Lightweight modulation identification method and system based on sparse graph construction, medium, equipment and product
A lightweight modulation recognition method constructed using sparse graphs, combined with convolutional neural networks and graph neural networks, solves the problems of high computational resource consumption and high power consumption in edge devices, achieving high-precision and high-efficiency modulation recognition.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ARMY ENG UNIV OF PLA
- Filing Date
- 2026-02-11
- Publication Date
- 2026-05-15
AI Technical Summary
Existing technologies for deploying automatic modulation recognition algorithms in edge devices suffer from high computational resource consumption, large latency, and high power consumption, making it difficult to achieve fast inference and real-time recognition.
A lightweight modulation recognition method based on sparse graph construction is adopted, which combines local feature capture of convolutional neural network with global relation modeling of graph neural network. Modulation recognition is achieved through feature extraction, graph convolution processing and classifier.
It significantly improves the accuracy and computational efficiency of automatic modulation recognition in complex environments, reduces computational load and power consumption, and is suitable for resource-constrained edge devices.
Smart Images
Figure CN122053320A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of communication and artificial intelligence signal processing technology, and in particular to a lightweight modulation recognition method, system, medium, device and product based on sparse graph construction. Background Technology
[0002] With the large-scale deployment of IoT devices, low-power edge computing nodes play a core role in intelligent communication systems. Automatic modulation identification technology is of great significance for accurately identifying the modulation mode of received signals in real time in applications such as unmanned systems, wireless sensor networks, electronic warfare, and secure communications.
[0003] Traditional automatic modulation recognition methods mainly include likelihood ratio-based recognition methods and manual feature-based recognition methods. Among them, the likelihood ratio method requires a large amount of prior information and a complex calculation process, making it unsuitable for low-computing-power devices; although the manual feature method has a lower computational cost, the model has poor adaptability, and the recognition accuracy drops significantly when the communication environment and system structure change.
[0004] In recent years, deep learning has brought a new end-to-end optimization model to modulation recognition, directly extracting and classifying features from raw signals through architectures such as convolutional neural networks, recurrent neural networks, and Transformers. However, these models often have a large number of parameters and high computational complexity, making them difficult to deploy in IoT edge devices with limited power consumption and memory.
[0005] Therefore, there is an urgent need for a lightweight deep learning modulation recognition scheme that balances high accuracy and computational efficiency to achieve fast inference and real-time recognition in resource-constrained edge devices. Summary of the Invention
[0006] The purpose of this invention is to overcome the problems of high computational resource consumption, large latency, and high power consumption when deploying automatic modulation recognition algorithms in edge devices in the prior art. It provides a lightweight modulation recognition method, system, medium, device, and product based on sparse graph construction, which combines the advantages of local feature capture of convolutional neural networks with the global relationship modeling capability of graph neural networks, and significantly improves the accuracy and computational efficiency of automatic modulation recognition in complex environments.
[0007] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution:
[0008] In a first aspect, the present invention provides a lightweight modulation recognition method based on sparse graph construction, comprising:
[0009] The received wireless communication signal is preprocessed to obtain a complex domain signal;
[0010] The complex domain signal is modulated and identified using a lightweight modulation recognition model:
[0011] The complex domain signal is subjected to local feature extraction using a feature extraction module to obtain local spatiotemporal features coupled with amplitude and phase.
[0012] The local spatiotemporal features are reconstructed in dimensions to obtain reconstructed local spatiotemporal features;
[0013] Based on the reconstructed local spatiotemporal features, graph convolution processing and feature enhancement are performed using a graph convolutional network based on a multi-connection strategy to obtain enhanced features;
[0014] The enhanced features are classified using a classifier to obtain the modulation type.
[0015] Optionally, the step of using the feature extraction module to extract local features from the complex domain signal to obtain local spatiotemporal features coupled with amplitude and phase includes:
[0016] Complex features of the complex domain signal are extracted using complex-valued convolution CVConv. ;
[0017] For the complex features After normalization, activation is performed using the GELU activation layer to obtain activation features. ;
[0018] The activation features were processed using the PPW module. Feature extraction is performed to obtain features. ;
[0019] Feature extraction using complex-valued convolution CVConv Complex characteristics ;
[0020] use Each PPW module for the complex feature Feature extraction is performed to obtain local spatiotemporal features. ;in, This is the default value.
[0021] Optionally, the complex features It can be obtained through the following formula:
[0022] ,
[0023] ,
[0024] in, Represents the real part convolution kernel. Represents the imaginary part of the convolution kernel. Represents a signal in the complex field. Represents the real part of a signal in the complex field. This represents the imaginary part of a signal in the complex field;
[0025] The features It can be obtained through the following formula:
[0026] ,
[0027] in, As a preset value, Indicates activation features Total length, Indicates activation features From starting position to position Features Indicates activation features From position Features at the termination position Indicates activation features consecutive representative segments, Indicates activation features The passive paragraph, Represents partial convolutional features. This indicates a splicing operation. The convolution kernel represents the spatial convolution. The kernel of a point convolution. This represents batch normalization, and GELU represents the activation function. This represents network regularization operations.
[0028] Optionally, the graph convolutional network based on a multi-connection strategy includes Each graph convolutional layer has an identical structure, comprising a Conditional Position Encoding (CPE) module, a Connectivity Graph Module (3CGC) module, a Maximum Relative Graph Convolution (MRConv) module, and a Feedforward Neural Network (FFN). This is the default value;
[0029] The data processing flow of the graph convolutional layer includes:
[0030] The input data is embedded using a Conditional Position Encoding (CPE) module to obtain the embedded features. ;
[0031] According to the embedded features Feature connection graphs are constructed using the 3CGC connection graph module. ;
[0032] The feature map is processed using the maximum relative graph convolution module MRConv. Perform graph convolution processing to obtain graph convolution features. ;
[0033] The graph convolution features are processed using a feedforward neural network (FFN). Perform feature enhancement to obtain enhanced features. .
[0034] Optionally, the step based on the embedded features Feature connection graphs are constructed using the 3CGC connection graph module. ,include:
[0035] Initialization and the embedded features Zero matrices of the same size are used as neighbor difference features. ;
[0036] The embedded features The last row of elements is shifted to the top, and the rest of the elements are shifted downwards. Construct a downward translation matrix. ;
[0037] Calculate the translation matrix With embedded features The difference ;
[0038] The neighbor difference features and Compare elements and replace the neighbor differences with the maximum value. The element at the corresponding position;
[0039] The embedded features Shift the rightmost two columns to the left, and shift the rest of the column to the right to construct a right translation matrix. ;
[0040] Calculate the right translation matrix With embedded features The difference ;
[0041] The neighbor difference features and Compare elements and replace the neighbor differences with the maximum value. The element at the corresponding position;
[0042] The embedded features Shift the leftmost two columns to the right, and shift the rest of the column to the left to construct a left translation matrix. ;
[0043] Calculate the left translation matrix With embedded features The difference ;
[0044] The neighbor difference features and Compare elements and replace the neighbor differences with the maximum value. The element at the corresponding position;
[0045] Neighbor differences With embedded features By concatenating along the channel dimension, a feature connectivity map is obtained. .
[0046] Optionally, the embedded features It can be obtained through the following formula:
[0047] ,
[0048] in, Indicates input data, This represents depthwise separable convolution;
[0049] The graph convolution features It can be obtained through the following formula:
[0050] ,
[0051] in, and For feature index, Represents graph convolution features The One characteristic, Characteristic connection graph The One characteristic, express The set of neighborhood features Characteristic connection graph The One characteristic.
[0052] The enhanced features It can be obtained through the following formula:
[0053] ,
[0054] in, This represents batch normalization, and GELU represents the activation function. This represents convolution.
[0055] Secondly, the present invention provides a lightweight modulation recognition system based on sparse graphs, comprising:
[0056] The data preprocessing module is used to: preprocess the received wireless communication signal to obtain a complex domain signal;
[0057] The modulation recognition module is used to: perform modulation recognition on the complex domain signal using a lightweight modulation recognition model.
[0058] The complex domain signal is subjected to local feature extraction using a feature extraction module to obtain local spatiotemporal features coupled with amplitude and phase.
[0059] The local spatiotemporal features are reconstructed in dimensions to obtain reconstructed local spatiotemporal features;
[0060] Based on the reconstructed local spatiotemporal features, graph convolution processing and feature enhancement are performed using a graph convolutional network based on a multi-connection strategy to obtain enhanced features;
[0061] The enhanced features are classified using a classifier to obtain the modulation type.
[0062] Thirdly, the present invention provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the steps of the lightweight modulation recognition method based on sparse graph construction as described in any of the first aspects.
[0063] Fourthly, the present invention provides a computer device, comprising:
[0064] Memory, used to store computer instructions;
[0065] A processor for executing the computer instructions to implement the steps of the lightweight modulation recognition method based on sparse graph construction as described in any of the first aspects.
[0066] Fifthly, the present invention provides a computer program product, including computer instructions that, when executed by a processor, implement the steps of the lightweight modulation recognition method based on sparse graph construction as described in any of the first aspects.
[0067] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: A lightweight yet high-performance AMR (Automatic Modulation Recognition) model, namely the lightweight modulation recognition model LCGNet, is proposed. LCGNet effectively combines Convolutional Neural Networks (CNNs) and Graph Neural Networks (GNNs). Through a collaborative mechanism, LCGNet enhances signal processing capabilities and improves analysis accuracy and robustness. This innovative combination not only improves feature extraction efficiency but also enhances signal relationship modeling, allowing the model to perform better in complex signal processing tasks. To improve the inference speed of GNNs, a novel graph construction algorithm, 3CGC, is proposed. This algorithm optimizes the graph structure, reduces the connection complexity between nodes, and significantly reduces the computational load during graph construction. Attached Figure Description
[0068] Figure 1 This is a schematic diagram of a lightweight modulation recognition method based on sparse graph construction, provided in an embodiment of the present invention. Detailed Implementation
[0069] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments and specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations thereof. In the absence of conflict, the embodiments and technical features in the embodiments can be combined with each other.
[0070] It should be noted that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.
[0071] Example 1
[0072] This invention discloses a lightweight modulation recognition method based on sparse graph construction, with reference to... Figure 1 As shown, the specific steps include the following:
[0073] S1, perform data preprocessing on the received wireless communication signal to obtain a complex domain signal;
[0074] S2, Modulation recognition of the complex domain signal is performed using a lightweight modulation recognition model:
[0075] S2.1, The feature extraction module is used to extract local features from the complex domain signal to obtain the local spatiotemporal features of amplitude and phase coupling;
[0076] S2.2, Reconstruct the dimensions of the local spatiotemporal features to obtain the reconstructed local spatiotemporal features;
[0077] S2.3, Based on the reconstructed local spatiotemporal features, graph convolution processing and feature enhancement are performed using a graph convolutional network based on a multi-connection strategy to obtain enhanced features;
[0078] S2.4, The enhanced features are classified using a classifier to obtain the modulation type.
[0079] Specifically, in step S1, the received wireless communication signal is first preprocessed to convert the original I / Q signal into a complex domain representation suitable for feature extraction.
[0080] Most current deep automatic modulation recognition methods suffer from large parameter counts, slow inference speeds, and difficulty in deployment at edge environments. This embodiment proposes a lightweight yet high-performance AMR network model, namely the lightweight modulation recognition model LCGNet, and a novel graph construction algorithm, 3CGC. This algorithm optimizes the graph structure, reduces the connection complexity between nodes, and significantly reduces the computational load during graph construction.
[0081] In step S2.1, this embodiment uses complex-valued CNN, partial convolution, and point convolution to design a convolution-based feature extraction module to capture local features of the input data. These different convolutions collectively facilitate the signal extraction process and improve the inference speed of the model. The step of using the feature extraction module to extract local features of the complex domain signal to obtain amplitude-phase coupled local spatiotemporal features includes:
[0082] Complex features of the complex domain signal are extracted using complex-valued convolution CVConv. :
[0083] ,
[0084] ,
[0085] in, Represents the real part convolution kernel. Represents the imaginary part of the convolution kernel. Represents a signal in the complex field. Represents the real part of a signal in the complex field. This represents the imaginary part of a signal in the complex field;
[0086] For the complex features After normalization, activation is performed using the GELU activation layer to obtain activation features. GELU provides smooth nonlinear transformations, which helps alleviate the gradient vanishing problem.
[0087] The activation features were processed using the PPW module. Feature extraction is performed to obtain features. :
[0088] ,
[0089] in, As a preset value, Indicates activation features Total length, Indicates activation features From starting position to position Features Indicates activation features From position Features at the termination position Indicates activation features consecutive representative segments, Indicates activation features The passive paragraph, Represents partial convolutional features. This indicates a splicing operation. The convolution kernel represents the spatial convolution. The kernel of a point convolution. This represents batch normalization, and GELU represents the activation function. Indicates network regularization operation;
[0090] Feature extraction using complex-valued convolution CVConv Complex characteristics ;
[0091] use Each PPW module for the complex feature Feature extraction is performed to obtain local spatiotemporal features. ;in, This is the default value.
[0092] In step S2.3, the graph convolutional network based on the multi-connection strategy includes Each graph convolutional layer has an identical structure, comprising a Conditional Position Encoding (CPE) module, a Connectivity Graph Module (3CGC) module, a Maximum Relative Graph Convolution (MRConv) module, and a Feedforward Neural Network (FFN). This is the default value;
[0093] The data processing flow of the graph convolutional layer includes:
[0094] The input data is embedded using a Conditional Position Encoding (CPE) module to obtain the embedded features. :
[0095] ,
[0096] in, Indicates input data, This represents depthwise separable convolution;
[0097] According to the embedded features Feature connection graphs are constructed using the 3CGC connection graph module. ;
[0098] The feature map is processed using the maximum relative graph convolution module MRConv. Perform graph convolution processing to obtain graph convolution features. :
[0099] ,
[0100] in, and For feature index, Represents graph convolution features The One characteristic, Characteristic connection graph The One characteristic, express The set of neighborhood features Characteristic connection graph The One characteristic.
[0101] The graph convolutional features are processed using a feed-forward neural network (FFN). Perform feature enhancement to obtain enhanced features. :
[0102] ,
[0103] in, This represents batch normalization, and GELU represents the activation function. This represents convolution.
[0104] This embodiment utilizes a connection graph module to transform signal samples into graph-structured data. It calculates the adjacency matrix between nodes using a three-connection condition, and performs global relation aggregation and updating of features. Assuming the embedded features... Represented as: Treating each feature value as a node yields a set of nodes: For each node, generate three edges according to fixed rules: 1. Shift two positions to the right; 2. Shift two positions to the left (if it goes out of bounds, loop back to the beginning of the sequence); 3. Point to the node directly below it. For example, node... The three sides point to respectively , and .
[0105] According to the embedded features Feature connection graphs are constructed using the 3CGC connection graph module. ,include:
[0106] Initialization and the embedded features Zero matrices of the same size are used as neighbor difference features. ;
[0107] The embedded features The last row of elements is shifted to the top, and the rest of the elements are shifted downwards. Construct a downward translation matrix. ;
[0108] Calculate the translation matrix With embedded features The difference ;
[0109] The neighbor difference features and Compare elements and replace the neighbor differences with the maximum value. The element at the corresponding position;
[0110] The embedded features Shift the rightmost two columns to the left, and shift the rest of the column to the right to construct a right translation matrix. ;
[0111] Calculate the right translation matrix With embedded features The difference ;
[0112] The neighbor difference features and Compare elements and replace the neighbor differences with the maximum value. The element at the corresponding position;
[0113] The embedded features Shift the leftmost two columns to the right, and shift the rest of the column to the left to construct a left translation matrix. ;
[0114] Calculate the left translation matrix With embedded features The difference ;
[0115] The neighbor difference features and Compare elements and replace the neighbor differences with the maximum value. The element at the corresponding position;
[0116] Neighbor differences With embedded features By concatenating along the channel dimension, a feature connectivity map is obtained. .
[0117] To verify the effectiveness of the LCGNet proposed in this embodiment, it was compared with current classic deep learning methods and state-of-the-art algorithms, namely MCLDNN (Multi-channel Convolutional Long Short-term Deep Neural Network), LSTM2 (Long Short-term Memory 2), TLDNN (Transformer-LSTM deep neural network) and PGNet (Partial convolutional graph neural networks).
[0118] In the experimental setup, the hyperparameters of the PPW module... Set to 2, the number of layers in the graph convolutional layer. Set to 7. When calculating inference time, the batch size is 128 for the RML2016.10a and RML22 datasets, and 2048 for the RML2018.01a dataset. The output feature dimension of the CFE module on the RML2016.10a, RML22, and RML2018.01a datasets is... They were set to 36, 36, and 40 respectively.
[0119] The RML2016.10a dataset has more than one class than RML22 (with AM-SSB added), with an SNR coverage of -20 dB to 18 dB (2 dB step size), generating a total of 220,000 samples. Each SNR point contains 1,000 samples. The RML22 dataset is an important dataset for real-world scenarios, correcting errors in RML2016.10a and improving dataset generation methods. The RML22 dataset contains 10 classes of modulated signals (BPSK, QPSK, 8PSK, PAM4, QAM16, QAM64, GFSK, CPFSK, WBFM, AM-DSB), with a SNR coverage of -20 dB to 20 dB (2 dB step size), generating a total of 420,000 samples. Each SNR point contains 2,000 samples. The RML2018.01a dataset covers 24 modulation types: OOK, 4ASK, 8ASK, BPSK, QPSK, 8PSK, 16 / 32PSK, 16 / 32 / 64 / 128APSK, 16 / 32 / 64 / 128 / 256QAM, AM-SSB-WC, AM-SSB-SC, AM-DSB-SC, AM-DSB-SC, FM, GMSK, and OQPSK. The signal-to-noise ratio (SNR) ranges from -20 dB to 30 dB in 2 dB steps, for a total of 26 levels. For each modulation type, there are 4096 IQ samples per SNR level, for a total of 2,555,904 samples (24 × 26 × 4096). Each sample is a complex sequence of size 1024 × 2.
[0120] The accuracy of LCGNet's predictions on the RML22 and RML2016.10a datasets compared with other benchmark models is shown in Table 1 below, and the accuracy of LCGNet's predictions on the RML2018.01a dataset compared with other benchmark models is shown in Table 2 below.
[0121] Table 1. Task performance on the RML22 and RML2016.10a datasets.
[0122]
[0123] Table 1. Task performance on the RML2018.01a dataset.
[0124]
[0125] As shown in Tables 1 and 2, LCGNet demonstrates unique advantages across all datasets, with a particularly significant improvement on the more challenging RML2018.01a dataset. Specifically, on RML2016.10a and RML22, LCGNet's parameter count is 0.018M and 0.071M, respectively, and its FLOPs are 1.92M and 8.374M, significantly lower than TLDNN (0.245M / 7.89M) and MCLDNN (0.369M / 41.88M). Even with substantial model compression, LCGNet still achieves accuracies of 63.40% and 72.68% on the two datasets, with a particularly pronounced advantage on RML22. On RML2018.01a (longer signal), LCGNet's advantage is further amplified: the training time is only 61.67s, achieving an order-of-magnitude speedup compared to PGNet (634.90s) and LSMT2 (218.18s). This is thanks to the efficient use of computational resources by the 3CGC algorithm, especially its outstanding performance in long sequence scenarios. In terms of inference time, LCGNet significantly outperforms PGNet (0.062ms) with 0.006ms, and is on par with other lightweight methods.
[0126] In summary, LCGNet achieves both accelerated training and inference on the longer RML2018 dataset while maintaining high accuracy, fully validating the efficiency of the 3CGC algorithm in processing long sequences. Experimental results demonstrate that LCGNet exhibits superior overall performance in complex signal processing tasks, especially in long signal scenarios.
[0127] Example 2:
[0128] Based on the same inventive concept as Embodiment 1, this embodiment of the invention discloses a lightweight modulation recognition system based on sparse graph construction, comprising:
[0129] The data preprocessing module is used to: preprocess the received wireless communication signal to obtain a complex domain signal;
[0130] The modulation recognition module is used to: perform modulation recognition on the complex domain signal using a lightweight modulation recognition model.
[0131] The complex domain signal is subjected to local feature extraction using a feature extraction module to obtain local spatiotemporal features coupled with amplitude and phase.
[0132] The local spatiotemporal features are reconstructed in dimensions to obtain reconstructed local spatiotemporal features;
[0133] Based on the reconstructed local spatiotemporal features, graph convolution processing and feature enhancement are performed using a graph convolutional network based on a multi-connection strategy to obtain enhanced features;
[0134] The enhanced features are classified using a classifier to obtain the modulation type.
[0135] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 1, and will not be repeated here.
[0136] Example 3:
[0137] This embodiment provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the steps of the lightweight modulation recognition method based on sparse graph construction as described in any of the embodiments.
[0138] Example 4:
[0139] This embodiment provides a computer device, including:
[0140] Memory, used to store computer instructions;
[0141] A processor for executing the computer instructions to implement the steps of the lightweight modulation recognition method based on sparse graph construction as described in any one of Embodiment 1.
[0142] Example 5:
[0143] This embodiment provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the lightweight modulation recognition method based on sparse graph construction as described in any one of Embodiment 1.
[0144] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0145] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.
[0146] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0147] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0148] 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 lightweight modulation recognition method based on sparse graph construction, characterized in that, include: The received wireless communication signal is preprocessed to obtain a complex domain signal; The complex domain signal is modulated and identified using a lightweight modulation recognition model: The complex domain signal is subjected to local feature extraction using a feature extraction module to obtain local spatiotemporal features coupled with amplitude and phase. The local spatiotemporal features are reconstructed in dimensions to obtain reconstructed local spatiotemporal features; Based on the reconstructed local spatiotemporal features, graph convolution processing and feature enhancement are performed using a graph convolutional network based on a multi-connection strategy to obtain enhanced features; The enhanced features are classified using a classifier to obtain the modulation type.
2. The lightweight modulation recognition method based on sparse graph construction according to claim 1, characterized in that, The step of using a feature extraction module to extract local features from the complex domain signal to obtain local spatiotemporal features coupled with amplitude and phase includes: Complex features of the complex domain signal are extracted using complex-valued convolution CVConv. ; For the complex features After normalization, activation is performed using the GELU activation layer to obtain activation features. ; The activation features were processed using the PPW module. Feature extraction is performed to obtain features. ; Feature extraction using complex-valued convolution CVConv Complex characteristics ; use Each PPW module for the complex feature Feature extraction is performed to obtain local spatiotemporal features. ;in, This is the default value.
3. The lightweight modulation recognition method based on sparse graph construction according to claim 2, characterized in that, The complex feature It can be obtained through the following formula: ， ， in, Represents the real part convolution kernel. Represents the imaginary part of the convolution kernel. Represents a signal in the complex field. Represents the real part of a signal in the complex field. This represents the imaginary part of a signal in the complex field; The features It can be obtained through the following formula: ， in, As a preset value, Indicates activation features Total length, Indicates activation features From starting position to position Features Indicates activation features From position Features at the termination position Indicates activation features consecutive representative segments, Indicates activation features The passive paragraph, Represents partial convolutional features. This indicates a splicing operation. The convolution kernel represents the spatial convolution. The kernel of a point convolution. This represents batch normalization, and GELU represents the activation function. This represents network regularization operations.
4. The lightweight modulation recognition method based on sparse graph construction according to claim 1, characterized in that, The graph convolutional network based on the multi-connection strategy includes Each graph convolutional layer has an identical structure, comprising a Conditional Position Encoding (CPE) module, a Connectivity Graph Module (3CGC) module, a Maximum Relative Graph Convolution (MRConv) module, and a Feedforward Neural Network (FFN). This is the default value; The data processing flow of the graph convolutional layer includes: The input data is embedded using a Conditional Position Encoding (CPE) module to obtain the embedded features. ; According to the embedded features Feature connection graphs are constructed using the 3CGC connection graph module. ; The feature map is processed using the maximum relative graph convolution module MRConv. Perform graph convolution processing to obtain graph convolution features. ; The graph convolution features are processed using a feedforward neural network (FFN). Perform feature enhancement to obtain enhanced features. .
5. The lightweight modulation recognition method based on sparse graph construction according to claim 4, characterized in that, The based on the embedded features Feature connection graphs are constructed using the 3CGC connection graph module. ,include: Initialization and the embedded features Zero matrices of the same size are used as neighbor difference features. ; The embedded features The last row of elements is shifted to the top, and the rest of the elements are shifted downwards. Construct a downward translation matrix. ; Calculate the translation matrix With embedded features The difference ; The neighbor difference features and Compare elements and replace the neighbor differences with the maximum value. The element at the corresponding position; The embedded features Shift the rightmost two columns to the left, and shift the rest of the column to the right to construct a right translation matrix. ; Calculate the right translation matrix With embedded features The difference ; The neighbor difference features and Compare elements and replace the neighbor differences with the maximum value. The element at the corresponding position; The embedded features Shift the leftmost two columns to the right, and shift the rest of the column to the left to construct a left translation matrix. ; Calculate the left translation matrix With embedded features The difference ; The neighbor difference features and Compare elements and replace the neighbor differences with the maximum value. The element at the corresponding position; Neighbor differences With embedded features By concatenating along the channel dimension, a feature connectivity map is obtained. .
6. The lightweight modulation recognition method based on sparse graph construction according to claim 4, characterized in that, The embedded features It can be obtained through the following formula: ， in, Indicates input data, This represents depthwise separable convolution; The graph convolution features It can be obtained through the following formula: ， in, and For feature index, Represents graph convolution features The One characteristic, Characteristic connection graph The One characteristic, express The set of neighborhood features Characteristic connection graph The One characteristic. The enhanced features It can be obtained through the following formula: ， in, This represents batch normalization, and GELU represents the activation function. This represents convolution.
7. A lightweight modulation recognition system based on sparse graph construction, characterized in that, include: The data preprocessing module is used to: preprocess the received wireless communication signal to obtain a complex domain signal; The modulation recognition module is used to: perform modulation recognition on the complex domain signal using a lightweight modulation recognition model. The complex domain signal is subjected to local feature extraction using a feature extraction module to obtain local spatiotemporal features coupled with amplitude and phase. The local spatiotemporal features are reconstructed in dimensions to obtain reconstructed local spatiotemporal features; Based on the reconstructed local spatiotemporal features, graph convolution processing and feature enhancement are performed using a graph convolutional network based on a multi-connection strategy to obtain enhanced features; The enhanced features are classified using a classifier to obtain the modulation type.
8. A computer-readable storage medium storing computer instructions thereon, characterized in that, When the computer instruction is executed by the processor, it implements the steps of the lightweight modulation recognition method based on sparse graph construction as described in any one of claims 1-6.
9. A computer device, characterized in that, include: Memory, used to store computer instructions; A processor for executing the computer instructions to implement the steps of the lightweight modulation recognition method based on sparse graph construction as described in any one of claims 1-6.
10. A computer program product comprising computer instructions, characterized in that, When executed by a processor, the computer instructions implement the steps of the lightweight modulation recognition method based on sparse graph construction as described in any one of claims 1-6.