Communication radiation source target confrontation decision-making method based on unsupervised learning

Abstract

The invention discloses a communication radiation source target confrontation decision-making method based on unsupervised learning. The method comprises the following steps of obtaining an IQ signal of a communication signal and performing signal preprocessing; training a contrast predictive coding CPC model based on the IQ data after signal preprocessing, and extracting IQ data features; inputting the IQ data features into a ResNet network for closed set individual classification, and outputting closed set scores; correcting the closed set score based on an OpenMax algorithm to obtain an open set score, and identifying a radiation source individual according to the open set score; constructing a target adversarial decision model based on a GBDT algorithm, extracting individual features of a communication radiation source and interference adversarial parameter features, and constructing a data set to train the target adversarial decision model; and inputting individual characteristic parameters of the communication radiation source to the trained target adversarial decision model, and outputting interference adversarial parameters. According to the method, the identification accuracy of the communication signals is improved, and the robustness in the face of signals of unknown categories is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of communication radiation source identification, and particularly relates to an adversarial decision method for communication radiation source targets based on unsupervised learning. Background Art

[0002] Individual identification of communication radiation sources is an important research topic in the field of communication countermeasures recently. By measuring the differences reflected by the transmitter in the signal, the signal and the transmitter are associated to determine which communication device the signal comes from, so as to realize device tracking. A major bottleneck in the development of individual identification technology for communication radiation sources is the difficulty in obtaining a sufficient amount of communication signals labeled with individual information. How to train a deep learning model with good generalization ability in the absence of sufficient labeled samples is a problem worthy of in-depth study.

[0003] For example, in the method for individual identification of open-set radiation sources based on deep learning with the patent publication number CN111914919A, the inter-class difference features of the training set are extracted through a convolutional neural network to generate a closed-set activation vector CSAV for known set classification, and the intra-class common features are used to calculate the known class reference vector, that is, the average activation vector MAV, and a Weibull model is constructed to establish an overall quantization model of known information; in the test stage, the open-set activation vector OSAV is calculated through the Weibull cumulative distribution function CDF, and the open-set activation vector OSAV is used to quantitatively represent the specific features of the test sample different from the known classes, and the open-set probability of the sample is estimated, but its robustness is poor.

[0004] In the current increasingly complex electromagnetic environment, there may be a large number of unknown communication signals in the same area. Traditional deep learning methods can only identify the categories that appear in the training set, and samples that do not appear in the training set are usually identified as known categories with high confidence, which obviously cannot meet the requirements of individual identification of communication radiation sources in the current complex electromagnetic environment and cannot realize the automation of communication radiation source target confrontation. Summary of the Invention

[0005] To overcome the defects and deficiencies of the existing technology, the present invention provides a communication radiation source target adversarial decision-making method based on unsupervised learning. The present invention combines unsupervised learning and open-set recognition, uses an unsupervised Contrastive Predictive Coding (CPC) model to obtain the IQ data features of communication signals, sends the IQ data features into a ResNet network for closed-set individual classification, and finally uses OpenMax to correct the obtained closed-set scores to obtain their open-set scores. The radiation source individuals are distinguished according to the open-set scores. After identifying the communication radiation source individuals, a target adversarial decision-making model based on the GBDT (decision tree) algorithm is constructed, the communication radiation source individual features and interference adversarial parameter features are extracted, and a data set is constructed to train the target adversarial decision-making model to achieve the automation of communication radiation source target confrontation. The present invention utilizes a large number of communication signals without labeled current task individual information to obtain IQ data features for improving the generalization performance of the model, so that a good recognition effect can be obtained even in the case of only a small number of samples with labeled individual information.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] The present invention provides a communication radiation source target adversarial decision-making method based on unsupervised learning, including the following steps:

[0008] Obtain the IQ signal of the communication signal and perform signal preprocessing on the IQ signal;

[0009] Train a contrastive predictive coding CPC model based on the IQ data after signal preprocessing and extract IQ data features;

[0010] Input the IQ data features into a ResNet network for closed-set individual classification and output closed-set scores;

[0011] Based on the OpenMax algorithm, correct the closed-set scores to obtain open-set scores, and identify the radiation source individuals according to the open-set scores;

[0012] After identifying the communication radiation source individuals, construct a target adversarial decision-making model based on the GBDT algorithm, extract the communication radiation source individual features and interference adversarial parameter features, and construct a data set to train the target adversarial decision-making model;

[0013] Input the communication radiation source individual feature parameters into the trained target adversarial decision-making model and output interference adversarial parameters.

[0014] As a preferred technical solution, the signal preprocessing of the IQ signal specifically includes:

[0015] Perform signal preprocessing on the IQ signal by using data normalization operation to convert the data into a distribution with a mean of 0 and a standard deviation of 1.

[0016] As a preferred technical solution, a contrast prediction coding (CPC) model is trained based on the IQ data after signal preprocessing, specifically including:

[0017] Construct a contrast prediction coding (CPC) model, and map the IQ data after signal preprocessing into latent features, expressed as:

[0018] z t = gen c (x t )

[0019] where g enc represents the encoder, x t represents the IQ data after signal preprocessing, and z t represents the latent features;

[0020] Generate context features based on an autoregressive model, expressed as:

[0021] ct = ga r (z ≤t )

[0022] where c t represents the context features, g ar represents the autoregressive model, and z ≤t represents the encoder hidden space representation;

[0023] Construct an InfoNCE loss to optimize the contrast prediction coding (CPC) model.

[0024] As a preferred technical solution, the InfoNCE loss is expressed as:

[0025]

[0026] where L N represents the InfoNCE loss, E represents the expectation, X represents the sample set composed of positive and negative samples, (x t+k , c t ) is a positive sample pair, (x j , c t ) can be used as a negative sample pair, x j represents the current moment, represents the similarity between the prediction of the context feature c t and the future true value x t+k ;

[0027] As a preferred technical solution, input the IQ data features into a ResNet network for closed-set individual classification, specifically including:

[0028] The IQ data features pass through an initial convolutional layer, a max pooling layer, a residual block, an average pooling layer, and a fully connected layer. The number of output channels of the fully connected layer is equal to the number of categories for closed-set individual classification. The fully connected layer maps the learned features to category probabilities. The output of the fully connected layer passes through the softmax function to convert the original category scores into a probability distribution, obtaining the prediction probability for each category and outputting a closed-set score to determine the communication radiation source individual category to which the input IQ data features belong.

[0029] As a preferred technical solution, correct the closed-set score based on the OpenMax algorithm to obtain an open-set score, and identify the radiation source individual according to the open-set score. Specifically, it includes:

[0030] Input the known individual category samples into the ResNet network to obtain the activation vector AV of the samples;

[0031] Retain the correctly classified activation vectors to obtain the set AV i ={AV i1 ,AV i2 ,...,AV im}, where AV im represents that among the m samples in the i-th class of training samples, m samples are recognized as the i-th class by the closed-set classification network;

[0032] Calculate the mean MAV i of the set AV i as the centroid of the i-th class of samples;

[0033] Calculate the distance D i from AV i1 ,AV i2 ,...,AV im in the set AV i to the centroid;

[0034] Fit the maximum value distribution in the distance D i based on the Weibull distribution, and the obtained fitting result is the cumulative distribution function CDF of Weibull;

[0035] Input the sample to be predicted into the ResNet network to obtain the score vector AV x ,expressed as:

[0036] AV x ={Score 1 ,Score 2 ,...,Score K}

[0037] where K is the number of known classes, and Score K represents the target similarity score;

[0038] Calculate the score vector AV x to the centroid of each known class (MAV 1 , MAV 2 ,..., MAV k ), and obtain the distances {D x1 , D x2 ,.., D xk};

[0039] Input the distance D xj into the Weibull distribution model of the j-th class to obtain the probability that the predicted sample does not belong to the j-th class, and calculate the corrected weight of the score Score j of the j-th class, which is expressed as:

[0040] w j = 1 - w_score(D xj )

[0041] where w j represents the corrected weight, and w_score(D xj ) represents the probability that the predicted sample does not belong to the j-th class;

[0042] Based on the corrected weight, the corrected score of the j-th class is:

[0043] Score′ j = Score j × w j

[0044] Calculate the score of the unknown class as:

[0045] Score unknown = Score 1 × (1 - w 1 ) + Score 2 × (1 - w 2 ) + …, Score k × (1 - w k )

[0046] Construct a new score vector as: {Score′ 1 , Score′ 2 ,..., Score′ K , Score unknwon};

[0047] Based on SoftMax, map the new score vector {Score′ 1 , Score′ 2 ,..., Score′ K , Score unknwon} to the probabilities of each individual classification.

[0048] As a preferred technical solution, a target adversarial decision-making model based on the GBDT algorithm is constructed, individual characteristics of communication radiation sources and interference countermeasure parameter characteristics are extracted, and a dataset is constructed to train the target adversarial decision-making model, specifically including:

[0049] Construct a feature parameter set A based on the interference countermeasure parameter characteristics, divide the feature parameter set A into C fuzzy groups based on the FCM algorithm, and find the clustering center of each fuzzy group i = 1, 2, …, C, making the value of its value function J the smallest, and the obtained result is the final division result;

[0050] The initialization membership matrix constraint condition of the FCM algorithm is:

[0051]

[0052] where, 0 ≤ u ij ≤ 1;

[0053] The value function of the FCM algorithm is:

[0054]

[0055] where, u ij is the fuzzy membership degree and 0 ≤ u ij ≤ 1, m represents the weight coefficient and m ≥ 2, is the Euclidean distance between the i-th clustering center and the j-th data point, a j is the value of the feature parameter of each interference signal;

[0056] Construct an objective function, and the necessary conditions for making the value function of the FCM algorithm reach the minimum are:

[0057]

[0058] where, λ j (j = 1, 2, …, N) is the Lagrange multiplier, and by taking the derivative of all input parameters, two necessary conditions for making the value function J the smallest are respectively:

[0059]

[0060] Find the clustering center and the fuzzy membership degree u ij by iteration, and obtain the final membership matrix U. The final membership matrix U is a matrix of size C × N. Each column represents an interference signal, and the corresponding C membership degree values represent the membership degree of the interference signal to each clustering result. The highest membership degree represents that the interference signal belongs to the class it represents.

[0061] The present invention also provides an unsupervised learning-based communication radiation source target confrontation decision-making system for implementing the above-mentioned unsupervised learning-based communication radiation source target confrontation decision-making method. The system includes: an IQ signal acquisition module, a signal preprocessing module, a CPC model training module, an IQ data feature extraction module, a closed-set individual classification module, a radiation source individual identification module, a target confrontation decision-making model construction module, a target confrontation decision-making model training module, and an interference confrontation parameter output module;

[0062] The IQ signal acquisition module is used to acquire the IQ signal of the communication signal;

[0063] The signal preprocessing module is used to perform signal preprocessing on the IQ signal;

[0064] The CPC model training module is used to train a contrastive predictive coding (CPC) model based on the IQ data after signal preprocessing;

[0065] The IQ data feature extraction module is used to extract IQ data features based on the trained contrastive predictive coding (CPC) model;

[0066] The closed-set individual classification module is used to input the IQ data features into a ResNet network for closed-set individual classification and output a closed-set score;

[0067] The radiation source individual identification module is used to correct the closed-set score based on the OpenMax algorithm to obtain an open-set score and identify the radiation source individual according to the open-set score;

[0068] The target confrontation decision-making model construction module is used to construct a target confrontation decision-making model based on the GBDT algorithm;

[0069] The target confrontation decision-making model training module is used to extract communication radiation source individual features and interference confrontation parameter features and construct a data set to train the target confrontation decision-making model;

[0070] The interference confrontation parameter output module is used to input the communication radiation source individual feature parameters into the trained target confrontation decision-making model and output interference confrontation parameters.

[0071] The present invention also provides a computer-readable storage medium storing a program, and when the program is executed by a processor, it implements the above-mentioned unsupervised learning-based communication radiation source target confrontation decision-making method.

[0072] The present invention also provides a computer device, including a processor and a memory for storing a program executable by the processor. When the processor executes the program stored in the memory, it implements the above-mentioned unsupervised learning-based communication radiation source target confrontation decision-making method.

[0073] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0074] (1) The present invention combines unsupervised learning and open-set recognition. The unsupervised contrastive predictive coding (CPC) model is used to extract the IQ data features of communication signals. The IQ data features are sent into a ResNet network for closed-set individual classification. The OpenMax algorithm is used to correct the obtained closed-set scores to obtain their open-set scores. The radiation source individuals are distinguished according to the open-set scores. After identifying the communication radiation source individuals, a target adversarial decision-making model based on the GBDT (decision tree) algorithm is constructed to extract the communication radiation source individual features and interference adversarial parameter features, and a dataset is constructed to train the target adversarial decision-making model to achieve the automation of communication radiation source target confrontation. The present invention uses a large number of communication signals without labeled current task individual information to obtain IQ data features to improve the generalization performance of the model, so that a good recognition effect can be obtained even in the case of only a small number of labeled individual information samples, which not only improves the recognition accuracy of communication signals, but also enhances the robustness when facing unknown category signals.

[0075] (2) The present invention uses the unsupervised contrastive predictive coding (CPC) model to extract the IQ data features of communication signals. The contrastive predictive coding (CPC) model can automatically learn the general features of signals from a large amount of unlabeled data without manual annotation, reducing the cost and complexity of data preprocessing;

[0076] (3) The present invention sends the IQ data features into a ResNet network for closed-set individual classification. The deep residual structure of the ResNet network effectively solves the problems of gradient disappearance and gradient explosion in the training of deep networks and improves the classification accuracy.

[0077] (4) The present invention uses the OpenMax algorithm to correct the obtained closed-set scores to obtain their open-set scores, which can take into account the similarity between categories and adjust the scores, thereby improving the recognition robustness in the open-set environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 is a schematic flow chart of the communication radiation source target adversarial decision-making method based on unsupervised learning of the present invention;

[0079] Figure 2 is a schematic architecture diagram of the CPC network model of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0080] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0081] Example 1

[0082] As Figure 1 shown, this embodiment provides an unsupervised learning-based communication radiation source target adversarial decision-making method, including the following steps:

[0083] S1: Obtain the IQ signal of the communication signal and perform signal preprocessing on the IQ signal so that the subsequent network can better process the signal data;

[0084] In this embodiment, data normalization operation is used to perform signal preprocessing on the IQ signal, and the data is converted into a distribution with a mean of 0 and a standard deviation of 1.

[0085] In this embodiment, the data normalization scheme used is the z-score method, and the formula of the z-score method is as follows:

[0086]

[0087] where x represents the IQ signal, μ represents the mean of the IQ signal, σ represents the standard deviation of the IQ signal, represents the signal processed by the z-score method, and the Z-score method can convert the original value into data with a unified mean of 0 and a standard deviation of 1, which is beneficial to the training of the subsequent neural network;

[0088] S2: Use the preprocessed IQ data of the unlabeled signal to train the contrastive predictive coding CPC model and extract the IQ data features;

[0089] As Figure 2 shown, construct the contrastive predictive coding CPC model architecture. First, use an encoder gen c to map the preprocessed and normalized IQ data x t into a string of latent features z t = g enc (x t ), and then use an autoregressive g ar to summarize the parts of all the encoder hidden space representations z ≤t in the latent space and generate a context feature c t = g ar (z ≤t );

[0090] In order to obtain the prediction result z t+k k moments later for the current context feature c, it is necessary to maximize the mutual information between the input sequence x t+k k moments later and the current context feature c so that the predicted is adjacent to the true value of the subsequent i moments z t+iAs similar as possible. According to formula (2), there is no way to control the joint distribution probability p(x, c) of x and c in the mutual information. Therefore, a concept of density ratio is proposed, as shown in formula (3):

[0091]

[0092] where p represents the predicted probability value under the corresponding feature, represents the prediction of the context feature c. and the future true value x t+k The degree of similarity is directly represented by the linear matrix W 1 , W 2 ,..., W k multiplying c t to be used as the predicted value (i.e., W t c t ), and the time values of adjacent K frames are the true values. The cosine similarity is used to measure the similarity, and the following function is obtained to approximate the density ratio:

[0093]

[0094] where T represents the time;

[0095] The InfoNCE Loss is used to optimize the model, and the definition of the InfoNCE Loss is as follows:

[0096]

[0097] where E represents the expectation, X = {x 1 , x 2 ,..., x N} is a set of samples, (x t+k , c t ) can be used as a positive sample pair, (x j , c t ) can be used as a negative sample pair, x j represents the current time. Among them, the positive sample comes from the sample that is k steps away from the current context c t , and the negative sample comes from the sample randomly selected from the sequence.

[0098] Therefore, in order to optimize this loss, the numerator should be as large as possible and the denominator should be as small as possible. This also meets the requirement for the mutual information between x and c, that is, the mutual information between positive sample pairs is larger, and the mutual information between negative sample pairs is smaller. Optimizing this loss is actually maximizing the mutual information between x t+k and c t .

[0099] S3: Input the IQ data features into the ResNet network for closed-set individual classification;

[0100] In this embodiment, the residual blocks inside the ResNet network use skip connections, which alleviates the problem of vanishing gradients caused by increasing depth in deep neural networks. The ResNet network makes a reference to the input of each layer and learns to form a residual function instead of learning some functions without a reference. This residual function is easier to optimize and can greatly deepen the network layers. The characteristics of ResNet are easy to optimize and can improve the accuracy by increasing a considerable depth.

[0101] In this embodiment, after the IQ data features are input into the ResNet network, the process of the ResNet network for closed-set individual classification can be divided into the following steps:

[0102] 1. Initial Convolutional Layer (conv1): First, the input IQ data features pass through the first convolutional layer of ResNet. This convolutional layer usually includes a convolution operation, a batch normalization operation, and a ReLU activation function. This step converts the input feature map from the original dimension to a lower-dimensional feature representation, while adding non-linearity, preparing for subsequent classification tasks.

[0103] 2. Max Pooling Layer: Next, the feature map passes through a max pooling layer. This operation can further reduce the spatial dimension of the feature map while retaining the most important feature information, enhancing the invariance of the model to input changes.

[0104] 3. Residual Blocks: Then, the feature map passes through multiple residual blocks (Residual Blocks). Each residual block consists of several convolutional layers, batch normalization layers, and ReLU activation functions. These residual blocks solve the problem of vanishing gradients in deep networks through skip connections, allowing gradients to flow directly through the network, enabling the network to be trained deeper.

[0105] 4. Feature Extraction: In the residual blocks, each Bottleneck structure further extracts and processes the input features. These structures usually include three convolutional layers: a 1x1 convolution for dimensionality reduction, a 3x3 convolution for feature extraction, and another 1x1 convolution for dimensionality increase. These operations enable the network to learn deeper feature representations.

[0106] 5. Average Pooling Layer: After passing through all residual blocks, the feature map passes through an average pooling layer. This operation calculates the average value of each channel of the feature map, further reducing the spatial dimension of the features in preparation for the fully connected layer;

[0107] 6. Fully Connected Layer: Finally, the features passing through the average pooling layer are flattened and passed through a fully connected layer (also known as a dense layer). The number of output channels of this layer is equal to the number of classes in the classification task. This fully connected layer maps the learned features to class probabilities;

[0108] 7. Classification Output: Finally, the output of the fully connected layer passes through a softmax function, which converts the original class scores into a probability distribution, thus obtaining the predicted probability for each class.

[0109] Through the above steps, the ResNet network can perform closed-set individual classification on the input IQ data features, that is, identify which class the input data belongs to within a known set of classes.

[0110] S4: Use the OpenMax algorithm to correct the closed-set scores to obtain their open-set scores, and distinguish radiation source individuals based on their open-set scores, enabling individual identification in the open-set scenario. Identify individuals not belonging to the training set as another class, specifically including:

[0111] S41: Obtain the distance D;

[0112] In this embodiment, for the pre-trained closed-set classification network (i.e., the ResNet network) in the previous step, taking the i-th known individual class as an example, all training samples of the i-th class are input into the closed-set classification network to obtain the activation vectors AV (Activation Vector) of these samples. The activation vector is the vector before softmax, and the AVs correctly classified by the closed-set classification network are retained (the ones with classification failures are not saved). Denote the set of retained AVs as AV i ={AV i1 ,AV i2 ,...,AV im}, where AV im indicates that there are m samples in the i-th class of training samples that are recognized as the i-th class by the closed-set classification network.

[0113] Then, use AV i to calculate its mean MAV i (Mean Activation Vector). MAV i is the centroid of the i-th class of samples.

[0114] Then use AV i ={AV i1 , AV i2 ,..., AV im} in the AV i1 , AV i2 ,..., AV im Calculate their distances to the centroid MAV i , denoted as D i ={D i1 , D i2 ,…, D im};

[0115] S42: Fit the maximum value distribution of D;

[0116] In this embodiment, the Weibull distribution is used to fit the maximum value distribution in D i , and the result of the fit is the cumulative distribution function CDF of Weibull.

[0117] S43: Correct the score;

[0118] For the sample to be predicted, first use the closed-set classification network to obtain its AV, denoted as AV x ={Score 1 , Score 2 ,..., Score K}, where K is the number of known class types, and Score K represents the target similarity score.

[0119] Then calculate the distances from AV x to the centroids of each known class (MAV 1 , MAV 2 ,..., MAV k ), {D x1 , D x2 ,.., D xk}.

[0120] Assume that the score of the j-th class in AV x needs to be corrected. Input D j into the Weibull distribution model CDF of the j-th class for output. This output represents the probability that the predicted sample does not belong to the j-th class, denoted as w_score(D xj ). Then subtract this output from 1 to get the probability that the sample belongs to the j-th class, denoted as w xj =1 - w_score(D j ), and use w xj as the score of the j-th class Score j j ​The corrected weight value is sufficient, that is, the score of the j-th category after correction is Score′ j = Score j × w j .

[0121] The scores of other known categories are corrected in the same way. The score of the unknown category is:

[0122] Score unknown = Score 1 × (1 - w 1 ) + Score 2 × (1 - w 2 ) + …, Score k × (1 - w k ).

[0123] In summary, a new score vector is obtained, expressed as:

[0124] {Score′ 1 , Score′ 2 ,..., Score′ K , Score unknwon};

[0125] S44: Map to classification probabilities;

[0126] Map the scores {Score′ 1 , Score′ 2 ,..., Score′ K , Score unknwon} obtained in the previous section to each classification probability using SoftMax.

[0127] When the maximum classification probability is obtained for the unknown category, or the maximum category classification probability is less than a certain threshold, it is recognized as the unknown category.

[0128] S5: After identifying the individual of the communication radiation source, construct an objective adversarial decision model based on the GBDT (Gradient Boosting Decision Tree) algorithm, extract the individual characteristics of the communication radiation source and the interference adversarial parameter characteristics, and construct a data set to train the objective adversarial decision model to achieve the automation of the communication radiation source objective confrontation;

[0129] In this embodiment, the decision tree can process a large amount of data in a short time, and at the same time has relatively simple requirements for the prepared data. However, the decision tree needs to preprocess the data, and the selection of its threshold requires experience or analysis. To solve this problem, fuzzy clustering is added in the process of generating the decision tree in this embodiment to achieve automatic preprocessing and classification of the data. At the same time, the Xie-Beni index is used to determine the number of branches of the decision tree, thus realizing the automated design of the overall process of the radiation source interference decision.

[0130] For the interference signal feature parameter set, a feature parameter α is selected. The values of the feature parameter α of all interference signals form the feature parameter set A. In the feature parameter set A, N is the total number of samples, that is, a j ∈A, j = 1, 2, …, N, a j is the value of the feature parameter of each interference signal. The fuzzy c - means clustering FCM algorithm uses fuzzy grouping to divide the feature parameter set A into C fuzzy groups and find the clustering center of each fuzzy group to minimize the value of its cost function J, and the resulting result is the final division result. The initialization membership matrix constraint condition of the FCM algorithm is:

[0131]

[0132] where 0 ≤ u ij ≤ 1. The membership matrix U describes the membership degree of the interference signal to the divided classes. Then the cost function of the FCM algorithm can be obtained as:

[0133]

[0134] where u ij is the fuzzy membership degree and 0 ≤ u ij ≤ 1, m represents the weight coefficient and m ≥ 2, is the Euclidean distance between the i - th clustering center and the j - th data point. In radar interference recognition, single - attribute clustering is used. At this time

[0135] Construct the following new objective function. The necessary conditions for minimizing the cost function of the FCM algorithm can be obtained as:

[0136]

[0137] In the formula, λ j (j = 1, 2, …, N) is the Lagrange multiplier. Taking the derivative of all the input parameters of the above formula, the two necessary conditions for minimizing the cost function J can be obtained as:

[0138]

[0139] The clustering center is obtained through iteration and the value of the fuzzy membership degree u ij . Finally, the membership matrix U is a matrix of size C × N. Each column represents an interference signal, and the corresponding C membership degree values represent the membership degree of the interference signal to each clustering result. The highest membership degree represents that the interference signal belongs to the class it represents.

[0140] It can be seen from the above two necessary conditions that this algorithm is an iterative algorithm, and its steps are as follows:

[0141] (1) Initialize the membership matrix U with random numbers and normalize it according to the constraint conditions;

[0142] (2) Calculate each cluster center according to the above formula

[0143] (3) Calculate the value of the value function according to the above formula, compare it with the value of the value function obtained last time. If the change amount is less than a certain threshold ε, the algorithm stops; otherwise, continue to execute step (4);

[0144] (4) Calculate a new membership matrix U according to the above formula and return to step (2).

[0145] Finally, based on the target adversarial decision-making model of the GBDT algorithm, input the characteristic parameters of the communication radiation source and output the interference countermeasure parameters to realize the automation of the communication radiation source target adversarial decision-making.

[0146] Embodiment 2

[0147] This embodiment provides a communication radiation source target adversarial decision-making system based on unsupervised learning, which is used to implement the communication radiation source target adversarial decision-making method based on unsupervised learning in Embodiment 1 above. The system includes: an IQ signal acquisition module, a signal preprocessing module, a CPC model training module, an IQ data feature extraction module, a closed-set individual classification module, a radiation source individual recognition module, a target adversarial decision-making model construction module, a target adversarial decision-making model training module, and an interference countermeasure parameter output module;

[0148] In this embodiment, the IQ signal acquisition module is used to acquire the IQ signal of the communication signal;

[0149] In this embodiment, the signal preprocessing module is used to perform signal preprocessing on the IQ signal;

[0150] In this embodiment, the CPC model training module is used to train a contrast prediction coding CPC model based on the IQ data after signal preprocessing;

[0151] In this embodiment, the IQ data feature extraction module is used to extract IQ data features based on the trained contrast prediction coding CPC model;

[0152] In this embodiment, the closed-set individual classification module is used to input the IQ data features into the ResNet network for closed-set individual classification and output a closed-set score;

[0153] In this embodiment, the radiation source individual recognition module is used to correct the closed-set score based on the OpenMax algorithm to obtain an open-set score, and identify the radiation source individual according to the open-set score;

[0154] In this embodiment, the target adversarial decision model construction module is used to construct a target adversarial decision model based on the GBDT algorithm;

[0155] In this embodiment, the target adversarial decision model training module is used to extract the individual characteristics of communication radiation sources and the characteristics of interference countermeasure parameters, and construct a dataset to train the target adversarial decision model;

[0156] In this embodiment, the interference countermeasure parameter output module is used to input the individual characteristic parameters of communication radiation sources into the trained target adversarial decision model and output the interference countermeasure parameters.

[0157] Embodiment 3

[0158] This embodiment provides a storage medium, which can be a storage medium such as ROM, RAM, disk, or optical disc. The storage medium stores one or more programs, and when the programs are executed by a processor, the method for target adversarial decision-making of communication radiation sources based on unsupervised learning in Embodiment 1 is implemented.

[0159] Embodiment 4

[0160] This embodiment provides a computing device, which can be a desktop computer, a laptop computer, a smart phone, a PDA handheld terminal, a tablet computer, or other terminal devices with a display function. The computing device includes a processor and a memory. The memory stores one or more programs, and when the processor executes the programs stored in the memory, the method for target adversarial decision-making of communication radiation sources based on unsupervised learning in Embodiment 1 is implemented.

[0161] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and shall be included in the protection scope of the present invention.

Claims

1. A communication radiation source target confrontation decision method based on unsupervised learning, characterized in that: The steps include: Acquire an IQ signal of a communication signal, and perform signal preprocessing on the IQ signal; Based on the IQ data after signal preprocessing, the comparative prediction coding CPC model is trained to extract the IQ data features; Input the IQ data features into the ResNet network for closed-set individual classification and output the closed-set score; Based on the OpenMax algorithm, the closed set score is corrected to obtain the open set score, and the radiation source individual is identified according to the open set score; After identifying the individual communication radiation source, a target confrontation decision model based on the GBDT algorithm is constructed to extract the individual characteristics of the communication radiation source and the interference confrontation parameter characteristics, and a data set is constructed to train the target confrontation decision model; Input the individual characteristic parameters of the communication radiation source into the trained target confrontation decision model and output the interference confrontation parameters.

2. The communication radiation source target confrontation decision method based on unsupervised learning according to claim 1 is characterized in that: The signal preprocessing of the IQ signal specifically includes: The data normalization operation is used to preprocess the IQ signal and convert the data into a distribution with a mean of 0 and a standard deviation of 1.

3. The communication radiation source target confrontation decision method based on unsupervised learning according to claim 1 is characterized in that: The CPC model is trained based on the IQ data after signal preprocessing, specifically including: Construct a comparative predictive coding CPC model to map the IQ data after signal preprocessing into potential features, which is expressed as: z t =g enc (x t ) Among them, g enc represents the encoder, x t represents the IQ data after signal preprocessing, z t Indicates potential characteristics; Generate context features based on the autoregressor, expressed as: c t =g ar (z ≤t ) Among them, c t represents the context feature, g ar represents the autoregressor, z ≤t represents the encoder latent space representation; Construct the InfoNCE loss-optimized comparative predictive coding CPC model.

4. The communication radiation source target confrontation decision method based on unsupervised learning according to claim 3 is characterized in that: The InfoNCE loss is expressed as: Among them, L N represents InfoNCE loss, E represents expectation, X represents the sample set consisting of positive samples and negative samples, (x t+k ,c t ) is a positive sample pair, (x j ,c t ) can be used as a negative sample pair, x j Indicates the current moment, Represents context feature c t The prediction and future true value x t+k degree of similarity.

5. The communication radiation source target confrontation decision method based on unsupervised learning according to claim 1 is characterized in that: Input the IQ data features into the ResNet network for closed-set individual classification, including: The IQ data features pass through the initial convolution layer, the maximum pooling layer, the residual block, the average pooling layer and the fully connected layer. The number of output channels of the fully connected layer is equal to the number of categories of the closed set individual classification. The fully connected layer maps the learned features to the category probabilities. The output of the fully connected layer is converted into a probability distribution through a softmax function to obtain the predicted probability of each category, and the closed set score is output to determine the individual category of the communication radiation source to which the input IQ data features correspond.

6. The communication radiation source target confrontation decision method based on unsupervised learning according to claim 1 is characterized in that: Based on the OpenMax algorithm, the closed set score is corrected to obtain the open set score, and the radiation source individual is identified according to the open set score, including: Input the known individual category samples into the ResNet network to obtain the activation vector AV of the samples; Keep the activation vector of the correct classification and get the set AV i ={AV i1 ,AV i2 ,…,AV im }, where AV im It means that among the training samples of the i-th category, m samples are identified as the i-th category by the closed-set classification network; Calculate the set AV i The mean MAV i , as the centroid of the i-th class sample; Calculate the set AV i Medium AV i1 ,AV i2 ,…,AV im Distance to the centroid D i ; Fitting distance D based on Weibull distribution i The maximum value distribution in the fitting result is the Weibull cumulative distribution function CDF; Input the sample to be predicted into the ResNet network to obtain the score vector AV x , expressed as: OF x ={Score1,Score2,…,Score K } Among them, K is the number of known classes, Score K represents the target similarity score; Calculate the score vector AV x To each known class centroid (MAV1, MAV2, ..., MAV k )'s distance {D x1 ,D x2 ,..,D xk }; The distance D xj Input into the Weibull distribution model of the jth class, obtain the probability that the predicted sample does not belong to the jth class, and calculate the score of the jth class j The modified weight is expressed as: w j =1-w_score(D xj ) Among them, w j Indicates the correction weight, w_score(D xj ) represents the probability that the predicted sample does not belong to the jth class; The corrected score of the jth category based on the corrected weight is: Score′ j =Score j ×w j The score for the unknown class is calculated as: Score unknown =Score1×(1-w1)+Score2×(1-w2)+…,Score k ×(1-w k ) The new score vector is constructed as: {Score′1, Score′2,…, Score′ K ,Score unknwon } Based on SoftMax, the new score vector {Score′1, Score′2,…, Score′ K ,Score unknwon } is mapped to the probability of each individual classification.

7. The communication radiation source target confrontation decision method based on unsupervised learning according to claim 1 is characterized in that: Construct a target confrontation decision model based on the GBDT algorithm, extract the individual characteristics of the communication radiation source and the interference confrontation parameter characteristics, and construct a data set to train the target confrontation decision model, including: Construct a feature parameter set A based on the interference countermeasure parameter characteristics, divide the feature parameter set A into C fuzzy groups based on the FCM algorithm, and find the cluster center of each fuzzy group i=1,2,…,C, so that the value of its value function J is minimized, and the result is the final partition result; The initial membership matrix constraints of the FCM algorithm are: Where 0≤u ij ≤1; The cost function of the FCM algorithm is: Among them, u ij is a fuzzy membership and 0≤u ij ≤1, m represents the weight coefficient and m≥2, is the Euclidean distance between the i-th cluster center and the j-th data point, a j is the value of the characteristic parameter of each interference signal; Construct the objective function and find the necessary conditions for the value function of the FCM algorithm to reach the minimum value: Among them, λ j (j=1,2,…,N) is the Lagrange multiplier. Taking the derivative of all input parameters, we get the two necessary conditions to minimize the value function J: Find the cluster center by iteration and fuzzy membership u ij The final membership matrix U is a matrix of size C×N. Each column represents an interference signal, and its corresponding C membership values ​​indicate the degree of membership of the interference signal to each clustering result. The highest degree of membership means that the interference signal belongs to the class it represents.

8. A communication radiation source target confrontation decision system based on unsupervised learning, characterized in that: The communication radiation source target confrontation decision method based on unsupervised learning is used to implement any one of claims 1-7, the system comprising: an IQ signal acquisition module, a signal preprocessing module, a CPC model training module, an IQ data feature extraction module, a closed set individual classification module, a radiation source individual identification module, a target confrontation decision model construction module, a target confrontation decision model training module, and an interference confrontation parameter output module; The IQ signal acquisition module is used to acquire the IQ signal of the communication signal; The signal preprocessing module is used to perform signal preprocessing on the IQ signal; The CPC model training module is used to train a comparative prediction coding CPC model based on the IQ data after signal preprocessing; The IQ data feature extraction module is used to extract IQ data features based on the trained comparative predictive coding CPC model; The closed set individual classification module is used to input the IQ data features into the ResNet network for closed set individual classification and output a closed set score; The radiation source individual identification module is used to correct the closed set score based on the OpenMax algorithm to obtain the open set score, and identify the radiation source individual according to the open set score; The target confrontation decision model construction module is used to construct a target confrontation decision model based on the GBDT algorithm; The target confrontation decision model training module is used to extract the individual characteristics of the communication radiation source and the interference confrontation parameter characteristics, and construct a data set to train the target confrontation decision model; The interference confrontation parameter output module is used to input the individual characteristic parameters of the communication radiation source into the trained target confrontation decision model and output the interference confrontation parameters.

9. A computer-readable storage medium storing a program, characterized in that: When the program is executed by a processor, the communication radiation source target confrontation decision method based on unsupervised learning as described in any one of claims 1 to 7 is implemented.

10. A computer device comprising a processor and a memory for storing a program executable by the processor, characterized in that: When the processor executes the program stored in the memory, the communication radiation source target confrontation decision method based on unsupervised learning as described in any one of claims 1 to 7 is implemented.

Images