A communication signal open set identification method based on three-channel time-frequency fusion

Abstract

The present application relates to the technical field of communication signal open set identification, and particularly relates to a communication signal open set identification method based on three-channel time-frequency fusion; the method comprises the following steps: adopting an end-to-end data preprocessing procedure to pre-process signals, automatically mapping original I / Q data into a time-frequency diagram, an I channel and a Q channel three-way input; adopting a three-channel parallel feature extraction structure to synchronously learn multi-scale features of time domain and frequency domain, and to splice and fuse the features for classification and detection; through open set identification, combining OpenMax and Weibull distribution tail fitting, introducing a probability calibration strategy, dynamically detecting and efficiently rejecting unknown modulation types; verifying the results, comparing with various mainstream models, verifying the comprehensive advantages in classification accuracy and unknown detection rate; through the above manner, unknown signal detection capability is improved, unknown samples can be dynamically detected and rejected, and the safety and robustness of the system are enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of open set recognition of communication signals, and in particular to a communication signal open set recognition method based on three-channel time-frequency fusion. BACKGROUND

[0002] With the rapid evolution of wireless communication technology, various analog and digital modulation methods are intertwined and superimposed in the same frequency band, and communication channels are often simultaneously affected by multiple interferences such as multipath fading, carrier frequency offset, and non-Gaussian noise. In such a complex reception environment, if the receiving end cannot rely on any prior information, it is necessary to realize real-time, stable and high-precision recognition of signal modulation methods to ensure the reliability and security of applications such as cognitive radio, spectrum monitoring, electronic countermeasures and edge intelligence. Since the end of the 20th century, automatic modulation recognition (AMR) technology has become a research hotspot. Traditional AMR methods are often based on closed set classification (CSC) assumption, which can judge a limited number of categories in the training set.

[0003] The AMR model based on closed set assumption often misclassifies with high confidence when facing modulation methods not seen in the training set, lacks a mechanism for dynamically detecting and rejecting unknown samples, seriously affects the safety and robustness of the system, and leads to insufficient unknown signal detection capability in open set scenarios. SUMMARY

[0004] The purpose of the present application is to provide a communication signal open set recognition method based on three-channel time-frequency fusion, which aims to solve the technical problem that the existing AMR model based on closed set assumption often misclassifies with high confidence when facing modulation methods not seen in the training set, lacks a mechanism for dynamically detecting and rejecting unknown samples, seriously affects the safety and robustness of the system, and leads to insufficient unknown signal detection capability in open set scenarios.

[0005] To achieve the above-mentioned purpose, the present application adopts a communication signal open set recognition method based on three-channel time-frequency fusion, which comprises the following steps:

[0006] An end-to-end data preprocessing process is used for signal preprocessing, which automatically maps the original I / Q data to a time-frequency graph, I channel and Q channel three-way input;

[0007] A three-channel parallel feature extraction structure is used to simultaneously learn multi-scale features in time and frequency domains, and the features are fused by splicing for classification and detection;

[0008] Through open set recognition, OpenMax and Weibull distribution tail fitting are combined, a probability calibration strategy is introduced, and dynamic detection and efficient rejection of unknown modulation types are realized.

[0009] The results are verified, and compared with various mainstream models, the comprehensive advantages in classification accuracy and unknown detection rate are verified.

[0010] Further, in the step of using an end-to-end data preprocessing process to preprocess the signal, the original I / Q data is automatically mapped to a time-frequency graph, an I channel and a Q channel three-way input:

[0011] The received radio frequency signal is sampled and the quadrature down-conversion processing is completed in the digital domain;

[0012] The and signals are filtered through a low-pass filter to obtain band-limited baseband signals I(n) and Q(n);

[0013] The co-directional component and the quadrature component are filtered and decimated in the digital domain to form a complex sequence X(n);

[0014] The time-frequency graph is obtained by performing short-time Fourier transform analysis on X(n).

[0015] Among them, in the step of using a three-channel parallel feature extraction structure to simultaneously learn the multi-scale features of the time domain and the frequency domain, and through splicing and fusion for classification and detection:

[0016] The I, Q waveform and time-frequency Figure Three channel are input in parallel, so that the network can simultaneously capture multi-source information in the time domain and the frequency domain in each convolution, and after the preprocessing process framework, the time-frequency graph matrix X(x, f), the I sequence and the Q sequence are obtained respectively, and the I / Q sequence is copied as a matrix along the frequency dimension to extend:

[0017]

[0018] The final input tensor is constructed as:

[0019]

[0020] The above splicing ensures that the convolution kernel simultaneously accesses the frequency drift in the propagation process, the modulation spectrum line change and the energy distribution dynamics on the time-frequency graph channel, and the instantaneous phase jitter, amplitude fluctuation and phase difference of the signal in the time domain on the I / Q channel in the same receptive field. The convolution output of the three-channel fusion:

[0021]

[0022] Among them , is the coordinate of the output feature map in the spatial dimension, is the number of output channels, is the size of the convolution kernel, is an input channel index, 、 is an offset of the convolution kernel in the height and width direction, is a convolution kernel weight corresponding to the th output channel, is a bias term of the th output channel, and t is the number of points in the time domain.

[0023] Further, in the step of dynamic detection and efficient rejection of unknown modulation types, OpenMax is combined with Weibull distribution tail fitting by open set identification, and a probability calibration strategy is introduced:

[0024] The preprocessed input samples are sent one by one into the trained deep neural network model, and forward propagation calculation is performed along each layer of the network;

[0025] Batch normalization is performed on the feature matrix to eliminate the difference in feature scale between different samples;

[0026] According to the true label of each sample, the feature matrix is divided by category, and the arithmetic average activation vector of each category, i.e., the mean vector, is calculated;

[0027] L2 normalization is applied to each mean vector to normalize its amplitude to 1, obtaining the unit norm class center vector of each category;

[0028] The center vector is used as the basis for subsequent extreme value modeling and Weibull fitting;

[0029] In the training stage, the intermediate features of each known category are aggregated to calculate the category center, and the cosine distance of each sample to its own center is obtained;

[0030] In the test stage, the distance between the test sample and each category center is calculated, and the decay coefficient is obtained according to the corresponding Weibull model, and the original Softmax probability is decayed according to the coefficient, and the accumulated part of the decay is taken as the probability of the unknown category. The probability is redistributed to generate an Openmax output containing the probability of the unknown class;

[0031] By setting a threshold , the highest Openmax probability is judged. If it is higher than the threshold, the sample is classified as an unknown category, otherwise it is classified according to the highest probability category. In the open set scenario, the model performance is evaluated comprehensively by accuracy, recall rate and F1 index.

[0032] Further, in the step of using the center vector as the basis for subsequent extreme value modeling and Weibull fitting:

[0033] The calculation is as follows:

[0034]

[0035] wherein, is the feature vector, is the module of the feature vector, is the regularized feature vector.

[0036] Further, in the training phase, the intermediate features of each known class are aggregated to calculate the class center, and the step of obtaining the cosine distance of each sample to the center to which it belongs is:

[0037] The calculation is as follows:

[0038]

[0039] wherein, is the feature vector of the ith sample, is the center vector of the kth class, is the cosine distance calculated and dynamically determines the proportion of samples used to fit the tail part according to the variance of the distance distribution, while calculating the tail sample size Tailsize,

[0040]

[0041] wherein, is the variance, is a certain sample, is the mean of the class sample, and n is the total number of samples,

[0042]

[0043] When the variance is greater than 0.5, an additional part of the tail proportion is increased to capture more extreme values when the distribution is more dispersed,

[0044]

[0045] wherein, indicates that x is rounded down, and the tail proportion is guaranteed to be no less than the minimum threshold ,

[0046] The first T largest distances are taken to fit the Weibull distribution parameters k and The maximum likelihood estimation is adopted, and the probability density function of the Weibull distribution is as follows:

[0047]

[0048] Further, k is the shape parameter, wherein, x is the cosine distance of each sample to the center vector.

[0049] wherein, in the test phase, the distance between the test sample and the center of each category is calculated, the decay coefficient is obtained according to the corresponding Weibull model , and the original Softmax probability is decayed according to the corresponding coefficient, and the part decayed is summed up as the probability of the unknown category.

[0050] The unknown category probability score is calculated as follows:

[0051]

[0052] wherein, is the decay coefficient, is the probability score of each category in the closed set recognition of each sample, and K is the number of known category signals,

[0053] The determination method is as follows:

[0054] .

[0055] Further, in the step of verifying the results and comparing with various mainstream models, the comprehensive advantages in classification accuracy and unknown detection rate are verified:

[0056] Compared with the traditional mainstream CNN model and the residual network ResNet model with I / Q input;

[0057] Comparing with the LSTM model input through the time-frequency graph after the I / Q input through the CNN;

[0058] The closed set recognition accuracy, known category accuracy, unknown category accuracy and overall accuracy are used to compare the actual verification results.

[0059] wherein, in the step of comparing the actual verification results using the closed set recognition accuracy, known category accuracy, unknown category accuracy and overall accuracy:

[0060] The closed set recognition accuracy refers to the classification accuracy of all known categories on the test set after training with known category signals, the known category recognition rate and the unknown category recognition rate are evaluated by adding unknown categories after training with only known categories, which respectively measure the recognition performance of the model for known categories and unknown categories, and the overall accuracy is the average of the known category recognition rate and the unknown category recognition rate.

[0061] The application discloses a communication signal open set identification method based on three-channel time-frequency fusion. BRIEF DESCRIPTION OF DRAWINGS

[0062] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0063] Figure 1 is a step flow chart of the communication signal open set identification method based on three-channel time-frequency fusion of the present application.

[0064] Figure 2 is a step flow chart of S100 of the present application.

[0065] Figure 3 is a step flow chart of S300 of the present application.

[0066] Figure 4 is a step flow chart of S400 of the present application.

[0067] Figure 5 is a data preprocessing flow chart of the present application.

[0068] Figure 6 is a three-channel input schematic diagram of the present application.

[0069] Figure 7 is an open set identification flow chart of the present application.

[0070] Figure 8 is a confusion matrix diagram of the known category of the present application.

[0071] Figure 9is an open set identification confusion matrix diagram of the present application. DETAILED DESCRIPTION

[0072] Exemplary embodiments will be described in detail with reference to the drawings, of which examples are shown. In the following description, the same numbers are used to denote the same elements throughout the several views. The embodiments described in the following exemplary embodiments are not meant to represent all embodiments in accordance with the present application.

[0073] The terminology used in the present application is for the purpose of describing particular embodiments only and is not intended to be limiting of the present application. As used in the description of the application and the appended claims, the singular forms "a", "an" and "the" are intended to include plural forms as well, unless the context clearly indicates otherwise. It also will be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0074] It is to be understood that the singular forms "a", "an", and "the" include plural referents unless the context clearly dictates otherwise. It is to be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0075] Referring to Figures 1-9 The present application provides a three-channel time-frequency fusion-based communication signal open set identification method, comprising the following steps:

[0076] S100: The signal is preprocessed by using an end-to-end data preprocessing process, and the original I / Q data is automatically mapped into a time-frequency diagram, an I channel and a Q channel three-way input.

[0077] In the present embodiment, the signal is preprocessed by using an end-to-end data preprocessing process, and the original I / Q data is automatically mapped into a time-frequency diagram, an I channel and a Q channel three-way input. The specific process is as follows:

[0078] S101: The received radio frequency signal is sampled and the quadrature down-conversion processing is completed in the digital domain;

[0079] S102: The and signals are passed through a low-pass filter to obtain band-limited baseband signals I(n) and Q(n);

[0080] S103: The co-directional component and the quadrature component are filtered and decimated in the digital domain, and are combined into a complex sequence X(n);

[0081] S104: Short-time Fourier transform analysis is performed on X(n) to obtain a time-frequency graph.

[0082] In the above process, first, the signal is emitted by the transmitting end using orthogonal up-conversion modulation, transmitted through the channel, and received. The received radio frequency signal is sampled and orthogonal down-conversion processing is completed in the digital domain. Subsequently, the I and Q signals are respectively filtered through low-pass filters to obtain band-limited baseband signals and After filtering and decimation of the co-directional component and the quadrature component in the digital domain, they are combined into a complex sequence . Short-time Fourier transform analysis is performed on to obtain a time-frequency graph, as shown in Figure 5 .

[0083] S200: A three-channel parallel feature extraction structure is adopted to synchronously learn multi-scale features in the time domain and the frequency domain, and through splicing fusion, it is used for classification and detection.

[0084] In this embodiment, a three-channel parallel input schematic diagram is shown in Figure 6 . A three-channel parallel feature extraction structure is adopted to synchronously learn multi-scale features in the time domain and the frequency domain, and through splicing fusion, it is used for classification and detection. In the above process, I and Q waveforms and time-frequency Figure Three channels are input in parallel, so that the network can simultaneously capture multi-source information in the time domain and the frequency domain in each convolution. After the pre-processing flow framework, the time-frequency graph matrix X(x, f), the I sequence, and the Q sequence are obtained respectively. The I / Q sequence is copied as a matrix along the frequency dimension for extension:

[0085]

[0086] The final input tensor is constructed as:

[0087]

[0088] The above splicing ensures that the convolution kernel simultaneously accesses the frequency drift, modulation spectrum line change, and energy distribution dynamics in the propagation process on the time-frequency graph channel and the instantaneous phase jitter, amplitude fluctuation, and phase difference of the signal in the time domain on the I / Q channel in the same receptive field. The convolution output of the three-channel fusion is:

[0089]

[0090] wherein , are the coordinates of the output feature map in the spatial dimension, is the number of output channels, is the size of the convolution kernel,​ For input channel index, , This refers to the offset of the convolution kernel in the height and width directions. For the first The convolutional kernel weights corresponding to each output channel For the first The bias term for each output channel, where t is the number of points in the time domain.

[0091] S300: By using open set recognition, combining OpenMax with Weibull distribution tail fitting, and introducing a probabilistic calibration strategy, it enables dynamic detection and efficient rejection of unknown modulation types.

[0092] In this embodiment, by using open set identification, OpenMax is combined with Weibull distribution tail fitting, and a probability calibration strategy is introduced to dynamically detect and efficiently reject unknown modulation types. The specific process is as follows:

[0093] S301: Feed the preprocessed input samples one by one into the trained deep neural network model and perform forward propagation calculations along each layer of the network;

[0094] S302: Perform batch standardization on the feature matrix to eliminate differences in feature scale between different samples;

[0095] S303: Based on the true label of each sample, divide the feature matrix into categories and calculate the arithmetic mean activation vector of each category, i.e., the mean vector.

[0096] S304: Apply L2 normalization to each mean vector to normalize its magnitude to 1, and obtain the unit norm class center vectors of each category;

[0097] S305: Use the center vector as the basis for subsequent extreme value modeling and Weibull fitting;

[0098] S306: During the training phase, the intermediate features of each known category are aggregated to calculate the category center and the cosine distance from each sample to its respective center is obtained.

[0099] S307: During the testing phase, calculate the distance between the test sample and the center of each category, and derive the attenuation coefficient based on the corresponding Weibull model. The original Softmax probability is decayed according to the corresponding coefficient, and the decayed part is summed up as the probability of the unknown class. The probability is redistributed to generate an Openmax output containing the probability of the unknown class.

[0100] S308: By setting a threshold The highest Openmax probability is determined, and if it is higher than the threshold, the sample is classified as unknown, otherwise, it is classified according to the highest probability category, and the model performance is evaluated in the open set scenario by accuracy, recall rate and F1 index.

[0101] In the above process, the open set recognition flowchart is as shown in Figure 7 First, the preprocessed input sample is sent into the trained deep neural network model one by one, and forward propagation calculation is performed along each layer of the network. Specifically, we usually select the second-to-last layer as the output end, extract the activation values of all samples from this layer, and splice them into the original feature matrix. Then, batch normalization is performed on the feature matrix to eliminate the differences in feature scales between different samples. Then, according to the true label of each sample, the feature matrix is divided by category, and the arithmetic average activation vector of each category, i.e. the mean vector, is calculated. Finally, L2 normalization is applied to each mean vector to normalize its amplitude to 1, thereby obtaining the unit norm "class center vector" of each category. These center vectors will serve as the basis for subsequent extreme value modeling and Weibull fitting, and the calculation method is as follows:

[0102]

[0103] wherein, is the feature vector, is the norm of the feature vector, is the feature vector after regularization.

[0104] In the training stage, the intermediate features of each known category are aggregated to calculate the category center and obtain the cosine distance of each sample to its center, and the calculation method is as follows:

[0105]

[0106] wherein, is the feature vector of the i-th sample, is the center vector of the k-th category, is the cosine distance calculated and according to the distance distribution, the proportion of samples used to fit the tail is dynamically determined, and the number of tail samples Tailsize is calculated,

[0107]

[0108] wherein, is the variance, is a certain sample, is the mean of the samples of this category, and n is the total number of samples,

[0109]

[0110] When the variance greater than 0.5 means that the tail ratio is additionally increased, so as to capture more extreme values when the distribution is more dispersed,

[0111] wherein, represents x down rounding, ensuring that the tail ratio is not lower than the minimum threshold ,

[0112] Take the first T maximum distance to fit Weibull distribution parameters k and , using maximum likelihood estimation, the probability density function of Weibull distribution is as follows:

[0113]

[0114] wherein, k is the shape parameter, is the scale parameter, and x is the cosine distance of each sample to the center vector,

[0115] In the test phase, the distance between the test sample and the center of each category is calculated, and the decay coefficient is obtained according to the corresponding Weibull model , and the original Softmax probability is decayed according to the coefficient, and the part decayed is summed up as the probability of unknown category, so as to realize the redistribution of probability and generate Openmax output containing the probability of unknown class. The unknown category probability score is calculated as follows:

[0116]

[0117] wherein, is the decay coefficient, is the probability score of each sample in the closed set recognition, and each category, K is the number of known categories,

[0118] Finally, the determination method is as follows:

[0119]

[0120] By setting a threshold , the highest Openmax probability is judged. If it is higher than the threshold, the sample is classified as an unknown category, otherwise, it is classified according to the highest probability category, and the model performance is evaluated in the open set scene with accuracy, recall rate and F1 index.

[0121] S400: result verification, compared with a variety of mainstream models, verify the comprehensive advantage in classification accuracy and unknown detection rate.

[0122] In this embodiment, the result verification is performed, compared with a variety of mainstream models, verify the comprehensive advantage in classification accuracy and unknown detection rate, the specific process is:

[0123] S401: Compared with the traditional mainstream model CNN model and the residual network ResNet model with I / Q input;

[0124] S402: With I / Q input, the LSTM model is compared after passing through the CNN and the time-frequency graph input;

[0125] S403: The actual verification result comparison is carried out using the closed set recognition accuracy, known class accuracy, unknown class accuracy and overall accuracy.

[0126] In the above process, the recognition method is compared with the traditional mainstream model (I / Q input (i.e. double channel) CNN model, I / Q input residual network ResNet model and I / Q input first, then through the CNN and then through the time-frequency graph input LSTM model, the actual verification result comparison is carried out using the closed set recognition accuracy, known class accuracy, unknown class accuracy and overall accuracy, wherein the closed set recognition accuracy refers to using known class signals for training, and evaluating the classification accuracy of the model on all known classes on the test set; the known class recognition rate and the unknown class recognition rate are evaluated by adding "unknown class" after training only with known classes, respectively measuring the recognition performance of the model on known classes and unknown classes; the overall accuracy is the average of the known class recognition rate and the unknown class recognition rate. In the above manner, the accuracy is obviously improved.

[0127] Further, see the following table, which shows the comparison results of the three-channel model of the present text and the existing double-channel method (double-channel refers to using I / Q data as input), the closed set recognition accuracy refers to using known class signals for training, and evaluating the classification accuracy of the model on all known classes on the test set; the known class recognition rate and the unknown class recognition rate are evaluated by adding "unknown class" after training only with known classes, respectively measuring the recognition performance of the model on known classes and unknown classes; the overall accuracy is the average of the known class recognition rate and the unknown class recognition rate.

[0128] Under the double-channel setting, the closed set recognition rate of the double-channel CNN is 80.97%, the known class recognition rate is 51.58%, the unknown class recognition rate is 34.66%, and the overall accuracy is 49.47%; the double-channel ResNet reaches 84.03%, 72.32%, 97.73% and 75.50% respectively; the CNN+LSTM can reach 90.38% on the closed set task, but the unknown class recognition rate is only 80.40%, and the overall accuracy is 75.25%.

[0129] After introducing the third channel of time-frequency graph, the three-channel CNN has a significant improvement in all indicators, with the closed set recognition rate increasing to 91.38%, the known class recognition rate increasing to 83.44%, the unknown class recognition rate jumping to 84.38%, and the overall accuracy reaching 83.56%, which is better than the above three double-channel methods; the three-channel ResNet also surpasses its double-channel version with 90.30%, 80.84%, 74.72% and 80.08%, indicating that the introduction of the time-frequency graph channel not only enhances the ability to distinguish known classes, but also significantly improves the open set recognition performance of unknown classes. In summary, after adding the time-frequency graph channel, the time and frequency domain features can be extracted at the same time, thereby significantly improving the stability and accuracy of the model in closed set and open set tasks.

[0130] Model Closed set identification accuracy Known class accuracy Unknown class accuracy Overall accuracy Dual-channel CNN 80.97% 51.58% 34.66% 49.47% Dual-channel ResNet 84.03% 72.32% 97.73% 75.50% Tri-channel CNN 91.38% 83.44% 84.38% 83.56% Tri-channel ResNet 90.30% 80.84% 74.72% 80.08% CNN+LSTM 90.38% 74.51% 80.40% 75.25%

[0131] Further, see Figure 8 and Figure 9 , such as Figure 8 is the confusion matrix of known classes, and the average diagonal accuracy reaches 99.2%, indicating that the model can almost completely correctly identify each type of modulation signal. Among them, the maximum confusion occurs when BPSK is misjudged as QPSK, and the proportion is only 1.8%. Even between the most confusing classes, the model can maintain an identification accuracy of more than 98%; Figure 9 is the open set recognition confusion matrix, and the proportion of known samples correctly judged as "known" is 98.4%, and the proportion of unknown samples correctly identified as "unknown" is 97.1%. Among them, the proportion of known samples misjudged as unknown is 1.6%, and the proportion of unknown samples misjudged as known is 2.9%, indicating that the model not only maintains a high recognition rate for known classes, but also has strong unknown class detection ability.

[0132] Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. The specification and examples given herein are intended as illustrative only and not intended to limit the scope of the present application. Certain features of the application are described above as belonging to a single embodiment, but some features of the application can be employed independently of other features and the application can be modified by combining some features with or without the inclusion of other features.

[0133] It should be understood that the application is not limited to the precise construction that has been described above and shown in the accompanying drawings and that various modifications and changes can be effected therein by those skilled in the art without departing from the scope of the application.

Claims

1. A method for open set identification of communication signals based on three-channel time-frequency fusion, characterized in that, Comprise the following steps: Adopt end-to-end data preprocessing process to preprocess signals, automatically map original I / Q data to time-frequency graph, I channel and Q channel three-way input; Adopt three-channel parallel feature extraction structure, simultaneously learn multi-scale features of time domain and frequency domain, and fuse through splicing for classification and detection; Through open set identification, combine OpenMax and Weibull distribution tail fitting, introduce probability calibration strategy, dynamically detect and efficiently reject unknown modulation types; Verify the results, compare with various mainstream models, verify the comprehensive advantages in classification accuracy and unknown detection rate; In the step of through open set identification, combining OpenMax and Weibull distribution tail fitting, introducing probability calibration strategy, dynamically detecting and efficiently rejecting unknown modulation types: Send the preprocessed input samples one by one into the trained deep neural network model, perform forward propagation calculation along each layer of the network, select the penultimate layer as the output end, extract the activation values of all samples from this layer, and splice them into the original feature matrix; Batch normalize the original feature matrix to eliminate the differences in feature scales between different samples; According to the true labels of each sample, divide the original feature matrix by category, and calculate the arithmetic mean activation vector of the features of each category, i.e. the mean vector; Apply L2 norm to each mean vector to normalize its amplitude to 1, and obtain the unit norm class center vector of each category; Use the center vector as the basis for subsequent extreme value modeling and Weibull fitting; In the training phase, aggregate the intermediate features of each known category to calculate the category center, and obtain the cosine distance of each sample to its own center; In the test phase, the distance between the test sample and the center of each category is calculated, and the attenuation coefficient is obtained according to the corresponding Weibull model , and the original Softmax probability is attenuated according to the coefficient, and the part attenuated is summed up as the probability of the unknown category. The probability is redistributed to generate an Openmax output containing the probability of the unknown category. By setting a threshold The highest Openmax probability is determined. If it is higher than the threshold, the sample is classified as an unknown class. Otherwise, it is classified according to the highest probability class. In the open set scenario, the model performance is evaluated comprehensively in terms of accuracy, recall rate, and F1 index. 2.The three-channel time-frequency fusion based communication signal open-set identification method of claim 1, wherein, In the step of adopting end-to-end data preprocessing process to preprocess signals, automatically map original I / Q data to time-frequency graph, I channel and Q channel three-way input: Sample the received radio frequency signal and complete quadrature down-conversion processing in the digital domain; The and The signal is passed through a low pass filter to obtain band-limited baseband signals I(n) and Q(n). Complete filtering and decimation of co-directional components and quadrature components in the digital domain to combine into a complex sequence X(n); Perform short-time Fourier transform analysis on X(n) to obtain a time-frequency graph. 3.The three-channel time-frequency fusion based communication signal open-set identification method of claim 1, wherein, In the step of adopting three-channel parallel feature extraction structure, simultaneously learning multi-scale features of time domain and frequency domain, and fusing through splicing for classification and detection: Adopt I, Q waveform and time-frequency graph three-channel parallel input, so that the network can simultaneously capture time domain and frequency domain multi-source information in each convolution, and after the pre-processing process framework, obtain time-frequency graph matrix X(x, f), I sequence and Q sequence, respectively, and extend the I / Q sequence to a matrix along the frequency dimension: The final input tensor is constructed as: The above splicing ensures that the convolution kernel simultaneously accesses the frequency drift in the propagation process, the modulation spectrum line change and the energy distribution dynamics on the time-frequency graph channel, and the instantaneous phase jitter, amplitude fluctuation and phase difference of the signal in the time domain on the I / Q channel in the same receptive field, and the convolution output of the three-channel fusion: wherein 、 is the coordinate of the output feature map in the spatial dimension, is the number of output channels, is the size of the convolution kernel, is the input channel index, 、 is the offset of the convolution kernel in the height and width direction, is the convolution kernel weight corresponding to the th output channel, is the bias term of the th output channel, and t is the number of points in the time domain. 4.The three-channel time-frequency fusion based communication signal open-set identification method of claim 1, wherein, In the step of using the center vector as the basis for subsequent extreme value modeling and Weibull fitting: The calculation method is as follows: wherein, is the eigenvector, is the norm of the eigenvector, is the regularized eigenvector.

5. The method of claim 4, wherein the method is based on a three-channel time- frequency fusion-based communication signal open-set identification. In the training phase, the step of aggregating the intermediate features of each known class to calculate the class center and obtaining the cosine distance of each sample to the center to which it belongs is as follows: The calculation method is as follows: wherein, is a feature vector of the i-th sample, is a center vector of the k-th class, is the cosine distance calculated and dynamically determines the proportion of samples used to fit the tail according to the variance of the distance distribution, while calculating the tail sample size Tailsize, wherein, is the variance, is a certain sample, is the mean of this class of samples, n is the total number of samples, When the variance greater than 0.5 indicates an additional increase in the tail fraction to capture more extreme values when the distribution is more spread out, wherein, represents x rounded down, ensuring that the tail proportion is not lower than the minimum threshold , The first T maximum distance fitting Weibull distribution parameters k and The probability density function of Weibull distribution is shown as follows by using maximum likelihood estimation: where k is a shape parameter, is a scale parameter, and x is the cosine distance of each sample to the center vector. 6.The three-channel time-frequency fusion based out-of-set detection method of communication signals according to claim 1, wherein, In the test phase, the distance between the test sample and the center of each category is calculated, and the attenuation coefficient is obtained according to the corresponding Weibull model , and the original Softmax probability is attenuated according to the coefficient, and the part attenuated is summed up as the probability of the unknown category. In the step of probability redistribution to generate the Openmax output containing the probability of the unknown category: The unknown class probability score calculation method is as follows: wherein, is a decay coefficient, is a probability score for each class per sample in closed set identification, K is the number of known classes of signals, The determination method is as follows: 。 7. The method of open-set identification of communication signals based on three-channel time-frequency fusion according to claim 1, characterized in that, In the step of performing result verification and comparing with various mainstream models to verify the comprehensive advantages in classification accuracy and unknown detection rate: Compared with the traditional mainstream model CNN model and the residual network ResNet model with I / Q input; Comparing with the LSTM model input through the time-frequency graph after the CNN; The actual verification result comparison is carried out by using the closed set recognition accuracy, known class accuracy, unknown class accuracy and overall accuracy.

8. The communication signal open-set discrimination method based on three-channel time-frequency fusion according to claim 7, wherein, In the step of using closed set recognition accuracy, known class accuracy, unknown class accuracy and overall accuracy to compare the actual verification results: The closed set recognition accuracy refers to using known class signals for training and evaluating the classification accuracy of the model for all known classes on the test set, the known class recognition rate and the unknown class recognition rate are evaluated by adding unknown classes after training only with known classes, which respectively measure the recognition performance of the model for known classes and unknown classes, and the overall accuracy is the average of the known class recognition rate and the unknown class recognition rate.

Images