A Communication Signal Modulation Classification Method Based on Aggregate Residual Transform Network
By introducing a coordinate attention mechanism and a convolutional shortcut connection module into the aggregated residual transform network, the problem of insufficient recognition performance of existing models under low signal-to-noise ratio is solved, and higher modulation recognition accuracy and generalization performance are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-05
- Publication Date
- 2026-03-10
AI Technical Summary
Existing residual network models fail to effectively utilize the long-range dependencies and sequential correlations of signals in modulation recognition, and suffer from insufficient network expressive power and a large number of parameters, resulting in decreased recognition performance under low signal-to-noise ratio conditions.
An aggregated residual transform network is adopted, and a coordinate attention mechanism module is introduced to capture the long-range dependencies and sequence correlations of the signal. The network's expressive power is enhanced by a convolutional shortcut connection module, and the generalization performance is improved by combining a convolutional global average pooling module.
Higher recognition accuracy and generalization performance were achieved under low signal-to-noise ratio conditions, improving the network's expressive power and focus on features, and reducing the risk of overfitting.
Smart Images
Figure CN116628566B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application mainly relates to the field of automatic modulation classification of communication signals, and particularly relates to a communication signal modulation classification method based on an aggregated residual transform network. BACKGROUND
[0002] Automatic modulation classification (AMC) is a link between signal detection and demodulation. In the field of communication, the modulation category needs to be known to help identify abnormal signals and monitor the spectrum to rationally allocate spectrum resources, and radar signal identification and other applications.
[0003] Traditional modulation recognition technology can be roughly divided into likelihood estimation-based recognition methods and feature extraction-based recognition methods, but both have shortcomings. The former needs to calculate the likelihood function of the received signal, and according to the Bayesian minimum error criterion, the likelihood ratio is compared with the preset threshold to make a decision. The likelihood-based method has high computational complexity and high requirements for prior knowledge. The feature-based recognition method extracts expert features from the received signal, and then determines the modulation type of the signal through a classifier. In extracting features, it is greatly influenced by expert experience, and it is difficult to meet the recognition needs under adverse factors such as noise interference and signal fading.
[0004] With the development of new generation communication technology, deep learning models that extract signal features in an end-to-end manner are gradually applied to modulation recognition. There are several models that use deep learning methods to identify modulation types. First, the AMC model based on convolutional neural network (CNN) uses the excellent spatial feature extraction capability of CNN to achieve better recognition results than methods that rely on expert experience to extract features in some scenarios. With the increase of network layers, residual structure is added to reduce the gradient vanishing during back propagation, but blindly deepening or widening the network will increase the parameter complexity and network design difficulty, which may reduce the performance gain of the existing model and reduce the signal recognition performance under low signal-to-noise ratio conditions. Second, due to the time sequence detail features of the signal and the dependence between adjacent sequence data, the AMC network based on recurrent neural network (RNN) also appears. The time sequence characteristics of the signal are used to learn the original signal features, time domain amplitude and phase information, etc. However, since the RNN model only focuses on the time sequence characteristics of the signal, the performance decreases under low signal-to-noise ratio conditions. Third, the AMC hybrid model that focuses on both spatial and temporal features of the signal appears, which contains a deep learning model with convolutional and long short-term memory (LSTM) layers. The related modules of the convolutional neural network with the recurrent neural network improve the classification accuracy and can extract spatial and temporal features, but the calculation cost is greatly increased, more memory is consumed, and more time is spent, which is not conducive to deploying the model on mobile devices. In addition, AMC models can also be divided according to the input form of the data. One class is to input the original I / Q signal data as input and directly process the original I / Q signal. The other class is to input the preprocessed signal into the network to extract features.
[0005] At present, the AMC model based on residual network does not significantly enhance the expression ability of the network, and there is still a lot of room for improvement. The reasons are mainly as follows: (1) The attention mechanism module used at present mainly focuses on local relationships, ignoring the long-range dependence and sequence correlation of signals as sequence data. (2) Only the identity shortcut in the original residual structure is used, and the full extraction of shallow features and the improvement of network expression ability are ignored. (3) The parameter amount of the classification module is large, and the association between the feature map and the category is not close enough.
[0006] Therefore, the present application proposes a modulation classification method based on an aggregated residual transform network to solve the above problems. SUMMARY
[0007] The purpose of this invention is to provide a modulation classification method based on aggregated residual transform network to solve the problems existing in the above-mentioned residual network model. This invention takes I / Q signal (I / Q signal refers to in-phase orthogonal signal, which is obtained by splitting the actual communication signal into two paths, one multiplied with the carrier and the other multiplied with the carrier signal shifted by 90°, thus obtaining two signals with a 90° phase difference) as input. Based on the long-range dependency relationship and the correlation between the sequence, a coordinate attention mechanism module is introduced to extract effective features from the temporal and spatial perspectives of the signal. A convolutional shortcut connection module is added to improve the network's expressive power. A convolutional global average pooling classification module is used to obtain global information and improve generalization ability, achieving higher accuracy and generalization performance in low signal-to-noise ratio environments.
[0008] The technical solution adopted in this invention is:
[0009] A communication signal modulation classification method based on aggregated residual transform network is characterized in that the aggregated residual transform network modulation classification model mainly consists of an aggregated residual group convolution module, a coordinate attention mechanism module, a convolutional shortcut connection module, and a convolutional global average pooling classification module for final classification, including the following steps:
[0010] S1: Sample the captured communication signal, where the standard deviation of the sampling rate offset is set to 0.01Hz, the maximum sampling rate offset is 50Hz, the standard deviation of the carrier frequency offset is 0.01Hz, the maximum carrier frequency offset is 500Hz, the number of sine waves used in frequency selective fading is 8, the sampling rate is set to 200kHz, the delay is set to [0.0, 0.9, 1.7], the amplitude corresponding to each delay time is [1, 0.8, 0.3], and each signal is represented as a dual-path in-phase orthogonal signal with 128 sampling points (i.e., in-phase orthogonal signal, with a data dimension of 128*2), forming an IQ dual-path time domain signal. Since the input feature maps of residual networks and coordinate attention mechanisms need to have three dimensions—length, width, and channels—while the original IQ dual-channel time-domain signals only contain data with two dimensions—length and width—a channel dimension is added to each IQ dual-channel time-domain signal to meet the network's input requirements. This makes the input dimension of each data point (128, 2, 1), and the signal is divided into a training set and a validation set with true class labels. During training, the validation set does not participate in the gradient descent process, but it is used to test the model's performance and adjust the parameters. The signal to be predicted is unlabeled.
[0011] S2: Based on step S1, the sample data in the training set is preprocessed with zero padding and two-dimensional convolution operations, and then input into the aggregated residual group convolution module for feature extraction. The aggregated residual group convolution module contains 16 identical convolution structure branches. The input data is first split into 16 low-dimensional tensors through 1*1 convolution, and then 3*3 convolution operations are performed in parallel on each branch to transform the data dimension. Finally, 1*1 convolution is used again to increase the dimension and restore it to the same channel dimension as the input. Then, these 16 branches are merged as the output result of the aggregated residual group convolution module.
[0012] S3: Based on step S2, the feature data obtained after the aggregation residual group convolution module in step S2 is sent to the coordinate attention mechanism module. After the two specific steps of coordinate information embedding and coordinate attention generation, feature data with embedded coordinate attention weights will be output, capturing the long-range dependency relationship, inter-channel relationship and location information of the feature space of the sequence data, and strengthening the attention to effective features.
[0013] S4: Based on steps S2 and S3, the feature data obtained after passing through the convolution module of aggregated residual groups in step S2 is sent to the convolution shortcut connection (as a cross-layer fast connection) module, and the feature data with coordinate attention weights output by step S3 is added to the result output by the convolution shortcut connection. The convolution shortcut connection is composed of two-dimensional convolutions with a kernel size of 1*1 and a stride of 1.
[0014] S5: Stack the modules from steps S2 to S4 four times. These four layers are processed sequentially, with the first layer processed first and the fourth layer processed last. The basic structure of these four layers of aggregated residual group convolutional modules remains unchanged, except that the number of convolutional kernels in the group convolutional modules increases from the first to the fourth layer. That is, the number of convolutional kernels in the aggregated residual group convolutional module in the first layer is the fewest, and then increases layer by layer, with the number of convolutional kernels in the aggregated residual group convolutional module in the fourth layer being the most (the number of convolutional kernels is set to 32 in the first layer, 64 in the second layer, 128 in the third layer, and 256 in the fourth layer). This four-layer structure constitutes... Four feature extraction units are constructed to fully extract and fuse the shallow and deep features of the data. The input data is then processed sequentially by the above four feature extraction modules, and the final output (feature map) is sent to the convolutional global average pooling classification module. The length dimension (H dimension) and width dimension (W dimension) of the feature map are multiplied and merged. After transforming the data shape, it is input into a one-dimensional convolution with a kernel size of 3 to reduce the dimension. Then, it passes through a global average pooling layer to prevent overfitting. The output is converted into modulation classification probability through the Softmax activation function, and the highest probability is taken as the signal modulation category recognized by the model.
[0015] S6: The training set is trained using the aggregated residual transform network constructed in steps S2 to S5 and optimized using the validation set. The classification cross-entropy loss function and Adam optimizer are used. After training for a certain number of rounds, the optimal weight file is obtained. Then, the aggregated residual transform network optimized by this training is used to predict and classify the test set signal to be identified. The output is the predicted signal modulation category.
[0016] The classification cross-entropy loss function is expressed as: in It is in the form of an N-dimensional vector (N is the number of categories), which refers to the probability that the prediction of the aggregated residual transform network output is each modulation category, and y is the true category that has been labeled after one-hot transformation.
[0017] In the application phase, for a communication signal A to be identified, the in-phase quadrature signal corresponding to the communication signal A is generated and the optimized aggregate residual transform network is used for prediction and classification, and the signal modulation category corresponding to the communication signal A is output.
[0018] The present invention also provides a server, characterized in that it includes a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing the steps of the above-described method.
[0019] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the above-described method.
[0020] The beneficial effects of this invention are:
[0021] (1) This invention realizes the automatic classification of modulation modes of communication signals. The proposed modulation classification method based on the aggregation residual transform network adds a convolution shortcut connection module, which enhances the expressive power of the network while preventing gradient vanishing, and integrates and extracts shallow and deep features.
[0022] (2) This invention also takes into account the sequential nature of signals and introduces a coordinate attention mechanism module to capture the long-range dependencies and correlations of sequential data, the relationships between channels, and the location information of the feature space, thereby strengthening the focus on effective features. To prevent overfitting, the ordinary fully connected layer is replaced with a convolutional global average pooling classification module, which strengthens the connection and mapping between feature maps and modulation categories, obtains global information, and improves generalization ability.
[0023] (3) Through extensive experimental verification on standard datasets, it is shown that the method proposed in this invention can achieve higher recognition accuracy and generalization performance when facing complex scenarios with low signal-to-noise ratio. Attached Figure Description
[0024] Figure 1 This is a flowchart of an automatic modulation classification method based on aggregated residual transform networks.
[0025] Figure 2 This is a diagram of the architecture of an automatic modulation classification model based on an aggregated residual transform network.
[0026] Figure 3 The graph shows the changes in recognition accuracy of different modulation recognition network models under different signal-to-noise ratios. Detailed Implementation
[0027] The modulation classification method based on aggregated residual transform network provided by the present invention will be described in detail below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can be arranged and designed in various different configurations. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.
[0028] like Figure 1 The diagram shows a flowchart of the training and prediction process using the automatic modulation classification method based on aggregated residual transform network proposed in this invention. Figure 2The diagram illustrates an automatic modulation recognition model architecture based on an aggregated residual transform network. This model primarily comprises an aggregated residual group convolution module, a coordinate attention mechanism, convolutional shortcut connections, and a convolutional global average pooling classification module. The model input layer sets the channel dimension for the sampled I / Q in-phase orthogonal dual-channel time-domain signal with dimension (128,2), making the input dimension of each signal sample (128,2,1) the input dimension. First, preprocessing is performed. The input I / Q sequence data undergoes zero-padding (3*3) to preserve edge information before being fed into a Conv2D convolutional layer (consisting of 16 convolutional kernels of size 7*7). Preprocessing is then completed again through zero-padding and max pooling with a pooling window size of 3 and a stride of 2. Next, the preprocessed data is input into the aggregated residual group convolutional module (which embeds 16 identical convolutional structure branches in parallel) for feature extraction. This includes a first layer of 1*1 2D convolutional dimensionality reduction, a second layer of 3*3 2D grouped convolutional structures that perform convolutional transformations on the low-dimensional embedded parts of the first layer's output, and a third layer of 2D convolutional dimensionality increase. The data is then merged to ensure the output maintains the same data dimension as the input. The data processed by the aggregated residual group convolutional module is then sent to the coordinate attention mechanism module to capture features more important for modulation recognition. Weighted feature maps are obtained using coordinate information embedding and coordinate attention generation. Next, the feature data output from the group convolutional module is fed into the convolutional shortcut connection module for further feature extraction, enhancing the network's representation of features. The output of this module is then added to the weighted feature map output by the coordinate attention mechanism module. This feature extraction module is stacked four times to form a four-layer feature extraction unit, with the number of convolutional kernels doubling in each layer. Therefore, the number of output channels after each layer becomes 32, 64, 128, and 256, respectively. Finally, the classification module multiplies and merges the length and width dimensions of the feature map output after the four-layer feature extraction module, changing it to a data dimension of (105, 256). Then, a one-dimensional convolution (composed of convolution kernels with a kernel size of 3*1 and the number of output channels equal to the total number of modulation categories) is used to reduce the feature map to the specified dimension. Then, the global average pooling module is used to obtain global information and reduce overfitting. Finally, the Softmax activation function is used to normalize the multi-class output and the maximum probability value is used as the classification result.
[0029] The aggregated residual group convolution module adds zero padding to the proposed model architecture to prevent edge information loss, and incorporates one-dimensional convolution and max pooling operations to preprocess the input data, which is beneficial for subsequent feature extraction. The input data enters the aggregated residual group convolution module, which stacks 16 identical topological branches in parallel. These 16 topological structures can be further subdivided into three sub-layers. The first layer uses 1*1 two-dimensional convolution to reduce the dimensionality, dividing the input into 16 low-dimensional embeddings. The second layer uses a 3*3 two-dimensional convolution structure with grouped convolutions to perform convolution transformations on the low-dimensional embeddings output from the first layer. The third layer increases the dimensionality of the concatenation results from the previous grouped convolutions, and then merges these 16 branches as the output of the group convolution module.
[0030] The coordinate attention mechanism module sends the feature data output from the aggregated residual group convolution module to the coordinate attention mechanism module, enabling it to more accurately focus on effective features related to modulation recognition, enhance useful features, and suppress interference information. The coordinate attention mechanism consists of two steps: coordinate information embedding and coordinate attention generation. In the coordinate information embedding step, global average pooling is performed along both the horizontal and vertical directions, aggregating the global feature map information extracted by the aggregated residual group convolution module into two feature maps that perceive different directions, respectively. In the coordinate attention generation step, the two feature maps with different perceptual information are first concatenated to obtain a single feature map. This map is then processed through a 1x1 convolution transformation, batch normalization, and non-linear activation to generate an intermediate feature map with horizontal and vertical positional information. This intermediate feature map is then transformed into a coordinate attention map with the same number of channels as the input through a 1x1 convolution transformation and a sigmoid activation function. Multiplying the coordinate attention map with the input feature map yields the feature data with embedded coordinate attention weights.
[0031] Convolutional Shortcut Connection Module: To further extract shallow features and enhance their expressive power, the feature data output from the aggregated residual group convolutional module is sent to the convolutional shortcut connection module. This module adds more weight parameters to express more detailed information. The convolutional shortcut connection module consists of two-dimensional convolutions with a kernel size of 1*1 and a stride of 1. It achieves the fusion of extracted shallow and deep signal features and adds its output to the weighted feature map output by the coordinate attention mechanism module, fully utilizing the feature information extracted from each layer. This not only covers the original solution space, avoiding network degradation, but also introduces more parameters that can be optimized, improving the network's expressive power.
[0032] The convolutional global average pooling classification module: This module stacks the aforementioned feature extraction module four times to form a four-layer feature extraction unit. The input data receives sufficient feature extraction, and the output feature map is fed into the classification module. First, the length and width dimensions of the feature map are multiplied and merged to adjust the dimensions suitable for one-dimensional convolution operations. Then, a one-dimensional convolution with a kernel size of 3 is used to identify local patterns in the sequence data, and the number of output feature channels is transformed into the total number of modulation categories. Next, global average pooling is used for dimensionality reduction, effectively suppressing overfitting and strengthening the correlation between feature maps and categories. Finally, the Softmax activation function is used to obtain the probability that the I / Q signal belongs to each modulation type.
[0033] Specific implementation examples:
[0034] This paper takes the radio signal standard dataset RML2016.10a as the research object, compares the automatic modulation classification method proposed in this invention with other models, analyzes the modulation recognition accuracy of different models under different signal-to-noise ratios, and experimental results show that the proposed method improves the classification accuracy and generalization.
[0035] The dataset used in this case study was generated by simulating the propagation characteristics of signals in harsh environments, taking into account factors such as center frequency offset, multipath fading, and Gaussian white noise. The signal-to-noise ratio (SNR) of the generated signals ranges from -20dB to 18dB, with intervals of 2dB. This dataset contains 11 different modulation categories, including 8 digital modulation types (BPSK, 8PSK, CPFSK, GFSK, PAM4, QAM16, QAM64, QPSK) and 3 analog modulation types (AM-DSB, AM-SSB, WBFM), totaling 220,000 samples. Each sample is represented by a dual-channel I / Q time-domain signal with 128 sampling points. Signals with different SNRs and modulation types were all shuffled and randomly selected in an 8:1:1 ratio as the training, validation, and test sets for the experiment.
[0036] All training and testing processes used the same computer hardware configuration (CPU: Intel i7-11800, GPU: GeForce RTX 3080), and all experiments were implemented using Python 3.8 and a Keras GPU backed by Tensorflow-gpu 2.8.1. The proposed network was trained and optimized using the cross-entropy loss function and the Adam optimizer, trained from scratch over 300 iterations with a batch size of 200 and an initial learning rate of 0.001. The learning rate was dynamically updated by monitoring the validation set loss; if the validation set loss did not decrease after 5 training iterations, the learning rate was reduced. Early stopping was also employed to prevent overfitting; training was terminated early if the validation set loss did not decrease after 10 iterations, resulting in better generalization ability of the model.
[0037] To verify the recognition capability and generalization of the method proposed in this invention, this case study will compare it with other AMC networks, including MCLDNN, LSTM, DAE, PET-CGDNN, ResNet, IC-AMCNET, and CGDNET networks based on this dataset. Figure 3 As shown, the method proposed in this invention achieves higher classification accuracy than other models. Table 1 presents a performance comparison of the proposed method and other models in terms of specific classification accuracy and inference time. In terms of classification accuracy, the proposed method achieves a higher global average recognition accuracy of approximately 63.83% under different signal-to-noise ratios (SNRs), exceeding the state-of-the-art MCLDNN by approximately 1.49%. Furthermore, the highest recognition accuracy is approximately 94.45%, exceeding the state-of-the-art MCLDNN by approximately 0.63%. In medium-to-high SNR (10dB–18dB) and low SNR (0dB–8dB) environments, the average accuracy of the proposed method is approximately 93.53% and 92.47%, respectively, both the highest. This indicates that the proposed model has excellent low SNR adaptability and robustness. When noise interference is more severe, low SNR can make it difficult to extract signal features, thus seriously affecting modulation recognition results. However, the proposed method still outperforms other networks in the SNR range of -20dB to 0dB. Even at low signal-to-noise ratios, the proposed method still extracts effective features, demonstrating significant performance improvements. In terms of inference time, the proposed method has a shorter inference time than MCLDNN, falling within an acceptable range. Furthermore, the improvement in modulation recognition accuracy at various signal-to-noise ratios far outweighs the slight increase in inference time. Therefore, the proposed method combines high recognition accuracy with acceptable inference time, making it advantageous for signal recognition applications in real-world channel environments.
[0038] Table 1 Performance Comparison of Different Models
[0039]
[0040] In summary, the automatic modulation classification method based on aggregated residual transform network proposed in this invention achieves excellent classification performance by integrating an aggregated residual group convolution module, a coordinate attention mechanism module, a convolutional shortcut connection module, and a convolutionally fully connected classification module. Furthermore, comparative experiments were conducted to verify the performance of the proposed model, demonstrating superior classification performance and generalization ability compared to other models on standard datasets. The average recognition accuracy of the proposed method is also higher than other networks at low signal-to-noise ratios (-20dB to 0dB). Therefore, compared to existing models, the method proposed in this invention exhibits superior classification performance, generalization ability, low signal-to-noise ratio adaptability, and anti-confusion capability.
[0041] The parts of this invention not described in detail are well-known to those skilled in the art.
[0042] The above are merely preferred embodiments of the present invention. The description of the specific embodiments above is intended to illustrate the technical solution of the present invention, and the present invention is not limited to the embodiments described above. For those skilled in the art, the present invention can have various variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention are within the protection scope of the present invention.
Claims
1. A communication signal modulation classification method based on a polyphase residual transformation network, comprising the steps of: 1) sampling each modulation category of communication signal to generate a corresponding in-phase / quadrature (I / Q) signal, and then adding a channel dimension to each I / Q signal to convert it into the input data form of the polyphase residual transformation network to generate a training sample and obtain a training set; 2) training the polyphase residual transformation network using the training samples in the training set, wherein the polyphase residual transformation network comprises a plurality of feature extraction units and a convolutional global average pooling classification module, each layer of the feature extraction unit comprises a polyphase residual group convolution module, a coordinate attention mechanism module and a convolution shortcut connection module, each layer of the feature extraction unit processes the input data to generate a feature map, the method comprising: the polyphase residual group convolution module extracts features from the input data and sends the extracted feature data to the coordinate attention mechanism module to capture more important features for modulation recognition, and then generates a feature map with weights by using coordinate information embedding and coordinate attention; the feature data output by the group convolution module is sent to the convolution shortcut connection module for further feature extraction, which enhances the expression of the network to the features, and then the output result is added to the weighted feature map output by the coordinate attention mechanism module to obtain a feature map; the number of convolution kernels in the polyphase residual group convolution module in different feature extraction units is different; the feature map generated by the last layer of the feature extraction unit is sent to the convolutional global average pooling classification module to obtain the predicted signal modulation category of the training sample; then the loss value is calculated according to the predicted signal modulation category of the training sample and the real category label, and the polyphase residual transformation network is optimized according to the loss value; 3) for a communication signal A to be identified, the corresponding I / Q signal of the communication signal A is generated and input into the optimized polyphase residual transformation network for prediction classification, and the corresponding signal modulation category of the communication signal A is output.
2. The method of claim 1, wherein, The convolutional global average pooling classification module reduces the dimension by multiplying the length dimension and the width dimension of the input feature map and then inputting it into a one-dimensional convolution, and then obtains the predicted signal modulation classification probability of the training sample by passing through a global average pooling layer and a Softmax activation function, and takes the highest probability as the predicted signal modulation category of the training sample.
3. The method of claim 1, wherein, The coordinate attention mechanism module performs coordinate information embedding and coordinate attention generation on the input feature data to obtain feature data with embedded coordinate attention weights.
4. The method of claim 3, wherein, The method for obtaining the feature data embedded with the coordinate attention weight by the coordinate attention mechanism module is: firstly, the input feature data is subjected to global average pooling along the horizontal and vertical directions respectively to obtain two feature maps of different perception directions; then the obtained two feature maps of different perception directions are connected, and then subjected to 1*1 convolution transformation, batch normalization and nonlinear activation processing to generate an intermediate feature map with horizontal and vertical direction position information, the intermediate feature map is transformed into a coordinate attention map with the same channel number as the input feature data through 1*1 convolution transformation and sigmoid activation function; then the coordinate attention map is multiplied with the input feature data to obtain the feature data embedded with the coordinate attention weight.
5. The method of claim 1, wherein, The aggregation residual group convolution module includes 16 groups of the same convolution structure branches, the input data is first split into 16 groups of low-dimensional tensors through 1*1 convolution, then the 3*3 convolution operation is performed on each convolution structure branch in parallel to convert the data dimension, and finally the 1*1 convolution is used again to restore the channel dimension consistent with the input, and then the output results of the 16 convolution structure branches are combined as the output results of the aggregation residual group convolution module.
6. The method according to any one of claims 1 to 5, characterized in that, The aggregation residual transformation network includes four feature extraction units, the number of convolution kernels in the aggregation residual group convolution module in the i+1 layer feature extraction unit is twice the number of convolution kernels in the aggregation residual group convolution module in the i layer feature extraction unit.
7. The method according to any one of claims 1 to 5, characterized in that, In step 2), the in-phase and quadrature signal data in the training sample are first subjected to zero padding and two-dimensional convolution preprocessing, and then input to the aggregation residual group convolution module for feature extraction.
8. The method of any one of claims 1-5, wherein, The loss function for calculating the loss value is a classification cross-entropy loss function , wherein is a predicted signal modulation category of the training sample, is a real category label of the training sample.
9. A server, characterized by The computer program is configured to be executed by the processor, and the computer program includes instructions for executing each step of the method of any one of claims 1 to 8.
10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is configured to be executed by the processor, and the computer program includes instructions for executing each step of the method of any one of claims 1 to 8.