Electromagnetic signal automatic modulation identification method and system during test based on time-frequency fusion

Abstract

The invention provides an electromagnetic signal automatic modulation identification method and system during testing based on time-frequency fusion. The method comprises the steps of firstly constructing an electromagnetic signal data set; secondly, optimizing a kernel function to realize time-frequency feature modeling, and constructing a double-input path; constructing a deep neural network again, inputting the preprocessed signal sample into a time-domain branch for processing, analyzing a time-frequency spectrogram through the time-frequency branch, and introducing a channel attention mechanism to realize feature fusion; secondly, freezing a time-frequency branch, only training a time-domain branch, outputting a logit value from an input signal through a neural network, and smoothing logit distribution into probability distribution through a scaling variable; carrying out sample slicing processing on an input signal, and realizing convergence by fusing a prediction result; and finally, optimizing affine parameters of the batch normalization layer to complete identification of the electromagnetic signal modulation category. According to the method, the recognition accuracy of the model in a complex wireless environment is improved, so that the electromagnetic signal can still be stably and reliably recognized under the condition of low SNR (Signal to Noise Ratio).

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of signal recognition, and in particular to a test-time electromagnetic signal automatic modulation recognition method and system based on time-frequency fusion. BACKGROUND

[0002] With the rapid development of wireless communication technology, the types of electromagnetic signals are increasing, and they exhibit highly dynamic and complex variation characteristics in multiple dimensions such as frequency, time, space, and power, resulting in an increasingly complex and variable overall electromagnetic environment. The time-varying nature of the wireless channel makes the signal prone to fading, noise interference, and other influences during transmission, greatly increasing the challenge of information transmission. At the same time, the dynamic nature and high uncertainty of electromagnetic signals also pose higher requirements for the security and stability of information systems. In this context, modulation mode recognition technology for communication signals is particularly important in the field of communication. Automatic modulation recognition, as one of the key technologies in wireless communication systems, its accuracy is directly related to subsequent demodulation, resource allocation, interference management, and even communication security.

[0003] Traditional modulation recognition methods rely heavily on manually constructed features, such as high-order statistics, cyclic spectrum, instantaneous amplitude and phase, etc., but such methods are difficult to adapt to feature changes in highly dynamic and non-ideal channel environments, and have high dependence on engineering experience, strong subjectivity, and poor generalization ability. With the development of deep learning technology, modulation recognition models based on neural networks have gradually become mainstream. In particular, deep convolutional networks with I / Q vectors as input can automatically extract complex signal features and improve recognition accuracy.

[0004] However, during the training process of deep learning models, a large amount of labeled data is crucial to improving model accuracy, and there are problems such as high data collection cost and high cost of manual labeling in electromagnetic signal data collection and labeling. When the deep learning model faces inconsistent distribution of training data and test data (such as different channel conditions, different devices, and different signal-to-noise ratio environments), the model performance will drop sharply, and the reasons include wireless channel time-varying, channel estimation error, and inconsistent front-end and back-end gains. These problems limit the generalization ability of the model.

[0005] Currently, some research attempts to use domain adaptation methods to alleviate the problem of data distribution deviation, such as transfer learning and domain adaptation, but these methods usually require labeled data from the target domain or additional training stages, limiting their feasibility in practical applications. Therefore, how to dynamically adjust the model at test time to improve its adaptability to unseen data has become a problem that needs to be solved in the field of wireless electromagnetic signal modulation recognition. SUMMARY

[0006] The present application aims to solve the problems of the existing radio electromagnetic signal modulation mode recognition method, which needs a large amount of labeled data in training and the recognition rate decreases under low SNR conditions, and proposes a test time electromagnetic signal automatic modulation recognition method and system based on time-frequency fusion.

[0007] In a first aspect, the present application provides a test time electromagnetic signal automatic modulation recognition method based on time-frequency fusion, comprising the following steps:

[0008] Constructing an electromagnetic signal dataset in multiple modulation modes and multiple SNR environments;

[0009] Optimizing the kernel function to realize time-frequency feature modeling, and constructing a double-input channel, while pre-processing the I / Q two-way signals of the electromagnetic signal to effectively input information into the model, the double-input channel comprising a time-frequency branch and a time-domain branch;

[0010] Constructing a deep neural network, inputting the pre-processed signal samples into the time-domain branch for processing, and analyzing the time-frequency spectrum through the time-frequency branch, introducing a channel attention mechanism to realize double-branch feature fusion, and finally constructing a modulation recognition model with enhanced discrimination ability;

[0011] Freezing the time-frequency branch in the test phase, training only the time-domain branch, inputting the signal into the deep neural network to output the logit value, and smoothing the high-confidence logit distribution into a probability distribution through scaling variables;

[0012] Sample slicing processing is performed on the input signal in the test phase, a starting point is randomly selected based on the probability distribution or sampling ratio to generate a perturbed sample, convergence is realized by fusing the prediction results of each segment, and finally the convergence result is output and the loss is calculated;

[0013] Based on the convergence result, the affine parameters of the batch normalization layer are optimized, the model parameters are updated by minimizing the loss function, and the automatic recognition of the electromagnetic signal modulation class is completed.

[0014] In a second aspect, the present application provides a test time electromagnetic signal automatic modulation recognition system based on time-frequency fusion, comprising:

[0015] A dataset construction module is used to construct an electromagnetic signal dataset in multiple modulation modes and multiple SNR environments, and to pre-process the original signal;

[0016] A double-input channel module is used to optimize the kernel function to realize time-frequency feature modeling, and to construct a double-input channel, while pre-processing the I / Q two-way signals of the electromagnetic signal to effectively input information into the model, the double-input channel comprising a time-frequency branch and a time-domain branch;

[0017] The double-flow attention fusion module is used for constructing a deep neural network, inputting the preprocessed signal sample into a time domain branch for processing, and analyzing a time-frequency spectrum through a time-frequency branch, introducing a channel attention mechanism to realize double-branch feature fusion, and finally constructing a modulation recognition model with enhanced discrimination ability.

[0018] The probability output module freezes the time-frequency branch in the test stage, trains only the time domain branch, inputs the signal into the deep neural network to output a logit value, and smoothes the high-confidence logit distribution into a probability distribution through a scaling variable.

[0019] The convergence module is used for sample slicing processing on the input signal in the test stage, randomly selects a starting point based on the probability distribution or a sampling ratio to generate a perturbation sample, realizes convergence by fusing the prediction results of each segment, finally outputs a convergence result and calculates a loss.

[0020] The modulation recognition module is used for optimizing the affine parameters of the batch normalization layer based on the convergence result, updating the model parameters by minimizing the loss function, and completing automatic recognition of the electromagnetic signal modulation class.

[0021] In a third aspect, the present application also provides a computer device, comprising a memory for storing a computer program, and a processor for executing the computer program to realize the steps of any of the above methods.

[0022] In a fourth aspect, the present application also provides a computer readable storage medium, which stores a computer program, wherein the computer program is executed by a processor to realize the steps of any of the above methods.

[0023] The beneficial effects of the present application are as follows:

[0024] First, an innovative time-frequency feature modeling method is introduced to improve the analysis ability of key modulation modes. The present application designs a time-frequency graph generation module based on spline function optimization, which can extract high-resolution time-frequency distribution information from the original I / Q signal as a supplementary information stream input into the deep neural network. The module introduces a double-flow structure in the training stage: the time domain branch learns the amplitude and phase evolution mode of the signal, and the time-frequency branch models the non-stationary and modulation transient characteristics. The collaborative learning of the two branches significantly enhances the discrimination ability of the model for diversified modulation signals.

[0025] Second, the present application smoothes the prediction results of the original sample. The sharp distribution lacks adaptability to low-confidence samples (such as noise or uncommon categories). By smoothing the output distribution, the model can more effectively generalize on a wide range of input data and prevent poor performance when facing new samples or unfamiliar data distribution.

[0026] Thirdly, the application enhances the strategy by introducing a sliding time window. The method divides the original signal in the time domain, extracts multiple short-time segments from the continuous I / Q signal to enhance the modeling ability of the model for local modulation structure, uses multiple perspectives to improve robustness, constructs enhanced samples, adapts to local disturbance and time drift, adjusts and learns itself under various different data distributions, improves the adaptability to new samples, and significantly improves the recognition accuracy of the model in complex wireless environments, so that the model can still recognize electromagnetic signals stably and reliably under low SNR conditions. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 a flowchart of an embodiment of the present application;

[0028] Figure 2 a model framework diagram of an embodiment of the present application;

[0029] Figure 3 an adaptive framework diagram during testing of an embodiment of the present application;

[0030] Figure 4 a baseline model and the model recognition performance diagram of an embodiment of the present application. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0032] It should be noted that the terms "first", "second", etc. used in the present application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "include" and "have" used in the present application and any variations thereof are intended to cover non-exclusive inclusion. The term "multiple" used in the present application means two and more than two. The term "and / or" used in the present application means one of the options or any combination of multiple options.

[0033] In one embodiment, as shown in Figure 1 a time-frequency fusion-based automatic modulation recognition method for electromagnetic signals in testing is provided. Through a training-test structure decoupling mechanism, a double-branch architecture is used to fuse I / Q signals and time-frequency distribution features in the training stage, on the one hand, the traditional I / Q time domain information is retained, and on the other hand, the time-frequency evolution features representing the signal are introduced to obtain stronger discrimination ability. In the present embodiment, the method comprises the following steps:

[0034] Step S1: Establishing a data set;

[0035] Further, in the embodiments of the present application, the electromagnetic signal dataset is collected and prepared, and the dataset needs to contain multiple different modulation modes and simulate different SNR environments.

[0036] Step S2: optimize the kernel function design to model the time-frequency features and construct a double-input channel;

[0037] Further, in the embodiments of the present application, a double-input channel is first constructed: a time-frequency branch: time-frequency distribution feature enhancement based on segmented spline optimization is performed on each signal segment to generate a corresponding two-dimensional time-frequency spectrum, and an input branch is used for feature fusion; a time domain branch: a standard complex signal is transformed into a shape transformation format as a standard neural network input.

[0038] Among them, the time-frequency branch uses a segmented spline optimization-based time-frequency distribution feature enhancement module. This module can convert the I / Q signal in complex form into a two-dimensional time-frequency graph representation. The neural network introduces an efficient channel attention mechanism to enhance the model's ability to perceive key channel features, and cooperates with the convolution structure to effectively extract significant features in the two-dimensional time-frequency graph. In the test stage, the time-frequency branch is frozen, and only the I / Q channel is used to complete efficient adaptation combined with a self-supervised distillation mechanism, ensuring fast deployment and robust prediction in resource-constrained scenarios. The present application can simultaneously capture the significant patterns of the modulation signal in the time and frequency dimensions, improving the model's ability to distinguish modulation modes, and exhibiting stronger anti-interference and discrimination ability, especially in low signal-to-noise ratio and channel uncertainty scenarios.

[0039] Among them, the empirical formula is used for preprocessing the electromagnetic signal I / Q two-channel signal in step S2:

[0040]

[0041]

[0042] Among them, represents the original complex electromagnetic data, represents the normalized complex signal sample. Among them and represent the real and imaginary parts of the complex signal, represents the imaginary unit, and the Wigner-Ville (WVD) distribution is calculated:

[0043]

[0044] Among them is a complex signal, is a conjugate function, is a delay, is a frequency variable, is a time variable, represents the signal energy density at time and frequency on the time-frequency plane.

[0045] where, due to the cross-term interference of the original WVD, the application adopts an optimized kernel function for smoothing, obtaining a general expression:

[0046]

[0047] where, represents the time-frequency distribution after smoothing using the kernel function, and the variables and are delay and frequency offset, respectively, for constructing joint time-frequency filtering, where is the ambiguity function in the time-delay-frequency-shift domain, and the formula is as follows:

[0048]

[0049] Further, to effectively suppress the cross-term in the WVD distribution and improve the energy concentration of the time-frequency distribution, the application proposes a two-dimensional kernel function form constructed based on piecewise splines, which is composed of the weighted combination of bivariate B-spline basis functions, and its expression is as follows:

[0050]

[0051] where, is the th B-spline basis function, is the kernel function coefficient to be optimized, and the design and optimization of the kernel function are completed by maximizing the concentration index, and finally the constructed kernel function is applied as the time-frequency graph input.

[0052] Step S3: dual-flow feature extraction and attention enhancement training;

[0053] In the embodiments of the application, a deep neural network structure is constructed, and an efficient channel attention mechanism module is introduced on the basis thereof to enhance the perception ability of the model to key modulation features. The preprocessed signal samples are input into the network for supervised training, and the original complex electromagnetic data are input as the time domain branch output, and the generated time-frequency feature graph is input as the time-frequency branch output for attention weighted fusion to obtain a basic model with improved discrimination ability.

[0054] Specifically, the efficient channel attention mechanism can guide the model to focus on key channel features of the signal without introducing a large number of parameters, thereby improving the overall recognition ability, and the formula is as follows:

[0055]

[0056] wherein, is the input feature map, is the number of channels; represents global average pooling for each channel; is a one-dimensional convolution operation (without bias); is the input feature of the th channel; is the weighted output feature.

[0057] Further, the adjusted network is used for dual-flow feature extraction and fusion, and the original complex electromagnetic data is input as an IQ branch output , and the other branch is a time-frequency graph output , which is further processed by WVD distribution to obtain time-frequency enhanced features , is the fusion weight, and attention weighting fusion is performed, and the formula is as follows:

[0058]

[0059] wherein, represents Sigmoid, is a learnable parameter.

[0060] As shown in Figure 2 , the model feature extractor is used as a backbone network in the adaptive framework during ATSD testing, and the structure includes a plurality of basic modules composed of one-dimensional convolution layers (1D Conv), batch normalization (Batch Normalization) and ReLU activation units, and a lightweight channel attention mechanism (Efficient Channel Attention, ECA) is introduced between key layers to enhance the perception ability of key spectral features. The network tail generates a fixed-dimensional discriminative feature vector through global average pooling (Global Average Pooling) and a fully connected layer (Fully Connected Layer). Each convolution layer is labeled, for example, Conv1d-64, Conv1d-128, and Conv1d-256, representing that the number of channels of the convolution layer gradually increases (64 → 128 → 256).

[0061] Step S4: in the test phase, the time-frequency branch is frozen, and the time-domain branch is improved;

[0062] In the embodiments of the present application, after the input signal passes through the neural network, the logit value of each category is obtained, which is the prediction value of the network for the signal category. Since the high-confidence samples have a sharp logit distribution, the present application introduces a scaling variable Calculate the predicted probability of the smooth distribution The formula is as follows:

[0063]

[0064]

[0065]

[0066] in, Indicates the first The sample at the th The original logit output for each category, This indicates that after temperature scaling, the first... The sample at the th Normalized logit for each category, and These are the mean and standard deviation of the logit, respectively. It's a scaling parameter. This means mapping the scaled logit to the positive probability space. The formula represents the normalization of the scaled logit to obtain the predicted probability of the class. This represents the number of categories.

[0067] Step S5: Perform segmentation and slicing processing on the original test input signal;

[0068] Specifically, the length is L The original signal passes through a length of l The window is slid-trimmed with a step size of [value missing]. s This process yields multiple signal segments with local time-domain characteristics. Based on a preset probability distribution or sampling ratio, a portion of the window's starting point is randomly selected to construct perturbation samples (e.g., noise addition, time-domain / frequency-domain transformation, geometric transformation, etc.). Each segment is predicted individually, and then the results are fused to obtain a new prediction. As a teacher signal, through communication The final loss value was calculated. .

[0069] Specifically, enhanced samples are constructed and fused with multiple fragments, with a sliding window length of [value missing]. w The sliding step size is s For each sample signal The signal is segmented to obtain several signal segments. The formula is as follows:

[0070]

[0071] in, K For the total number of windows, each Inherit the modulation label of the mother sequence Data augmentation for the test phase, the input signal is divided into multiple segments by sliding segmentation , and perturbations are added to the input electromagnetic signal samples as augmented samples to input the model respectively to obtain the prediction probability of each segment Fuse the prediction probabilities of all segments, and the formula is as follows:

[0072]

[0073] Among them, the augmented sample data is constructed and input into the model to obtain the final teacher signal , guide the distribution of the current scaled sample Converge to keep it stable.

[0074] The following formula is the self-supervised distillation loss, which guides the prediction distribution of the scaled sample to optimize the prediction distribution of the model in the test phase and improve its generalization ability:

[0075]

[0076] Among them, represents the number of samples during testing, represents the self-supervised distillation loss function, which is used to guide the prediction distribution of the model in the test phase to be more stable.

[0077] Step S6: Optimize the affine parameters of the batch normalization layer And .

[0078] The method of the present application does not retrain the entire neural network, but only optimizes the affine parameters of the batch normalization layer. These parameters are updated by minimizing the loss function, and the model after updating the parameters judges the input to realize the identification of the electromagnetic signal modulation class. The formula is as follows:

[0079]

[0080] Among them is the affine parameter of the BN layer, is the learning rate, that is, the loss function obtained.

[0081] In order to clearly illustrate the present application, the following will be described in detail in conjunction with specific embodiments.

[0082] The embodiment takes the radio electromagnetic signal modulation mode recognition as an example to verify the effectiveness of the test electromagnetic signal automatic modulation recognition method based on time-frequency fusion proposed by the application. The training data contains 11 types of radio electromagnetic signals in the SNR range of 14dB-16dB in the RML2016.10A data set, and the test data uses the same 11 types of signals as the training set, but the SNR range is in a low SNR (0dB to 14dB) environment to simulate the complex signal propagation in the real electromagnetic environment.

[0083] Firstly, the original signal data is converted into a data format conforming to the input of the neural network model through dimension transformation, so that it can adapt to the training and inference of the deep learning model. After training, the test data is predicted. The test self-adaptive optimization process is as shown in Figure 3 The figure shows the test self-adaptive optimization framework of the application, the original logits output after forward propagation, and by setting a hyperparameter T (temperature coefficient), its function is to scale the distribution of logits, so that the activation function is more “smooth”. As can be seen from the figure, the color becomes clearer and deeper blue, which means that the logits have been optimized to a more discriminative distribution, which helps the model to output more reliable predictions. In this process, only the affine parameters of the batch normalization layer are dynamically adjusted in the test stage, without the need to retrain the entire model, thereby effectively reducing the computational overhead and improving the real-time performance.

[0084] The test data is predicted, and the results are as shown in Figure 4 The figure shows the confusion matrix results of the baseline model (as shown in Figure 4 (a)) and the method of the application (as shown in Figure 4 (b)), respectively. Compared with the recognition accuracy of 31.518% of the baseline model, the method of the application achieves a recognition accuracy of 63.400% under the condition of 0dB, which significantly improves the recognition performance of the wireless signal modulation mode in the low signal-to-noise ratio environment.

[0085] The application embodiment also provides a test electromagnetic signal automatic modulation recognition system based on time-frequency fusion, comprising:

[0086] A data set construction module is configured to construct an electromagnetic signal data set in a multi-modulation mode and a multi-SNR environment, and to perform preprocessing operations on the original signal;

[0087] A double-input channel module is configured to optimize the kernel function to realize time-frequency feature modeling, and to construct a double-input channel while preprocessing the I / Q two-way signals of the electromagnetic signal to effectively input information into the model. The double-input channel includes a time-frequency branch and a time-domain branch.

[0088] The dual-flow attention fusion module is used for constructing a deep neural network, inputting the preprocessed signal sample into a time domain branch for processing, and analyzing a time-frequency spectrum through a time-frequency branch, introducing a channel attention mechanism to realize dual-branch feature fusion, and finally constructing a modulation recognition model with enhanced discrimination ability.

[0089] The probability output module freezes the time-frequency branch in the test stage, trains only the time domain branch, inputs a signal into the deep neural network to output a logit value, and smoothes a high-confidence logit distribution into a probability distribution through a scaling variable.

[0090] The convergence module is used for sample slicing processing on the input signal in the test stage, randomly selecting a starting point based on the probability distribution or a sampling ratio to generate a perturbation sample, realizing convergence through fusion of a prediction result of each segment, finally outputting a convergence result and calculating a loss.

[0091] The modulation recognition module is used for optimizing affine parameters of a batch normalization layer based on the convergence result, updating model parameters by minimizing a loss function, and completing automatic recognition of a modulation class of an electromagnetic signal.

[0092] Embodiments of the present application also provide a computer device including a memory and a processor, and the memory stores a computer program, wherein the computer program is set to execute steps in any of the above methods when running.

[0093] Embodiments of the present application also provide a computer readable storage medium, which stores a computer program, wherein the computer program is set to execute steps in any of the above methods when running.

[0094] In an example embodiment, the above computer readable storage medium can include, but is not limited to, a U disk, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store computer programs.

[0095] The above is only a preferred embodiment of the present application, and does not limit the present application in any form. Although the present application has been disclosed as above, it is not intended to limit the present application. Any skilled person in the art can make some changes or modifications to the above disclosed technical content without departing from the scope of the technical solution of the present application, and make equivalent embodiments with equivalent changes, as long as they do not depart from the technical solution of the present application, and are within the spirit and principles of the present application. Any simple modification, equivalent replacement and improvement of the above embodiments, which are within the scope of the present application, are also within the protection scope of the present application.

Claims

1. A method for automatic modulation recognition of electromagnetic signals in time domain based on time-frequency fusion, characterized in that, The method comprises the following steps: A data set of electromagnetic signals in multiple modulation modes and multiple SNR environments is constructed; A time-frequency feature model is established by optimizing a kernel function, and a double-input channel is constructed, and I / Q signals of the electromagnetic signals are preprocessed to effectively input information into the model, the double-input channel comprising a time-frequency branch and a time-domain branch; A deep neural network is constructed, the preprocessed signal samples are input into the time-domain branch for processing, and a channel attention mechanism is introduced to realize double-branch feature fusion by analyzing a time-frequency spectrum through the time-frequency branch, and finally a modulation recognition model with enhanced discrimination ability is constructed; In the test stage, the time-frequency branch is frozen, only the time-domain branch is trained, the input signal is output as a logit value through the deep neural network, and the high-confidence logit distribution is smoothed into a probability distribution through scaling variables; Sample slicing processing is performed on the input signal in the test stage, a starting point is randomly selected based on the probability distribution or sampling ratio to generate a perturbed sample, convergence is realized by fusing the prediction results of each segment, and finally the convergence result is output and loss is calculated; Based on the convergence result, the affine parameters of the batch normalization layer are optimized, the model parameters are updated by minimizing the loss function, and automatic recognition of the modulation class of the electromagnetic signal is completed.

2. The method of claim 1, wherein, The preprocessing converts the original signal into a standard complex format through an empirical formula, and forms an input format suitable for the structure of the deep neural network through dimension transformation.

3. The test-time electromagnetic signal automatic modulation identification method based on time-frequency fusion according to claim 1 or 2, characterized in that, The kernel function is composed of a weighted combination of a bivariate B-spline basis function, which can suppress the cross-term interference in the WVD distribution and improve the energy concentration of the time-frequency distribution.

4. The method of claim 1, wherein, The time-frequency branch enhances signal features through segmented spline optimization and generates a corresponding two-dimensional time-frequency spectrum, and the input branch is used for feature fusion. The time-domain branch converts the complex signal into a standard input format for processing by the deep neural network.

5. The test-time electromagnetic signal automatic modulation identification method based on time-frequency fusion according to claim 4, characterized in that, The sample slice is a signal segment with local time-domain features obtained by slidingly intercepting the original signal with a length of L using a window with a length of l at a step of s .

6. The test-time electromagnetic signal automatic modulation identification method based on time-frequency fusion according to claim 1 or 5, characterized in that, The perturbed sample includes noise addition, time / frequency domain transformation, and geometric transformation.

7. The method of claim 1, wherein, The feature extractor used in the test stage is composed of a one-dimensional convolution layer, a batch normalization, and a ReLU activation unit, which strengthens the perception ability of key spectral features through a lightweight channel attention mechanism, and finally outputs a fixed-dimensional feature vector through global average pooling and a fully connected layer.

8. The automatic modulation identification system of electromagnetic signal in test time based on time-frequency fusion, characterized in that, The method comprises the following steps: A data set module is used to construct a data set of electromagnetic signals in multiple modulation modes and multiple SNR environments, and to perform preprocessing operations on the original signals; A double-input channel module is used to optimize a kernel function to establish a time-frequency feature model, and to construct a double-input channel, and to preprocess I / Q signals of the electromagnetic signals to effectively input information into the model, the double-input channel comprising a time-frequency branch and a time-domain branch; A double-flow attention fusion module is used to construct a deep neural network, the preprocessed signal samples are input into the time-domain branch for processing, and a channel attention mechanism is introduced to realize double-branch feature fusion by analyzing a time-frequency spectrum through the time-frequency branch, and finally a modulation recognition model with enhanced discrimination ability is constructed; The probability output module freezes the time-frequency branch in the test phase, trains only the time-domain branch, inputs a signal into the deep neural network to output a logit value, and smoothes a high-confidence logit distribution into a probability distribution through a scaling variable; The convergence module is configured to perform sample slicing processing on the input signal in the test phase, randomly selects a starting point based on the probability distribution or a sampling ratio to generate a perturbation sample, realizes convergence by fusing a prediction result of each segment, finally outputs a convergence result, and calculates a loss; The modulation identification module is configured to optimize affine parameters of a batch normalization layer based on the convergence result, updates model parameters by minimizing a loss function, and completes automatic identification of a modulation category of the electromagnetic signal. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to implement the steps of the method in any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method in any one of claims 1 to 7.

Images