An open set recognition method for modulated signals based on spatial distribution feature extraction network

Abstract

The present invention discloses an open set recognition method for modulated signals based on a spatial distribution feature extraction network, comprising: obtaining a modulated signal to be recognized; inputting the modulated signal to be recognized into a trained spatial distribution feature extraction network to extract a spatial distribution feature vector to be recognized; using the spatial distribution feature vector to be recognized to determine whether the modulated signal to be recognized is an unknown class signal or a known class signal; and using the class corresponding to the minimum value of all elements in the spatial distribution feature vector of the modulated signal to be recognized that is determined to be a known class signal as the class of the modulated signal to be recognized, thereby completing the recognition of the modulated signal to be recognized. The recognition method of the present invention can effectively extract features in an open environment and can effectively identify the class of modulated signals in the field of spectrum monitoring, breaking the closed world assumption of the prior art.

Description

Technical Field

[0001] The present invention belongs to the technical field of modulation signals, and in particular relates to an open set recognition method for modulation signals based on a spatial distribution feature extraction network. Background Art

[0002] Modulation signal recognition technology is used to identify the modulation type of unknown signals and is widely used in signal detection, spectrum sharing, and interference identification. There are two traditional methods for modulation signal recognition technology, namely likelihood-based methods and feature-based methods. The likelihood-based method calculates the likelihood function of the received signal, estimates the unknown parameters, and then compares them with a pre-set threshold to obtain the signal type. This method is optimal in minimizing the probability of misclassification, but it is accompanied by high computational complexity and requires more prior knowledge. The method based on artificially designed features extracts the features of the received signal and then compares the features with the theoretical values. However, this method relies on artificially designed features and requires additional domain knowledge.

[0003] In recent years, deep learning has made significant progress in the field of modulated signal recognition thanks to its powerful automatic feature extraction capabilities. Compared with traditional methods, modulation signal recognition methods based on deep learning achieve higher efficiency and accuracy. Currently, researchers have also proposed many methods for modulation signal recognition based on deep learning, such as convolutional neural networks and long short-term memory networks. However, these methods can only recognize signal types that were present during the training phase (known classes), which is called closed-set recognition. For signal types that did not appear during the training phase (unknown classes), closed-set recognition methods mistakenly classify them as known classes, which limits their application in the real world.

[0004] Most of the above methods require feature extraction from a training set containing hundreds of samples per modulation type to achieve a certain level of recognition accuracy. However, in practice, some modulation types have only a few samples, significantly reducing the recognition accuracy of the algorithm and rendering it ineffective. Therefore, developing a method for efficient feature extraction and recognition in open environments is of great significance in the field of spectrum monitoring. Summary of the Invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a modulation signal open set recognition method based on a spatial distribution feature extraction network.

[0006] The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0007] A modulation signal open set recognition method based on a spatial distribution feature extraction network, the recognition method comprising:

[0008] Obtaining a modulation signal to be identified;

[0009] Inputting the modulated signal to be identified into a trained spatial distribution feature extraction network, mapping the modulated signal to be identified from a signal space to a feature space through the trained spatial distribution feature extraction network to obtain an original feature vector to be identified, and extracting the spatial distribution feature vector to be identified based on the original feature vector to be identified;

[0010] Determine whether the minimum value of all elements in the spatial distribution feature vector to be identified is greater than a preset threshold; if so, determine the modulated signal to be identified as an unknown class signal; if not, determine the modulated signal to be identified as a known class signal;

[0011] The category corresponding to the minimum value of all elements in the spatial distribution feature vector of the modulation signal to be identified, which is determined to be a known class signal, is used as the category of the modulation signal to be identified, thereby completing the identification of the modulation signal to be identified.

[0012] Optionally, the spatial distribution feature extraction network includes K1 convolution modules, K2 fully connected layers, a spatial distribution feature extraction layer and a classifier based on Euclidean distance, which are connected in sequence.

[0013] Optionally, each of the convolution modules includes a convolution layer, a ReLu activation layer, a batch normalization layer and a maximum pooling layer connected in sequence.

[0014] Optionally, K1 is 8, K2 is 2, the number of convolution kernels of the 8 convolution layers in the 8 convolution modules are 8, 16, 8, 32, 64, 128, 128, and 128, respectively, the convolution kernel sizes of the 8 convolution layers are all 3×3, the padding is 2, and the step size is 1; the padding of the maximum pooling layer is 2; the 2 fully connected layers include 128 nodes and M nodes, respectively.

[0015] Optionally, the training method of the spatial distribution feature extraction network includes:

[0016] S1.1. Obtain a training sample set, wherein the training sample set includes M types of sample modulation signals, each type of the sample modulation signals having several different signal-to-noise ratios, and each signal-to-noise ratio of each type of the sample modulation signals includes m1 signals;

[0017] S1.2. The signal sequence corresponding to the sample modulation signal is sequentially input into the spatial distribution feature extraction network, the spatial distribution feature extraction network outputs the corresponding prediction category, and the spatial distribution feature extraction network is trained by the stochastic gradient descent algorithm to obtain the trained spatial distribution feature extraction network.

[0018] Optionally, step S1.2 includes:

[0019] S1.21. Inputting the signal sequence corresponding to the sample modulation signal into the K1 convolution modules of the spatial distribution feature extraction network in sequence to map the sample modulation signal from the signal space to the feature space to obtain an M-dimensional original feature vector;

[0020] S1.22. Input the M-dimensional original feature vector into the K2 fully connected layers to convert the original feature vector into a one-dimensional feature vector;

[0021] S1.23. Input the one-dimensional feature vector to a spatial distribution feature extraction layer to calculate the Euclidean distance between the one-dimensional feature vector and the class center vector of each class of sample modulation signals, and obtain a first spatial distribution feature vector of the sample modulation signal based on the Euclidean distance between the one-dimensional feature vector and the class center vector of each class of modulation signals;

[0022] S1.24. Inputting the first spatial distribution feature vector into the Euclidean distance-based classifier to select the category of the class center vector corresponding to the shortest Euclidean distance in the first spatial distribution feature vector as the classification result of the sample modulation signal, and predicting the probability of the sample modulation signal belonging to each category based on the label prediction model;

[0023] S1.25. Calculate the hybrid loss function for the current number of iterations based on the classification result and the probability, and update the parameters of the spatial distribution feature extraction network using a stochastic gradient descent algorithm to obtain an updated spatial distribution feature extraction network.

[0024] S1.26. Determine whether the maximum number of iterations has been reached. If so, use the updated spatial distribution feature extraction network obtained in step S1.25 as the trained spatial distribution feature extraction network. If not, repeat steps S1.21 to S1.26 until the maximum number of iterations has been reached to obtain the trained spatial distribution feature extraction network.

[0025] Optionally, the label prediction model is:

[0026]

[0027] Among them, p ED (y=k|x) is the probability, x is the sample modulation signal, is the first spatial distribution eigenvector, is the class center vector of the k-th class sample modulation signal, k=1,2,....,M, is the class center vector of the modulation signal of the i-th class sample.

[0028] Optionally, the hybrid loss function is:

[0029] L Hbrid =L EDCE +λ·L Max-Min

[0030] Among them, L Hbrid is the mixed loss function, L EDCE is the cross entropy loss function based on Euclidean distance, L Max-Min is the spatial distribution loss function, λ is a hyperparameter, λ∈[0,1];

[0031] L EDCE =-logp ED (y=k|x)

[0032]

[0033] Among them, N i is the number of signals in the i-th sample modulation signal, is the one-dimensional feature vector output by the p-th signal in the i-th type sample modulation signal through the K2 fully connected layers in the spatial distribution feature extraction network, is the class center vector of the modulation signal of the j-th class sample.

[0034] Optionally, after obtaining the trained spatial distribution feature extraction network, the method further includes:

[0035] S2.1. Obtain a training test set, where the training test set includes N types of test modulation signals, each type of the test modulation signals having several different signal-to-noise ratios, and each signal-to-noise ratio of each type of the test modulation signals includes n signals;

[0036] S2.2. Sequentially input the signal sequence corresponding to the test modulation signal into the trained spatial distribution feature extraction network to obtain a second spatial distribution feature vector output by the spatial distribution feature extraction layer in the trained spatial distribution feature extraction network;

[0037] S2.3. Determine whether the minimum value of all elements in the second spatial distribution feature vector is greater than a preset threshold; if so, determine the test modulated signal as an unknown signal; if not, determine the test modulated signal as a known signal;

[0038] S2.4. The category corresponding to the minimum value of all elements in the second spatial distribution feature vector of the test modulation signal determined to be a known class signal is used as the category of the test modulation signal.

[0039] Optionally, the calculation formula of the preset threshold is:

[0040]

[0041] Among them, τ is the preset threshold, N l is the number of signals in the lth type of modulation signal, is the one-dimensional feature vector output by the K2 fully connected layers in the trained spatial distribution feature extraction network of the qth signal in the lth type of modulation signal, is the class center vector of the lth class modulation signal.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] The recognition method of the present invention provides a spatial distribution feature extraction network, which is a deep feature extraction network. The spatial distribution feature extraction network can map the modulated signal from the original signal space to the feature space and extract the spatial distribution features. The extracted spatial distribution features are then compared with the preset threshold to determine which are unknown class signals and which are known class signals, thereby classifying the known class signals to obtain the recognition results of the modulated signals. This method can effectively extract features in an open environment and effectively identify the categories of modulated signals in the field of spectrum monitoring, breaking the closed world assumption of the prior art.

[0044] This paper uses a designed hybrid loss function to train a spatially distributed feature extraction network, making samples of the same category more compact in the feature space and samples of different categories more separated, thereby achieving more efficient training. Furthermore, the spatially distributed feature extraction network introduces a Euclidean distance-based classifier to reject signals of unknown classes, breaking the closed-world assumption of existing techniques.

[0045] The present invention will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 1 is a flow chart of a method for open-set recognition of modulated signals based on a spatial distribution feature extraction network provided by an embodiment of the present invention;

[0047] Figure 2 is a schematic diagram of a spatial distribution feature extraction network provided by an embodiment of the present invention;

[0048] Figure 3 It is a schematic diagram of a training and testing method for a spatial distribution feature extraction network provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0050] Example 1

[0051] Most existing modulation signal recognition methods are based on the assumption of a closed world. The basic idea is: for the modulation signal, first design a neural network such as a convolutional neural network as a feature extractor, and use the Softmax nonlinear function as the final output layer; then use the cross-entropy loss function to optimize the network parameters; finally, use the category corresponding to the maximum value of the Softmax function output as the recognition result. The existing technology lacks the robustness to handle unknowns outside of the category in an open environment, and cannot meet the requirements of practical applications in a constantly changing environment. When untrained samples are mixed into the test set, the existing technology mistakenly classifies them as samples of known categories. The purpose of the present invention is to provide an open set recognition method for modulation signals based on a spatially distributed feature extraction network, which solves the problem that traditional recognition methods cannot handle open-world recognition well, and improves recognition performance and robustness.

[0052] See Figure 1 , Figure 1 1 is a flow chart of a method for identifying an open set of modulated signals based on a spatial distribution feature extraction network provided by an embodiment of the present invention. The present invention provides a method for identifying an open set of modulated signals based on a spatial distribution feature extraction network, the method comprising:

[0053] Step 1: Obtain the modulated signal to be identified.

[0054] Step 2: Input the modulated signal to be identified into the trained spatial distribution feature extraction network, so as to map the modulated signal to be identified from the signal space to the feature space through the trained spatial distribution feature extraction network to obtain the original feature vector to be identified, and extract the spatial distribution feature vector of the modulated signal to be identified based on the original feature vector to be identified.

[0055] Specifically, the modulated signal to be identified is input into a trained spatial distribution feature extraction network. The trained spatial distribution feature extraction network first maps the modulated signal to be identified from the signal space to the feature space to obtain the original feature vector to be identified. Then, the original feature vector to be identified is converted into a one-dimensional feature vector to be identified. Then, the Euclidean distance between the one-dimensional feature vector to be identified and the class center vector of each type of modulation signal is calculated, and the spatial distribution feature vector to be identified is obtained based on the Euclidean distance between the original feature vector to be identified and the class center vector of each type of modulation signal. A sample library is established, and the sample library contains various analog modulation signals for calculating the Euclidean distance with the one-dimensional feature vector to be identified.

[0056] Step 3: Determine whether the minimum value of all elements in the spatial distribution feature vector to be identified is greater than a preset threshold. If so, the modulated signal to be identified is determined to be an unknown class signal. If not, the modulated signal to be identified is determined to be a known class signal.

[0057] Step 4: The category corresponding to the minimum value of all elements in the spatial distribution feature vector of the modulation signal to be identified that is determined to be a known class signal is taken as the category of the modulation signal to be identified, thereby completing the identification of the modulation signal to be identified.

[0058] In a specific embodiment, see Figure 2 ,The spatial distribution feature extraction network includes K1 convolution modules, K2 fully connected layers, ,the spatial distribution feature extraction layer and the Euclidean distance based classifier,which are connected in sequence.

[0059] Optionally, each convolution module contains a convolution layer, a ReLu activation layer, a batch normalization layer, and a maximum pooling layer connected in sequence.

[0060] Furthermore, K1 is 8, K2 is 2, and the number of convolution kernels of the 8 convolution layers in the 8 convolution modules are 8, 16, 8, 32, 64, 128, 128, and 128, respectively. The convolution kernel size of the 8 convolution layers is 3×3, the padding is 2, and the stride is 1; the padding of the maximum pooling layer is 2; the two fully connected layers include 128 nodes and M nodes, respectively.

[0061] The present invention also provides a training method for a spatial distribution feature extraction network, see Figure 3 , the training method of the spatial distribution feature extraction network includes:

[0062] S1.1. Obtain a training sample set, where the training sample set includes M types of sample modulation signals, each type of sample modulation signal has several different signal-to-noise ratios, and each signal-to-noise ratio of each type of sample modulation signal includes m1 signals.

[0063] For example, the sample modulation signals in the training sample set are selected from the RadioML dataset. Each type of sample modulation signal contains 15 different signal-to-noise ratios, which are -14dB, -12dB, -10dB, -8dB, -6dB, -4dB, -2dB, 0dB, 2dB, 4dB, 6dB, 8dB, 10dB, 12dB, and 14dB.

[0064] S1.2. The signal sequence corresponding to the sample modulation signal is sequentially input into the spatial distribution feature extraction network. The spatial distribution feature extraction network outputs the corresponding prediction category, and the spatial distribution feature extraction network is trained by the stochastic gradient descent algorithm to obtain a trained spatial distribution feature extraction network.

[0065] Here, the maximum number of iterations is initially set to T, T≥100, and the initial iteration number t=0. When the iteration number reaches the maximum number of iterations T, the training ends and a trained spatial distribution feature extraction network is obtained.

[0066] In an optional embodiment, step S1.2 includes:

[0067] S1.21. Input the signal sequence corresponding to the sample modulation signal into the K1 convolution modules of the spatial distribution feature extraction network in sequence to map the sample modulation signal from the signal space to the feature space to obtain the M-dimensional original feature vector.

[0068] S1.22. Input the M-dimensional original feature vector into K2 fully connected layers to convert the original feature vector into a one-dimensional feature vector;

[0069] S1.23. Input the one-dimensional feature vector into the spatial distribution feature extraction layer to calculate the Euclidean distance between the one-dimensional feature vector and the class center vector of each type of sample modulation signal, and obtain the first spatial distribution feature vector of the sample modulation signal based on the Euclidean distance between the one-dimensional feature vector and the class center vector of each type of modulation signal.

[0070] Here, the Euclidean distance between the one-dimensional feature vector and the class center vector of the i-th class sample modulation signal is 1≤i≤M.

[0071] After obtaining the Euclidean distance between the one-dimensional feature vector and the class center vector of each class of sample modulation signals, the Euclidean distance between the one-dimensional feature vector and the class center vector of each class of sample modulation signals can be used as an element in the first spatial distribution feature vector, thereby obtaining the first spatial distribution feature vector. The first spatial distribution feature vector is expressed as:

[0072] S1.24. Input the first spatial distribution feature vector into the classifier based on Euclidean distance to select the category of the class center vector corresponding to the shortest Euclidean distance in the first spatial distribution feature vector as the classification result of the sample modulated signal, and predict the probability of the sample modulated signal being each category based on the label prediction model.

[0073] Here, the label prediction model is:

[0074]

[0075] Among them, p ED (y=k|x) is the probability, x is the sample modulation signal, is the first spatial distribution eigenvector, is the class center vector of the k-th class sample modulation signal, k=1,2,....,M, is the class center vector of the modulation signal of the i-th class sample.

[0076] S1.25. Based on the classification results and probabilities, calculate the mixed loss function of the current number of iterations, and update the parameters of the spatial distribution feature extraction network through the stochastic gradient descent algorithm to obtain the updated spatial distribution feature extraction network.

[0077] Here, the hybrid loss function is:

[0078] L Hbrid =L EDCE +λ·L Max-Min

[0079] Among them, L Hbrid is the mixed loss function, L EDCE is the cross entropy loss function based on Euclidean distance, L EDCE The output of the spatial distribution feature extraction layer is used as the input feature of the loss function, L Max-Min is the spatial distribution loss function, λ is a hyperparameter, λ∈[0,1];

[0080] L EDCE =-logp ED (y=k|x)

[0081]

[0082] Among them, N i is the number of signals in the i-th sample modulation signal, is the one-dimensional feature vector output by the p-th signal in the i-th type sample modulation signal through the K2 fully connected layers in the spatial distribution feature extraction network, is the class center vector of the j-th class sample modulation signal, 1≤j≤M, is the average intra-class distance of the modulated signal of the i-th class sample, is the distance between the modulation signal of the i-th sample and the modulation signal of the j-th sample.

[0083] This embodiment adopts the hybrid loss function as described above, and updates parameters such as the weights of the convolution module in the spatial distribution feature extraction network and the class center vector of the spatial distribution feature extraction layer through the stochastic gradient descent algorithm to obtain an updated spatial distribution feature extraction network.

[0084] S1.26. Determine whether the maximum number of iterations has been reached. If so, use the updated spatial distribution feature extraction network obtained in step S1.25 as the trained spatial distribution feature extraction network. If not, repeat steps S1.21 to S1.26 until the maximum number of iterations has been reached to obtain a trained spatial distribution feature extraction network.

[0085] After obtaining the trained spatial distribution feature extraction network, the present invention also uses the training test set to test the trained spatial distribution feature extraction network, that is, using the trained spatial distribution feature extraction network to classify the signal category of each sample in the training test set, specifically including:

[0086] S2.1. Obtain a training test set, where the training test set includes N types of test modulation signals, each type of test modulation signal has several different signal-to-noise ratios, and each signal-to-noise ratio of each type of test modulation signal includes n signals.

[0087] S2.2. Input the signal sequence corresponding to the test modulation signal into the trained spatial distribution feature extraction network in sequence to obtain the second spatial distribution feature vector output by the spatial distribution feature extraction layer in the trained spatial distribution feature extraction network.

[0088] S2.3. Determine whether the minimum value of all elements in the second spatial distribution feature vector is greater than a preset threshold. If so, determine the test modulated signal as an unknown type signal; if not, determine the test modulated signal as a known type signal.

[0089] Here, the calculation formula of the preset threshold is:

[0090]

[0091] Among them, τ is the preset threshold, N l is the number of signals in the lth type of modulation signal, is the one-dimensional feature vector output by the K2 fully connected layers in the trained spatial distribution feature extraction network of the qth signal in the lth type of modulation signal, is the class center vector of the modulation signal of the lth class sample.

[0092] S2.4. The category corresponding to the minimum value of all elements in the second spatial distribution feature vector of the test modulation signal determined to be a known class signal is used as the category of the test modulation signal.

[0093] The recognition method of the present invention provides a spatial distribution feature extraction network, which is a deep feature extraction network. The spatial distribution feature extraction network can map the modulated signal from the original signal space to the feature space and extract the spatial distribution features. The extracted spatial distribution features are then compared with the preset threshold to determine which are unknown class signals and which are known class signals, thereby classifying the known class signals to obtain the recognition results of the modulated signals. This method can effectively extract features in an open environment and effectively identify the categories of modulated signals in the field of spectrum monitoring, breaking the closed world assumption of the prior art.

[0094] This paper uses a designed hybrid loss function to train a spatially distributed feature extraction network, making samples of the same category more compact in the feature space and samples of different categories more separated, thereby achieving more efficient training. Furthermore, the spatially distributed feature extraction network introduces a Euclidean distance-based classifier to reject signals of unknown classes, breaking the closed-world assumption of existing techniques.

[0095] It should be noted that the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of technical features indicated. Therefore, features defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0096] In the description of this specification, the reference terms "one embodiment," "some embodiments," "example," "specific example," or "some examples" mean that the specific features or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described can be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification.

[0097] Although the present invention has been described herein with reference to various embodiments, those skilled in the art may understand and implement other variations of the disclosed embodiments by reviewing the drawings and the disclosure in the process of implementing the claimed invention. In the specification, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple components or steps. The fact that certain measures are described in different embodiments does not mean that these measures cannot be combined to produce good results.

[0098] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A modulated signal open set recognition method based on a spatial distribution feature extraction network, characterized in that: The identification method comprises: Obtaining a modulation signal to be identified; The modulated signal to be identified is input into a trained spatial distribution feature extraction network, so that the modulated signal to be identified is mapped from the signal space to the feature space through the trained spatial distribution feature extraction network to obtain the original feature vector to be identified, and the spatial distribution feature vector to be identified is extracted according to the original feature vector to be identified; the spatial distribution feature extraction network includes K1 convolution modules, K2 fully connected layers, spatial distribution feature extraction layers and a classifier based on Euclidean distance connected in sequence, and each of the convolution modules includes a convolution layer, a ReLu activation layer, a batch normalization layer and a maximum pooling layer connected in sequence; K1 is 8, K2 is 2, and the number of convolution kernels of the 8 convolution layers in the 8 convolution modules is 8, 16, 8, 32, 64, 128, 128, and 128, respectively. The convolution kernel size of the 8 convolution layers is 3×3, the padding is 2, and the step size is 1; the padding of the maximum pooling layer is 2; the 2 fully connected layers include 128 nodes and M nodes respectively; Determine whether the minimum value of all elements in the spatial distribution feature vector to be identified is greater than a preset threshold; if so, determine the modulated signal to be identified as an unknown class signal; if not, determine the modulated signal to be identified as a known class signal; The class corresponding to the minimum value of all elements in the spatial distribution feature vector of the modulation signal to be identified, which is determined to be a known class signal, is used as the class of the modulation signal to be identified, thereby completing the identification of the modulation signal to be identified; The training method of the spatial distribution feature extraction network includes: S1.

1. Obtain a training sample set, wherein the training sample set includes M types of sample modulation signals, each type of the sample modulation signals having several different signal-to-noise ratios, and each signal-to-noise ratio of each type of the sample modulation signals includes m1 signals; S1.

2. Sequentially inputting the signal sequence corresponding to the sample modulation signal into the spatial distribution feature extraction network, the spatial distribution feature extraction network outputting the corresponding predicted category, predicting the probability of the sample modulation signal belonging to each category based on the label prediction model, and calculating the mixed loss function for the current number of iterations based on the classification result and the probability, and training the spatial distribution feature extraction network using a stochastic gradient descent algorithm to obtain the trained spatial distribution feature extraction network; The label prediction model is: Among them, p ED (y=k|x) is the probability, x is the sample modulation signal, is the first spatial distribution eigenvector, is the class center vector of the k-th class sample modulation signal, k=1,2,....,M, is the class center vector of the modulation signal of the i-th class sample; The hybrid loss function is: L Hbrid =L EDCE +λ·L Max-Min Among them, L Hbrid is the mixed loss function, L EDCE is the cross entropy loss function based on Euclidean distance, L Max-Min is the spatial distribution loss function, λ is a hyperparameter, λ∈[0,1]; L EDCE =-logp ED (y=k|x) Among them, N i is the number of signals in the i-th sample modulation signal, is the one-dimensional feature vector output by the p-th signal in the i-th type sample modulation signal through the K2 fully connected layers in the spatial distribution feature extraction network, is the class center vector of the modulation signal of the j-th class sample.

2. The modulation signal open set recognition method based on the spatial distribution feature extraction network according to claim 1 is characterized in that: Step S1.2 includes: S1.

21. Inputting the signal sequence corresponding to the sample modulation signal into the K1 convolution modules of the spatial distribution feature extraction network in sequence to map the sample modulation signal from the signal space to the feature space to obtain an M-dimensional original feature vector; S1.

22. Input the M-dimensional original feature vector into the K2 fully connected layers to convert the original feature vector into a one-dimensional feature vector; S1.

23. Input the one-dimensional feature vector to a spatial distribution feature extraction layer to calculate the Euclidean distance between the one-dimensional feature vector and the class center vector of each class of sample modulation signals, and obtain a first spatial distribution feature vector of the sample modulation signal based on the Euclidean distance between the one-dimensional feature vector and the class center vector of each class of modulation signals; S1.

24. Inputting the first spatial distribution feature vector into the Euclidean distance-based classifier to select the category of the class center vector corresponding to the shortest Euclidean distance in the first spatial distribution feature vector as the classification result of the sample modulation signal, and predicting the probability of the sample modulation signal belonging to each category based on the label prediction model; S1.

25. Calculate the hybrid loss function for the current number of iterations based on the classification result and the probability, and update the parameters of the spatial distribution feature extraction network using a stochastic gradient descent algorithm to obtain an updated spatial distribution feature extraction network. S1.

26. Determine whether the maximum number of iterations has been reached. If so, use the updated spatial distribution feature extraction network obtained in step S1.25 as the trained spatial distribution feature extraction network. If not, repeat steps S1.21 to S1.26 until the maximum number of iterations has been reached to obtain the trained spatial distribution feature extraction network.

3. The modulation signal open set recognition method based on the spatial distribution feature extraction network according to claim 1 is characterized in that: After obtaining the trained spatial distribution feature extraction network, the method further includes: S2.

1. Obtain a training test set, where the training test set includes N types of test modulation signals, each type of the test modulation signals having several different signal-to-noise ratios, and each signal-to-noise ratio of each type of the test modulation signals includes n signals; S2.

2. Sequentially input the signal sequence corresponding to the test modulation signal into the trained spatial distribution feature extraction network to obtain a second spatial distribution feature vector output by the spatial distribution feature extraction layer in the trained spatial distribution feature extraction network; S2.

3. Determine whether the minimum value of all elements in the second spatial distribution feature vector is greater than a preset threshold; if so, determine the test modulated signal as an unknown signal; if not, determine the test modulated signal as a known signal; S2.

4. The category corresponding to the minimum value of all elements in the second spatial distribution feature vector of the test modulation signal determined to be a known class signal is used as the category of the test modulation signal.

4. The modulation signal open set recognition method based on the spatial distribution feature extraction network according to claim 1 or 3, characterized in that: The calculation formula of the preset threshold is: Among them, τ is the preset threshold, N l is the number of signals in the lth type of modulation signal, is the one-dimensional feature vector output by the K2 fully connected layers in the trained spatial distribution feature extraction network of the qth signal in the lth type of modulation signal, is the class center vector of the lth class modulation signal.

Images