Signal modulation recognition method and system based on time-diagram hybrid input and feature dynamic fusion network

Abstract

The application discloses an observation signal modulation identification method and system, and belongs to the technical field of signal processing. The method comprises the following steps: acquiring an observation signal; wherein the observation signal comprises a real part signal and an imaginary part signal; performing horizontal visual graph transformation on the real part signal and the imaginary part signal respectively to obtain corresponding graph adjacency matrices; filtering the real part signal and the imaginary part signal based on the graph adjacency matrices to obtain graph filtering complex signals; inputting the observation signal and the graph filtering complex signals into a single-flow data feature extraction network to synchronously perform feature extraction, and obtaining time domain features and graph domain filtering features respectively; inputting the time domain features and the graph domain filtering features into a feature dynamic fusion network to perform feature dynamic fusion, and obtaining finally fused features; and inputting the finally fused features into a classifier to obtain a modulation mode of the observation signal. The application improves the identification accuracy of the modulation mode.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, and in particular to a signal modulation recognition method and system based on time-map hybrid input and feature dynamic fusion network. Background Technology

[0002] With the increasing demand for adaptive modulation and coding and intelligent signal demodulation in B5G (Beyond B5G) / 6G systems, blind modulation identification has become a key supporting technology for non-cooperative communication and intelligent spectrum management at the receiver. Modulation identification aims to automatically determine the modulation type of the transmitted signal using only observed signal samples, and its accuracy directly affects the efficiency of subsequent demodulation, information extraction, and interference management. Typically, the modulation identification problem can be modeled as a pattern recognition task based on feature extraction and classification decisions. The core lies in extracting discriminative features from a limited number of received samples that are highly separable from the modulation type and less affected by channel distortion, and designing a low-complexity classifier structure. However, in non-cooperative receiving scenarios, improving the accuracy of identifying the modulation type of the transmitted signal has always been a challenge in engineering applications and theoretical research, specifically manifested in the following ways:

[0003] (1) Traditional feature-based recognition algorithms perform poorly in multi-class scenarios. With the increase of signal modulation methods, it is very difficult to define a threshold to distinguish multiple modulation methods in multi-class scenarios.

[0004] (2) Traditional deep learning networks have low recognition accuracy at low signal-to-noise ratios. Traditional deep learning networks, such as convolutional neural networks and transfer learning networks, are susceptible to noise and have low recognition accuracy at low signal-to-noise ratios. Summary of the Invention

[0005] The purpose of this invention is to provide a signal modulation recognition method and system based on time-map hybrid input and dynamic feature fusion network. By extracting features from the graph signal and the time-domain signal respectively, and then dynamically fusing the features, the modulation mode of the observed signal can be identified. This invention improves the recognition accuracy. This invention is achieved through the following technical solutions.

[0006] In a first aspect, the present invention provides a method comprising:

[0007] Acquiring observation signals in cognitive radio application scenarios ; wherein, the observation signal Including real part signal and imaginary part signal ;

[0008] real part signal and imaginary part signal Perform horizontal visible graph transformations to obtain the corresponding graph adjacency matrices. , ;

[0009] Based on graph adjacency matrix For real part signal and imaginary part signal Filtering is performed to obtain the graph-filtered complex signal. ;

[0010] Observation signal Sum-of-Charts Filtering Complex Signals The input data is simultaneously extracted into a single-stream data feature extraction network to obtain temporal features. and graph domain filtering features ;

[0011] Time domain features and graph domain filtering features The input is fed into a feature dynamic fusion network for dynamic feature fusion to obtain the final fused features;

[0012] The final fused features are input into the classifier to obtain the modulation scheme of the observed signal.

[0013] Optionally, the single-stream data feature extraction network includes a sequentially connected multi-scale feature extraction module, a residual module, an attention mechanism module, and a gated recurrent module;

[0014] The single-stream data feature extraction network performs the following operations: It processes the observed signal... Sum-of-Charts Filtering Complex Signals Simultaneously, the input is fed into the multi-scale feature extraction module for feature extraction at different scales, resulting in multi-scale output features. Multi-scale output features The input is fed into the residual module for residual mapping to obtain the residual output features. Residual output characteristics Adaptive average pooling and dimension adjustment are performed to obtain dimension-adjusted features. Features after dimension adjustment The input is fed into the attention mechanism module for attention feature refinement to obtain the attention mechanism transformation features. Transform the attention mechanism into features The input is fed into the gated loop module for updating and resetting, yielding the time-domain features. or graph domain filtering features .

[0015] In this invention, the single-stream data feature extraction network is a self-designed network based on the multi-scale feature extraction module, residual module, attention mechanism module and gated recurrent module, and is also a custom name.

[0016] Optionally, the multi-scale output features are calculated using the following formula:

[0017] ,

[0018] ,

[0019] ,

[0020] In the formula, This represents multi-scale output features. to These represent features at different scales after convolution with kernel sizes of 1×1, 3×3, 5×5, and 7×7, respectively. This represents the features obtained by splicing features from different scales. Indicates batch normalization, This represents the activation function. This represents the observed signal or graph-filtered complex signal input to the multi-scale feature extraction module.

[0021] Optionally, the residual output features and the dimension-adjusted feature expressions are as follows:

[0022] ,

[0023] ,

[0024] In the formula, Indicates the residual output characteristics. This represents the features after adaptive average pooling. This represents the features after dimensionality adjustment. Indicates adaptive average pooling. This represents one-dimensional convolution.

[0025] Optional, attention mechanism transformation features It is obtained by calculation using the following formula:

[0026] ,

[0027] ,

[0028] In the formula, Indicates channel attention weights. and All represent the weights of the fully connected network. This represents the Sigmoid activation function. It represents the Hadamah accumulation. This represents the channel attention output features. Represents spatial attention weights. This indicates max pooling. This represents the changing characteristics of the attention mechanism.

[0029] Optionally, the attention mechanism can be transformed into features. The input is fed into the gated loop module for updating and resetting, yielding the time-domain features. or graph domain filtering features Including: If If the observed signal is input to the multi-scale feature extraction module, then the attention mechanism will transform the features. The input is fed into the gated loop module for updating and resetting, yielding the time-domain features. The expression is as follows:

[0030] ,

[0031] In the formula, Representing time-domain features, This indicates update and reset operations for the gated loop module;

[0032] like If the input is a graph-filtered complex signal to the multi-scale feature extraction module, then the attention mechanism will transform the features. The input is fed into the gated loop module for updating and resetting, resulting in graph domain filtering features. The expression is as follows:

[0033] ,

[0034] In the formula, Indicates graph domain filtering characteristics;

[0035] The formula for calculating the complex signal by graph filtering is as follows:

[0036] ,

[0037] In the formula, This represents a graph-filtered complex signal. and These are the real part signal and the imaginary part signal, respectively. and These are the graph adjacency matrices corresponding to the real and imaginary parts of the signal, respectively.

[0038] Optionally, the feature dynamic fusion network performs the following operation: converting temporal features... and graph domain filtering features The input is fed into a cross-attention mechanism for effective feature focusing, yielding temporal features. In graph domain filtering features Guided updated features and graph domain filtering features In the time domain characteristics Guided updated features Multilayer perceptron is used to enhance the two types of features obtained, resulting in enhanced features. ;

[0039] Dynamically learn feature fusion weights and fuse the two types of features based on these weights to obtain the fused features. Residual connections are used to fuse the fused features and the enhanced features to obtain the final fused features. .

[0040] Optional, time-domain features In graph domain filtering features Guided updated features and graph domain filtering features In the time domain characteristics The updated feature expressions are as follows:

[0041] ,

[0042] In the formula, Time-domain features In graph domain filtering features Guided by updated features, Graph domain filtering features In the time domain characteristics Guided by updated features;

[0043] The enhanced feature expression is as follows:

[0044] ,

[0045] In the formula, Indicates the enhanced features, This represents a multilayer perceptron;

[0046] The feature fusion weight expression is as follows:

[0047] ,

[0048] In the formula, Indicates the feature fusion weights. Indicates the normalized weights. Indicates the weights of the fully connected components;

[0049] The formula for calculating the fused features is as follows:

[0050] ,

[0051] In the formula, Indicates the characteristics after fusion. Representing features respectively The weight coefficients, and satisfying Among them, weight and weight All based on weight Calculated;

[0052] The final fused feature calculation formula is as follows:

[0053] ,

[0054] In the formula, This indicates the final characteristics after fusion.

[0055] Optionally, the classifier includes the calculation of classification scores and predicted categories, with the following formulas:

[0056] ,

[0057] In the formula, Represents the classification weight matrix. Represents the number of categories, Representing dimension, This represents the bias vector. Indicates the category score. This indicates the predicted category of the classifier, i.e., the predicted modulation method of the observed signal.

[0058] Secondly, the present invention provides a signal modulation recognition system based on time-map hybrid input and feature dynamic fusion network, characterized in that it includes:

[0059] The acquisition module in signal cognitive radio applications is used to acquire observed signals. ; wherein, the observation signal Including real part signal and imaginary part signal ;

[0060] The matrix generation module is used to generate the real part of the signal. and imaginary part signal Perform horizontal visible graph transformations to obtain the corresponding graph adjacency matrices. , ;

[0061] The filtering module is used for graph adjacency matrix-based filtering. For real part signal and imaginary part signal Filtering is performed to obtain the graph-filtered complex signal. ;

[0062] The feature extraction module is used to extract the observed signal. Sum-of-Charts Filtering Complex Signals The input data is simultaneously extracted into a single-stream data feature extraction network to obtain temporal features. and graph domain filtering features ;

[0063] The feature fusion module is used to fuse temporal features. and graph domain filtering features The input is fed into a feature dynamic fusion network for dynamic feature fusion to obtain the final fused features;

[0064] The classification module is used to input the final fused features into the classifier to obtain the modulation scheme of the observed signal.

[0065] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0066] This invention introduces a signal modulation recognition method based on time-graph hybrid input and dynamic feature fusion network. By performing a horizontal visual transformation on the real and imaginary parts of the observed signal, a graph adjacency matrix of the real and imaginary parts is obtained, thus yielding a graph-filtered complex signal. This graph-filtered complex signal can then be input into a single-stream data feature extraction network, enabling the network to extract the topological features of the observed signal. This overcomes the limitation of traditional neural networks, which can only extract statistical features of the observed signal. Dynamic weights are obtained by fusing and normalizing the time-domain features with the graph-domain filtered features. Dynamic weights are used to dynamically adjust two features during learning. and Fusion weights and This increases the weight of more effective features during fusion, thereby improving the recognition accuracy of different modulation modes of the observed signal. Through the cross-attention mechanism and residual connection set by the feature dynamic fusion network, more effective features are focused on during multi-feature fusion, further improving the recognition accuracy of different modulation modes of the observed signal. Attached Figure Description

[0067] Figure 1 The diagram shown is a schematic flowchart of a signal modulation recognition method based on time-map hybrid input and feature dynamic fusion network in one embodiment of the present invention.

[0068] Figure 2 The diagram shown is a schematic diagram of the complex signal extraction process of the graph filtering in one embodiment of the present invention;

[0069] Figure 3 The diagram shown is a schematic diagram of a single-stream data feature extraction network structure in one embodiment of the present invention;

[0070] Figure 4 The diagram shown is a schematic diagram of a feature dynamic fusion network structure in one embodiment of the present invention;

[0071] Figure 5 The diagram shown is a comparison of the average recognition accuracy of the recognition method described in this invention and the traditional deep learning network recognition method in one embodiment of the invention.

[0072] Figure 6 The diagram shown is a comparison of the recognition accuracy of a 2DCNN network under different modulation signals in one embodiment of the present invention.

[0073] Figure 7 The diagram shown is a comparison of the recognition accuracy of the GRU network under different modulation signals in one embodiment of the present invention.

[0074] Figure 8 The diagram shown is a comparison of the recognition accuracy of a 1DCNN network under different modulation signals in one embodiment of the present invention.

[0075] Figure 9 The diagram shown is a comparison of the recognition accuracy of the Transfer network under different modulation signals in one embodiment of the present invention.

[0076] Figure 10 The diagram shown is a comparison of the recognition accuracy of the recognition method described in this invention under different modulation signals in one embodiment of the invention. Detailed Implementation

[0077] The following description, in conjunction with the accompanying drawings and specific embodiments, provides further details. In this description, it should be understood that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature.

[0078] Example 1 This embodiment introduces a signal modulation recognition method based on time-map hybrid input and feature dynamic fusion network, such as... Figure 1 As shown, it includes the following steps:

[0079] Acquiring observation signals in cognitive radio application scenarios ; wherein, the observation signal Including real part signal and imaginary part signal ;

[0080] real part signal and imaginary part signal Perform horizontal visible graph transformations to obtain the corresponding graph adjacency matrices. , ;

[0081] Based on graph adjacency matrix For real part signal and imaginary part signal Filtering is performed to obtain the graph-filtered complex signal. ;

[0082] Observation signal Sum-of-Charts Filtering Complex Signals The input data is simultaneously extracted into a single-stream data feature extraction network to obtain temporal features. and graph domain filtering features ;

[0083] Time domain features and graph domain filtering features The input is fed into a feature dynamic fusion network for dynamic feature fusion to obtain the final fused features;

[0084] The final fused features are input into the classifier to obtain the modulation scheme of the observed signal.

[0085] Example 2 Based on Example 1, this example introduces a specific implementation process of a signal modulation recognition method based on time-map hybrid input and feature dynamic fusion network, which includes the following:

[0086] I. Calculation

[0087] Step 1: Calculate the graph-filtered complex signal

[0088] In one specific embodiment of the present invention, a flowchart illustrating the computational graph filtering of complex signals is shown below. Figure 2 As shown. First, the observed signal is acquired. The observed signal is a time-domain signal, and its real part and imaginary part are defined as follows: For the real part signal and imaginary part signal Perform a horizontal visual graph transformation to obtain the corresponding graph adjacency matrix. Based on horizontal visibility Figure 1 The properties of a mapping allow us to directly define the real and imaginary parts of the signal as graphical signals, and then filter the graphical signals to obtain the graphical filtered complex signal. The calculation formula is as follows:

[0089] .

[0090] Step 2: Calculate time-domain features and graph-domain filtering features.

[0091] In one specific embodiment of the present invention, the observed signal and the graph-filtered complex signal are combined. Simultaneously, the data is input into a single-stream data feature extraction network, where features are extracted synchronously. Specifically, time-domain features are extracted from the observed signal, and graph-filtered complex signals are extracted to obtain graph-domain filtered features.

[0092] In a specific embodiment of the present invention, a schematic diagram of the structure of a single-stream data feature extraction network is shown below. Figure 3 As shown, it includes a sequentially connected multi-scale feature extraction module, residual module, attention mechanism module, and gated loop module; it performs the following operations: [The text abruptly ends here, likely due to an incomplete translation or a formatting error.] Sum-of-Charts Filtering Complex Signals Simultaneously, the input is fed into the multi-scale feature extraction module for feature extraction at different scales, resulting in multi-scale output features. Multi-scale output features The input is fed into the residual module for residual mapping to obtain the residual output features. Residual output characteristics Adaptive average pooling and dimension adjustment are performed to obtain dimension-adjusted features. Features after dimension adjustment The input is fed into the attention mechanism module for attention feature refinement to obtain the attention mechanism transformation features. Transform the attention mechanism into features The input is fed into the gated loop module for updating and resetting, yielding the time-domain features. or graph domain filtering features .

[0093] In one specific embodiment of the present invention, the multi-scale output features are calculated using the following formula:

[0094] ,

[0095] ,

[0096] ,

[0097] In the formula, This represents multi-scale output features. to These represent features at different scales after convolution with kernel sizes of 1×1, 3×3, 5×5, and 7×7, respectively. This represents the features obtained by splicing features from different scales. Indicates batch normalization, This represents the activation function. This represents the observed signal or graph-filtered complex signal input to the multi-scale feature extraction module.

[0098] In one specific embodiment of the present invention, the residual output features and the dimension-adjusted feature expressions are as follows:

[0099] ,

[0100] ,

[0101] In the formula, Indicates the residual output characteristics. This represents the features after adaptive average pooling. This represents the features after dimensionality adjustment. Indicates adaptive average pooling. This represents one-dimensional convolution.

[0102] In one specific embodiment of the present invention, the attention mechanism transforms features. It is obtained by calculation using the following formula:

[0103] ,

[0104] ,

[0105] In the formula, Indicates channel attention weights. and All represent the weights of the fully connected network. This represents the Sigmoid activation function. It represents the Hadamah accumulation. This represents the channel attention output features. Represents spatial attention weights. This indicates max pooling. This represents the changing characteristics of the attention mechanism.

[0106] In one specific embodiment of the present invention, the attention mechanism is transformed into a feature. The input is fed into the gated loop module for updating and resetting, yielding the time-domain features. or graph domain filtering features Including: If If the observed signal is input to the multi-scale feature extraction module, then the attention mechanism will transform the features. The input is fed into the gated loop module for updating and resetting, yielding the time-domain features. The expression is as follows:

[0107] ,

[0108] In the formula, This indicates update and reset operations for the gated loop module;

[0109] like If the input is a graph-filtered complex signal to the multi-scale feature extraction module, then the attention mechanism will transform the features. The input is fed into the gated loop module for updating and resetting, resulting in graph domain filtering features. The expression is as follows:

[0110] .

[0111] Step 3, calculate the final fused features

[0112] In one specific embodiment of the present invention, a feature dynamic fusion network is used to process temporal features. and graph domain filtering features The features are then fused to obtain the final fused features. A schematic diagram of the feature dynamic fusion network structure is shown below. Figure 4 As shown, perform the following operations: Convert time-domain features and graph domain filtering features The input is fed into a cross-attention mechanism for effective feature focusing, yielding temporal features. In graph domain filtering features Guided updated features and graph domain filtering features In the time domain characteristics Guided updated features Multilayer perceptron is used to enhance the two types of features obtained, resulting in enhanced features. ;

[0113] Dynamically learn feature fusion weights and fuse the two types of features based on these weights to obtain the fused features. Residual connections are used to fuse the fused features and the enhanced features to obtain the final fused features. .

[0114] In one specific embodiment of the present invention, time-domain features In graph domain filtering features Guided updated features and graph domain filtering features In the time domain characteristics The updated feature expressions are as follows:

[0115] ,

[0116] In the formula, Time-domain features In graph domain filtering features Guided by updated features, Graph domain filtering features In the time domain characteristics Guided by updated features;

[0117] The enhanced feature expression is as follows:

[0118] ,

[0119] In the formula, Indicates the enhanced features, This represents a multilayer perceptron;

[0120] The feature fusion weight expression is as follows:

[0121] ,

[0122] In the formula, Indicates the feature fusion weights. Indicates the normalized weights. Indicates the weights of the fully connected components;

[0123] The formula for calculating the fused features is as follows:

[0124] ,

[0125] In the formula, Indicates the characteristics after fusion. Representing features respectively The weight coefficients, and satisfying Among them, weight and weight All based on weight Calculated;

[0126] The final fused feature calculation formula is as follows:

[0127] ,

[0128] In the formula, This indicates the final characteristics after fusion.

[0129] Step 4, Classification

[0130] In one specific embodiment of the present invention, the classifier includes the calculation of classification score and predicted category, and the calculation formulas are as follows:

[0131] ,

[0132] In the formula, Represents the classification weight matrix. Represents the number of categories, Representing dimension, This represents the bias vector. Indicates the category score. This indicates the predicted category of the classifier, i.e., the predicted modulation method of the observed signal.

[0133] The final fused features The signal is input into a classifier, which can then be used to predict the modulation scheme of the observed signal.

[0134] II. Experiment

[0135] In one specific embodiment of the present invention, the publicly available dataset RML2018.01a is used for verification. This dataset contains 24 modulation signals, including 4-ary Amplitude Shift Keying (4ASK), 8-ary Phase Shift Keying (8PSK), 64-ary Quadrature Amplitude Modulation (64QAM), Double-Sideband Suppressed Carrier Amplitude Modulation (AM-DSB-SC), Single-Sideband Suppressed Carrier Amplitude Modulation (AM-SSB-SC), BPSK, Frequency Modulation (FM), Gaussian Filtered Minimum Shift Keying (GMSK), and QPSK. A modulation identification task is performed on 9 of the listed modulation signals.

[0136] Table 1. Recognition accuracy of different algorithm models

[0137] Table 1 shows a comparison of the overall recognition accuracy of the method described in this invention with existing traditional deep learning networks on the RML2018.01a dataset. The method proposed in this invention is compared with existing classic recognition methods including 2DCNN, GRU, 1DCNN, and Transfer. GRU is a network built based on GRU modules. As can be seen from the table, the method described in this invention achieves higher recognition accuracy and F1 score than other models on the same dataset, demonstrating better recognition performance. Although the inference time is slightly longer than other methods, it is measured in microseconds and is almost negligible. Acc in Table 1 represents the recognition accuracy.

[0138] Figure 5 The figure shows the overall recognition accuracy of different models under different SNR conditions. Figure 5 It can be seen that at low SNR ( The proposed method significantly outperforms 2DCNN, 1DCNN, and Transfer models, but shows little difference from the GRU model. This is because the GRU model effectively captures the dependencies in time series data, thus exhibiting better recognition performance. With the increase of SNR (…), Under high SNR conditions, the proposed method in this invention significantly outperforms the GRU model in recognition performance. Therefore, the proposed model in this invention exhibits superior overall performance.

[0139] Figures 6-9 These are schematic diagrams illustrating the recognition accuracy of traditional deep learning networks for different modulation schemes. Figure 10 This diagram illustrates the recognition accuracy of the method of the present invention for different modulation schemes. The horizontal axis represents the signal-to-noise ratio, and the vertical axis represents the recognition accuracy. The legend in the upper right corner shows nine different modulation schemes. (Combined with...) Figures 6-10 As can be seen, compared with 2DCNN and 1DCNN at low signal-to-noise ratios, the method proposed in this invention has a significant advantage in initial recognition accuracy for FM and QPSK signals. Although the initial recognition accuracy is not outstanding for some modulated signals, as the signal-to-noise ratio increases to 0dB, the recognition accuracy of the proposed method for all nine modulated signals approaches 100%, while other models do not achieve complete recognition for some modulated signals. Therefore, the method proposed in this invention can achieve complete recognition of multiple modulation methods more quickly. Comparison shows that the recognition performance of this invention is superior to existing traditional deep learning models, demonstrating the superiority of this invention.

[0140] Example 3 This embodiment describes an observation signal modulation identification system, including:

[0141] The signal acquisition module is used to acquire observation signals in cognitive radio application scenarios. ; wherein, the observation signal Including real part signal and imaginary part signal ;

[0142] The matrix generation module is used to generate the real part of the signal. and imaginary part signal Perform horizontal visible graph transformations to obtain the corresponding graph adjacency matrices. , ;

[0143] The filtering module is used for graph adjacency matrix-based filtering. For real part signal and imaginary part signal Filtering is performed to obtain the graph-filtered complex signal. ;

[0144] The feature extraction module is used to extract the observed signal. Sum-of-Charts Filtering Complex Signals The input data is simultaneously extracted into a single-stream data feature extraction network to obtain temporal features. and graph domain filtering features ;

[0145] The feature fusion module is used to fuse temporal features. and graph domain filtering features The input is fed into a feature dynamic fusion network for dynamic feature fusion to obtain the final fused features;

[0146] The classification module is used to input the final fused features into the classifier to obtain the modulation scheme of the observed signal.

[0147] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for identifying modulation of observed signals, characterized in that, include: Acquire observation signals in cognitive radio application scenarios; wherein the observation signals include real part signals and imaginary part signals; The real and imaginary parts of the signal are transformed by a horizontal visual diagram to obtain the corresponding graph adjacency matrix. Filtering the real and imaginary parts of the signal based on the graph adjacency matrix yields a graph-filtered complex signal. The observed signal and the graph-filtered complex signal are input into a single-stream data feature extraction network for simultaneous feature extraction, resulting in time-domain features and graph-domain filtered features, respectively. Temporal features and graph domain filtered features are input into a feature dynamic fusion network to perform feature dynamic fusion and obtain the final fused features. The final fused features are input into the classifier to obtain the modulation scheme of the observed signal.

2. The signal modulation recognition method based on time-map hybrid input and feature dynamic fusion network according to claim 1, characterized in that, The single-stream data feature extraction network consists of a sequentially connected multi-scale feature extraction module, a residual module, an attention mechanism module, and a gated recurrent module; The single-stream data feature extraction network performs the following operations: simultaneously inputting the observed signal and the graph-filtered complex signal into the multi-scale feature extraction module for feature extraction at different scales to obtain multi-scale output features; inputting the multi-scale output features into the residual module for residual mapping to obtain residual output features; performing adaptive average pooling and dimension adjustment on the residual output features to obtain dimension-adjusted features; inputting the dimension-adjusted features into the attention mechanism module for attention feature purification to obtain attention mechanism transformed features; and inputting the attention mechanism transformed features into the gated loop module for updating and resetting to obtain temporal features or graph-domain filtered features.

3. The signal modulation recognition method based on time-map hybrid input and feature dynamic fusion network according to claim 2, characterized in that, Multi-scale output features are calculated using the following formula: ， ， ， In the formula, This represents multi-scale output features. to These represent features at different scales after convolution with kernel sizes of 1×1, 3×3, 5×5, and 7×7, respectively. This represents the features obtained by splicing features from different scales. Indicates batch normalization, This represents the activation function. This represents the observed signal or graph-filtered complex signal input to the multi-scale feature extraction module.

4. The signal modulation recognition method based on time-map hybrid input and feature dynamic fusion network according to claim 3, characterized in that, The residual output features and the dimension-adjusted feature expressions are as follows: ， ， In the formula, Indicates the residual output characteristics. This represents the features after adaptive average pooling. This represents the features after dimensionality adjustment. Indicates adaptive average pooling. This represents one-dimensional convolution.

5. The signal modulation recognition method based on time-map hybrid input and feature dynamic fusion network according to claim 4, characterized in that, Attention mechanism transformation features It is obtained by calculation using the following formula: ， ， In the formula, Indicates channel attention weights. and All represent the weights of the fully connected network. This represents the Sigmoid activation function. It represents the Hadamah accumulation. This represents the channel attention output features. Represents spatial attention weights. This indicates max pooling. This represents the transformation characteristics of the attention mechanism.

6. The signal modulation recognition method based on time-map hybrid input and feature dynamic fusion network according to claim 5, characterized in that, Transform the attention mechanism The input is fed into the gated loop module for updating and resetting, yielding the time-domain features. or graph domain filtering features Including: If If the observed signal is input to the multi-scale feature extraction module, then the attention mechanism will transform the features. The input is fed into the gated loop module for updating and resetting, yielding the time-domain features. The expression is as follows: ， In the formula, Representing time-domain features, This indicates update and reset operations for the gated loop module; like If the input is a graph-filtered complex signal to the multi-scale feature extraction module, then the attention mechanism will transform the features. The input is fed into the gated loop module for updating and resetting, resulting in graph domain filtering features. The expression is as follows: ， In the formula, Indicates graph domain filtering characteristics; The formula for calculating the complex signal by graph filtering is as follows: ， In the formula, This represents a graph-filtered complex signal. and These are the real part signal and the imaginary part signal, respectively. and These are the graph adjacency matrices corresponding to the real and imaginary parts of the signal, respectively.

7. The signal modulation recognition method based on time-map hybrid input and feature dynamic fusion network according to claim 6, characterized in that, The feature dynamic fusion network performs the following operations: inputting temporal features and graph domain filtered features into a cross-attention mechanism for effective feature focusing, obtaining the features updated by the temporal features guided by the graph domain filtered features, and the features updated by the graph domain filtered features guided by the temporal features, respectively; and using a multilayer perceptron to enhance the two types of features to obtain the enhanced features. The feature fusion weights are dynamically learned and the two features are fused based on the feature fusion weights to obtain the fused features. The fused features and the enhanced features are then fused together using residual connections to obtain the final fused features.

8. The signal modulation recognition method based on time-map hybrid input and feature dynamic fusion network according to claim 7, characterized in that, Temporal characteristics In graph domain filtering features Guided updated features and graph domain filtering features In the time domain characteristics The updated feature expressions are as follows: ， In the formula, Time-domain features In graph domain filtering features Guided by updated features, Graph domain filtering features In the time domain characteristics Guided by updated features; The enhanced feature expression is as follows: ， In the formula, Indicates the enhanced features, Represents a multilayer perceptron; The feature fusion weight expression is as follows: ， In the formula, Indicates the feature fusion weights. Indicates the normalized weights. Indicates the weights of the fully connected components; The formula for calculating the fused features is as follows: ， In the formula, Indicates the characteristics after fusion. Representing features respectively The weight coefficients, and satisfying Among them, weight and weight All based on weight Calculated; The final fused feature calculation formula is as follows: ， In the formula, This indicates the final characteristics after fusion.

9. The signal modulation recognition method based on time-map hybrid input and feature dynamic fusion network according to claim 8, characterized in that, The classifier includes the calculation of classification score and predicted class, and the calculation formulas are as follows: ， In the formula, Represents the classification weight matrix. Represents the number of categories, Representing dimension, This represents the bias vector. Indicates the category score. This indicates the predicted category of the classifier, i.e., the predicted modulation method of the observed signal.

10. A signal modulation recognition system based on time-map hybrid input and feature dynamic fusion network, characterized in that, include: A signal acquisition module is used to acquire observation signals in cognitive radio application scenarios; wherein, the observation signals include real part signals and imaginary part signals; The matrix generation module is used to perform horizontal visual transformations on the real and imaginary parts of the signal respectively to obtain the corresponding graph adjacency matrix. The filtering module is used to filter the real and imaginary signals based on the graph adjacency matrix to obtain the graph-filtered complex signal; The feature extraction module is used to input the observed signal and the graph-filtered complex signal into the single-stream data feature extraction network for simultaneous feature extraction, thereby obtaining time-domain features. and graph domain filtering features; The feature fusion module is used to fuse temporal features. The graph domain filtering features are input into the feature dynamic fusion network for dynamic feature fusion to obtain the final fused features; The classification module is used to input the final fused features into the classifier to obtain the modulation scheme of the observed signal.

Images