Open set modulation identification method and system based on comparative learning and OpenMax

Abstract

The invention relates to an open set modulation identification method and system based on comparative learning and OpenMax, and belongs to the technical field of artificial intelligence. Comprising the following steps: acquiring a modulation signal through signal acquisition, and segmenting the modulation signal to obtain segmented signals; inputting the segmented signals into a network model for training, extracting deep signal features, realizing original classification of a known modulation category, and obtaining an original classification probability; the network model comprises a convolutional network and a long-short term memory network; signal depth modulation features are obtained through a trained network model, a distance set is constructed, and classification and identification of all modulation categories including unknown categories are completed by using OpenMax. According to the invention, high-precision classification of unknown categories and known categories is realized.

Description

Technical Field

[0001] The present invention relates to an open-set modulation recognition method and system based on contrast learning and OpenMax, belonging to the technical field of artificial intelligence. Background Art

[0002] With the rapid development of communication technology, various electromagnetic devices are increasing day by day, the electromagnetic environment is becoming more and more complex, various interferences are becoming more and more serious, and the types of signal modulation are gradually diversified. These complex electromagnetic environments pose higher requirements and more severe challenges to existing signal recognition technologies, especially signal recognition technologies in the open set.

[0003] Through end-to-end learning, deep learning realizes the autonomous extraction of deep modulation features from signal data or primary signal features, reduces the dependence on expert knowledge, and reduces the demand for feature engineering. Deep learning has been applied in many fields such as computer vision and natural language processing, and has achieved good results. At the same time, the application of deep learning in signal modulation recognition is becoming more and more mature, and the recognition accuracy is increasing day by day, which is significantly better than traditional algorithms.

[0004] However, the current modulation recognition algorithms based on deep learning mainly solve the problem that the number of modulation types in the test set and the training set is the same. In actual applications, it is difficult to collect samples of all modulation types. As a result, when there are new modulation types in the test set, the modulation types that do not appear in the training set are misclassified into the categories in the training set, resulting in classification errors. Summary of the Invention

[0005] Aiming at the current complex electromagnetic environment and the increasingly diversified modulation methods, it is difficult for the training set to realize the cleaning and annotation of all modulation type data. The purpose of the present invention is to propose an open-set modulation recognition method based on contrast learning and OpenMax to solve the problems in the above background art.

[0006] For the open-set modulation recognition problem, the present invention first uses a convolutional network to extract the deep modulation features of the original orthogonal signal, and then uses a supervised contrast loss to maximize the feature similarity between samples of the same category and minimize the feature similarity between samples of different categories, so as to learn a more discriminative feature representation. Finally, OpenMax completes the category recognition of known categories and the classification of known categories and unknown categories. Under open-set conditions, the method of the present invention takes into account the category recognition rate of known categories and the classification accuracy of known categories and unknown categories.

[0007] The present invention also provides an open-set signal modulation recognition system based on contrast learning and OpenMax.

[0008] To achieve the above object, the present invention provides the following technical solutions: An open-set modulation recognition method based on contrastive learning and OpenMax, including: Acquire modulation signals through signal acquisition and segment them to obtain segmented signals; Input the segmented signals into a network model for training, extract deep signal features, achieve the original classification of known modulation categories, and obtain the original classification probability; the network model includes a convolutional network and a long short-term memory network; Obtain the signal deep modulation features through the trained network model, construct a distance set, and use OpenMax to complete the classification and recognition of all modulation categories including unknown classes.

[0009] Preferably according to the present invention, signal acquisition includes: Collect 15 types of original modulation signals from the actual environment using a spectrum analyzer; the 15 types of original modulation signals include: {4QAM, 16QAM, 32QAM, 64QAM, BPSK, 8PSK, 16PSK, 32PSK, 2FSK, 4ASK, 4ASK, 8ASK, 16APSK, AM, FM}.

[0010] Preferably according to the present invention, segment the 15 types of original modulation signals collected.

[0011] Preferably according to the present invention, the convolutional neural network and the long short-term memory network perform deep modulation feature extraction on the input time series signals, i.e., the segmented signals, through self-learning, and compress the data to a low-dimensional space through 2 convolutional layers and 2 long short-term memory networks, so as to complete the recognition of each known modulation type and the classification of known and unknown modulation types.

[0012] Preferably according to the present invention, obtain the signal deep modulation features through the trained network model; including: Process the segmented signals through the trained convolutional neural network to extract the spatio-temporal features of the signals. In the feature extraction stage, the processing process is shown in formula (1): (1); In the formula, is the spatio-temporal feature output by the convolutional neural network, W represents the convolutional kernel weights, b is the bias, is the activation function of the convolutional layer; i, j respectively represent the current i-th row and j-th column; m, n represents the convolutional kernel size, representing the length and width of the convolutional kernel respectively, l represents the number of channels, k represents the number of convolutional kernels; Then, flatten the extracted feature F into F’ , F ’ After passing through the long short - term memory network, the temporal modulation features of the signal, that is, the signal depth modulation features, are extracted. The processing process is shown in formula (2): (2); In formula (2), G is the extracted temporal modulation feature, is LSTM the activation function of layer; g represents the feature mapping relationship, b and c represent bias vectors, and F’(t), that is F’, after expansion F ’ is a one - dimensional sequence, t represents the position point of the sequence, represents the connection weight from hidden to hidden, U represents the connection weight from input to hidden layer, V represents the connection weight from hidden layer to output layer, h(t) represents the state of the hidden layer at time t, and h(t - 1) represents the state of the hidden layer at time t - 1.

[0013] According to the preference of the present invention, the loss function of the network model is the sum of the cross - entropy loss function and the supervised contrast loss function. The calculation formulas are shown in formulas (3), (4), and (5) respectively: (3); (4); (5); In the formula, N represents the number of samples of the modulation signal, K represents the number of modulation categories, and represent the true label and the predicted label of the sample respectively, P(i) represents the number of samples with the same label as sample i , represents the feature vector of sample i , represents all the sample sets except i in the sample, represents the classification loss, represents the input vector, represents the bias vector, represents the weight matrix, is the supervised contrast loss, represents the feature vector of other samples in A except sample i, represents the temperature parameter, represents the weight.

[0014] Preferably according to the present invention, a distance set is constructed, and OpenMax is used to complete the classification and recognition of all modulation categories including unknown categories; including: Step 1: Calculate the Softmax probability corresponding to each category of the trained network model, that is: (6); In formula (6), S C represents the output probability value corresponding to each category, K represents the number of categories, W T yi represents the weight of the i-th category, x i represents the input of the i-th category, b yi represents the bias of the i-th category; Step 2: Calculate the mean of the activation vectors of all training samples in each category to obtain the mean activation vector MAV (Mean Activation Vector, MAV) of this category. The mathematical expression is as follows:

[0015] (7); Among them, AV m represents the c activation vector corresponding to the m th input sample of the C C-th category, N c represents the total number of samples in the C-th category, and MAV is denoted as the center vector of the C-th category; Step 3: Calculate the distance between the activation vectors of all correctly classified samples in this category and the MAV of this category, using the Euclidean distance, to form the distance set of this category. The distance calculation formula is: In formula (8), MAV k is denoted as the center vector of the K-th category, AV m represents the C activation vector corresponding to the m th input sample of the K-th category; Step 4: Use the Weibull distribution in extreme value theory to fit the distance set of each category; the probability density function of the Weibull distribution is: In formula (9), is the scale parameter, is the shape parameter; x represents the input; The cumulative distribution function corresponding to the Weibull distribution is: (10); Calculate the Weibull distribution corresponding to the classification according to formula (9) and formula (10); Step Five: Calculate the reduced probability corresponding to each category , and the calculation formula is as follows: (11); In formula (11), S c (x) is S C , d cx represents the distance, is the scale parameter, is the shape parameter; Step Six: Calculate the probability of the unknown category, that is, the probability of the unknown category is the original Softmax probability minus the reduced probability, and the calculation formula is: (12); In formula (12), C is the number of categories; Step Seven: Normalize the probabilities of all categories to obtain the classification probabilities of all categories , and the specific processing is as follows: (13).

[0016] A computer device, comprising a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the steps of the open-set modulation recognition method based on contrast learning and OpenMax are implemented.

[0017] A computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the open-set modulation recognition method based on contrast learning and OpenMax are implemented.

[0018] An open-set modulation recognition system based on contrast learning and OpenMax, comprising: A signal training unit, configured to: obtain a modulation signal through signal acquisition and segment it to obtain segmented signals; input the segmented signals into a network model for training, extract deep signal features, implement the original classification of known modulation categories, and obtain the original classification probability; the network model includes a convolutional network and a long short-term memory network; A signal testing unit, configured to: obtain the deep modulation features of the signal through the trained network model, construct a distance set, and use OpenMax to complete the classification and recognition of all modulation categories including unknown categories.

[0019] The beneficial effects of the present invention are: First, the primary objective of the present invention is to address the problem that in practical applications, the modulation types are complex and variable, making it difficult to construct an annotated dataset that includes all modulation types. To effectively solve this problem, the present invention adopts an innovative method, which combines a convolutional neural network, supervised contrastive learning, and OpenMax to achieve high-precision classification of unknown and known classes.

[0020] Second, although deep learning has achieved good results in the field of modulation recognition, it mainly focuses on closed-set modulation signal recognition algorithms. However, in practical applications, the modulation methods are diverse and it is difficult to establish a dataset that includes all modulation types. The innovation of the present invention lies in the successful application of OpenMax in the field of electromagnetic signal modulation recognition, bringing new possibilities and opportunities to this field.

[0021] Third, in terms of optimization, to expand the differences between different classes and increase the aggregation degree among samples of the same class, the present invention innovatively uses supervised contrastive loss to improve the recognition accuracy of known modulation classes and widen the boundary between known and unknown classes.

[0022] Fourth, the present invention is not only applicable to the processing of electromagnetic signal modulation recognition but can also be widely applied to other fields. For example, in the field of abnormal signal detection, it can be used to detect the presence of illegal drone signals, thereby ensuring electromagnetic security and reliable data transmission. This multi-field applicability makes this invention have broad market potential and is expected to provide solutions to signal processing problems in various fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a schematic flowchart of the open-set modulation recognition method based on contrastive learning and OpenMax of the present invention; Figure 2 It is a framework diagram of the open-set modulation recognition method based on contrastive learning and OpenMax of the present invention; Figure 3 It is a schematic structural diagram of the network model of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0024] To have a clearer understanding of the technical features of the present invention, the following describes the specific embodiments of the present invention in conjunction with the accompanying drawings. The technical solutions in the present invention are clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0025] To facilitate the understanding of the present invention, the present invention will be described more comprehensively below with reference to the relevant drawings. Several embodiments of the present invention are given. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive.

[0026] Embodiment 1 An open-set modulation recognition method based on contrastive learning and OpenMax, as Figure 1 and Figure 2 shown, includes: Obtain the modulation signal through signal acquisition and segment it to obtain segmented signals; Input the segmented signals into the network model for training, extract deep signal features, achieve the original classification of known modulation categories, and obtain the original classification probability; the network model includes a convolutional network and a long short-term memory network; Obtain the signal deep modulation features through the trained network model, construct a distance set, and use OpenMax to complete the classification and recognition of all modulation categories including unknown categories. The present invention can detect and identify signals of different modulation types, which is crucial for accurate signal analysis. OpenMax, that is, the OpenMax algorithm, is an open-set recognition method based on a deep neural network. It realizes the effective recognition of unknown categories by post-processing the activation vectors output by the model. The core idea of this algorithm is to use the statistical characteristics of known categories to infer the existence of unknown categories.

[0027] The present invention collects the modulation signal through a spectrum analyzer or USRP to obtain the baseband signal, that is, the in-phase signal and the quadrature signal; uses the baseband signal as the input feature of the convolutional neural network, and uses the convolutional neural network to learn and extract the deep modulation features of the input baseband signal, expands the classification distance between different categories through the supervised contrastive loss, and completes the classification probability of each known category and the position category through OpenMax. USRP, Universal Software Radio Peripheral, is a general software radio peripheral that controls and realizes various functions of a wireless communication system through software such as GNURadio.

[0028] Embodiment 2 An open-set modulation recognition method based on contrastive learning and OpenMax, the difference is that: Signal acquisition, including: Collect 15 types of original modulated signals from the actual environment using a spectrum analyzer; the 15 types of original modulated signals include: {4QAM, 16QAM, 32QAM, 64QAM, BPSK, 8PSK, 16PSK, 32PSK, 2FSK, 4ASK, 4ASK, 8ASK, 16APSK, AM, FM}.

[0029] including , collect the original noisy modulated signals from the actual application environment. During the acquisition of the required signal data, the settings for the signal generator are as follows: The signal generator and the spectrum analyzer are set to the same center frequency. r[n] represents the signal received by the receiver, I[n] represents the in-phase component of the received signal, Q[n] represents the quadrature component of the received signal, and n represents the signal sample point.

[0030] Segment the 15 types of original modulated signals collected. The length of each sample is , denoted as ; N represents the signal length. For the collected signal data, every 1000 points are divided into one sample;

[0031] The convolutional neural network and the long short-term memory network perform deep modulation feature extraction on the input time series signal, that is, the segmented signal (1000×2), through self-learning, and compress the data to a low-dimensional space through 2 convolutional layers and 2 long short-term memory networks, so as to complete the recognition of each known modulation type and the classification of known modulation types and unknown modulation types.

[0032] As Figure 3 shown, the network model includes 1 input layer (Input Layer), 2 convolutional layers (Convolutional Layer) {Conv1, Conv2}, 2 long short-term memory networks (Long Short-Term Memory, LSTM) {LSTM1, LSTM2} and 1 fully connected layer (Dense). The detailed parameters of the network model structure are shown in Table 1 specifically.

[0033] Table 1 Details of the detailed parameters of the network model structure;

[0034] The training, validation and testing processes based on contrastive learning and OpenMax are specifically as follows: Input: Training dataset ; Output: Classification probabilities of each known modulation type and unknown modulation type .

[0035] The dataset Divided into a training set, a validation set, and a test set; Based on the constructed compression model, set the batch size of the batch processing samples to 256, the number of iterations epoch to 100, and select Adam as the optimizer; Set the learning rate update and early stopping strategy for training, monitor the loss function val_loss of the validation set during the training process. When 5 training epochs have passed without the model performance improving, the learning rate will automatically be scaled to 0.1 times the original, and the minimum value of the learning rate is set to 10 -6 When the validation set loss function has not decreased by 0.001 in 15 training epochs, the training is terminated early; Randomly initialize the weights and biases of all layers of the open-set modulation recognition network based on contrast learning and OpenMax, including the convolutional neural network and the long short-term memory network; For epoch in 150 Input the training set data; Calculate the network output features; Calculate the original classification probability of the known modulation type output by the network; The loss function loss of the training set and the loss function val_loss of the validation set; Solve the parameter gradients and update the weights and biases of each layer; Save the optimal modulation recognition model and its parameters.

[0036] Obtain the signal deep modulation features through the trained network model, including: Process the segmented signal through the trained convolutional neural network to extract the spatio-temporal features of the signal. In the feature extraction stage, the processing process is shown in formula (1): (1); In the formula, is the spatio-temporal feature output by the convolutional neural network, W represents the convolutional kernel weights, b is the bias, is the activation function of the convolutional layer; i, j respectively represent the current i-th row and j-th column; m, n represents the convolutional kernel size, representing the length and width of the convolutional kernel respectively, l represents the number of channels, k represents the number of convolutional kernels; Then, flatten the extracted feature F into F ’ , F ’After passing through the long short-term memory network, the temporal modulation features of the signal, i.e., the signal depth modulation features, are extracted. The processing process is shown in Equation (2): (2); In Equation (2), G is the extracted temporal modulation feature, is LSTM the activation function of the layer; g represents the feature mapping relationship, b and c represent the bias vectors, and F’(t), i.e., F’, after expansion F ’ is a one-dimensional sequence, t represents the position point of the sequence, represents the hidden-to-hidden connection weight, U represents the input-to-hidden connection weight, V represents the hidden-to-output connection weight, h(t) represents the hidden layer state at time t, and h(t - 1) represents the hidden layer state at time t - 1.

[0037] According to the training data and the label data y , the optimization objective function of the algorithm, that is, the loss function, of the network model is the sum of the cross-entropy loss function and the supervised contrastive loss function. The calculation formulas are shown in Equations (3), (4), and (5) respectively: (3); (4); (5); In the formula, N represents the number of samples of the modulation signal, K represents the number of modulation categories, and represent the true label and the predicted label of the sample respectively, P(i) represents the number of samples with the same label as the sample i , represents the feature vector of the sample i , represents all the sample sets except i in the sample, represents the classification loss, represents the input vector, represents the bias vector, represents the weight matrix, is the supervised contrastive loss, represents the feature vector of other samples except sample i in A, represents the temperature parameter, represents the weight.

[0038] After the algorithm model is trained based on the training set, labels, and loss function, it further adjusts the output probability by introducing OpenMax function, that is, parameterize the SoftMax function. By reducing the Softmax scores of each known class one by one and dividing their differences among the unknown classes. Construct a distance set and use OpenMax to complete the classification and recognition of all modulated classes including the unknown class; including:

[0039] Step 1: Calculate the Softmax probability corresponding to each class of the trained network model, that is: (6); In formula (6), S C represents the output probability value corresponding to each class, K represents the number of classes, W T yi represents the weight of the i-th class, x i represents the input of the i-th class, b yi represents the bias of the i-th class; Step 2: Calculate the mean of the activation vectors of all training samples for each class to obtain the mean activation vector MAV (Mean Activation Vector, MAV) of this class. MAV represents the central position of this class in the feature space, and the mathematical expression is as follows:

[0040] (7); Among them, AV m represents the c th m input sample corresponding to the C C-th class, N c represents the total number of samples in the C-th class, and MAV is denoted as the central vector of the C-th class; (8); In formula (8), MAV k is denoted as the central vector of the K-th class, AV m represents the C th m input sample corresponding to the Step 4: Use the Weibull distribution in extreme value theory to fit the distance set for each category; The Weibull distribution is a probability distribution used to describe extreme value events and can well characterize the extreme values in the distance set. The probability density function of the Weibull distribution is:

[0041] (9); In formula (9), is the scale parameter, is the shape parameter; x represents the input; The cumulative distribution function corresponding to the Weibull distribution is: (10); Calculate the Weibull distribution corresponding to the classification according to formula (9) and formula (10); Step 5: Calculate the reduced probability corresponding to each category , and the calculation formula is as follows: (11); In formula (11), S c (x) is S C , d cx represents the distance, is the scale parameter, is the shape parameter; Step 6: Calculate the probability of the unknown category, that is, the probability of the unknown category is the difference between the original Softmax probability and the reduced probability, and the calculation formula is: (12); In formula (12), C is the number of categories; Step 7: Normalize the probabilities of all categories to obtain the classification probabilities of all categories , and the specific processing is as follows: (13).

[0042] Example 3 A computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the open-set modulation recognition method based on contrast learning and OpenMax described in Example 1 or 2.

[0043] Example 4 A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the steps of the open-set modulation recognition method based on contrast learning and OpenMax described in Example 1 or 2.

[0044] Embodiment 5 An open-set modulation recognition system based on contrastive learning and OpenMax, comprising: A signal training unit, configured to: obtain a modulation signal through signal acquisition and segment it to obtain segmented signals; input the segmented signals into a network model for training, extract deep signal features, achieve the original classification of known modulation categories, and obtain original classification probabilities; the network model includes a convolutional network and a long short-term memory network; A signal testing unit, configured to: obtain signal deep modulation features through the trained network model, construct a distance set, and use OpenMax to complete the classification and recognition of all modulation categories including unknown categories.

Claims

1. An open-set modulation recognition method based on contrastive learning and OpenMax, characterized in that Including: Obtain a modulated signal through signal acquisition and segment it to obtain segmented signals; Input the segmented signals into a network model for training, extract deep signal features, achieve the original classification of known modulation categories, and obtain the original classification probability; the network model includes a convolutional network and a long short-term memory network; Obtain the signal deep modulation features through the trained network model, construct a distance set, and use OpenMax to complete the classification and recognition of all modulation categories including unknown categories.

2. The open-set modulation recognition method based on contrastive learning and OpenMax according to claim 1, wherein Signal acquisition, including: Use a spectrum analyzer to collect 15 types of original modulation signals from the actual environment; the 15 types of original modulation signals include: {4QAM, 16QAM, 32QAM, 64QAM, BPSK, 8PSK, 16PSK, 32PSK, 2FSK, 4ASK, 4ASK, 8ASK, 16APSK, AM, FM}.

3. The open-set modulation recognition method based on contrastive learning and OpenMax according to claim 2, characterized in that, Segment the 15 types of original modulation signals collected.

4. The open-set modulation recognition method based on contrastive learning and OpenMax according to claim 1, characterized in that The convolutional neural network and the long short-term memory network perform deep modulation feature extraction on the input time series signals, i.e., segmented signals, through self-learning, and compress the data to a low-dimensional space through 2 convolutional layers and 2 long short-term memory networks, so as to complete the recognition of each known modulation type and the classification of known and unknown modulation types.

5. The open-set modulation recognition method based on contrastive learning and OpenMax according to claim 4, characterized in that Obtain the signal deep modulation features through the trained network model; including: Process the segmented signals through the trained convolutional neural network to extract the spatio-temporal features of the signals. In the feature extraction stage, the processing process is as shown in formula (1): (1)； Wherein, is the spatio-temporal feature output by the convolutional neural network, W represents the convolutional kernel weights, b is the bias, is the activation function of the convolutional layer; i, j respectively represent the current row i and column j; m, n represents the convolutional kernel size, representing the length and width of the convolutional kernel respectively, l represents the number of channels, k represents the number of convolutional kernels; Then, the extracted feature F is flattened into F ’ , F ’ After passing through the long short-term memory network, the temporal modulation feature of the signal, that is, the signal depth modulation feature, is extracted. The processing process is shown in Equation (2): (2)； In formula (2), G is the extracted time-series modulation feature, is LSTM the activation function of the layer; g represents the feature mapping relationship, b and c represent the bias vectors, and F’(t), that is, F’, after expansion F ’ is a one-dimensional sequence, t represents the position point of the sequence, represents the connection weight from hidden to hidden, U represents the connection weight from input to the hidden layer, V represents the connection weight from the hidden layer to the output layer, h(t) represents the state of the hidden layer at time t, and h(t - 1) represents the state of the hidden layer at time t - 1.

6. The open-set modulation recognition method based on contrastive learning and OpenMax according to claim 4, characterized in that Loss function of the network model It is the sum of the cross-entropy loss function and the supervised contrastive loss function, and the calculation formulas are shown in Formulas (3), (4), and (5) respectively: (3)； (4)； (5)； In the formula, N represents the number of samples of the modulation signal, K represents the number of modulation categories, and respectively represent the true label and the predicted label of the sample, P(i) represents the number of samples with the same label as the sample i , represents the sample i 's feature vector, represents all the sample sets except i in the sample, represents the classification loss, represents the input vector, represents the bias vector, represents the weight matrix, supervised contrastive loss, represents the feature vectors of other samples in A except sample i, represents the temperature parameter, represents the weight.

7. The open-set modulation recognition method based on contrastive learning and OpenMax according to claim 4, wherein Construct a distance set and use OpenMax to complete the classification and recognition of all modulation categories including unknown categories; including: Step 1: Calculate the Softmax probability corresponding to each category of the trained network model, i.e.: （6）； In formula (6), S C represents the output probability value corresponding to each category, K represents the number of categories, and W T yi represents the weight of the i-th category, x i represents the input of the i-th category, and b yi represents the bias of the i-th category; Step 2: Calculate the mean of the activation vectors of all training samples in each category to obtain the mean activation vector MAV of this category. The mathematical expression is as follows: （7）； in, AV m Indicates c Class m The activation vector corresponding to the input sample, N C Represents the total number of samples in category C, MAV c Denote as the center vector of the Cth class; Step 3: Calculate the distance between the activation vectors of all correctly classified samples in this category and the MAV of this category, using the Euclidean distance, to form the distance set of this category. The distance calculation formula is: （8）； In formula (8), MAV k is denoted as the center vector of the K-th class, AV m represents the C activation vector corresponding to the m th input sample of the class; Step 4: Use the Weibull distribution in extreme value theory to fit the distance set of each category; the probability density function of the Weibull distribution is: （9）； In formula (9), is the scale parameter, is the shape parameter; x represents the input; The cumulative distribution function corresponding to the Weibull distribution is: （10）； Calculate the Weibull distribution corresponding to the classification according to formula (9) and formula (10); Step Five: Calculate the reduced probability corresponding to each category , and the calculation formula is as follows: （11）； In formula (11), S c (x), that is, S C , d cx represents the distance, is the scale parameter, is the shape parameter; Step 6: Calculate the probability of the unknown category, that is, the probability of the unknown category is the original Softmax difference between the probability and the reduced probability, and the calculation formula is: （12）； In formula (12), C is the number of categories; Step Seven: Normalize the probabilities of all categories to obtain the classification probabilities of all categories , and the specific processing is as follows: （13）。 8. An open-set modulation recognition system based on contrastive learning and OpenMax, characterized in that Including: A signal training unit, configured to: obtain a modulated signal through signal acquisition and segment it to obtain segmented signals; Input the segmented signals into a network model for training, extract deep signal features, achieve the original classification of known modulation categories, and obtain the original classification probability; the network model includes a convolutional network and a long short-term memory network; A signal testing unit, configured to: obtain the signal deep modulation features through the trained network model, construct a distance set, and use OpenMax to complete the classification and recognition of all modulation categories including unknown categories.

Images