Communication interference signal open set identification method, system, equipment and medium

Abstract

The invention discloses a communication interference signal open set identification method, system and device and a medium. The method comprises the following steps: carrying out short-time Fourier transform on a communication interference signal to obtain a time-frequency graph; performing multi-scale feature representation on the time-frequency diagram by using a Res2Net-based multi-scale feature extraction network; constructing a projection layer to carry out hyperspherical embedding on the multi-scale features, carrying out modeling by utilizing vMF distribution, and designing a compactness and dispersity loss function; multi-scale features after modeling are input into an OpenMax layer for open set recognition, output of a projection layer is used as an activation vector to improve OpenMax, Euclidean distance is replaced with cosine distance, input sample unknown is calculated based on an extreme value theory, category probability distribution is adjusted, and some abnormal samples can be classified into unknown categories; the Res2Net more flexibly captures multi-scale information than the convolutional network; the separation of inter-class features and the compactness of intra-class features are effectively enhanced through the hyperspherical vMF distribution; the OpenMax layer endows the network with unknown category identification capability; the method has better performance and wider universality; the system, the equipment and the medium are used for realizing the communication interference signal open set identification method.

Description

Technical Field

[0001] The present invention belongs to the field of communication technologies, and particularly relates to a method, a system, a device and a medium for identifying open-set communication interference signals. Background Art

[0002] Modern wireless communication systems are widely used in multiple fields such as military, civilian, and aerospace. Their stability and security directly affect the reliability of communication. Identifying communication interference signals is a core link in wireless communication anti-jamming technology. Its goal is to accurately determine the type and parameters of interference by analyzing the signal characteristics in the electromagnetic environment, providing a decision-making basis for subsequent anti-jamming strategies. Due to the unpredictability of the actual electromagnetic environment, interference signals often have characteristics such as time-variation, non-linearity, and randomness. Some interference signals may not even have appeared in previous training data, that is, they have open-set characteristics. Traditional methods based on classification and pattern matching are difficult to deal with such unknown interferences. Therefore, exploring a cognitive mechanism for unknown interferences and combining data-driven open-set recognition technologies to improve the system's adaptability to complex electromagnetic environments has become one of the key technical challenges in the current communication anti-jamming field.

[0003] Researchers at home and abroad have classified the methods for open-set recognition problems into discriminative and generative categories. Among them, the discriminative method is to judge whether a sample belongs to a known or unknown category by establishing discriminative indicators such as empirical thresholds or distance thresholds, and the generative method is to generate data through a generative network to simulate unknown classes for training.

[0004] Regarding the discriminant method, Yoshihashi et al. proposed a classification-reconstruction learning (CROSR) method for open-set recognition, which combines the classification task and the reconstruction task, learns more generalized feature representations through a deep hierarchical reconstruction network, and uses Weibull distribution modeling for unknown detection (Yoshihashi R, Shao W, Kawakami R, et al. Classification-Reconstruction Learning for Open-Set Recognition[C]. 2019 IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019: 4011-4020.). Sun et al. proposed a conditional Gaussian distribution learning method for open-set recognition (Sun X, Yang Z, Zhang C, et al. Conditional gaussian distribution learning for open set recognition[C]. Proceedings of the IEEE / CVF conference on computer vision and pattern recognition. 2020: 13480-13489.). Although the above generative methods have certain open-set recognition capabilities, the training of these methods is based on the closed-set assumption, and the open-set space is not specially considered.

[0005] Regarding generative methods, Neal et al. proposed an open-set recognition method based on counterfactual images, which uses generative adversarial networks to generate samples that are close to known classes but do not belong to any training classes, thereby enhancing the training data and converting open-set recognition into a multi-classification problem (Neal L, Olson M, Fern X, et al. Open set learning with counterfactual images[C]. Proceedings of the European conference on computer vision (ECCV). 2018:613-628.). Schlachter proposed an open-set recognition method based on Intra-Class Splitting. By dividing the data of known classes into typical samples and atypical samples, and using the latter to simulate unknown classes, the open-set problem is converted into an (N+1)-class classification problem. In addition, closed-set regularization is introduced to ensure the accuracy of the classifier on known classes (Schlachter P, Liao Y, Yang B. Open-set recognition using intra-class splitting[C]. 2019 27th European signal processing conference (EUSIPCO). IEEE, 2019:1-5.). The above methods help the model identify unknown classes by generating samples of unknown classes, but the generated samples do not fully represent the data of unknown classes.

[0006] Through the above analysis, the problems and defects of the existing technologies are as follows:

[0007] (1) The existing communication interference signal recognition methods mainly focus on closed-set recognition. However, with the continuous evolution of wireless communication interference technologies, new interference signals emerge in an endless stream, which poses great challenges to traditional closed-set recognition methods. Especially in a highly dynamic and complex electromagnetic environment, many interference signals may not appear during the training process, resulting in recognition errors.

[0008] (2) In the existing open-set recognition methods, discriminative methods are trained based on the closed-set assumption and do not fully consider the differences between known-class data and unknown-class data. Therefore, when unknown interference signals or new communication environments appear, they cannot effectively identify unknown classes, leading to a significant decline in recognition accuracy. Generative models learn the data distribution by generating samples and are used for open-set recognition or adversarial sample generation. However, the generated samples often cannot fully reflect the complexity and diversity of unknown interference signals. Especially in the field of communication interference, in the face of a complex and uncertain interference environment, the performance of generative methods may not be ideal. Summary of the Invention

[0009] In order to overcome the problems existing in the above-mentioned prior art, the purpose of the present invention is to provide a method, system, device and medium for open-set recognition of communication interference signals; perform short-time Fourier transform on the communication interference signals to obtain a time-frequency diagram; use a multi-scale feature extraction network based on Res2Net to perform multi-scale feature characterization on the time-frequency diagram; perform hypersphere embedding on the multi-scale features through constructing a projection layer, model the features using the von Mises-Fisher (vMF) distribution under the hypersphere, and design a compactness loss function and a dispersion loss function; perform open-set recognition using the OpenMax layer, improve OpenMax using the output of the projection layer as the activation vector, and replace the Euclidean distance in the OpenMax algorithm with the cosine distance under the hypersphere to calculate the unknownness of the input sample based on extreme value theory, and readjust the class probability distribution so that some samples with high confidence but anomalies can be classified as unknown classes; having higher recognition accuracy and stronger versatility.

[0010] In order to achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0011] A method for open-set recognition of communication interference signals, characterized in that the method includes:

[0012] Step S1: Perform short-time Fourier transform on the communication interference signals to obtain a time-frequency diagram, and use the obtained time-frequency diagram as the input of the multi-scale feature extraction network;

[0013] Step S2: Use a multi-scale feature extraction network based on Res2Net to perform multi-scale feature characterization on the time-frequency diagram obtained in Step S1;

[0014] Step S3: Perform hypersphere embedding on the multi-scale features in Step S2 through constructing a projection layer, model the features using the vMF distribution under the hypersphere, and design a compactness loss function and a dispersion loss function;

[0015] Step S4: Input the multi-scale features modeled in Step S3 into the OpenMax layer for open-set recognition, improve OpenMax using the output of the projection layer as the activation vector, and replace the Euclidean distance in the OpenMax algorithm with the cosine distance under the hypersphere to calculate the unknownness of the input sample based on extreme value theory, and readjust the class probability distribution so that some samples with high confidence but anomalies can be classified as unknown classes.

[0016] The specific method of Step S1 is as follows:

[0017] The short-time Fourier transform of the communication interference signals is expressed as:

[0018]

[0019] where \(x(t)\) is the source signal, is the window function, and STFT x (t, f) is the time-frequency diagram.

[0020] The specific method of step S2 is as follows:

[0021] Input the time-frequency diagram STFT x (t, f) obtained in step S1 into the multi-scale feature extraction network \(g(\cdot)\) to obtain the multi-scale feature \(g(\text{STFT} x (t, f))\);

[0022] The multi-scale feature extraction network is composed of six Res2Net networks connected in series; the input of the first layer of the Res2Net network is the time-frequency diagram, and the output is the multi-scale feature. The input of each layer of the Res2Net network from the second layer to the sixth layer is the multi-scale feature output by the previous layer, and the output is a deeper multi-scale feature; the input of each layer of the Res2Net network, after being convolved by conv1D(k = 1), where \(k\) is the convolution kernel length, divides the multi-scale feature of \(D\) channels into \(s\) multi-scale feature subsets, denoted by \(x i where \(i\in\{1, 2, \cdots, s\}\); each multi-scale feature subset \(x i has the same spatial size and the number of channels is \(D / s\); except for \(x1\), each multi-scale feature subset \(x i has a corresponding conv1D(k = 3) convolution, denoted as \(K i ()\); let \(y i represent the output of \(K i ()\); the multi-scale feature subset \(x i is added to the output of \(K i-1 ()\), and then input into \(K i ()\); the conv1D(k = 3) convolution of \(x1\) is omitted, and the output \(y i can be expressed as:

[0023]

[0024] The specific method of step S3 is as follows:

[0025] Step S3.1: Use the projection layer \(h(\cdot)\) to map the multi-scale feature \(g(\text{STFT} x (t, f))\) obtained in step S2 to a low-dimensional feature space to generate a feature representation and further perform unit normalization on it to finally obtain the normalized embedded feature \(z\):

[0026]

[0027] Among them, denotes the 2-norm of;

[0028] such that the embedded features lie on a unit hypersphere:

[0029]

[0030] where S d is the unit hypersphere;

[0031] Step S3.2: Model the distribution of the embedded features on the hypersphere using the vMF distribution. For the unit vector in the communication interference signal class k, its probability density function is defined as p d as:

[0032]

[0033] where μ k is the class center vector with unit norm, κ≥0 represents the magnitude of the distribution compactness, the distribution density around the mean direction μ k , Z d (κ) is the normalization factor; under the vMF distribution, the conditional probability that an embedded vector z is assigned to class k' is:

[0034]

[0035] where K is the number of known classes of communication interference signals;

[0036] Step S3.3: Design two key loss functions based on the vMF distribution on the hypersphere: the compactness loss and the dispersion loss;

[0037] The optimization objective of the compactness loss is to make each class of communication interference signal samples input during training more closely surround the class center in the hypersphere embedding space, improve the within-class compactness of each class of communication interference signals, and make all communication interference signal samples of the same class more similar to each other; the compactness loss is expressed as:

[0038]

[0039] where z i denotes the hypersphere embedding of the sample, μ k'(i) is the center vector of the class to which the sample belongs, τ is the temperature parameter used to control the smoothness of the probability distribution, and N is the number of communication interference signal samples;

[0040] The optimization objective of the dispersion loss is to make the input different-class communication interference signal samples during training as far apart as possible in the hypersphere embedding space, reducing the cosine similarity between the class centers of different communication interference signal classes; the dispersion loss is expressed as:

[0041]

[0042] where, μ j and μ k represent the class center vectors of class j and class k respectively.

[0043] The specific method in step S4 is as follows:

[0044] Step S4.1: Use a multi-layer perceptron to construct an OpenMax layer. The multi-layer perceptron is composed of a fully connected layer, a ReLU activation layer, a fully connected layer, and a Softmax layer stacked in sequence;

[0045] Step S4.2: Use the multi-scale feature extraction network, projection layer, and OpenMax layer described in step S2, step S3.1, and step S4.1 to be connected in sequence to construct an open-set recognition network for communication interference signals;

[0046] Step S4.3: Use the simulated communication interference signals to train the open-set recognition network for communication interference signals described in step S4.2 to obtain a trained network. The optimization objective during training is to add the compactness loss and dispersion loss described in step S3.3 on the basis of the cross-entropy loss. Its overall loss function is expressed as:

[0047]

[0048] where, λ d and λ c are hyperparameters, which are used to control the weights of the dispersion loss and the compactness loss respectively, is the cross-entropy loss;

[0049] Step S4.4: Input the communication interference signals into the trained network for classification, and use the output feature vector of the projection layer as the activation vector to improve the OpenMax algorithm. The activation vector AV of the correctly classified communication interference signal is expressed as:

[0050]

[0051] where, z is the hypersphere embedding vector normalized by the projection layer, indicating the position of the communication interference signal sample in the hypersphere space;

[0052] Step S4.5: Calculate the mean activation vector of each type of communication interference signal sample; the mean activation vector of each type of communication interference signal is obtained by calculating the mean value in the activation vector AV of the correctly classified communication interference signals in Step S4.4, and the mean activation vector MAV is expressed as:

[0053] MAV = mean(AV) = mean(z)

[0054] where mean() represents the operation of calculating the mean value;

[0055] Step S4.6: Calculate the similarity between the activation vector AV of each correctly classified communication interference signal and the mean activation vector MAV of the corresponding type of communication interference signal sample. The cosine distance under the hypersphere is used to replace the Euclidean distance in the original OpenMax algorithm, and the cosine distance d cos is expressed as:

[0056]

[0057] Since the projection layer has normalized the embedded feature z and the class center vector μ c the calculation of the cosine distance d cos is simplified to:

[0058]

[0059] Step S4.7: Based on the extreme value theory (EVT), fit the Weibull distribution to the cosine distance d of each type of communication interference signal sample. The probability density distribution function p and the cumulative distribution function P under the Weibull distribution are expressed as: cos where λ and κ are the scale parameter and the shape parameter respectively, and x is the communication interference signal sample;

[0060]

[0061] where λ and κ are the scale parameter and the shape parameter respectively, and x is the communication interference signal sample;

[0062] Based on the extreme value theory (EVT), the process of fitting the Weibull distribution to the cosine distance d of each type of communication interference signal sample specifically includes: sorting the cosine distances calculated in Step S4.6 from largest to smallest, selecting the top η largest cosine distances to fit the Weibull distribution of each type of communication interference signal, and using the maximum likelihood method to estimate the scale parameter λ and the shape parameter κ. Then the maximum likelihood function L is: cos where λ and κ are the scale parameter and the shape parameter respectively, and x is the communication interference signal sample;

[0063]

[0064] Based on the classical maximum likelihood estimation, estimate the parameters and As follows:

[0065]

[0066] Step S4.8: Save the Weibull distribution of each type of communication interference signal fitted in step S4.7 and the mean activation vector MAV of each type of communication interference signal calculated in step S4.5 for subsequent testing;

[0067] Step S4.9: Input the communication interference signal for testing into the network trained in step S4.3, and calculate the activation vector AV of the input communication interference signal according to the method in step S4.4. This activation vector AV represents the position of the sample in the hypersphere embedding space;

[0068] Step S4.10: Calculate the cosine distance d using the activation vector AV obtained in step S4.9 and the mean activation vector MAV saved in step S4.8 cos ;

[0069]

[0070] Step S4.11: Calculate the correction factor of the score value of the sample x for each category, which is used to adjust the score vector SV of the sample x. Among them, the score vector SV is the output vector of the second fully connected layer in step S4.1, and the score value of each category is the element corresponding to this category in the score vector SV. The correction factor is calculated using the Weibull distribution of each type of communication interference signal saved in step S4.8:

[0071]

[0072] Among them, ω k is the correction factor of the sample x for the k-th score value in the score vector. If the sample x is close to the center of category k (ω k ≈1), the score value of category k remains unchanged; if the sample x deviates from the center of category k (ω k <<1), the score value of category k is weakened;

[0073] Step S4.12: Input the communication interference signal sample to be tested into the network trained in step S4.3, and use the output vector of the second fully connected layer of the OpenMax layer as the score vector; reallocate the score value of this sample for each category using the correction factor calculated in step S4.11, and introduce an additional score value for the unknown category:

[0074] SV k = ω k SV k , k = 1, 2,..., K

[0075]

[0076] Among them, SV k is the score value of the known class k after being corrected, and (1 - ω k ) represents the suppressed part of class k. These weakened class score values are aggregated and given to the "unknown class" score value SV K+1 ;

[0077] Step S4.13: Calculate the corrected class probability distribution: OpenMax uses the corrected score value SV k and the unknown class score value SV K+1 to perform normalization calculation to obtain the corrected probability P OpenMax :

[0078]

[0079] Let y * = argmax j P OpenMax (y = j|x), that is, the class corresponding to the maximum probability is y * , and OpenMax finally classifies according to y * and P(y = y * |x): If y * = K + 1 or P(y = y * |x) < θ, where K + 1 is the unknown class number of the communication interference signal, and θ is the threshold, it is determined as the unknown class and the classification is rejected; otherwise, it is determined as the known class y * = k', k' ∈ {1, 2,..., K}, where k' ∈ {1, 2,..., K} is the known class number of the communication interference signal; complete the open-set recognition of the communication interference signal.

[0080] A system for an open-set recognition method of a communication interference signal, characterized by comprising:

[0081] A short-time Fourier transform module, which is used to perform short-time Fourier transform on the communication interference signal in step S1, and the obtained time-frequency diagram is used as the network input;

[0082] A multi-scale feature extraction network based on Res2Net, which is used to perform multi-scale feature characterization on the time-frequency diagram obtained in step S1 by using the multi-scale feature extraction network based on Res2Net in step S2;

[0083] A projection layer based on the hyperspherical vMF distribution, which is used to perform hyperspherical embedding on the multi-scale features in step S3, model the features using the vMF distribution under the hypersphere, and design a compactness loss function and a dispersion loss function;

[0084] The classification layer based on improved OpenMax is used to input the multi-scale features modeled in step S3 into the OpenMax layer for open-set recognition in step S4. The output of the projection layer is used as the activation vector to improve OpenMax, and the cosine distance under the hypersphere is used to replace the Euclidean distance in the OpenMax algorithm to calculate the unknownness of the input samples, and the class probability distribution is readjusted so that some samples with high confidence but anomalies can be classified as unknown classes.

[0085] An apparatus for an open-set recognition method of communication interference signals, comprising: a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, it can implement the open-set recognition method of communication interference signals described in any one of the above.

[0086] A computer storage medium for receiving a user input program, wherein the computer program stored in the storage medium can perform open-set recognition of communication interference signals based on the open-set recognition method of communication interference signals described in any one of the above technical solutions.

[0087] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0088] The present invention proposes a multi-scale feature extraction network based on Res2Net, and the convolutional connections at multiple levels can better capture the multi-scale information in the signal. This design can effectively improve the processing ability of complex signals in the signal separation task. Especially when processing signals with non-linear and multi-scale features, it can extract useful feature information more flexibly and accurately.

[0089] The present invention proposes a compactness loss and a separability loss based on the hyperspherical vMF distribution, where the compactness loss promotes the proximity of samples of the same class to the class prototype, and the dispersibility loss is used to increase the angular distance between different class prototypes. The joint optimization of these two loss functions enables the known class samples to form a tight clustering, while the unknown class samples will be far away from all known classes, thereby improving the performance of open-set recognition.

[0090] The present invention improves the OpenMax open-set recognition method by replacing the Euclidean distance with the cosine distance and using the output of the projection layer as the activation vector. This improved OpenMax open-set recognition method makes full use of the characteristics of the hyperspherical embedding space, ensures the scale consistency of the features, eliminates the noise interference caused by scale changes, and thus improves the training stability and inference accuracy of the model.

[0091] In summary, compared with the prior art, the present invention explores the importance of effectively extracting multi-scale features and standardizing the embedding space in the open-set recognition of communication interference signals; the present invention can be used for recognition in any scenario involving the recognition of unknown communication interference signals, especially in a highly dynamic and complex electromagnetic environment; based on the compactness loss and separability loss of the hyperspherical vMF distribution, a more obvious distinction is established between known and unknown classes, and it can better handle unknown classes compared with traditional methods; the improved OpenMax open-set recognition method significantly enhances the training stability and inference accuracy by replacing the Euclidean distance with the cosine distance and using the output of the projection layer as the activation vector; the present invention has higher recognition accuracy and stronger versatility. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] Figure 1 is a flowchart of a method, system, medium, and device for open-set recognition of communication interference signals provided by an embodiment of the present invention.

[0093] Figure 2 is a multi-scale feature extraction network based on Res2Net in the method for open-set recognition of communication interference signals provided by an embodiment of the present invention.

[0094] Figure 3 is a schematic diagram of the simulation experiment results of the recognition accuracy of the method for open-set recognition of communication interference signals provided by an embodiment of the present invention and other open-set recognition methods under different signal-to-interference-plus-noise ratios.

[0095] Figure 4 is a schematic diagram of the simulation experiment results of the recognition accuracy of the method for open-set recognition of communication interference signals provided by an embodiment of the present invention and other open-set recognition methods under different opening degrees.

[0096] Figure 5 is a visualization diagram of the feature space of the method for open-set recognition of communication interference signals provided by an embodiment of the present invention when using different loss functions, where Figure 5 (a) is a visualization diagram of the feature space using the compactness loss and dispersibility loss proposed by the present invention, Figure 5 (b) is a visualization diagram of the feature space using the traditional cross-entropy. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0097] 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 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.

[0098] Aiming at the problems existing in the prior art, the present invention provides a method, medium, and device for open-set recognition of communication interference signals, which will be described in detail below with reference to the accompanying drawings.

[0099] For the communication interference signal open-set recognition method, system, medium and device provided by the present invention, those of ordinary skill in the art can also implement using other steps. Figure 1 The communication interference signal open-set recognition method, medium and device provided by the present invention are only a specific embodiment.

[0100] As Figure 1 shown, the communication interference signal open-set recognition method provided by the embodiment of the present invention has the following specific steps:

[0101] Step S1: Perform short-time Fourier transform on the communication interference signal to obtain a time-frequency diagram, and use the obtained time-frequency diagram as the input of the multi-scale feature extraction network. The specific process is as follows:

[0102] Performing short-time Fourier transform on the communication interference signal is expressed as:

[0103]

[0104] where x(t) is the source signal, is the window function, and STFT x (t, f) is the time-frequency diagram.

[0105] Step S2: Use the multi-scale feature extraction network based on Res2Net to perform multi-scale feature characterization on the time-frequency diagram obtained in step S1. The specific steps are as follows

[0106] Input the time-frequency diagram STFT x (t, f) obtained in step S1 into the multi-scale feature extraction network g(·) to obtain the multi-scale feature g(STFT x (t, f));

[0107] The multi-scale feature extraction network is composed of six Res2Net networks connected in series. As Figure 2 shown; the input of the first layer of the Res2Net network is the time-frequency diagram, and the output is the multi-scale feature. The input of each layer of the Res2Net network from the second layer to the sixth layer is the multi-scale feature output by the previous layer, and the output is a deeper multi-scale feature; the input of each layer of the Res2Net network, after being convolved by conv1D(k = 1), where k is the convolution kernel length, evenly divides the multi-scale feature of D channels into s multi-scale feature subsets, denoted by x i , where i ∈ {1, 2,..., s}; each multi-scale feature subset x i has the same spatial size, and the number of channels is D / s; except for x1, each multi-scale feature subset x i has a corresponding conv1D(k = 3) convolution, denoted as K i (); denoted by y i to represent Ki Output of (); multi-scale feature subset x i Add to K i-1 Add the output of () and then input K i (); omit the conv1D(k = 3) convolution of x1, then the output y i Can be expressed as:

[0108]

[0109] Each conv1D(k = 3) convolution operator may sense feature information from all feature subsets {x j , j ≤ i}.

[0110] Step S3: Perform hypersphere embedding on the multi-scale features of Step S2 by constructing a projection layer, model the features using the vMF distribution under the hypersphere, and design a compactness loss function and a dispersion loss function. The specific process is as follows:

[0111] Step S3.1: Use the projection layer h(·) to map the multi-scale features g(STFT x (t, f)) obtained in Step S2 to a low-dimensional feature space to generate a feature representation Further perform unitization processing on it to finally obtain the normalized embedded feature z:

[0112]

[0113] Among them, Represents The 2-norm of;

[0114] Make the embedded feature lie on a unit hypersphere:

[0115]

[0116] Among them, S d Is the unit hypersphere;

[0117] Step S3.2: Use the von Mises-Fisher (vMF) distribution to model the distribution of the embedded features on the hypersphere. For the unit vector in the communication interference signal category k Its probability density function is defined as p d As:

[0118]

[0119] Among them, μ k Is the category center vector with unit norm, κ ≥ 0 represents the size of the distribution compactness, and the distribution density around the mean direction μ k , Z d(κ) is a normalization factor; under the vMF distribution, the conditional probability that an embedded vector z is assigned to class k' is as follows:

[0120]

[0121] where K is the number of known classes of communication interference signals;

[0122] Step S3.3: Design two key loss functions based on the von Mises-Fisher (vMF) distribution on the hypersphere: compactness loss and dispersion loss;

[0123] The optimization objective of the compactness loss is to make each class of communication interference signal samples input during training closer to the class center of that class in the hypersphere embedding space, improve the within-class compactness of each class of communication interference signals, and make all communication interference signal samples of the same class more similar to each other; the compactness loss is expressed as:

[0124]

[0125] where z i represents the hypersphere embedding of the sample, μ k'(i) is the center vector of the class to which the sample belongs, τ is the temperature parameter used to control the smoothness of the probability distribution, and N is the number of communication interference signal samples;

[0126] The optimization objective of the dispersion loss is to make the samples of different classes of communication interference signals input during training as far apart as possible in the hypersphere embedding space and reduce the cosine similarity between the class centers of different communication interference signals; the dispersion loss is expressed as:

[0127]

[0128] where μ j and μ k represent the class center vectors of class j and class k respectively;

[0129] Step S4: Input the multi-scale features modeled in Step S3 into the OpenMax layer for open-set recognition, use the output of the projection layer as the activation vector to improve OpenMax, and replace the Euclidean distance in the OpenMax algorithm with the cosine distance under the hypersphere to calculate the unknownness of the input sample based on the extreme value theory, and readjust the class probability distribution so that some samples with high confidence but anomalies can be classified as unknown classes. The specific process is as follows:

[0130] Step S4.1: Construct an OpenMax layer using a multi-layer perceptron. The multi-layer perceptron is composed of a fully connected layer, a ReLU activation layer, a fully connected layer, and a Softmax layer stacked in sequence;

[0131] Step S4.2: Connect the multi-scale feature extraction network, projection layer, and OpenMax layer described in Step S2, Step S3.1, and Step S4.1 in sequence to construct an open-set recognition network for communication interference signals.

[0132] Step S4.3: Train the open-set recognition network for communication interference signals described in Step S4.2 using a simulated communication interference signal training set to obtain a trained network. The optimization objective during training is to add the compactness loss and dispersion loss described in Step S3.3 to the cross-entropy loss. Its overall loss function is expressed as:

[0133]

[0134] where λ d and λ c are hyperparameters used to control the weights of the dispersion loss and compactness loss respectively, is the cross-entropy loss.

[0135] Step S4.4: Input the communication interference signal training set into the trained network for classification. Use the output feature vector of the projection layer as the activation vector to improve the OpenMax algorithm. The activation vector AV is expressed as:

[0136]

[0137] where z is the hypersphere embedding vector normalized by the projection layer, representing the position of the communication interference signal sample in the hypersphere space;

[0138] Step S4.5: Calculate the mean activation vector of each class of communication interference signal samples; the mean activation vector of each class of communication interference signal is obtained by calculating the mean of the activation vectors AV of the correctly classified communication interference signals in Step S4.4. The mean activation vector MAV is expressed as:

[0139] MAV = mean(AV) = mean(z)

[0140] where mean() represents the operation of calculating the mean;

[0141] Step S4.6: Calculate the similarity between the activation vector AV of each correctly classified communication interference signal and the mean activation vector MAV of the corresponding class of this communication interference signal sample. Replace the Euclidean distance in the original OpenMax algorithm with the cosine distance under the hypersphere. Then the cosine distance dcos It is expressed as:

[0142]

[0143] Since the projection layer has embedded the feature z and the category center vector μ c Normalized, the cosine distance d cos The calculation is simplified to

[0144]

[0145] Step S4.7: Based on the extreme value theory (EVT), the cosine distance d of each type of communication interference signal sample is calculated. cos Fitting Weibull distribution, the probability density distribution function p and cumulative distribution function P under Weibull distribution are expressed as:

[0146]

[0147] Among them, λ and κ are scale parameter and shape parameter respectively, and x is the communication interference signal sample;

[0148] The process of fitting the Weibull distribution specifically includes: sorting the cosine distances calculated in step S4.6 from large to small, selecting the first n largest cosine distances for fitting the Weibull distribution of each type of communication interference signal, and estimating the scale parameter λ and shape parameter κ using the maximum likelihood method. The maximum likelihood function L is:

[0149]

[0150] Based on the classical maximum likelihood estimation, the estimated parameters and as follows:

[0151]

[0152] Step S4.8: Save the Weibull distribution of each type of communication interference signal fitted in step S4.7 and the mean activation vector MAV of each type of communication interference signal calculated in step S4.5 for subsequent testing;

[0153] Step S4.9: During the test, the communication interference signal used for the test is input into the network trained in step S4.3, and the activation vector AV of the input communication interference signal is calculated according to the method of step S4.4, and the activation vector AV represents the position of the sample in the hypersphere embedding space;

[0154] Step S4.10: Calculate the cosine distance d using the activation vector AV obtained in step S4.9 and the mean activation vector MAV saved in step S4.8cos ;

[0155]

[0156] Step S4.11: To determine whether the communication interference signal sample belongs to an unknown category, calculate the correction factor of the sample x for each category score value, which is used to adjust the score vector SV of the sample x. Among them, the score vector SV is the output vector of the second fully connected layer in Step S4.1, and each category score value is the element corresponding to that category in the score vector SV. The correction factor is calculated using the Weibull distribution of each type of communication interference signal saved in Step S4.8:

[0157]

[0158] where ω k is the correction factor of the sample x for the k-th score value in the score vector. If the sample x is close to the center of category k (ω k ≈1), the score value of category k remains unchanged; if the sample x deviates from the center of category k (ω k <<1), the score value of category k is weakened;

[0159] Step S4.12: Input the communication interference signal sample to be tested into the network trained in Step S4.3, and use the output vector of the second fully connected layer of the OpenMax layer as the score vector; re-distribute the score values of this sample for each category using the correction factor calculated in Step S4.11, and introduce additional score values for the unknown category:

[0160] SV k = ω k SV k , k = 1, 2,..., K

[0161]

[0162] where SV k is the corrected score value of the known category k, and (1 - ω k ) represents the suppressed part of category k. These weakened category score values are aggregated and assigned to the "unknown category" score value SV K+1 ;

[0163] Step S4.13: Calculate the corrected category probability distribution: OpenMax uses the corrected score value SV k and the unknown category score value SV K+1 to perform normalization calculation to obtain the corrected probability P OpenMax :

[0164]

[0165] Let \(y\) * \(=\arg\max\) j \(P\) OpenMax (\(y = j|x\)), that is, the category corresponding to the maximum probability is \(y\) * , OpenMax finally classifies according to \(y\) * and \(P(y = y\) * \(|x)\): If \(y\) * \(= K + 1\) or \(P(y = y\) * \(|x)<\theta\), where \(K + 1\) is the unknown category number of the communication interference signal, \(\theta\) is the threshold, it is determined as the unknown category and the classification is rejected; otherwise it is determined as the known category \(y\) * \(= k'\), \(k'\in\{1,2,\cdots,K\}\), where \(k'\in\{1,2,\cdots,K\}\) is the known category number of the communication interference signal.

[0166] The technical effects of the present invention will be described in detail below in combination with simulation experiments.

[0167] To evaluate the performance of the present invention, simulation verification is carried out. The specific simulation conditions and parameters are as follows

[0168] The experiment was completed using a computer with a 16-core AMD EPYC 7502 processor, an NVIDIA RTX 4090 GPU, and 80G of memory. The operating system is ubuntu20.04, the software environment is Python3.10, and the deep learning framework Pytorch2.3.0 is used for network construction and training. During the training process, the momentum stochastic gradient descent optimizer is adopted, the momentum factor is 0.9, the learning rate is 0.5, and the attenuation coefficient is 10 -4 , \(\tau\) in the vMF distribution is 0.1, \(\lambda\) in the loss function d is 0.4, \(\lambda\) c is 0.8, the final threshold \(\theta\) is 0.6, the update factor \(\alpha\) in the prototype update is 0.5, the batch training size batch_size is 128, and the training iteration steps epoch is 100. The experiment selects multi-tone jamming (MTJ), linear frequency sweep jamming (LFM), noise frequency modulation jamming (NFM), partial band noise jamming (10 -415×1000 PBNJ) and binary phase shift keying interference (BPSK) are used as known classes for training. For unknown classes, single-tone interference (STJ) and sinusoidal frequency modulation interference (SFM) are selected. During testing, both known and unknown classes participate in the test. The sampling frequency of the interference signal is 100 MHz, the number of sampling points is 1024, the range of the jamming-to-noise ratio (JNR) is -10 - 18 dB, with values taken every 2 dB, and the added noise is Gaussian white noise. During the training process, 1000 samples of each interference signal are generated at each signal-to-noise ratio as training samples, that is, a total of 15×1000 training samples for each signal. During the testing process, 500 samples of each signal are generated at each signal-to-noise ratio as training samples, that is, a total of 15×500 test samples for each signal.

[0169] To effectively evaluate the performance of the open-set recognition algorithm, three performance evaluation metrics are introduced in the experiment:

[0170] The recognition accuracy of known samples AKS (Accuracy on Known Samples) is used to measure the recognition accuracy of known classes, and the recognition accuracy of unknown samples AUS (Accuracy on Unknown Samples) is used to measure the recognition accuracy of unknown classes. The mathematical expressions are as follows:

[0171]

[0172] where TK and FK represent the number of correctly classified and misclassified samples of known-class samples, and TU and FU represent the number of correctly classified and misclassified samples of unknown-class samples.

[0173] The normalized recognition accuracy (NA, Normalized Accuracy) is used to measure the overall performance of open-set recognition. The mathematical expression is as follows:

[0174] NA = λAKS + (1 - λ)AUS

[0175] where λ is used to measure the importance of the recognition performance of known and unknown classes in open-set recognition. Since in the actual open-set recognition of communication interference, there are more known classes, fewer unknown classes, and the number of known classes is increasing with the continuous research on unknown classes, so in subsequent simulations, It corresponds to the ratio of known classes to unknown classes in the experimental settings.

[0176] Figure 3Specifically shows the comparison of the recognition performance of the algorithm in this paper with other competing open-set recognition algorithms under different signal-to-noise ratios. Three comparison algorithms are selected in this paper: Softmax algorithm, openGan algorithm, and CPN algorithm. To fairly compare the advantages and disadvantages of the algorithms, the Softmax algorithm and CPN algorithm use the same feature extractor ResNet18 as the algorithm in this paper, while openGan uses the GAN network in its literature. Since the conventional Softmax does not have the ability of open-set recognition, in the comparative experiment, the signal prediction probability under 95% confidence is added as the threshold value to provide the open-set recognition ability. From Figure 3 it can be seen that only when the signal-to-noise ratio JNR <= -6 is extremely low, the recognition rates of the openGan algorithm and CPN algorithm are slightly higher than that of the algorithm in this paper. When JNR > -6, the correct rate of open-set recognition of the algorithm in this paper is better than the three comparison algorithms, and the correct rate of open-set recognition is 1.5% - 6.5% higher than that of the second-highest CPN algorithm. This shows that the algorithm in this paper combines the high-dimensional features extracted by ResNet and the vMF distribution modeling, and effectively optimizes the recognition accuracy of unknown classes by enhancing the inter-class discrimination and intra-class compactness and the introduction of the OpenMax layer.

[0177] Figure 4 Specifically shows the comparison of the recognition performance of the algorithm in this paper with other competing open-set recognition algorithms under different opening degrees. The opening degree is measured by the ratio of known classes to unknown classes, that is, K vU, where K is the number of known classes and U is the number of unknown classes. The performance index uses the average normalized recognition rate, that is, the mean value of the normalized recognition rates in the middle section of the signal-to-noise ratio used in the experiment, JNR = -2 - 10 dB, as the evaluation index. From Figure 4 it can be seen that the recognition rate of the algorithm in this paper is always higher than that of other algorithms under different opening degrees, and when there are only two known classes, the normalized recognition rate can also reach 86%. This shows that the algorithm in this paper can also achieve a high recognition correct rate in a complex environment, that is, when there are a large number of unknown-class communication interference signals in the environment, which is better than the other three algorithms.

[0178] Figure 5 Specifically shows the visual comparison of the embedding space between training optimization using the compactness loss and dispersion loss proposed by the algorithm in this paper and training optimization using the traditional cross-entropy. The T-SNE dimensionality reduction method is used to reduce the 128-dimensional features in the embedding space to obtain a two-dimensional feature visualization graph. From Figure 5It can be seen that, by adopting the vMF distribution based on hypersphere modeling and the method of compactness loss and dispersion loss, the sample feature distributions of the same class are more concentrated, and the distances between the sample feature distributions of different classes are farther, leaving more space for the open area and making the open set recognition more accurate. While by adopting the method of conventional cross-entropy loss, the sample feature distributions of the same class are more dispersed and the sample feature distributions between different classes are close, which makes the open area smaller and the samples falling between different classes cannot be correctly recognized. For example, when the sample is in the intersection area of partial band interference and noise frequency modulation interference in the figure, it cannot be correctly recognized. Therefore, compared with using the traditional cross-entropy loss for training optimization, adopting the compactness loss and dispersion loss proposed in this paper can obtain more discriminative embedding features.

[0179] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated design hardware. Those of ordinary skill in the art can understand that the above devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code is provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuits of programmable hardware devices such as very large scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable logic devices such as field programmable gate arrays, or can be implemented by software executed by various types of processors, or can be implemented by a combination of the above hardware circuits and software, such as firmware.

[0180] The above is only the specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present invention by those skilled in the art within the technical scope disclosed by the present invention shall be covered by the protection scope of the present invention.

Claims

1. A communication interference signal open set identification method, characterized in that: The following steps are involved: Step S1: Perform short-time Fourier transform on the communication interference signal to obtain a time-frequency graph, and use the obtained time-frequency graph as an input to a multi-scale feature extraction network; Step S2: Use a multi-scale feature extraction network based on Res2Net to perform multi-scale feature representation on the time-frequency graph obtained in step S1; Step S3: embed the multi-scale features of step S2 into a hypersphere by constructing a projection layer, model the features using the vMF distribution under the hypersphere, and design a compactness loss function and a dispersion loss function; Step S4: Input the multi-scale features modeled in step S3 into the OpenMax layer for open set recognition, use the output of the projection layer as the activation vector to improve OpenMax, and use the cosine distance under the hypersphere to replace the Euclidean distance in the OpenMax algorithm. Based on the extreme value theory, the unknown of the input samples is calculated, and the category probability distribution is readjusted so that some high-confidence but abnormal samples can be classified as unknown categories.

2. The communication interference signal open set identification method according to claim 1, characterized in that: The specific method of step S1 is: The short-time Fourier transform of the communication interference signal is expressed as: Where x(t) is the source signal, is the window function, STFT x (t,f) is the time-frequency diagram.

3. The communication interference signal open set identification method according to claim 1, characterized in that: The specific method of step S2 is: The time-frequency diagram STFT obtained in step S1 x (t,f) is input into the multi-scale feature extraction network g(·) to obtain the multi-scale feature g(STFT x (t,f)); The multi-scale feature extraction network is composed of six Res2Net networks connected by residual connection; the input of the first layer of Res2Net network is the time-frequency graph, and the output is the multi-scale feature; the input of each layer of Res2Net network from the second to the sixth layer is the multi-scale feature output of the previous layer, and the output is the multi-scale feature of the deeper layer; The input of each layer of Res2Net network is convolved by conv1D(k=1), where k is the length of the convolution kernel, and the multi-scale features of D channels are evenly divided into s multi-scale feature subsets, using x i Represents, where i∈{1,2,...,s}; each multi-scale feature subset x i have the same spatial size and the number of channels is D / s; except x1, each multi-scale feature subset x i There are corresponding conv1D (k = 3) convolutions, denoted by K i (); use y i K i Output of (); multi-scale feature subset x i With K i-1 () outputs are added and then input K i (); omitting the conv1D(k=3) convolution of x1, the output is y i It can be expressed as:

4. The communication interference signal open set identification method according to claim 1, characterized in that: The specific method of step S3 is: Step S3.1: Use the projection layer h(·) to transform the multi-scale features g(STFT x (t,f)) is mapped to a low-dimensional feature space to generate feature representation Further unitization is performed to finally obtain the normalized embedding feature z: in, express The 2-norm of ; Make the embedded features lie on a unit hypersphere: Among them, S d is the unit hypersphere; Step S3.2: Model the embedded feature distribution on the hypersphere using the von Mises-Fisher (vMF) distribution. For the unit vector in the communication interference signal category k, Its probability density function is defined as p d for: Among them, μ k is the category center vector with unit norm, κ≥0 indicates the size of the distribution density, around the mean direction μ k The distribution density, Z d (κ) is the normalization factor; the conditional probability that an embedding vector z is assigned to category k' under the von Mises-Fisher (vMF) distribution for: in, K is the number of known categories of communication interference signals; Step S3.3: Design two key loss functions based on the von Mises-Fisher (vMF) distribution of the hypersphere: compactness loss and dispersion loss; The optimization goal of the compactness loss is to make each type of communication interference signal sample input during training more closely surround the category center of the category in the hypersphere embedding space, improve the internal compactness of each type of communication interference signal category, and make all communication interference signal samples of the same category more similar to each other; compactness loss It is expressed as: Among them, z i represents the hyperspherical embedding of the sample, μ k'(i) is the center vector of the category to which the sample belongs, τ is the temperature parameter used to control the smoothness of the probability distribution, and N is the number of communication interference signal samples; The optimization goal of the dispersion loss is to make the different types of communication interference signal samples input during training distributed as far apart as possible in the hypersphere embedding space, and reduce the cosine similarity between the centers of different communication interference signal categories; dispersion loss It is expressed as: Among them, μ j and μ k Represent the category center vectors of category j and category k respectively.

5. The communication interference signal open set identification method according to claim 1, characterized in that: The specific method of step S4 is: Step S4.1: Use a multi-layer perceptron to construct an OpenMax layer. The multi-layer perceptron is composed of a fully connected layer, a ReLU activation layer, a fully connected layer, and a Softmax layer stacked in sequence; Step S4.2: Use the multi-scale feature extraction network, projection layer, and OpenMax layer described in step S2, step S3.1, and step S4.1 to sequentially connect and construct a communication interference signal open set recognition network; Step S4.3: Use the simulated communication interference signal to train the communication interference signal open set recognition network described in step S4.2 to obtain a trained network. The optimization goal during training is to add the compactness loss and dispersion loss described in step S3.3 to the cross entropy loss. The overall loss function is It is expressed as: Among them, λ d and λ c are hyperparameters, which are used to control the weights of dispersion loss and compactness loss, respectively. is the cross entropy loss; Step S4.4: Input the communication interference signal into the trained network for classification, use the output feature vector of the projection layer as the activation vector to improve the OpenMax algorithm, and the activation vector AV of the correctly classified communication interference signal is expressed as: Among them, z is the hypersphere embedding vector normalized by the projection layer, which represents the position of the communication interference signal sample in the hypersphere space; Step S4.5: Calculate the mean activation vector of each type of communication interference signal sample; the mean activation vector of each type of communication interference signal is obtained by performing mean calculation on the activation vector AV of the correctly classified communication interference signal in step S4.4, and the mean activation vector MAV is expressed as: MAV=mean(AV)=mean(z) Among them, mean() represents the mean operation; Step S4.6: Calculate the similarity between the activation vector AV of each correctly classified communication interference signal and the mean activation vector MAV of the corresponding category of its communication interference signal sample, and use the cosine distance under the hypersphere to replace the Euclidean distance in the original OpenMax algorithm. Then the cosine distance d cos It is expressed as: Since the projection layer has embedded the feature z and the category center vector μ c Normalized, the cosine distance d cos The calculation is simplified to: Step S4.7: Based on the extreme value theory (EVT), the cosine distance d of each type of communication interference signal sample is calculated. cos Fitting Weibull distribution, the probability density distribution function p and cumulative distribution function P under Weibull distribution are expressed as: Among them, λ and κ are scale parameter and shape parameter respectively, and x is the communication interference signal sample; Based on the extreme value theory (EVT), the cosine distance d of each type of communication interference signal sample is calculated. cos The process of fitting the Weibull distribution specifically includes: sorting the cosine distances calculated in step S4.6 from large to small, selecting the first n largest cosine distances for fitting the Weibull distribution of each type of communication interference signal, and estimating the scale parameter λ and shape parameter κ using the maximum likelihood method. The maximum likelihood function L is: Based on the classical maximum likelihood estimation, the estimated parameters and as follows: Step S4.8: Save the Weibull distribution of each type of communication interference signal fitted in step S4.7 and the mean activation vector MAV of each type of communication interference signal calculated in step S4.5 for subsequent testing; Step S4.9: Input the communication interference signal for testing into the network trained in step S4.3, and calculate the activation vector AV of the input communication interference signal according to the method of step S4.4, where the activation vector AV represents the position of the sample in the hypersphere embedding space; Step S4.10: Calculate the cosine distance d using the activation vector AV obtained in step S4.9 and the mean activation vector MAV saved in step S4.8 cos ; Step S4.11: Calculate the correction factor of sample x for each category score value, which is used to adjust the score vector SV of sample x, where the score vector SV is the output vector of the second fully connected layer in step S4.1, and each category score value is the element corresponding to the category in the score vector SV. The correction factor is calculated using the Weibull distribution of each type of communication interference signal saved in step S4.8: Among them, ω k is the correction factor of sample x for the kth score value in the score vector. If sample x is close to the center of category k (ω k ≈1), the score value of category k remains unchanged; if the sample x deviates from the center of category k (ω k <<1), then the score value of category k is weakened; Step S4.12: Input the communication interference signal sample to be tested into the network trained in step S4.3, and use the output vector of the second fully connected layer of the OpenMax layer as the score vector; use the correction factor calculated in step S4.11 to redistribute the score values ​​of the sample to each category, and introduce additional score values ​​for unknown categories: SV k =ω k SV k ,k=1,2,...,K Among them, SV k is the corrected score of the known class k, (1-ω k ) represents the suppressed part of category k, and these weakened category scores are summed up to give the "unknown category" score SV K+1 ; Step S4.13: Calculate the corrected category probability distribution: OpenMax uses the corrected score value SV k and unknown category score SV K+1 Perform normalization calculation to obtain the corrected probability P OpenMax : Let y * = arg max j P OpenMax (y=j|x), that is, the category with the maximum probability is y * , OpenMax finally based on y * and P(y=y * |x) to classify: if y * =K+1 or P(y=y * |x)<θ, where K+1 is the unknown category number of the communication interference signal, θ is the threshold, if it is determined to be an unknown class, the classification is rejected; otherwise, it is determined to be a known class y * =k', k'∈{1,2,...,K}, where k'∈{1,2,...,K} is the known category number of the communication interference signal; the open set identification of the communication interference signal is completed.

6. A system for the open set identification method of communication interference signals according to any one of claims 1 to 5, characterized in that: include: A short-time Fourier transform module, used to perform a short-time Fourier transform on the communication interference signal in step S1, and the obtained time-frequency diagram is used as a network input; A multi-scale feature extraction network based on Res2Net is used to perform multi-scale feature representation on the time-frequency graph obtained in step S1 using the multi-scale feature extraction network based on Res2Net in step S2; A projection layer based on a hypersphere vMF distribution is used to perform hypersphere embedding on the multi-scale features of step S2 in step S3, model the features using the vMF distribution under the hypersphere, and design a compactness loss function and a dispersion loss function; The classification layer based on the improved OpenMax is used to input the multi-scale features modeled in step S3 into the OpenMax layer for open set recognition in step S4, use the output of the projection layer as the activation vector to improve OpenMax, and use the cosine distance under the hypersphere to replace the Euclidean distance in the OpenMax algorithm. The unknown of the input samples is calculated based on the extreme value theory, and the category probability distribution is readjusted so that some high-confidence but abnormal samples can be classified as unknown categories.

7. The device of the communication interference signal open set identification method according to any one of claims 1 to 5, characterized in that: include: A memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, it can implement the communication interference signal open set identification method described in any one of claims 1 to 5.

8. A computer storage medium for receiving a user input program, characterized in that: When the computer program stored in the storage medium is executed by the processor, it can perform open-set identification of communication interference signals based on the open-set identification method of communication interference signals according to any one of claims 1 to 5.

Images