A Small-Sample Number of Signal Sources Detection Method Using Semi-Supervised Generative Adversarial Networks

Abstract

The present invention relates to a small-sample number of signal sources detection method applying a semi-supervised generative adversarial network, which includes calculating a sample sampling covariance matrix and its eigenvalues based on the output data of an array antenna; calculating corresponding linear shrinkage coefficients according to the obtained eigenvalues and preprocessing them; using the preprocessed linear shrinkage coefficients to train the semi-supervised generative adversarial network, and then using the trained neural network to predict the number of signal sources. Compared with the traditional eigenvalue-based classification, the method of the present invention uses the linear shrinkage coefficients as the input features of the network to generate more obvious classification features, so as to achieve good classification performance under the conditions of fewer labels and samples and lower signal-to-noise ratio; at the same time, the combination of semi-supervised learning and generative adversarial network is adopted to correct the classifier in an adversarial manner to achieve a performance close to that of full-supervised learning under a small amount of labeled data.

Description

Technical Field

[0001] The present invention belongs to the technical field of array signal processing, and particularly relates to a small-sample number-of-signal-sources detection method applying a semi-supervised generative adversarial network. Background Art

[0002] The detection of the number of signal sources plays an important role in the field of array signal processing such as radar, sonar, wireless communication, and vehicle networking, and is important prior information in many signal space parameter estimation algorithms. At present, many methods have been proposed to solve this problem. Among them, relatively well-known traditional methods include the Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), Minimum Description Length (MDL), Linear Shrinkage based MDL (LS-MDL), AIC based on Random Matrix Theory (RMT-AIC), heuristic Shrinkage Coefficient Detection (SCD heur ) and Two-Step Difference (TSD) based on the linear shrinkage coefficient and other algorithms. However, the above algorithms are all based on theoretical derivations and various assumptions, and their effects in practical applications are very limited.

[0003] In recent years, with the development of Artificial Intelligence (AI) technology, some source number detection algorithms based on Deep Learning (DL) have been proposed. These algorithms can learn the features of a large amount of data through a specially designed Deep Neural Network (DNN) to produce accurate estimates, and have strong robustness in the case of a small number of samples and low signal-to-noise ratio. These methods include eigenvalue-based regression networks and classification networks, which use DNN to learn the mapping relationship between the eigenvalues after forward and backward spatial smoothing techniques and the source number to solve the coherent source number estimation problem, and have achieved better estimation results than traditional methods, especially in the case of low signal-to-noise ratio. On this basis, a Logarithmic Eigenvalue-based Classification Network (LogECNet) is proposed to improve the classification performance of DNN, which uses logarithmic operations on eigenvalues to increase the relative distance between signal and noise eigenvalues. However, logarithmic eigenvalues do not provide a relatively large distance in the case of a small number of samples. In this case, the linear shrinkage coefficient has more obvious features than eigenvalues. In the ideal case, the shrinkage coefficients of its signal and noise are 0 and 1 respectively. Therefore, it is obviously more advantageous to select the linear shrinkage coefficient as the input of the neural network. In addition, most existing neural networks rely on large-scale labeled data for training to obtain excellent performance. However, labels are difficult to obtain in practical applications, and unlabeled data is often encountered. Therefore, a large number of inefficient and costly manual labeling operations are required, increasing the challenge of high-performance source number estimation. In this context, developing a source number estimation method with a small number of labeled data has important application value and practical needs. Accordingly, the present invention combines Semi-Supervised Learning (SSL) with Generative Adversarial Network (GAN) to effectively promote the training of the classifier by using unlabeled data, and can provide better detection performance in the case of a small number of labels. Summary of the Invention

[0004] In order to overcome the deficiencies of the existing deep learning-based source number estimation algorithms, the present invention proposes a small-sample source number detection method applying semi-supervised generative adversarial network in combination with the linear shrinkage coefficient under the general asymptotic system. This method is used to solve the problem of insufficient general applicability of the existing technology, and realizes the accurate estimation of the source number under the conditions of a large-scale array, a small number of samples, and only a small amount of labeled data.

[0005] The technical solution adopted by the present invention is a small-sample source number detection method applying a semi-supervised generative adversarial network, and the method comprises the following steps:

[0006] S1. Set d far-field sources, where the d far-field sources are incident on an array antenna composed of m antennas at angles q 1 ,…,q d , and obtain the output data of the array antenna at time t: x(t) = As(t) + n(t), where A represents an m×d-dimensional array steering matrix, s(t) represents a d×1-dimensional independently and identically distributed Gaussian signal, and n(t) represents an m×1 noise vector; sample the output data x(t) n times, collect n sampling samples and calculate the covariance matrix of the sampling samples, and its expression is: And calculate its descending eigenvalues l 1 ,…,l m , where m represents the number of array antennas;

[0007] S2. Based on the m descending eigenvalues calculated in step S1, calculate the corresponding linear shrinkage coefficients and perform preprocessing on them. The specific process is as follows:

[0008] S2.1. Based on the calculated m-th descending eigenvalue l i , i = 1,…,m, calculate the corresponding linear shrinkage coefficients: , where j = 0,…,m - 1. Since may be greater than 1, use as the final linear shrinkage coefficient;

[0009] S2.2. When the number of samples n is less than the number of array elements m, process the obtained in step S2.1 to keep the lengths of the linear shrinkage coefficients consistent, and its expression is: Then normalize the linear shrinkage coefficients to obtain a preprocessed linear shrinkage coefficient vector, and its expression is: where r represents an m-dimensional vector composed of linear shrinkage coefficients, 1 m represents an m-dimensional all-ones vector, and max(·) and min(·) respectively represent returning the maximum and minimum values in the vector;

[0010] S3. Based on the preprocessed linear shrinkage coefficients generated in step S2, generate a training data set, and use the training data set to train a semi-supervised generative adversarial network to obtain a trained semi-supervised generative adversarial network;

[0011] S4. Use the trained semi-supervised generative adversarial network to predict the number of source signals.

[0012] The beneficial effects of the present invention are as follows: By adopting the above method for detecting the number of small-sample signal sources using a semi-supervised generative adversarial network, a dataset is constructed by utilizing the structural characteristics of the shrinkage coefficient. Meanwhile, a suitable neural network is constructed and trained, and the linear shrinkage coefficient is used as the input feature of the network to generate more obvious classification features. This method realizes the accurate estimation of the number of signal sources under the conditions of a large-scale array, a small number of samples, and only a small amount of labeled data, achieving a performance close to that of full-supervised learning.

[0013] Preferably, the specific process of generating the training dataset is as follows: Based on the preprocessed linear shrinkage coefficient generated in step S2, a labeled linear shrinkage coefficient training dataset and an unlabeled linear shrinkage coefficient training dataset are generated. Among them, the labeled linear shrinkage coefficient training dataset is expressed as: The unlabeled linear shrinkage coefficient training dataset is expressed as: where W and J respectively represent the quantities of the corresponding datasets.

[0014] Preferably, in step S3, the semi-supervised generative adversarial network includes a discriminative network and a generative network. The discriminative network includes a classifier and a discriminator. The specific process of training the semi-supervised generative adversarial network using the training dataset is as follows: Set the number of training rounds, and input the training dataset into the semi-supervised generative adversarial network for training; in each round of training, the training method is: the generative network is trained once and the discriminative network is trained twice; among them, the training process of the discriminative network is: randomly extract M groups of labeled data from the training dataset W for training the classifier in the discriminative network, and then update the parameters of the classifier through the cross-entropy loss function; randomly extract M groups of unlabeled data from the training dataset J for training the discriminator in the discriminative network, and then update the parameters of the discriminator through the discriminator loss function; the training process of the generative network is: select M groups of m-dimensional random vectors subject to a Gaussian distribution, input the random vectors into the generator of the generative network, generate generated data after passing through the generator, input the generated data into the discriminator for adversarial training, and then update the parameters of the generative network through the loss function of the generative network.

[0015] Preferably, in step S3, the generative network includes an input layer A, a hidden layer A, and an output layer A; the generative network is a DNN network with m neurons in each layer; among them, the activation function of the hidden layer A is ReLU, and the activation function of the output layer A is Tanh. The loss function of the generative network is expressed as:

[0016] where D(·) and G(·) respectively represent the forward propagation of the discriminative network and the generative network, and the forward propagation of the discriminative network only considers the output of the discriminator.

[0017] Preferably, in step S3, the discrimination network includes an input layer B, a hidden layer B, and an output layer B. Among them, the discrimination network is divided into an upper branch and a lower branch. The upper branch consists of two fully connected layers and outputs the classification probability of the number of signal sources with K dimensions. The lower branch consists of one fully connected layer and outputs one-dimensional data. The activation function of the hidden layer B of the discrimination network is LeakyReLU with a negative axis slope of 0.2. The output layer of the upper branch uses the Softmax activation function. The loss function of the discrimination network can be expressed as: where α represents a non-negative hyperparameter, represents the network weights and biases; L sup represents the classification loss of supervised learning, which is expressed as: L unsup represents the discrimination loss of unsupervised learning, which is expressed as:

[0018] where,

[0019] Preferably, the specific process of predicting the number of signal sources using the trained semi-supervised generative adversarial network is as follows: Input the test set into the trained semi-supervised generative adversarial network to predict the number of signal sources. Use the linear shrinkage coefficient in the test set as the input of the discrimination network of the trained semi-supervised generative adversarial network. Its classifier outputs a probability distribution. According to the probability distribution, find the maximum probability. The index of the maximum probability corresponds to the estimated number of signal sources. Description of the Drawings

[0020] Figure 1 is the overall block diagram of the network model design of a small-sample number-of-signal-sources detection method applying a semi-supervised generative adversarial network according to the present invention;

[0021] Figure 2A is the structure diagram of the generative network of a small-sample number-of-signal-sources detection method applying a semi-supervised generative adversarial network according to the present invention;

[0022] Figure 2B is the structure diagram of the discrimination network of a small-sample number-of-signal-sources detection method applying a semi-supervised generative adversarial network according to the present invention;

[0023] Figure 3A is the comparison diagram of the accurate detection probability (Probability) of the number of signal sources varying with the signal-to-noise ratio (SNR) obtained by the number-of-signal-sources estimation method adopted in Scenario 1-1 in Simulation Experiment 1 described in the present invention;

[0024] Figure 3BComparison graph of the accurate detection probability (Probability) of the number of signal sources obtained by the signal source number estimation method adopted in Scenario 1-2 in Simulation Experiment 1 described in the present invention varying with the signal-to-noise ratio (SNR);

[0025] Figure 4A Comparison graph of the accurate detection probability (Probability) of the number of signal sources obtained by the signal source number estimation method adopted in Scenario 1-3 in Simulation Experiment 1 described in the present invention varying with the signal-to-noise ratio (SNR);

[0026] Figure 4B Comparison graph of the accurate detection probability (Probability) of the number of signal sources obtained by the signal source number estimation method adopted in Scenario 1-4 in Simulation Experiment 1 described in the present invention varying with the signal-to-noise ratio (SNR);

[0027] Figure 5A Comparison graph of the accurate detection probability (Probability) of the number of signal sources obtained by the signal source number estimation method adopted in Scenario 1-5 in Simulation Experiment 1 described in the present invention varying with the number of samples (Number of Samples);

[0028] Figure 5B Comparison graph of the accurate detection probability (Probability) of the number of signal sources obtained by the signal source number estimation method adopted in Scenario 1-6 in Simulation Experiment 1 described in the present invention varying with the number of samples (Number of Samples);

[0029] Figure 6A Comparison graph of the accurate detection probability (Probability) of the number of signal sources obtained by the signal source number estimation method adopted in Scenario 2-1 in Simulation Experiment 1 described in the present invention varying with the signal-to-noise ratio (SNR);

[0030] Figure 6B Comparison graph of the accurate detection probability (Probability) of the number of signal sources obtained by the signal source number estimation method adopted in Scenario 2-2 in Simulation Experiment 2 described in the present invention varying with the signal-to-noise ratio (SNR). Detailed implementation manners

[0031] The present invention will be further described below with reference to the accompanying drawings and in combination with specific implementation manners, so that those skilled in the art can implement it according to the text of the specification. The protection scope of the present invention is not limited to this specific implementation manner.

[0032] The present invention relates to a small - sample source number detection method applying a semi - supervised generative adversarial network. This method realizes accurate estimation of the source number under the conditions of a large - scale array, a small number of samples, and only a small amount of labeled data. The small number of samples mentioned in the present invention refers to a quantity ranging from 10 to 100, and the only a small amount of labeled data mentioned in the present invention means that the number of labeled data is less than 50% of the total data set. The method includes the following steps:

[0033] S1. Set d far - field sources, and the d far - field sources are incident on the array antenna composed of m antennas at angles θ 1 , …, θ d . Obtain the output data of the array antenna at time t: x(t)=As(t)+n(t), where A represents an m×d - dimensional array steering matrix, s(t) represents a d×1 - dimensional independent and identically - distributed Gaussian signal, and n(t) represents an m×1 noise vector; sample the output data x(t) n times, collect n sampling samples and calculate the covariance matrix of the sampling samples, and its expression is: And calculate its descending eigenvalues l 1 , …, l m , where m represents the number of array antennas;

[0034] S2. Based on the m descending eigenvalues calculated in step S1, calculate the corresponding linear shrinkage coefficients and pre - process them. The specific process is as follows:

[0035] S2.1. Based on the calculated m - th descending eigenvalue l i , i = 1, …, m, calculate the corresponding linear shrinkage coefficients: , where j = 0, …, m - 1. Since may be larger than 1, use as the final linear shrinkage coefficient;

[0036] S2.2. When the number of samples n is less than the number of array elements m, process the obtained in step S2.1 to keep the length of the linear shrinkage coefficients consistent, and its expression is: Then normalize the linear shrinkage coefficients to obtain the pre - processed linear shrinkage coefficient vector, and its expression is: where r represents an m - dimensional vector composed of linear shrinkage coefficients, 1 m represents an m - dimensional all - 1 vector, and max(·) and min(·) respectively represent returning the maximum value and the minimum value in the vector;

[0037] S3. Based on the pre - processed linear shrinkage coefficients generated in step S2, generate a training data set, and use the training data set for Figure 1Train the semi-supervised generative adversarial network shown to obtain the trained semi-supervised generative adversarial network;

[0038] S4. Use the trained semi-supervised generative adversarial network to predict the number of information sources.

[0039] Furthermore, the specific process of generating the training data set is as follows: Based on the preprocessed linear contraction coefficients generated in step S2, generate a labeled linear contraction coefficient training data set and an unlabeled linear contraction coefficient training data set. Among them, the labeled linear contraction coefficient training data set is expressed as: The unlabeled linear contraction coefficient training data set is expressed as: where W and J respectively represent the quantities of the corresponding data sets.

[0040] Furthermore, in step S3, the semi-supervised generative adversarial network includes a discriminator network and a generator network. The discriminator network includes a classifier and a discriminator. The specific process of training the semi-supervised generative adversarial network using the training data set is as follows: Set the number of training rounds, and input the training data set into the semi-supervised generative adversarial network for training; in each round of training, the training method is: the generator network is trained once and the discriminator network is trained twice; among them, as Figure 1 shown, the training process of the discriminator network is: randomly extract M groups of labeled data from the training data set W for training the classifier in the discriminator network, and then update the parameters of the classifier through the cross-entropy loss function; randomly extract M groups of unlabeled data from the training data set J for training the discriminator in the discriminator network, and then update the parameters of the discriminator through the discriminator loss function; as Figure 1 shown, the training process of the generator network is: select M groups of m-dimensional random vectors that follow a Gaussian distribution, input the random vectors into the generator of the generator network, generate generated data after passing through the generator, input the generated data into the discriminator for adversarial training, and then update the parameters of the generator network through the loss function of the generator network.

[0041] Furthermore, in step S3, as Figure 2A shown, the generator network includes an input layer A, a hidden layer A, and an output layer A; the generator network is a DNN network with m neurons in each layer; among them, the activation function of the hidden layer A is ReLU, the activation function of the output layer A is Tanh, and the input vector z is an m-dimensional random Gaussian vector that generates generated data through the generator network The loss function of the generator network is expressed as:

[0042] where \(D(\cdot)\) and \(G(\cdot)\) represent the forward propagations of the discriminator network and the generator network respectively, and the forward propagation of the discriminator network only considers the output of the discriminator.

[0043] Further, in step S3, as Figure 2B shown, the discriminator network includes an input layer B, a hidden layer B, and an output layer B; both the input layer B and the hidden layer B are composed of fully connected layers. Among them, the upper branch is composed of two fully connected layers to output the classification probability of the K-dimensional source number, and the lower branch is composed of one fully connected layer to output a one-dimensional data representing its authenticity, that is, the upper branch is a classifier and the lower branch is a discriminator; the activation function of the hidden layer B of the discriminator network is LeakyReLU with a negative axis slope of 0.2, and the output layer of the upper branch uses the Softmax activation function, and the input is the training data and the generated data After passing through the discriminator network, a K + 1-dimensional prediction vector is output The loss function of this network is divided into two parts. One part is the classification loss of supervised learning, which is expressed as: The other part is the discriminative loss of unsupervised learning, which is expressed as: where

[0044] Similarly, the forward propagation of the discriminator network only considers the output of the discriminator, and then parameter regularization is used to prevent the network from overfitting; therefore, the loss function of this network can be expressed as: where \(\alpha\) is a non-negative hyperparameter, is the network weights and biases.

[0045] Further, the specific process of using the trained semi-supervised generative adversarial network to predict the source number is as follows: The test set is input into the trained semi-supervised generative adversarial network to predict the source number. The linear shrinkage coefficient in the test set is used as the input of the discriminator network of the trained semi-supervised generative adversarial network. Its classifier outputs a probability distribution, and the maximum probability is found according to the probability distribution. The index of the maximum probability corresponds to the estimated source number.

[0046] Next, the performance of a small-sample source number detection method applying a semi-supervised generative adversarial network provided by the present invention is analyzed through simulation experiments, and the simulation process is all carried out in a Python environment.

[0047] Simulation parameter configuration: Consider a uniform linear array with \(m = 30\) and the element spacing is half wavelength. For the generated dataset, the number of samples \(n\) is randomly generated in the range of \([15, 65]\). When \(m>n\), the signal-to-noise ratio is randomly set in the range of \([-6, 10]\) dB, and when \(m < n\), the signal-to-noise ratio is randomly set in the range of \([-10, 10]\) dB. The number of sources is randomly generated in the range of \([1, 10]\), that is, \(K = 10\); at the same time, the source directions are randomly selected in the range of \([-60^{\circ}, 60^{\circ}]\). Based on this setting, a total of 8000 training samples of the preprocessed linear shrinkage coefficients are generated, of which 75% are used for training and 25% are used for validation. During the training process, the number of training epochs is 8000, and the regularization parameter \(\alpha=2\times10 -4 , and the batch size \(M = 32\) for each training, and the learning rate is \(3\times10 -4 is used to train the network. During the simulation process, the accurate detection probability obtained from 1000 independent Monte Carlo trials is used to evaluate the effectiveness of different algorithms.

[0048] Simulation Experiment 1: The accurate detection probability of the method of the present invention at different labeling rates and other algorithms varying with the signal-to-noise ratio (SNR) or the number of samples (Number of Samples)

[0049] Five far-field sources are incident on the uniform linear array from \(\{5^{\circ}, 10^{\circ}, 15^{\circ}, 20^{\circ}, 25^{\circ}\}\). Considering the following configurations, three different labeling rates of 10% (SSGAN1), 30% (SSGAN2), 50% (SSGAN3) and the fully supervised classification network based on linear shrinkage coefficients (LSCNet) are selected for comparison:

[0050] Scenario 1-1: \(m = 30\), \(n = 15\);

[0051] Scenario 1-2: \(m = 30\), \(n = 40\);

[0052] Scenario 1-3: \(m = 30\), \(n = 15\);

[0053] Scenario 1-4: \(m = 30\), \(n = 35\);

[0054] Scenario 1-5: \(m = 30\), \(SNR=-2\) dB;

[0055] Scenario 1-6: \(m = 30\), \(SNR=-8\) dB;

[0056] The curves of the accurate detection probability of the number of sources for Scenarios 1-1 to 1-2 are respectively as Figure 3A - 3BAs shown, it compares different deep learning methods. It can be seen that the method of the present invention can exhibit good estimation performance with a small number of labels, and a label rate of 30% can achieve performance close to that of fully supervised learning, and the performance based on the linear shrinkage coefficient is better than the method of logarithmic eigenvalues; in addition, Scenarios 1-3 to 1-6 are comparisons between the method of the present invention and traditional methods, and the curves of the accurate detection probability of the number of signal sources are respectively as Figure 4A - 5B shown. It can be seen that the method of the present invention exhibits excellent detection performance under different signal-to-noise ratios or numbers of samples, which fully proves the better general applicability and better signal source number estimation performance of the method of the present invention.

[0057] Simulation Experiment 2: Variation of the accurate detection probability of the method of the present invention and other algorithms with the signal-to-noise ratio (SNR)

[0058] Seven far-field sources are incident on a uniform linear array from {0°, 4°, 8°, 12°, 16°, 20°, 24°}:

[0059] Scenario 2-1: m = 50, n = 30;

[0060] Scenario 2-2: m = 50, n = 60;

[0061] The curves of the accurate detection probability of the number of signal sources for Scenarios 2-1 to 2-2 are respectively as Figure 6A - 2B shown. It can be seen that the method of the present invention can provide stable and competitive signal source number estimation performance under the conditions of a large-scale array with a small number of samples and a small number of labeled data, which fully proves the superiority of the method of the present invention.

Claims

1. A method for detecting the number of small sample sources using a semi-supervised generative adversarial network, characterized in that: The method comprises the following steps: S1, set d far-field sources, the d far-field sources are at angles q1, K, q d The signal is incident on an array antenna composed of m antennas, and the output data of the array antenna at time t is obtained: x(t)=As(t)+n(t), where A represents an m×d-dimensional array steering matrix, s(t) represents a d×1-dimensional independent and identically distributed Gaussian signal, and n(t) represents an m×1 noise vector; the output data x(t) is sampled n times, n sampled samples are collected, and the covariance matrix of the sampled samples is calculated, and the expression is: And calculate its descending eigenvalue l1,L,l m , where m represents the number of array antennas; S2. Based on the m descending eigenvalues ​​calculated in step S1, the corresponding linear shrinkage coefficient is calculated and preprocessed. The specific process is as follows: S2.1, based on the calculated mth descending eigenvalue l i , i=1,…,m, calculate the corresponding linear contraction coefficient: Where j = 0,…,m-1, due to May be greater than 1, using As the final linear shrinkage coefficient; S2.

2. When the number of samples n is less than the number of array elements m, the Processing is performed to keep the length of the linear shrinkage coefficient consistent, and its expression is: Then the linear shrinkage coefficient is normalized to obtain the preprocessed linear shrinkage coefficient vector, which is expressed as: Where r represents an m-dimensional vector consisting of linear shrinkage coefficients, 1 m represents an m-dimensional all-one vector, max(·) and min(·) represent the maximum and minimum values ​​in the returned vector respectively; S3, based on the preprocessed linear shrinkage coefficients generated in step S2, generating a training data set, and using the training data set to train a semi-supervised generative adversarial network to obtain a trained semi-supervised generative adversarial network; S4. Use the trained semi-supervised generative adversarial network to predict the number of sources.

2. According to claim 1, a method for detecting the number of small sample sources using a semi-supervised generative adversarial network is characterized in that: The specific process of generating the training data set is: based on the preprocessed linear shrinkage coefficient generated in step S2, a labeled linear shrinkage coefficient training data set and an unlabeled linear shrinkage coefficient training data set are generated, wherein the labeled linear shrinkage coefficient training data set is expressed as: The unlabeled linear shrinkage coefficient training dataset is represented as: Among them, W and J represent the number of corresponding datasets respectively.

3. According to claim 2, a method for detecting the number of small sample sources using a semi-supervised generative adversarial network is characterized in that: In step S3, the semi-supervised generative adversarial network includes a discriminant network and a generative network, the discriminant network includes a classifier and a discriminator, and the specific process of training the semi-supervised generative adversarial network using the training data set is: setting the number of training rounds, inputting the training data set into the semi-supervised generative adversarial network for training; In each round of training, the training method is: the generating network is trained once and the discriminating network is trained twice; wherein, the discriminating network training process is: randomly extracting M groups of labeled data from the training data set W for classifier training in the discriminating network, and then updating the classifier parameters through the cross entropy loss function; randomly extracting M groups of unlabeled data from the training data set J for discriminator training in the discriminating network, and then updating the discriminator parameters through the discriminator loss function; the generating network training process is: selecting M groups of m-dimensional random vectors that obey a Gaussian distribution, inputting the random vectors into the generator of the generating network, generating data after passing through the generator, inputting the generated data into the discriminator for adversarial training, and then updating the parameters of the generating network through the loss function of the generating network.

4. The method for detecting the number of small sample sources using a semi-supervised generative adversarial network according to claim 3, characterized in that: In step S3, the generation network includes an input layer A, a hidden layer A and an output layer A; the generation network is a DNN network with m neurons in each layer; wherein the activation function of the hidden layer A is ReLU, the activation function of the output layer A is Tanh, and the loss function of the generation network is expressed as: Where D(·) and G(·) represent the forward propagation of the discriminant network and the generative network, respectively, and the forward propagation of the discriminant network only considers the output of the discriminator.

5. The method for detecting the number of small sample sources using a semi-supervised generative adversarial network according to claim 4, characterized in that: In step S3, the discriminant network includes an input layer B, a hidden layer B and an output layer B; wherein the discriminant network is divided into an upper branch and a lower branch, the upper branch is composed of two fully connected layers and outputs the classification probability of the K-dimensional source number, and the lower branch is composed of one fully connected layer and outputs one-dimensional data; the activation function of the hidden layer B of the discriminant network is LeakyReLU with a negative axis slope of 0.2; the output layer of the upper branch adopts the Softmax activation function; the loss function of the loss function of the discriminant network can be expressed as: Among them, α represents a non-negative hyperparameter, Represents network weight and bias; L sup represents the classification loss of supervised learning, which is expressed as: L unsup represents the discriminative loss of unsupervised learning, which is expressed as: in, 6. The method for detecting the number of small sample sources using a semi-supervised generative adversarial network according to claim 5, characterized in that: In step S4, the specific process of using the trained semi-supervised generative adversarial network to predict the number of information sources is: inputting the test set into the trained semi-supervised generative adversarial network to predict the number of information sources, using the linear shrinkage coefficient in the test set as the input of the discriminant network of the trained semi-supervised generative adversarial network, and its classifier outputs a probability distribution, and finds the maximum probability according to the probability distribution, and the index of the maximum probability corresponds to the estimated number of information sources.

Images