes of terminal interference noise data; the output layer is composed of a softMax function, which is the probability value of the noise classifier to determine each noise type.

5. A multi-terminal interference generation method for a multi-discriminator generative adversarial network, Features: It includes the following steps: S1. Construct an interference noise sample data set according to the multi-terminal interference noise data generation requirements, wherein the interference noise sample data set includes at least terminal interference noise data, interference noise classification code, and noise classification label value; S2. Classify the interference noise sample data set according to the noise classification label value to obtain a number of small batches of interference sample noise data with the same noise classification label value, and all small batches of interference sample noise data constitute a training set; S3, inputting the small batch interference sample noise data in the training set into the generator to obtain terminal interference noise data, and inputting the terminal interference noise data into the discriminator corresponding to the noise classification label value of the small batch interference sample noise data to calculate the generator gradient parameter; S4, inputting the terminal interference noise data and the noise classification label value in the small batch interference sample noise data into the corresponding discriminator, and calculating the discriminator gradient parameter; S5. Update the parameters of the generator and the corresponding discriminator respectively according to the calculated generator gradient parameters and discriminator gradient parameters; S6. Use the generator with updated parameters to generate a set of terminal interference noise data, and classify the terminal interference noise data through the classifier. If the classification accuracy reaches the threshold, the training is completed, and the current generator is saved and used for interference data generation; if the classification accuracy does not reach the threshold, return to step S3 to repeat the training until the threshold is reached.

6. According to the multi-terminal interference generation method of a multi-discriminator generative adversarial network according to claim 5, Features: In step S3, the generator gradient parameters are calculated as follows: YGeneratedLoss=-crossentropy(YGenerated,T,'WeightsFormat','UC') lossG = YGeneratedLoss gradientsG = dlgradient(lossG, netG.Learnables,'RetainData',true,'EnableHigherDerivatives',true) where YGeneratedLoss represents the cross entropy value of the generated terminal interference noise data passing through the discriminator, YGenerated represents the terminal interference noise data, and T is the noise classification label value corresponding to YGenerated; 'WeightsFormat' represents the option for specifying the dimension of the weight parameter, 'UC' represents that the elements in the weight vector correspond to the channel dimension of the predicted value, and crossentropy() is the cross entropy calculation function; lossG represents the loss value of the terminal interference noise data generated by the generator obtained by the discriminator; gradientsG is the generator gradient parameter, netG.Learnables is the learnable parameter of the generator network, 'RetainData',true represents the use of intermediate data retained for calculating gradients for calculation, 'EnableHigherDerivatives',true represents the enablement of higher-order derivative calculations, and dlgradient() is the neural network gradient calculation function.

7. The multi-terminal interference generation method of a multi-discriminator generative adversarial network according to claim 6, Features: In step S4, the discriminator gradient parameters are calculated as: YRealLoss=-crossentropy(YReal,T,'WeightsFormat','UC') lossD=YRealLoss+(1-YGeneratedLoss) gradientsD = dlgradient(lossD, netD.Learnables,'RetainData',true,'EnableHigherDerivatives',true) where YRealLoss represents the cross entropy value of the real terminal interference noise data passing through the discriminator, YReal is the terminal interference noise data in the small batch interference noise sample data; T is the noise classification label value corresponding to YReal; lossD represents the loss value of the discriminator, netD.Learnables represents the learnable parameters of the discriminator network, crossentropy() is the cross entropy calculation function; dlgradient() is the neural network gradient calculation function.

8. The multi-terminal interference generation method of a multi-discriminator generative adversarial network according to claim 7, Features: In step S5, the parameters of the generator and discriminator are updated as follows: [netG, trailingAvgG, trailingAvgSqG] = adamupdate(netG, gradientsG, trailingAvgG, trailingAvgSqG, iteration, learnRate, gradientDecayFactor, squaredGradientDecayFactor) Among them, netG represents the update generator network, trailingAvgG represents the moving average of the parameter gradient, trailingAvgSqG represents the moving average of the square gradient of the parameter, iteration is the number of iterations, learnRate represents the global learning rate, gradientDecayFactor is the gradient decay factor, squaredGradientDecayFactor is the square gradient decay factor, and adamupdate() represents the parameter update calculation function for deep learning using the Adam algorithm.

Images