Modulated signal recognition method based on multi-domain mixed attention

Abstract

The present invention proposes a modulation signal recognition method based on multi-domain hybrid attention, which mainly solves the problems of incomplete feature extraction and low recognition accuracy under low signal-to-noise ratio in existing methods. The scheme includes: 1) preprocessing discrete complex signals to obtain I / Q / A / P / F sequences; 2) constructing a spatial feature extraction module to extract waveform features such as amplitude of the sequence; 3) using discrete cosine transform to obtain frequency domain features, and extracting feature information of different frequency components through the frequency domain attention module; 4) constructing a time feature extraction module to mine the dependency between spatial features and frequency domain features and perform deep fusion to obtain the time features of the modulation signal; 5) inputting the time features into the classifier, and outputting the recognition results of the modulation signal category through the classifier. The present invention can obtain the correlation and complementarity between features in different domains, enhance the stability and reliability of the model, and effectively improve the recognition accuracy of each modulation signal under low signal-to-noise ratio.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communication technology, and further relates to communication signal modulation recognition technology, specifically a modulation signal recognition method based on multi-domain mixed attention, which can be used to identify the modulation type of radio signals in electronic warfare. Background Art

[0002] Automatic modulation identification is a method in which the receiver extracts effective features from the acquired signal and further identifies the specific modulated signal by analyzing the extracted features. Automatic modulation identification is the first step in non-cooperative communication and is also the key to ensuring the normal operation of the communication system. It is currently widely used in military and civilian fields.

[0003] In the military field, automatic modulation identification can be used to monitor and detect radio signals, thereby obtaining information about enemy communication systems for military intelligence collection and combat command. In the civilian field, automatic modulation identification can be used to identify the modulation mode of radio communication equipment, realize automated spectrum monitoring and planning, and can also be used to evaluate the quality of radio signals and optimize the modulation mode to improve the effectiveness and reliability of the communication system.

[0004] As the demand for communication quality increases and the communication environment becomes more complex, modulation methods are becoming more and more diverse. The fifth generation of mobile communication technology has gradually been implemented from the research level to specific applications. When its core technology cognitive radio reconfigures spectrum resources, it is necessary to identify the user's modulation signal and perform further analysis to obtain the reset parameters, which places higher performance requirements on the automatic modulation recognition (AMR) technology. Therefore, a high-performance modulation recognition method for automatic modulation recognition technology can effectively promote the development of the next generation of mobile communication technology.

[0005] At present, traditional modulation recognition methods include likelihood ratio and feature extraction, both of which require the extraction of features of modulated signals and recognition through decision trees, support vector machines and other methods. The feature extraction process has high computational complexity and weak generalization ability, making it difficult to adapt to the increasingly complex communication environment. It is easily affected by noise, frequency offset and clock offset, resulting in a significant decrease in recognition performance under low signal-to-noise ratio conditions. With the rapid development of deep learning in recent years, scholars have successively applied this technology to the field of signal modulation recognition. In the literature [Zang K, Ma Z. Automatic modulation classification based on hierarchical recurrent neural networks with grouped auxiliary memory [J]. IEEE Access, 2020, 8: 213052-213061], Zang K and Ma Z proposed a hierarchical recurrent neural network structure with grouped auxiliary memory, which uses the IQ data of the signal as the input of the network and can capture the long-term dependency characteristics of the signal. The experimental results show that this method can handle variable-length signal inputs with different parameters and can effectively identify each modulated signal under high signal-to-noise ratio. In the literature [Liang R, Yang L, Wu S, et al. A three-stream cnn-lstmnetwork for automatic modulation classification [C] / / 2021 13th International Conference on Wireless Communications and Signal Processing (WCSP), 2021: 1-5], Liang R and Yang L et al. proposed a multi-branch modulation recognition method based on CNN and LSTM networks. The IQ data is preprocessed to obtain the amplitude characteristics, frequency characteristics and phase characteristics of the signal and use them as the input of the three branches respectively. The Convolutional Neural Networks (CNN) extracts the spatial characteristics of the signal, and the Long and Short Term Memory (LSTM) extracts the temporal characteristics of the signal. The experimental results show that this method can effectively improve the recognition accuracy of each modulation mode. Through an in-depth analysis of the mainstream algorithms in the field of modulation recognition, the method based on deep learning can achieve better recognition performance and generalization ability, does not require artificial feature selection, and can mine the deep characteristics of the modulated signal.However, the current network architecture used for automatic modulation recognition mainly adopts a combination of CNN and recurrent neural network (RNN). The adaptive features extracted by this method have certain redundancy. When the signal-to-noise ratio is low or there are many modulation types, the ability to characterize features needs to be improved. Summary of the invention

[0006] The purpose of the present invention is to propose a modulated signal recognition method based on multi-domain mixed attention in view of the deficiencies of the prior art, so as to solve the problems that the existing recognition methods have high computational complexity in extracting features, weak generalization ability, difficulty in adapting to increasingly complex communication environments, and susceptibility to noise, frequency offset and clock offset, resulting in a significant decrease in recognition performance under low signal-to-noise ratio conditions. First, the original IQ data is processed to obtain the time domain and frequency domain features of the signal, and then the local receptive field of CNN and the self-attention mechanism of the temporal feature extraction module Conformer are used to extract the feature dependencies within the data, and obtain the correlation and complementarity between features in different domains, thereby enhancing the stability and reliability of the signal recognition model and effectively improving the recognition accuracy of each modulated signal.

[0007] The present invention achieves the above-mentioned purpose by the following specific steps:

[0008] (1) Preprocess the discrete complex signals in the data set and convert them into time domain features, namely I / Q sequences and A / P / F sequences, and concatenate them to obtain a joint feature vector;

[0009] (2) constructing a spatial feature extraction module consisting of a convolution layer, a batch normalization layer, a ReLU activation function layer, and a maximum pooling layer, inputting the joint feature vector into the module, obtaining the waveform characteristics of the modulated signal and projecting it into a high-dimensional continuous space, and obtaining the spatial characteristics of the modulated signal;

[0010] (3) A frequency domain attention module is formed by a one-dimensional discrete cosine transform, two linear mapping layers, and a recalibration layer. The spatial features are input into the module, and the frequency domain attention module extracts the effective information in the frequency component of the modulated signal to obtain the frequency domain features.

[0011] (4) constructing a temporal feature extraction module by means of a feedforward connection module, a convolution module and a cross-domain feature fusion module, wherein each module adopts residual connection and layer normalization, and the cross-domain feature fusion module is used to extract cross-domain complementary features of the signal through a cross-attention mechanism; inputting spatial features and frequency domain features into the temporal feature extraction module, mining the dependency relationship between spatial features and frequency domain features through the temporal feature extraction module and performing deep fusion to obtain the temporal features of the modulated signal;

[0012] (5) The time feature is input into the classifier, and the classifier outputs the recognition result of the modulation signal category.

[0013] Compared with the prior art, the present invention has the following advantages

[0014] First, since the present invention extracts the spatial feature information between adjacent sampling points of the modulated signal through the local receptive field of CNN and realizes the adaptation with the time feature extraction module, the model is further enhanced to extract the spatial and temporal features of the modulated signal;

[0015] Second, the present invention uses one-dimensional discrete cosine transform to obtain comprehensive and rich frequency domain features of the modulation signal, and extracts the frequency components in the modulation signal that contribute most to recognition through the frequency domain attention mechanism, thereby effectively improving the recognition accuracy;

[0016] Third, the present invention uses the cross-attention mechanism to fully integrate the time domain and frequency domain features of the modulated signal, focusing on the correlation and complementarity between different domains, so as to learn the deep difference characteristics of the signal;

[0017] Fourth, since the present invention builds a modulation signal recognition network with a spatial feature extraction module, a frequency domain attention module and a time feature extraction module, the spatial feature extraction module extracts the spatial features between adjacent sampling points, the frequency domain attention module extracts more refined frequency domain features, and the time feature extraction module integrates the time domain and frequency domain features of the modulation signal, thereby effectively improving the recognition accuracy of each modulation signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Flow chart for realizing the method of the present invention;

[0019] Figure 2 It is a schematic diagram of the structure of the spatial feature extraction module of the present invention;

[0020] Figure 3 It is a schematic diagram of the frequency domain attention structure of the present invention;

[0021] Figure 4 It is a schematic diagram of the structure of the time feature extraction module of the present invention;

[0022] Figure 5 The following is a curve diagram showing the recognition accuracy of each modulation mode by the method of the present invention;

[0023] Figure 6 is the confusion matrix of the method of the present invention when SNR=-10dB;

[0024] Figure 7 is the confusion matrix of the method of the present invention when SNR=0dB;

[0025] Figure 8is the confusion matrix of the method of the present invention when SNR=18dB;

[0026] Fig. 9 This is a comparison chart of the recognition accuracy of the method of the present invention and other mainstream methods. DETAILED DESCRIPTION

[0027] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments:

[0028] Example 1: Reference Figure 1 The present invention proposes a modulation signal recognition method based on multi-domain mixed attention, which specifically includes the following steps:

[0029] Step 1. Preprocess the discrete complex signals in the data set, convert them into time domain features, namely I / Q sequences and A / P / F sequences, and concatenate them to obtain the joint feature vector, which is implemented as follows:

[0030] Assume the received signal is y n :

[0031] y n =y I (n)+jy Q (n);

[0032] Among them, y I (n) and y Q (n) respectively represent the in-phase component and the orthogonal component of the nth sample of the received signal, n=0, 1, ..., N-1, N represents the length of the received signal; j represents the imaginary part.

[0033] The I / Q sequence is expressed as:

[0034]

[0035] The A / P / F sequence is expressed as:

[0036]

[0037] The instantaneous amplitude expression is: The instantaneous phase expression is: The instantaneous frequency expression is y F (n) = y P (n)-y P (n-1).

[0038] Step 2. Construct a spatial feature extraction module consisting of a convolution layer, a batch normalization layer, a ReLU activation function layer, and a maximum pooling layer, input the joint feature vector into the module, obtain the waveform features such as the amplitude of the modulated signal, and project it into a high-dimensional continuous space to obtain the spatial features of the modulated signal. This module first uses three convolution layers to scan the joint feature vector to extract the spatial feature information between adjacent sampling points, then processes it through a batch normalization layer and a ReLU activation function layer, and finally extracts the detailed information of the modulated signal features through a maximum pooling layer and processes it through a random dropout layer, finally obtaining the output of the spatial feature extraction module. Where L represents the number of channels.

[0039] In this embodiment, the parameters of the spatial feature extraction module are set as follows: the convolution kernel adopts one-dimensional convolution, the input channels of the one-dimensional convolution are 5, namely I, Q, A, P and F, the number of output channel convolution kernels is 128, the size is 1×3, and the step size is 1; the format of the input data is (512,5,128), where 512 is the batch size and 128 is the number of sampling points of each data; the data is scanned with 128 convolution kernels of 1×3 size to extract the spatial features between adjacent sampling points, and the output matrix size of the one-dimensional convolution module is (512,128,128); after the convolution output, the texture information of the adjacent sampling points is extracted by the maximum pooling layer, the convolution kernel size is 1×3, and the step size is 1.

[0040] Step 3. A frequency domain attention module is constructed by a one-dimensional discrete cosine transform, two linear mapping layers, and a recalibration layer. The spatial features are input into the module. The frequency domain attention module extracts the effective information in the frequency component of the modulated signal to obtain the frequency domain features. The steps are as follows:

[0041] (3.1) The feature vector obtained by the spatial feature extraction module Split along the channel dimension to obtain C group features [X 0 ,X 1 ,...,X C-1 ];

[0042] (3.2) Perform one-dimensional discrete cosine transform (DCT) on each group of segmented features to extract the features of different frequency components of the signal; concatenate the calculated frequency components to obtain the result after DCT processing. In this embodiment, the frequency domain feature Freq corresponding to the i-th channel can be obtained according to the following formula: i :

[0043]

[0044] Among them, f i Yes X iThe corresponding DCT frequency components, the basis function of DCT is defined as The result after DCT processing can be obtained by splicing the calculated frequency components.

[0045] (3.3) The features extracted by DCT are reduced in dimension through a linear mapping layer, and then projected to the original dimension through a linear mapping layer. After Sigmoid normalization, a weighted feature vector is obtained. The input features are recalibrated in the channel dimension, and the obtained frequency domain features are applied to the input matrix respectively to obtain a frequency feature vector S with attention weight. f :

[0046] S f =σ(W2δ(W1·Freq)),

[0047] Among them, σ represents the Sigmoid activation function, δ represents the ReLU activation function, W1 represents the coefficient of dimensionality reduction of the fully connected layer, and W2 represents the coefficient of dimensionality increase of the fully connected layer.

[0048] Step 4. A temporal feature extraction module is constructed by a feedforward connection module, a convolution module and a cross-domain feature fusion module, wherein residual connection and layer normalization are used between each module. The cross-domain feature fusion module is used to extract cross-domain complementary features of the signal through a cross-attention mechanism, and input the spatial features and frequency domain features into the temporal feature extraction module. The temporal feature extraction module mines the dependency between the spatial features and the frequency domain features and performs deep fusion to obtain the temporal features of the modulated signal, which are more comprehensive and rich.

[0049] The temporal feature extraction module takes time domain features and frequency domain features as input. The specific structure is: residual connection -> feedforward connection 1 -> layer normalization -> residual connection -> convolution Conv1 -> layer normalization -> residual connection -> cross-domain feature fusion module -> layer normalization -> residual connection -> convolution Conv2 -> layer normalization -> residual connection -> feedforward connection 2 -> layer normalization.

[0050] The time domain and frequency domain features are passed through the i-layer time feature extraction module with the same structure, and the correlation between feature vectors at different positions is calculated according to the following formula:

[0051]

[0052] x Conv1 =x FFN1 +Conv1(x FFN1 )

[0053] x MHSA =x Conv1 +MHSA(x Conv1 )

[0054] x Conv2 =x MHSA +Conv2(x MHSA )

[0055]

[0056] Out i =LayerNorm(x FFN2 )

[0057] Where x represents the spatial feature extracted by the spatial feature extraction module; x FFN1 and x FFN2 Respectively represent the features extracted by the first feedforward module FFN1 and the second feedforward module FFN2, x Conv1 and x Conv2 Respectively represent the features extracted by the first convolution module Conv1 and the second convolution module Conv2, x MHSA Represents the features extracted by the multi-head self-attention module MHSA, LayerNorm represents layer normalization, Out i Represents the output of the i-th layer of the temporal feature extraction module.

[0058] In the above, the cross-domain complementary features of the signal are extracted through the cross-attention mechanism, as follows:

[0059] The time domain data is linearly mapped to construct the query vector Q α , the frequency domain data is linearly mapped to construct the key vector K β and V β , the cross attention module formula is expressed as:

[0060] MultiHead(Q α ,K β ,V β )=Concat(head1,head2,...,head i ,...,head h )W O ,

[0061] head i =Attention(QW i Q ,KW i K ,VW i V ),

[0062]

[0063] Among them, Concat represents the concatenation operation, head irepresents the attention value of the i-th head, W o is the weight matrix of the model, W i Q , W i K , W i V are the parameter matrices representing the query vector Q, key vector K, and value vector V respectively; represents the scale factor;

[0064] First, the similarity coefficient is obtained by performing a scaled dot product operation on Q and K. The attention value after the dot product has a large fluctuation range, which can easily cause problems such as gradient disappearance, so it needs to be divided by the scaling factor It can distract attention and enhance the generalization ability of the model. Secondly, the Softmax function is used to normalize the output to obtain the weight coefficient corresponding to the V value. This value can not only enhance the weight of the sampling points with high correlation in the signal, but also reduce the weight of the sampling points with low correlation. The weight coefficient is used to weight V to obtain the result of self-attention. Finally, multiple self-attention weights are spliced ​​and synthesized to obtain the features of multiple subspaces, which is the cross-attention value that integrates the time domain and frequency domain features.

[0065] Step 5. Input the time feature into the classifier, and output the recognition result of the modulation signal category through the classifier. Specifically, the high-dimensional features are first mapped to low dimensions through a one-dimensional convolution layer, that is, the number of input channels is 128, the number of output channels is 1, the convolution kernel size is 1×3, the step size is 1, and the features are flattened; secondly, feature mapping is performed through two fully connected layers to obtain the output result of the number of modulation signal categories to be identified; finally, normalization is performed through the Softmax function to obtain the recognition result of the final modulation signal and output it.

[0066] In this embodiment, the public data set RadioML2016.10b can be used to train and test the modulation signal recognition model of the present invention to obtain the final model used for detection.

[0067] Embodiment 2: The overall implementation steps of the identification method proposed in this embodiment are the same as those in Embodiment 1. Now, specific settings are given for the parameters therein to further describe the specific implementation process of the present invention:

[0068] Step a: Preprocess the discrete complex signals in the data set, convert them into time domain feature I / Q sequences and A / P / F sequences, and concatenate them to obtain a joint feature vector; the length of each sample in the data set is 128, so the dimension of the joint feature vector is 5×128. The main parameter settings of the data set are shown in Table 1:

[0069] Table 1. Main parameters of the RadioML2016.10b dataset

[0070]

[0071]

[0072] Step b: For the joint feature vector, the spatial feature extraction module mines the waveform features such as the amplitude of the modulated signal and projects the features into a high-dimensional continuous space:

[0073] Since the values ​​in the modulated signal belong to one-dimensional data, that is, discrete points sampled in time series, the convolution kernel in the model adopts one-dimensional convolution. The input channels of one-dimensional convolution are 5, namely I, Q, A, P and F, and the number of output channels, i.e., convolution kernels, is 128. The convolution kernel size is 1×3 and the step size is 1. The format of the input data is (512,5,128), where 512 is the batch size and 128 is the number of sampling points for each piece of data. The data is scanned with 128 1×3 convolution kernels to extract the spatial features between adjacent sampling points. The output matrix size of the one-dimensional convolution module is (512,128,128). After the convolution output, the texture information of adjacent sampling points is extracted by the maximum pooling layer. The convolution kernel size is 1×3 and the step size is 1. Finally, the feature vector is used as the input of the spatial feature extraction module.

[0074] Step c: Use discrete cosine transform to convert time domain features into frequency domain features, and extract effective information in the frequency components through the frequency domain attention module:

[0075] Step d: For the time domain features and frequency domain features of the signal, the dependency between feature vectors is mined through the time feature extraction module, and the time domain features and frequency domain features of the signal are deeply integrated using cross attention to extract comprehensive and rich features of the signal;

[0076] Step e: Output the probability distribution and recognition result of the modulation signal category through the classifier.

[0077] First, the high-dimensional features are mapped to low dimensions through a one-dimensional convolution layer, that is, the number of input channels is 128, the number of output channels is 1, the convolution kernel size is 1×3, the step size is 1, and the features are flattened. Secondly, feature mapping is performed through two fully connected layers to obtain the output result of the number of modulation signal categories to be identified. Finally, the Softmax function is used for normalization processing to obtain the final recognition result of the modulation signal.

[0078] Step f: Modulation signal recognition model training and verification:

[0079] The joint features of the signal obtained in step a are input into the modulation recognition model based on multi-domain hybrid attention for training. The proportion of the number of samples predicted correctly by the model to the total number of samples and the confusion matrix are used as evaluation indicators. The recognition accuracy and confusion matrix of various modulation signals under different model parameters are compared to determine the optimal model parameters. Finally, the optimal network parameters are verified as shown in Table 2:

[0080] Table 2 Optimal network parameters

[0081] Parameter name Parameter settings Number of encoder layers 8 Number of attention heads 8 Attention vector dimension 128 Training iterations (Epoch) 80 Batch Size 512 Learning Rate 0.001 Loss Function CrossEntropyLoss Optimizer Adam Dropout 0.5

[0082] Step g: Comparison of the effects of the modulation signal recognition model:

[0083] In order to demonstrate the feasibility of the method proposed in the patent, the comparative experiment is compared with VT-CNN, ResNet, LSTM, CLDNN, MCLDNN and Transformer models. Most of the above methods are based on CNN and LSTM models and their hybrid models for feature extraction and classification, and use I / Q sequence as the input of the model. Transformer uses the self-attention mechanism for feature extraction, and uses I / Q sequence and A / P / F sequence as the input of the model. Fig. 9 As shown, the method proposed in this patent achieves good recognition results at both low SNR and high SNR.

[0084] The method of the present invention has wide applications in both military and civilian fields. In the military field, it can be used as an important component of electronic jamming and electronic defense to monitor and scout radio signals, thereby obtaining information on enemy communication systems for military intelligence collection and combat command; in the civilian field, it can be used to identify the modulation mode of radio communication equipment, realize automated spectrum monitoring and planning, and can also be used to evaluate the quality of radio signals and optimize the modulation mode, so as to improve the effectiveness and reliability of the communication system.

[0085] The effect of the present invention is further described below in conjunction with simulation experiments.

[0086] 1. Experimental conditions:

[0087] The hardware platform of the simulation experiment of the present invention is: Intel(R) Core(TM) i5-12600K CPU@4.9GHz, running memory is 16GB, graphics card is NVIDIA GeForce RTX 3090Ti, video memory is 24GB;

[0088] The software platform of the simulation experiment of the present invention is: Linux Ubuntu18.04 operating system, the model is built based on the Pytorch framework, the Pytorch version is 1.12.1, and the CUDA version is 11.3.

[0089] 2. Experimental content:

[0090] Using public data sets, VT-CNN, ResNet, LSTM, CLDNN, MCLDNN and Transformer models and the modulation signal recognition model of the present invention are trained respectively, and the recognition performance of each model is tested to compare the performance differences between the mainstream recognition models and the modulation recognition method of the present invention.

[0091] 3. Evaluation indicators:

[0092] The evaluation indicators are divided into recognition accuracy and confusion matrix. Recognition accuracy indicates the proportion of samples predicted correctly by the model to the total number of samples. The confusion matrix is ​​represented by an n×n square matrix, where n represents the total number of classes, the horizontal axis represents the predicted classification results of the model, and the vertical axis represents the actual classification results.

[0093] 4. Experimental results and analysis:

[0094] Figure 5 The following is the recognition accuracy curve of the method of the present invention for 8 digital modulations and 2 analog modulation signals. It can be seen from the figure that the recognition accuracy of each modulation signal gradually increases with the increase of the signal-to-noise ratio. When the signal-to-noise ratio is -2dB, the recognition accuracy of other modulation modes except the analog signal WBFM can reach more than 82%. When the signal-to-noise ratio is greater than 0dB, the recognition accuracy of most modulation signals can reach more than 97%, which is relatively stable as a whole without obvious fluctuations.

[0095] Figure 6 , Figure 7 and Figure 8The confusion matrices of the method of the present invention when SNR=-10dB, 0dB, and 18dB are respectively; it can be seen that when SNR=-10dB, the signal is seriously distorted due to noise and the like, and the model cannot extract effective features, so most of the modulation signals cannot be effectively identified. When SNR=0dB, the recognition accuracy of most modulation signals can reach more than 92%, and the two analog modulation signals WBFM and AM-DSB are easily confused, with 61% of WBFM mistakenly identified as AM-DSB and 5% of AM-DSB mistakenly identified as WBFM. For high-order modulation modes 16QAM and 64QAM, the feature difference information is small, and confusion is easy to occur under low signal-to-noise ratio. Since the model integrates the time domain and frequency domain features of the modulation signal and further extracts the deep difference features of the signal, the confusion situation when the signal-to-noise ratio is 0dB is greatly improved, and the recognition accuracy of 16QAM and 64QAM is as high as 99%. When SNR=18dB, the recognition rate of most modulation signals can reach more than 96%. The confusion problem between the two analog modulation signals has been improved to a certain extent, and the recognition accuracy of 16QAM and 64QAM signals is as high as 99%.

[0096] Fig. 9 The figure is a comparison chart of the recognition accuracy of the method of the present invention and other mainstream methods. It can be seen from the figure that with the increase of the signal-to-noise ratio, the recognition accuracy of all models gradually increases. When the signal-to-noise ratio is higher than 0dB, the recognition accuracy is basically stable. The methods of the present invention all show higher recognition accuracy than other mainstream recognition models.

[0097] The above experimental results prove the correctness and effectiveness of the method proposed in the present invention.

[0098] Parts of the present invention that are not described in detail belong to common knowledge among those skilled in the art.

[0099] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. It is obvious that for professionals in this field, after understanding the content and principles of the present invention, they may make various modifications and changes in form and details without departing from the principles and structures of the present invention. However, these modifications and changes based on the ideas of the present invention are still within the scope of protection of the claims of the present invention.

Claims

1. A modulation signal recognition method based on multi-domain mixed attention, characterized in that: The specific steps include: (1) Preprocess the discrete complex signals in the data set and convert them into time domain features, namely I / Q sequences and A / P / F sequences, and concatenate them to obtain a joint feature vector; (2) constructing a spatial feature extraction module consisting of a convolution layer, a batch normalization layer, a ReLU activation function layer, and a maximum pooling layer, inputting the joint feature vector into the module, obtaining the waveform characteristics of the modulated signal and projecting it into a high-dimensional continuous space, and obtaining the spatial characteristics of the modulated signal; (3) A frequency domain attention module is formed by a one-dimensional discrete cosine transform, two linear mapping layers, and a recalibration layer. The spatial features are input into the module, and the frequency domain attention module extracts the effective information in the frequency component of the modulated signal to obtain the frequency domain features. The specific steps are as follows: (3.1) The feature vector obtained by the spatial feature extraction module Split along the channel dimension to obtain C group features [X 0 ,X 1 ,...,X C-1 ]; (3.2) Perform one-dimensional discrete cosine transform (DCT) on each group of segmented features to extract the features of different frequency components of the signal; The calculated frequency components are concatenated to obtain the result after DCT processing; (3.3) The features extracted by DCT are reduced in dimension through a linear mapping layer, and then projected to the original dimension through a linear mapping layer. After Sigmoid normalization, a weighted feature vector is obtained. The input features are recalibrated in the channel dimension, and the obtained frequency domain features are applied to the input matrix respectively to obtain a frequency feature vector S with attention weight. f : S f =σ(W2δ(W1·Freq)), Among them, σ represents the Sigmoid activation function, δ represents the ReLU activation function, W1 represents the coefficient of dimension reduction of the fully connected layer, and W2 represents the coefficient of dimension increase of the fully connected layer; (4) constructing a temporal feature extraction module by means of a feedforward connection module, a convolution module and a cross-domain feature fusion module, wherein each module adopts residual connection and layer normalization, and the cross-domain feature fusion module is used to extract cross-domain complementary features of the signal through a cross-attention mechanism; inputting spatial features and frequency domain features into the temporal feature extraction module, mining the dependency relationship between spatial features and frequency domain features through the temporal feature extraction module and performing deep fusion to obtain the temporal features of the modulated signal; The temporal feature extraction module takes the time domain features and the frequency domain features as inputs, and the specific structure is: residual connection->feedforward connection 1->layer normalization->residual connection->convolution Conv1->layer normalization->residual connection->cross-domain feature fusion module->layer normalization->residual connection->convolution Conv2->layer normalization->residual connection->feedforward connection 2->layer normalization; The time domain and frequency domain features are passed through the i-layer time feature extraction module with the same structure, and the correlation between feature vectors at different positions is calculated according to the following formula: x Conv1 =x FFN1 +Conv1(x FFN1 ) x MHSA =x Conv1 +MHSA(x Conv1 ) x Conv2 =x MHSA +Conv2(x MHSA ) Out i =LayerNorm(x FFN2 ) Where x represents the spatial feature extracted by the spatial feature extraction module; x FFN1 and x FFN2 Respectively represent the features extracted by the first feedforward module FFN1 and the second feedforward module FFN2, x Conv1 and x Conv2 Respectively represent the features extracted by the first convolution module Conv1 and the second convolution module Conv2, x MHSA Represents the features extracted by the multi-head self-attention module MHSA, LayerNorm represents layer normalization, Out i represents the output of the i-th layer of the temporal feature extraction module; The cross-domain complementary features of the signal are extracted through the cross-attention mechanism, as follows: The time domain data is linearly mapped to construct the query vector Q α , the frequency domain data is linearly mapped to construct the key vector K β and V β , the cross attention module formula is expressed as: MultiHead(Q α ,K β ,V β )=Concat(head1,head2,...,head i ,...,head h )W O , head i =Attention(QW i Q ,KW i K ,VW i V ), Among them, Concat represents the concatenation operation, head i represents the attention value of the i-th head, W o is the weight matrix of the model, W i Q , W i K , W i V are the parameter matrices representing the query vector Q, key vector K, and value vector V respectively; First, the similarity coefficient is obtained by performing a scaled dot product operation on Q and K. Second, the Softmax function is used to normalize the output to obtain the weight coefficient corresponding to the V value, and the weight coefficient is used to weight V to obtain the result of self-attention. Multiple self-attention weights are spliced ​​and synthesized to obtain the features of multiple subspaces, which is the cross-attention value that integrates the time domain and frequency domain features. (5) The time feature is input into the classifier, and the classifier outputs the recognition result of the modulation signal category.

2. The method according to claim 1, characterized in that: The step (1) is implemented as follows: Assume the received signal is y n : y n =y I (n)+you Q (a); Among them, y I (n) and y Q (n) respectively represent the in-phase component and the quadrature component of the nth sample of the received signal, n=0, 1, ..., N-1, N represents the length of the received signal; j represents the imaginary part; The I / Q sequence is expressed as: The A / P / F sequence is expressed as: The instantaneous amplitude expression is: The instantaneous phase expression is: The instantaneous frequency expression is y F (n) = y P (n)-y P (n-1).

3. The method according to claim 1, characterized in that: In step (2), the joint feature vector is input into the spatial feature extraction module, which first uses three convolutional layers to scan the joint feature vector to extract the spatial feature information between adjacent sampling points, then processes it through a batch normalization layer and a ReLU activation function layer, and finally extracts the detailed information of the modulation signal features through a maximum pooling layer and processes it through a random dropout layer, finally obtaining the output of the spatial feature extraction module. Where L represents the number of channels.

4. The method according to claim 3, characterized in that: The specific parameters of the spatial feature extraction module are set as follows: the convolution kernel adopts one-dimensional convolution, the input channels of the one-dimensional convolution are 5, namely I, Q, A, P and F, the number of output channel convolution kernels is 128, the size is 1×3, and the step size is 1; the format of the input data is (512,5,128), where 512 is the batch size and 128 is the number of sampling points of each data; the data is scanned with 128 1×3 convolution kernels to extract the spatial features between adjacent sampling points, and the output matrix size of the one-dimensional convolution module is (512,128,128); after the convolution output, the texture information of adjacent sampling points is extracted by the maximum pooling layer, the convolution kernel size is 1×3, and the step size is 1.

5. The method according to claim 1, characterized in that: The recognition result of the modulation signal category output by the classifier in step (5) is specifically as follows: first, the high-dimensional features are mapped to low dimensions through a one-dimensional convolution layer, that is, the number of input channels is 128, the number of output channels is 1, the convolution kernel size is 1×3, the step size is 1, and the features are flattened; secondly, feature mapping is performed through two fully connected layers to obtain the output result of the number of modulation signal categories to be identified; finally, normalization is performed through the Softmax function to obtain the recognition result of the final modulation signal and output it.

Images