Signal modulation recognition method and system based on GNN and hybrid expert architecture

Abstract

The application discloses a signal modulation recognition method and system based on GNN and a mixed expert architecture, and relates to the technical field of wireless signal modulation recognition. The steps are as follows: after an IQ signal is normalized, the IQ signal is cut into equal-length small segments, and the segmented signals are separated according to I channels, Q channels and IQ combined channels; through a routing mechanism, the most suitable adjacency matrix is selected for each signal segment from an expert library; the signal segments of each channel and the selected adjacency matrix are sent into corresponding graph neural networks to extract local features, the local features are subjected to feature remodeling and cross-channel fusion to form segment-level fusion features; the inter-segment time sequence dependence of the segment-level fusion features is captured by using a recurrent neural network, and finally, a signal modulation type prediction result is output by a classifier. The application combines the graph neural network and the expert mixed mechanism, guarantees the adaptability of the model to different signal modes, and improves the feature learning ability and recognition accuracy of the complex radio signal modulation type.

Description

Technical Field

[0001] This invention relates to the field of wireless signal modulation recognition technology, and in particular to a signal modulation recognition method and system based on GNN and hybrid expert architecture. Background Technology

[0002] Automatic Modulation Classification (AMC) is a core technology in the field of wireless communication signal processing. Its core task is to automatically identify the modulation scheme of received complex-valued radio signals (I / Q sequences) in the absence of prior signal information. This technology is a key support for cognitive radio dynamic spectrum access, military communication interference monitoring, and intelligent management of civilian spectrum resources. It plays an irreplaceable role in communication assurance in complex electromagnetic environments and is also an important prerequisite for achieving efficient signal demodulation and improved spectrum utilization. In existing technologies, traditional convolutional neural networks (CNNs) and recurrent neural networks (RNNs) for AMC struggle to adapt to the time-varying characteristics of signals; graph neural network models such as DTSG-Net, while achieving piecewise graph modeling, lack dynamic structure adjustment capabilities; and while Mixture of Experts (MoE) models possess pattern adaptation potential, they are not integrated with piecewise graph modeling. Therefore, for those skilled in the art, how to overcome the limitations of traditional models and achieve architectural fusion to improve AMC accuracy in complex electromagnetic environments is an urgent problem to be solved. Summary of the Invention

[0003] The purpose of this invention is to provide a signal modulation recognition method and system based on GNN and hybrid expert architecture. It integrates MoE structure with segmentation processing and GNN dynamic graph modeling, giving full play to the ability of GNN to capture flexible node relationships and the advantages of MoE dynamic adaptation. While ensuring the model's adaptability to different signal modes, it improves the feature learning ability and recognition accuracy of complex radio signal modulation types.

[0004] To achieve the above objectives, the present invention provides the following solution: a signal modulation recognition method based on GNN and hybrid expert architecture, the specific steps of which include the following:

[0005] S1. Obtain the IQ signal, normalize the IQ signal and then divide it into equal-length segments. The segmented signal is then separated into I channel, Q channel and IQ combined channel.

[0006] S2. The routing mechanism selects the most suitable adjacency matrix from the expert database for each signal segment, thereby achieving dynamic adaptation between the signal segment and the graph structure.

[0007] S3. Input each channel signal segment and the selected adjacency matrix into the corresponding graph neural network to extract local features, and perform feature reshaping and cross-channel fusion on the local features to form segment-level fusion features;

[0008] S4. The inter-segment temporal dependency of the segment-level fusion features is captured using a recurrent neural network, and the signal modulation type prediction result is finally output by a classifier.

[0009] Preferably, the normalization process for the IQ signal includes removing the mean and standardization to obtain a preprocessed signal. The preprocessed signal is then divided into equal-length segments using a non-overlapping sliding window, and the segmentation process satisfies the following: Where SN is the preset number of nodes within a segment, NS is the number of segments, L is the sampling length of a single sample signal, and the sliding step size is consistent with the preset number of nodes within a segment.

[0010] Preferably, the normalization process is as follows:

[0011] ;

[0012] in, Let be the element of the b-th sample, the l-th sampling point, and the c-th channel in the original signal; This is the mean of the channel; This represents the standard deviation of the channel. To avoid division by zero constants.

[0013] Preferably, the most suitable graph structure is selected for each signal segment by calculating the matching score between each channel signal segment and the adjacency matrix in the expert database. The specific steps are as follows:

[0014] For each channel's single-segment feature, it is mapped to dimension D through a channel-specific multilayer perceptron. At the same time, the adjacency matrix is ​​projected to dimension D through vectorization and a linear layer. The inner product of the two is used to obtain the initial matching score. Gaussian noise is added to the initial matching score to obtain the final matching score.

[0015] Based on the final matching score, for each signal segment of each channel, the expert index with the highest matching score is selected from the expert library through the argmax operation to form a routing index matrix;

[0016] Based on the routing index matrix, a batch-specific adjacency matrix is ​​indexed from the expert database. After dimensional rearrangement, it corresponds one-to-one with the corresponding signal segment in the batch and segment dimensions, providing dynamic graph structure input for the subsequent graph neural network.

[0017] Preferably, the specific steps for reshaping and cross-channel fusion of the local features to form segment-level fused features are as follows:

[0018] First, the features of each channel are reshaped into a structured representation of [batch size × number of segments × number of nodes within a segment × hidden dimension];

[0019] Subsequently, the channel features of the I channel, Q channel, and IQ combined channel are spliced ​​together at the channel dimension to obtain segment-level features;

[0020] Finally, the segment-level features are output by means pooling the nodes within the segment.

[0021] Preferably, the segment-level fusion features are input into a two-layer stacked recurrent neural network to capture long-term temporal correlations between segments and obtain global temporal features.

[0022] Preferably, the global temporal features are input into a dataset-specific classifier and mapped to a category space to complete signal recognition; the classification process is defined as follows:

[0023] ;

[0024] Where D represents the target dataset, For the corresponding dataset, a linear classifier, Let be the class prediction vector for the b-th sample, and C be the number of classes in the dataset. Finally, the class probability distribution is obtained through the softmax function.

[0025] Preferably, after predicting the modulation type of the output signal, the method further includes evaluating the accuracy of the prediction results using leave-one-out or cross-validation, and analyzing the adaptability of different expert graph structures to the signal mode.

[0026] On the other hand, a signal modulation recognition method based on GNN and hybrid expert architecture is provided, including an IQ signal acquisition and channel separation module, a graph structure adaptation module, a feature extraction and fusion module, and a modulation recognition module; wherein,

[0027] The IQ signal acquisition and channel separation module is used to acquire IQ signals, normalize the IQ signals and divide them into equal-length segments, and then separate the segmented signals according to I channel, Q channel and IQ combined channel;

[0028] The graph structure adaptation module is used to select the most suitable adjacency matrix from the expert database for each signal segment through a routing mechanism, thereby realizing the dynamic adaptation of the signal segment and the graph structure.

[0029] The feature extraction and fusion module is used to input each channel signal segment and the selected adjacency matrix into the corresponding graph neural network to extract local features, and to perform feature reshaping and cross-channel fusion on the local features to form segment-level fused features.

[0030] The modulation recognition module is used to capture the inter-segment temporal dependence of the segment-level fusion features using a recurrent neural network, and finally outputs the signal modulation type prediction result through a classifier.

[0031] Preferably, it also includes an evaluation module for evaluating the accuracy of the signal modulation type prediction results and analyzing the adaptability of different expert graph structures to the signal mode.

[0032] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0033] (1) It overcomes the limitations of existing models (such as CNN and RNN) in processing radio signals, which are difficult to adapt to the structural changes of the signal in different time periods due to the use of fixed structures (fixed convolution kernels or recurrent units), and the problem of insufficient modulation recognition accuracy caused by the inability to flexibly capture complex node relationships in the signal.

[0034] (2) By segmenting the original signal, the expert routing mechanism of MoE is used to dynamically select the most suitable graph structure for each signal segment. The GNN extracts intra-segment features and the RNN captures inter-segment temporal dependencies, giving full play to the GNN's ability to capture flexible node relationships and the advantages of MoE's dynamic adaptation. While ensuring the model's adaptability to different signal modes, the feature learning ability and recognition accuracy of complex radio signal modulation types are improved. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart of the method of the present invention;

[0037] Figure 2 This is a structural diagram of the model of the present invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] This invention uses the RML2016.10a dataset as experimental data. This dataset contains signals of 11 modulation types. The dataset is divided into training set and test set in a 4:1 ratio. The input signal is an IQ signal with shape (176000,128,2) (number of data, data length, number of input channels).

[0040] The purpose of this invention is to provide a signal modulation recognition method based on GNN and hybrid expert architecture, such as... Figure 1 As shown, the specific steps include the following:

[0041] S1. Obtain the IQ signal, normalize the IQ signal and then divide it into equal-length segments. The segmented signal is then separated into I channel, Q channel and IQ combined channel.

[0042] S2. The routing mechanism selects the most suitable adjacency matrix from the expert database for each signal segment, thereby achieving dynamic adaptation between the signal segment and the graph structure.

[0043] S3. Input each channel signal segment and the selected adjacency matrix into the corresponding graph neural network to extract local features, perform feature reshaping and cross-channel fusion on the local features to form segment-level fused features;

[0044] S4. The inter-segment temporal dependency of segment-level fusion features is captured by a recurrent neural network, and the signal modulation type prediction result is finally output by a classifier.

[0045] Furthermore, in S1, the acquired IQ signal is first normalized (mean removed and standardized), and then the original long time-series signal is divided into multiple equal-length segments through a segmentation operation (Unfold), forming a feature form of [batch, number of segments, number of nodes within a segment, number of channels], preparing for subsequent local feature extraction. Specifically, step S1 includes the following:

[0046] S11. Define the form of the input IQ signal. The input signal is a batch dual-channel timing signal containing an in-phase component I and a quadrature component Q, expressed as: Where B is the batch size and L is the sampling length of a single sample signal. and These are the in-phase and quadrature component matrices, respectively, with the signal dimension being [batch size × sampling length × number of channels].

[0047] S12. Normalize the IQ signal. This process eliminates signal strength differences by removing the mean and standardizing the signal. The process is defined as follows:

[0048] ;

[0049] in, Let be the element of the b-th sample, the l-th sampling point, and the c-th channel in the original signal; This is the mean of the channel; This represents the standard deviation of the channel. To avoid the small constant of division by zero, the signal dimension is maintained after normalization. .

[0050] S13. Segment the long time-series signal through a segmentation operation. An unfolded sliding window is used to segment the normalized long signal into equal-length segments. The segmentation process satisfies the following: Where SN is the preset number of nodes within a segment (window size), NS is the number of segments (number of windows), and the sliding step is the same as the window size (step=SN). After dimensionality rearrangement, the output feature form is: The dimensions are [batch size × number of segments × number of nodes within a segment × number of channels], which provide structured input for subsequent multi-channel local feature extraction.

[0051] S14. Separate multi-channel signal segments. From the segmented signal, separate single-channel and combined-channel signal segments, specifically:

[0052] Channel I signal segment: Extracts the in-phase component of the segmented signal, defined as... ,in The I component corresponding to the b-th sample, the n-th segment, and the s-th node is obtained by indexing the first dimension of the segmented signal channel: ;

[0053] Q-channel signal segment: Extracts the orthogonal components of the segmented signal, defined as... ,in The Q components at the corresponding positions are obtained by indexing the second dimension of the segmented signal channel: ;

[0054] IQ combined channel signal segment: The I and Q components are spliced ​​and flattened at the node dimension, defined as... The feature vector of the b-th sample and the n-th segment is Achieved through dimensional reshaping: .

[0055] In this embodiment, the specific operation process is as follows: The original input IQ signal (shape (176000,128,2)) is standardized, the mean of the time dimension is removed, and then scaled using the standard deviation with a 1e-5 perturbation. The output is still (176000,128,2). The signal is segmented using an unfold operation. Assuming the number of nodes in each segment is segment_nodes=32, the number of segments is NS=4. After dimension adjustment, the segmented signal is output as (176000,4,32,2). The I channel is separated to obtain x_i(176000,4,32), and the Q channel is obtained as x_q(176000,4,32). The I / Q channels are flattened and merged to obtain x_iq(176000,4,64).

[0056] Furthermore, step S2 includes the following:

[0057] S21. Define the expert database and the input to the routing module. Construct an expert database containing NE learnable adjacency matrices. And clearly define the characteristics of the routing input.

[0058] Expert database adjacency matrix: Adapted to dimensions based on channel type, I / Q single channel correspondence. IQ combination channel corresponding It includes topology types such as dense connectivity, sparse connectivity (random 20% of nodes), local connectivity (neighborhoods at different scales), and random connectivity, all of which are learnable parameters. Initial features are assigned through initialization strategies such as Xavier / Kaiming;

[0059] The input to the routing module is the single-segment feature vector of each channel, i.e., channel I. Q channel IQ combination channel .

[0060] S22. Calculate the matching score between the signal segment and the expert. The fit is calculated using the routing module, and dynamic noise is injected. First, the single-segment features of each channel are mapped to dimension D using a channel-specific MLP (two-layer linear transformation + ReLU), while simultaneously mapping the adjacency matrix... The initial matching score is obtained by vectorizing and projecting the vectorized data onto dimension D, and then performing the inner product of the two. ,in , This is the characteristic mapping function of the adjacency matrix.

[0061] To improve the robustness of the routing, trainable Gaussian noise needs to be added to the final score:

[0062] ;

[0063] in, Standard Gaussian noise, The non-negative learnable scale parameter. This is the final matching score.

[0064] S23. Select the optimal adjacency matrix. Based on the noisy matching score, select the optimal expert adjacency matrix for each signal segment: For each signal segment in each channel, select the expert index with the highest matching score from the expert database through the argmax operation: This forms a routing index matrix: I channel Q channel IQ combination channel ,element .

[0065] S24. Generate batch adjacency matrices. Select the corresponding adjacency matrix from the expert database based on the routing index and adjust its dimensions: For each channel, select the adjacency matrix from the expert database based on the index matrix. The middle index retrieves the batch-specific adjacency matrix: The dimensions need to be adapted to the GNN input: the I / Q single-channel adjacency matrix is... , The IQ combination channel is After dimensional rearrangement, it corresponds one-to-one with the corresponding signal segment in the batch and segment dimensions, providing dynamic graph structure input for subsequent GNN blocks.

[0066] In this embodiment, the specific operation process is as follows: The expert database initializes num_experts trainable adjacency matrices, including dense, sparse, local (1 / 2, 1 / 4, 1 / 8 scale) and random types. Each individual matrix has a shape of (32, 32), and the overall shape is (num_experts, 32, 32). Three NoisyRouters process x_i (input dimension 32), x_q (input dimension 32), and x_iq (input dimension 64) respectively, calculating logits and then taking the argmax to obtain the expert index (all with a shape of (176000, 4)). Figure 2 As shown, the adjacency matrix is ​​matched from the expert database based on the index to obtain the batch adjacency matrix of the three channels (all of which are (176000,4,32,32)).

[0067] Furthermore, step S3 includes the following:

[0068] S31. Extract local features using an intra-segment GNN. Input each channel signal segment and its corresponding batch adjacency matrix into a dedicated GNN block to capture the structured relationships between nodes within the segment. Specifically: Channel I signal segment Adjacency Matrix Input GNN block Output features Similarly, the Q channel signal segment Adjacency Matrix IQ combined channel signal segment Adjacency Matrix Input the corresponding GNN blocks respectively and Output features and Where H is the hidden layer dimension of the GNN output, the above process can be uniformly represented as: .

[0069] S32. Feature Reshaping and Cross-Channel Fusion. The local features output by the GNN are dimensionally adjusted and aggregated to form segment-level feature vectors. First, the features of each channel are reshaped into [batch size × number of segments × number of nodes within a segment × hidden dimension], i.e. , (Note: The node dimension of the IQ combined channel has been adapted to SN after passing through GNN); then, the three-channel features are concatenated along the channel dimension to obtain... Finally, by performing mean pooling on the nodes within each segment, segment-level features are output. ,Right now ,in This represents the fusion feature of the nth segment of the bth sample.

[0070] S33. Modeling temporal dependencies using inter-segment RNNs. Segment-level feature sequences are input into a two-layer stacked RNN module to capture long-term temporal correlations between segments. The first layer of the RNN... As input, output intermediate time series features The second layer RNN pairs Further processing involves taking the hidden state at the last time step as the global temporal feature. ,Right now , ,in, This indicates that the last element of the sequence is extracted.

[0071] S34. The classifier outputs the final result. Global temporal features are input into a dataset-specific classifier, mapped to the class space to complete signal recognition. The classification process is defined as follows:

[0072] ;

[0073] Where D represents the target dataset, For the corresponding dataset, a linear classifier, Let be the class prediction vector for the b-th sample, and C be the number of classes in the dataset. Finally, the class probability distribution is obtained through the softmax function.

[0074] In this embodiment, the specific operation process is as follows: x_i, x_q, and x_iq are reshaped into (176000×4,32,1), (176000×4,32,1), and (176000×4,32,2), respectively, and the adjacency matrix is ​​simultaneously reshaped into (176000×4,32,32). Three GNNBlocks (containing gnn_layers layer graph convolutions) process the three-channel data respectively, mapping them to the hidden dimension hidden_channels=64, and the output feature shape is (176000×4,32,64). After reshaping, the three-channel features are concatenated to obtain (176000,4,32,192), and the node dimension is compressed by mean(dim=2) to output the segment-level fused feature (176000,4,192).

[0075] Furthermore, step S4 includes the following:

[0076] S41. Load and shard multiple datasets. A distributed data parallelism strategy is adopted for multi-task scenarios, generating a global index sequence based on a fixed random seed to ensure deterministic shuffling. Data sets with insufficient length are padded (short padding reuses preceding data, long padding is completed by repeating the index), and then equal-length data blocks are obtained by subsampling according to process number. The output signal and label are [batch size × sampling length × number of channels], providing structured input for parallel training.

[0077] S42. Subsequent learning uses SNR (Simultaneous Range) samples. Initially, only high SNR samples are used for training, gradually expanding to low SNR samples as the process progresses to reduce the impact of noise. Multiple dataset tasks process different outputs through independent classifier heads.

[0078] S43, Optimizer and Learning Rate Scheduling Configuration. Cosine annealing scheduling with preheating is used in the initial stage. The learning rate increases linearly to the baseline value, i.e. Then it enters the annealing stage. It supports hot restart to enhance model generalization. The overall learning rate decay formula is as follows:

[0079] ;

[0080] S44. Model Training and Evaluation Loop. During training, batch data is forward-propagated to obtain prediction results. The total loss is the weighted sum of the cross-entropy losses of each dataset:

[0081] ;

[0082] in, For dataset weights, Cross-entropy loss is used to update model parameters (including GNN weights, RNN parameters, etc.) in reverse. Evaluation employs leave-one-out or cross-validation, calculating metrics such as accuracy on various datasets, and analyzing the adaptability of different expert graph structures to signal patterns, providing a basis for model optimization.

[0083] In this embodiment, the specific operation process is as follows: a two-layer RNN processes segment-level features. The first layer, RNN1 (input 192, hidden dimension 64), outputs (176000, 4, 64); the second layer, RNN2, takes the hidden state at the last time step to obtain the global feature (176000, 64). The multi-task classifier calls the linear layer of the corresponding dataset, outputting 11-dimensional logits(176000, 11) (corresponding to 11 modulation types). Training uses cross-entropy loss for optimization, and testing uses the argmax of logits to obtain the prediction result, with accuracy used to evaluate performance.

[0084] The above describes the specific implementation process of the DTSG_MoA model in this invention for modulation identification tasks on the RML2016.10a dataset. This invention combines a graph neural network (GNN) with a hybrid expert (MoE) mechanism to transform a one-dimensional time-series signal into a graph structure. It utilizes a routing module to dynamically select appropriate expert adjacency matrices for different signal segments, extracts spatial features through intra-segment GNNs, and captures temporal dependencies through inter-segment RNNs, fully exploiting the complex modulation characteristics of the signal. This allows it to maintain stable identification performance even in multi-dataset and complex channel environments, providing an efficient solution for radio signal modulation identification tasks.

[0085] On the other hand, a signal modulation recognition method based on GNN and hybrid expert architecture is provided, including an IQ signal acquisition and channel separation module, a graph structure adaptation module, a feature extraction and fusion module, and a modulation recognition module; wherein,

[0086] The IQ signal acquisition and channel separation module is used to acquire IQ signals, normalize the IQ signals and divide them into equal-length segments, and then separate the segmented signals into I channels, Q channels and IQ combined channels.

[0087] The graph structure adaptation module is used to select the most suitable adjacency matrix from the expert library for each signal segment through a routing mechanism, so as to realize the dynamic adaptation of the signal segment and the graph structure.

[0088] The feature extraction and fusion module is used to feed the signal segments of each channel and the selected adjacency matrix into the corresponding graph neural network to extract local features, and to perform feature reshaping and cross-channel fusion on the local features to form segment-level fused features.

[0089] The modulation recognition module is used to capture the inter-segment temporal dependence of segment-level fusion features using a recurrent neural network, and finally outputs the signal modulation type prediction result through a classifier.

[0090] It also includes an evaluation module for evaluating the accuracy of signal modulation type prediction results and analyzing the adaptability of different expert graph structures to signal modes.

[0091] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A signal modulation recognition method based on GNN and hybrid expert architecture, characterized in that, The specific steps include the following: S1. Obtain the IQ signal, normalize the IQ signal and then divide it into equal-length segments. The segmented signal is then separated into I channel, Q channel and IQ combined channel. S2. The routing mechanism selects the most suitable adjacency matrix from the expert database for each signal segment, thereby achieving dynamic adaptation between the signal segment and the graph structure. S3. Input each channel signal segment and the selected adjacency matrix into the corresponding graph neural network to extract local features, and perform feature reshaping and cross-channel fusion on the local features to form segment-level fusion features; S4. Use a recurrent neural network to capture the inter-segment temporal dependence of the segment-level fusion features, and finally output the signal modulation type prediction result through a classifier.

2. The signal modulation recognition method based on GNN and hybrid expert architecture according to claim 1, characterized in that, The normalization processing on the IQ signal comprises removing mean value and standardization to obtain a preprocessed signal, and the preprocessed signal is cut into equal-length small segments by using non-overlapping sliding windows, and the segmentation process satisfies: ; wherein, SN is a preset number of intra-segment nodes, NS is a number of segments, L is a single-sample signal sampling length, and a sliding step length is consistent with the preset number of intra-segment nodes.

3. The signal modulation recognition method based on GNN and hybrid expert architecture according to claim 2, characterized in that, The normalization process is as follows: ； wherein, is the element of the original signal at the bth sample, the lth sampling point, the cth channel; is the mean of the channel; is the standard deviation of the channel; is a constant to avoid division by zero.

4. The signal modulation recognition method based on GNN and hybrid expert architecture according to claim 1, characterized in that, By calculating the matching score between each channel signal segment and the adjacency matrix in the expert database, the most suitable graph structure is selected for each signal segment. The specific steps are as follows: For each channel's single-segment feature, it is mapped to dimension D through a channel-specific multilayer perceptron. At the same time, the adjacency matrix is ​​projected to dimension D through vectorization and a linear layer. The inner product of the two is used to obtain the initial matching score. Gaussian noise is added to the initial matching score to obtain the final matching score. Based on the final matching score, for each signal segment of each channel, the expert index with the highest matching score is selected from the expert library through the argmax operation to form a routing index matrix; Based on the routing index matrix, a batch-specific adjacency matrix is ​​indexed from the expert database. After dimensional rearrangement, it corresponds one-to-one with the corresponding signal segment in the batch and segment dimensions, providing dynamic graph structure input for the subsequent graph neural network.

5. The signal modulation recognition method based on GNN and hybrid expert architecture according to claim 1, characterized in that, The specific steps for reshaping and cross-channel fusion of the local features to form segment-level fused features are as follows: First, the characteristics of each channel are reshaped into a structured representation of: batch size × number of segments × number of nodes within a segment × hidden dimension; Subsequently, the channel features of the I channel, Q channel, and IQ combined channel are spliced ​​together at the channel dimension to obtain segment-level features; Finally, the segment-level features are output by means pooling the nodes within the segment.

6. The signal modulation recognition method based on GNN and hybrid expert architecture according to claim 1, characterized in that, The segment-level fusion features are input into a two-layer stacked recurrent neural network to capture long-term temporal correlations between segments and obtain global temporal features.

7. The signal modulation recognition method based on GNN and hybrid expert architecture according to claim 6, characterized in that, The global temporal features are input into a dataset-specific classifier and mapped to a category space to complete signal recognition; the classification process is defined as follows: ； Where D represents the target dataset, For the corresponding dataset, a linear classifier, Let be the class prediction vector for the b-th sample, and C be the number of classes in the dataset. Finally, the class probability distribution is obtained through the softmax function.

8. The signal modulation recognition method based on GNN and hybrid expert architecture according to claim 1, characterized in that, After predicting the modulation type of the output signal, the method also includes evaluating the accuracy of the prediction results using leave-one-out or cross-validation, and analyzing the adaptability of different expert graph structures to the signal mode.

9. A signal modulation recognition system based on GNN and hybrid expert architecture, characterized in that, It includes an IQ signal acquisition and channel separation module, a graph structure adaptation module, a feature extraction and fusion module, and a modulation recognition module; among which, The IQ signal acquisition and channel separation module is used to acquire IQ signals, normalize the IQ signals and divide them into equal-length segments, and then separate the segmented signals according to I channel, Q channel and IQ combined channel; The graph structure adaptation module is used to select the most suitable adjacency matrix from the expert database for each signal segment through a routing mechanism, thereby realizing the dynamic adaptation of the signal segment and the graph structure. The feature extraction and fusion module is used to input each channel signal segment and the selected adjacency matrix into the corresponding graph neural network to extract local features, and to perform feature reshaping and cross-channel fusion on the local features to form segment-level fused features. The modulation recognition module is used to capture the inter-segment temporal dependence of the segment-level fusion features using a recurrent neural network, and finally outputs the signal modulation type prediction result through a classifier.

10. A signal modulation recognition system based on GNN and hybrid expert architecture according to claim 9, characterized in that, It also includes an evaluation module for evaluating the accuracy of the signal modulation type prediction results and analyzing the adaptability of different expert graph structures to the signal mode.

Images