Modulation signal identification method based on variable convolution and convolution gated linear unit

Abstract

The invention relates to a modulation signal identification method based on a variable convolution and convolution gated linear unit. The method comprises the following steps: acquiring a modulation signal to be identified; inputting the modulation signal into a trained modulation signal identification network model, wherein the model comprises a stem network, a backbone network and a classification network; the stem network performs data preprocessing, preliminary feature extraction and redundancy suppression on the modulation signal to be identified to obtain a first feature; the backbone network performs deep feature extraction and adaptive screening on the first feature to obtain a second feature; wherein the backbone network comprises a plurality of basic modules and a multi-scale fusion module which are connected in sequence; each basic module comprises a deformable convolution layer and a convolution gating linear unit; and the classification network performs dimension compression and classification decision on the second feature, and outputs probability distribution of an identification result. The method can improve the recognition accuracy and robustness in a complex electromagnetic environment.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication and artificial intelligence technology, specifically relating to a method for identifying modulated signals based on variable convolution and convolution-gated linear units. Background Technology

[0002] Automatic modulation identification (EMI) is a key preprocessing technology in wireless communication systems, widely used in spectrum monitoring, cognitive radio, and electronic countermeasures. Its core task is to accurately classify the modulation scheme of received signals when the modulation parameters are unknown. Traditional identification methods are mainly divided into likelihood theory-based methods and manual feature-based methods: the former relies on strict statistical assumptions and complex likelihood ratio calculations, resulting in significant performance degradation in real-world, non-ideal environments; the latter requires expert experience to manually design features, has limited generalization ability, and is difficult to adapt to increasingly complex communication signal environments.

[0003] In real-world applications, such as dense urban wireless communication environments or complex electromagnetic countermeasures scenarios, received signals are often affected by multiple factors, including multipath fading, impulse noise, and interference from adjacent channels, exhibiting characteristics such as low signal-to-noise ratio, non-stationarity, and multi-scale aliasing. Signal features are unevenly distributed in the time-frequency domain, with local details and global structures coexisting. Traditional methods and existing deep learning methods both face problems such as decreased recognition accuracy and insufficient adaptability in these real-world physical environments.

[0004] In recent years, with the development of deep learning technology, modulation recognition methods based on neural networks have become a mainstream research direction. Convolutional neural networks, recurrent neural networks, and Transformer models have been successively introduced to automatically learn signal features in an end-to-end manner. In particular, the CNN-Transformer hybrid architecture combines, to some extent, the local perceptual advantages of convolution with the global modeling capabilities of self-attention. However, these methods still lack an active screening mechanism for key features and noise during feature fusion; convolution operations themselves easily introduce spatial and channel redundancy, leading to low computational efficiency; and single-scale convolution is difficult to effectively capture the multi-scale, non-stationary dynamic characteristics of signals in the time-frequency domain.

[0005] In summary, existing modulation recognition methods based on CNN-Transformer lack selectivity in feature fusion, have computational redundancy, and have limited adaptability to multi-scale signal features, resulting in insufficient recognition accuracy and robustness in complex electromagnetic environments. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this invention provides a method for identifying modulated signals based on variable convolution and convolution-gated linear units. The technical problem to be solved by this invention is achieved through the following technical solution: This invention provides a method for identifying modulated signals based on variable convolution and convolution-gated linear units, including: Acquire the modulation signal to be identified; The modulation signal to be identified is input into a trained modulation signal recognition network model, which includes a stem network, a trunk network, and a classification network. The stem network is used to perform data preprocessing, preliminary feature extraction, and redundancy suppression on the modulated signal to be identified, to obtain a first feature. The backbone network is used to perform deep feature extraction and adaptive filtering on the first feature to obtain the second feature; wherein, the backbone network includes several basic modules connected in sequence and a multi-scale fusion module; Each basic module includes a first normalized layer, a deformable convolutional layer, a second normalized layer, and a convolutionally gated linear unit connected in sequence; wherein, the deformable convolutional layer is used to dynamically adjust the sampling point position according to the input content to adaptively extract the structural features of the signal in the time-frequency domain; the convolutionally gated linear unit is used to generate spatially aware adaptive weights through the gated branches composed of convolutional layers to filter features; The multi-scale fusion module is used to receive the output features of the last basic module, extract features under different receptive fields through parallel multi-scale convolutional layers, and adaptively weight and fuse features of different scales using a channel attention mechanism to obtain the second feature. The classification network is used to perform dimensionality compression and classification decisions on the second feature, and outputs the probability distribution of the recognition result.

[0007] Compared with the prior art, the beneficial effects of the present invention are as follows: To address the shortcomings of existing CNN-Transformer-based modulation recognition methods, such as lack of selectivity in feature fusion, computational redundancy, and limited adaptability to multi-scale signal features, resulting in insufficient recognition accuracy and robustness in complex electromagnetic environments, this invention provides a modulation signal recognition method based on variable convolution and convolution-gated linear units. This method effectively improves recognition performance and robustness in complex electromagnetic environments through the collaborative work of the stem network, backbone network, and classification network. Specifically, this method first preprocesses and suppresses redundancy in the input modulation signal through the stem network to obtain cleaner preliminary features. Then, the backbone network performs deep processing on the preliminary features, where deformable convolutional layers dynamically adjust their receptive fields based on the local content of the input signal, adaptively focusing on and extracting discriminative structural features in the time-frequency domain. Next, convolutionally gated linear units utilize their unique gating mechanism to adaptively filter the extracted features based on spatial context, strengthening key information and suppressing irrelevant interference. Further, the multi-scale fusion module at the tail of the backbone network captures contextual information at different scales through a parallel structure and uses an attention mechanism to achieve intelligent feature fusion, thereby enhancing the model's perception and adaptation to the multi-scale characteristics of the signal. Finally, the classification network makes decisions on the fused high-level features and outputs the modulation type identification result. The method described in this invention significantly improves the adaptability of feature extraction, the selectivity of feature fusion, the ability to perceive multi-scale signals, and overall computational efficiency, thus exhibiting higher recognition accuracy and stronger robustness in complex electromagnetic environments with low signal-to-noise ratios and multiple interferences. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of data processing of the trained modulation signal recognition network model in the method proposed in this invention, provided in an embodiment of the invention; Figure 2 This is a schematic diagram of the model architecture of the stem network provided in an embodiment of the present invention; Figure 3 This is a data processing example diagram of the preprocessing module provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the model architecture of the spatial redundancy suppression unit provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the model architecture of the channel redundancy suppression unit provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the model architecture of deformable convolutional layers in each basic module provided in the embodiments of the present invention; Figure 7 This is a schematic diagram of the model architecture of the convolutional gated linear unit in each basic module provided in the embodiments of the present invention; Figure 8 This is a schematic diagram of the model architecture of the multi-scale fusion module provided in an embodiment of the present invention; Figure 9 This is a simulation comparison of the performance of various modulation signal recognition methods provided in the embodiments of the present invention and the method proposed in the present invention on the RML2016.10a dataset; Figure 10 These are simulation graphs showing the performance of various modulation signal recognition methods provided in the embodiments of the present invention and the method proposed in the present invention on the RML2018.01a dataset; Figure 11 These are simulation graphs showing the performance of various modulation signal recognition methods provided in the embodiments of the present invention and the method proposed in the present invention on the RML2016.10a dataset; Figure 12 The above are simulation diagrams showing the performance of various modulation signal recognition methods provided in the embodiments of the present invention and the method proposed in the present invention on the RML2018.01a dataset.

[0009] Explanation of icon numbers: Stem network, Basic Block, Multi-scale fusion module (SE-MSConv), Classifier network, Normalization layer (LN), Deformable convolution (DCN), Convolution gated linear unit (CGLU), Preprocessing module (PRC), Convolutional layer (Conv), Spatial redundancy suppression unit (SRSU), Channel redundancy suppression unit (CRSU), Batch normalization layer (BN), Connectivity layer (concat), Maxpool layer, Group convolution (GWC), Point convolution (PWC), Global average pooling layer (Avgpool), Interpolation layer (Interpolation), Fully connected layer (FC), Depthwise convolutional layer (DW Conv), Gaussian error linear unit (GELU), ReLU. Detailed Implementation

[0010] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0011] It should be noted that in the description of this invention, the terms "first," "second," etc., are only used to distinguish and refer to similar components or operations appearing in the specification for ease of description. They do not imply any limitation on order, quantity, or importance, nor do they suggest any necessary differences in structure or function between the mentioned components. For example, "first XXX" and "second XXX" in the following text only indicate that they appear in different locations on the network or play different roles, but their "XXX" themselves have the same structure or belong to the same type of operation.

[0012] The proposed signal recognition method based on variable convolution and convolution-gated linear unit modulation, as described in the accompanying drawings, will now be described in detail.

[0013] Figure 1 This is a schematic diagram illustrating the data processing of the trained modulation signal recognition network model in the method proposed in this invention, as provided in an embodiment of the invention. Combined with... Figure 1 Therefore, the method includes: S1: Acquire the modulation signal to be identified.

[0014] For example, radio electromagnetic wave signals in dense urban wireless communication environments or complex electromagnetic countermeasure scenarios are captured by a receiving antenna. After down-conversion and analog-to-digital conversion, a complex sampling sequence of baseband or intermediate frequency is obtained. The sequence is then orthogonally demodulated to separate the in-phase component I and the quadrature component Q, forming a two-dimensional vector. Based on the signal length, it is normalized into an I / Q signal matrix with a dimension of 2×128 or 2×1024, which serves as the direct input to the modulation signal recognition network model.

[0015] S2: Input the modulation signal to be identified into the trained modulation signal recognition network model. The trained modulation signal recognition network model includes: stem network, trunk network and classification network. The stem network is used to perform data preprocessing, preliminary feature extraction, and redundancy suppression on the modulation signal to be identified, thereby obtaining the first feature. The backbone network is used to perform deep feature extraction and adaptive filtering on the first feature to obtain the second feature; the backbone network consists of several basic modules connected in sequence and a multi-scale fusion module. Each basic module includes a first normalized layer, a deformable convolutional layer, a second normalized layer, and a convolutionally gated linear unit connected in sequence; wherein, the deformable convolutional layer is used to dynamically adjust the sampling point position according to the input content to adaptively extract the structural features of the signal in the time-frequency domain; the convolutionally gated linear unit is used to generate spatially aware adaptive weights through the gated branches composed of convolutional layers to filter features; The multi-scale fusion module receives the output features of the last basic module, extracts features from different receptive fields through parallel multi-scale convolutional layers, and adaptively weights and fuses features of different scales using a channel attention mechanism to obtain the second feature. The classification network is used to perform dimensionality compression and classification decisions on the second feature, and outputs the probability distribution of the recognition result.

[0016] Based on the structure and working principle of the modulation signal recognition network model, the data processing process of the trained modulation signal recognition network model to be recognized is described.

[0017] Figure 2This is a schematic diagram of the stem network model architecture provided in an embodiment of the present invention. For example... Figure 2 As shown, the stem network includes: a preprocessing module, a spatial redundancy suppression unit, and a channel redundancy suppression unit; wherein, the preprocessing module is used to perform dimensional unification and multi-scale initial feature extraction on the input modulated signal to be identified, and output a two-dimensional feature map with a unified spatial size; the spatial redundancy suppression unit and the channel redundancy suppression unit are used to suppress the redundant information of the input feature map in the spatial and channel dimensions in turn.

[0018] It should be noted that, in addition to the above structure, the stem network also includes: two 3×3 convolutional layers, two batch normalization (BN) layers, and one ReLU activation function layer. Specifically, the connection relationships of the structures in the stem network are (from left to right): Preprocessing Module (PRC), third 3×3 convolutional layer, Spatial Redundancy Suppression Unit (SRSU), Channel Redundancy Suppression Unit (CRSU), Batch Normalization Layer (BN), ReLU activation function layer, fourth 3×3 convolutional layer, and BN layer. Here, the two 3×3 convolutional layers are used to reduce the spatial size of the feature subsets. These two layers process the two segmented feature subsets respectively, reducing their size to one-quarter of the original input spatial size to achieve lightweight subsequent processing. The batch normalization layer and the ReLU activation function layer constitute the batch normalization and activation layer, used to perform batch normalization and nonlinear activation processing on the features output by the channel redundancy suppression unit.

[0019] It should be noted that the preprocessing module performs multi-scale initial feature extraction and dimensionality unification processing through its internal parallel convolutional kernels with different receptive fields, outputting a two-dimensional feature map with a uniform spatial size.

[0020] For example, Figure 3 This is an example diagram of data processing in the preprocessing module provided in this embodiment of the invention. For an input signal with dimensions (2, 128), a new dimension is added, resulting in (2, 1, 128). The resulting signal is then processed through three convolutions with kernel sizes of 1×3, 1×5, and 1×7, each with a stride of 1. Different kernel sizes allow the network to learn features of different ranges at the same level; smaller kernels capture local features, while larger kernels capture broader contextual information. The output sizes are (42, 1, 128), (43, 1, 128), and (43, 1, 128), respectively. The outputs are then concatenated to (128, 1, 128), and a dimension transformation is performed to obtain an output X with dimensions (1, 128, 128). The specific process is as follows: Figure 3As shown in (a) of the diagram; for an input signal with dimension (2, 1024), a dimension is added to become (2, 1, 1024), and the result still passes through three branches. The stride of each convolution is set to 2, and the 1×7 convolution is replaced with a max pooling operation with a stride of 2 to reduce computational cost. The output sizes are (171, 1, 512), (171, 1, 512), and (170, 1, 512). The outputs are then concatenated to (512, 1, 512), and a dimension transformation is performed to obtain an output with dimension (1, 512, 512). Finally, a max pooling operation with a stride of 4 is performed to obtain an output X with dimension (1, 128, 128). The specific process is as follows: Figure 3 As shown in (b) of the diagram.

[0021] Use two convolutional pairs with kernel size 3×3 and stride 2 to select features. Figure X Dimension transformation and feature extraction are performed. The first 3×3 convolutional layer transforms the dimensions of X into (32, 64, 64), thus achieving initial feature extraction. The second 3×3 convolutional layer is mainly used for downsampling, and the output dimensions become (64, 32, 32).

[0022] Figure 4 This is a schematic diagram of the model architecture of the spatial redundancy suppression unit provided in an embodiment of the present invention. Figure 4 As shown, the spatial redundancy suppression unit includes: a group normalization layer, a third product unit, a sigmoid activation layer, a first adder, a second adder, and a third connection layer. The group normalization layer is used to perform channel grouping normalization on the input weighted feature map to obtain a normalized feature map. The third product unit and the sigmoid activation layer constitute a weight generation module, which is used to calculate the importance weight of each spatial location using the normalized feature map. The first adder and the second adder constitute a feature separation module, which is used to divide the importance weight into importance weight and suppression weight according to a preset threshold, and multiply them by the original input feature map respectively to obtain a first feature subset and a second feature subset. The information richness of the first feature subset is greater than that of the second feature subset. The third connection layer concatenates the first feature subset and the second feature subset in the channel dimension to obtain the feature map after spatial redundancy suppression.

[0023] For example, in the spatial redundancy suppression unit, the feature map is passed through its group normalization layer for channel normalization, the spatial importance weight is calculated by the weight generation module, and then the feature separation module separates the features into two subsets with rich information and less information according to the threshold. Finally, the third connection layer performs segmentation, recombination and channel splicing on the two subsets to output the feature map after spatial redundancy suppression.

[0024] The formula for calculating the group normalization layer is as follows: ,in, , They are the mean and standard deviation. It is a very small positive constant. and The two weights are updated through backpropagation and gradient descent. A larger value indicates a richer feature map. Input features... Decomposed into multiple sub-features , , ..., Each sub-feature corresponds to a different channel. For each channel feature... Perform normalization to generate weights The calculation principle is as follows: Then, the normalized weights will be... Multiplying the input feature map by the normalized input feature map results in a reweighted feature map. After processing with the Sigmoid activation function, we obtain The calculation process is as follows: By selecting a threshold T , for input Process the elements. If elements greater than T are set to 1, then the importance weights are obtained. If elements greater than T are set to 0, then the suppression weights are obtained. Weighting Compared with the original input features Perform element-wise multiplication, that is: To obtain information-rich features ; suppress weights Compared with the original input features Perform element-wise multiplication, that is: Features with less information were obtained. .Will and The channels are divided equally to obtain and , and Then and The results are concatenated on the channels to obtain the output. .

[0025] Figure 5 This is a schematic diagram of the model architecture of the channel redundancy suppression unit provided in an embodiment of the present invention. Figure 5As shown, the channel redundancy suppression unit includes: a third 1×1 convolutional layer, a fourth 1×1 convolutional layer, a first group convolutional layer, a first point convolutional layer, a second point convolutional layer, a third adder, a fourth connection layer, a third global average pooling layer, and a second Softmax layer. The third and fourth 1×1 convolutional layers simultaneously perform channel dimensionality reduction on the feature maps after spatial redundancy suppression, obtaining a first channel dimensionality-reduced feature map and a second channel dimensionality-reduced feature map. The first group convolutional layer and the first point convolutional layer simultaneously perform group convolution and point-by-point convolution on the first channel dimensionality-reduced feature map, respectively extracting spatial features and restoring inter-channel information exchange, obtaining group convolutional features and first point convolutional features. Subsequently, the third adder is used to sum the features to obtain the channel information features. The process involves: 1) Convolutional layer performing pointwise convolution on the reduced-dimensional feature map of the second channel to generate second-point convolutional features; 2) Concatenating the second-point convolutional features and the feature map after spatial redundancy suppression along the channel dimension to obtain the output feature map of the fourth connection layer; 3) Global average pooling layer performing global average pooling on the channel information feature map and the output feature map of the fourth connection layer to obtain first and second global spatial information; 4) Softmax layer performing Softmax normalization on the first and second global spatial information to generate corresponding channel weights; and 5) Weighted fusion of the corresponding channel weights with the channel information feature map and the output feature map of the fourth connection layer to obtain the feature map after channel redundancy suppression.

[0026] For example, features after suppressing spatial redundancy Based on preset ratio α Divided into two parts and ,in α It's a hyperparameter, 0 < α <1. for α ×C channels, and For (1) α The input has 1×C channels, and then two 1×1 convolutional kernels are used to reduce the number of channels to half of the original number; the input... Group convolution (GWC) is performed, and 1×1 pointwise convolution (PWC) is used to compensate for the lack of inter-channel information exchange in group convolution. Then, the two sets of outputs are added together to obtain rich features. On another side road, Use 1×1 pointwise convolution to obtain shallow information, and then apply it to the channels. Additional features are obtained by splicing. For the transformed features and The global average pooling operation is performed, and the calculation process is as follows: To obtain global spatial information and ;Will and The softmax operation is performed to generate weights, and the calculation formula is as follows: Generate normalized weights for each channel. and Weighted guidance and Perform the summation and output the final result. .

[0027] at last, The second 3×3 convolutional layer further downsamples and transforms the channels, and then the output is stabilized by the final batch normalization layer to obtain the first feature, which is then sent to the subsequent backbone network for deep processing.

[0028] In one possible implementation, the backbone network comprises 12 sequentially connected basic modules, as well as a multi-scale fusion module connected to the last basic module.

[0029] Figure 6 This is a schematic diagram of the model architecture of deformable convolutional layers in each basic module provided in the embodiments of the present invention. For example... Figure 6 As shown, the deformable convolutional layer in each basic module includes: a first 3×3 convolutional layer, a first Softmax layer, a first connection layer, a first interpolation layer, and a second 3×3 convolutional layer. The first 3×3 convolutional layer receives the output features from the first normalization layer and outputs sampling point offsets Δx, Δy, and weight parameters Δm. The first Softmax layer normalizes the weight parameters Δm to obtain normalized weights. The first connection layer dynamically adjusts the sampling point coordinates of the output features from the first normalization layer based on the offsets Δx and Δy to obtain the offset sampling coordinates. The first interpolation layer uses bilinear interpolation to calculate the features at the offset sampling coordinates to obtain the interpolated features. The second 3×3 convolutional layer performs a weighted sum of the normalized weights and the interpolated features to obtain the output features of the deformable convolutional layer.

[0030] Figure 7 This is a schematic diagram of the model architecture of the convolutional gated linear unit in each basic module provided in the embodiments of the present invention. For example... Figure 7As shown, the convolutional gated linear unit in each basic module includes: a first fully connected layer, a second fully connected layer, a first deep convolutional layer, a Gaussian error linear unit, a first product unit, and a third fully connected layer. The first and second fully connected layers process the output features of the second normalized layer in parallel, generating main information flow features and gated intermediate features, respectively. The first deep convolutional layer performs spatial convolution on the gated intermediate features to incorporate local contextual information, obtaining the first convolutional features. The Gaussian error linear unit performs non-linear activation on the first convolutional features to generate gate weights. The first product unit performs element-wise multiplication of the gate weights and the main information flow features to achieve feature filtering, obtaining the filtered features. The third fully connected layer transforms the channel dimension of the filtered features to obtain the output features of the convolutional gated linear unit.

[0031] For example, the output feature X of the second normalization layer is first passed through two parallel fully connected layers, increasing the channel dimension to four times its original size: the main information flow branch generates a linear transformation result A, and the gated signal branch generates intermediate features B. The gated branch embeds a 3×3 deep convolutional layer before activation, enabling it to fuse local spatial context information when generating weights, resulting in a convolutional output C. Subsequently, the result D is obtained through a Gaussian error linear unit (GELU), whose activation function can be expressed as: ,in It is the cumulative distribution function (CDF) of the standard normal distribution. Finally, feature weighting is achieved by element-wise multiplication of A and D, and then the output X' is obtained by passing through a fully connected layer to reduce the number of channels to 1 / 4 of the previous layer.

[0032] Here, within each basic module, the input features are first stabilized by the first normalization layer, then extracted by the deformable convolutional layer, normalized by the second normalization layer, and then input into the convolutional gated linear unit for feature selection and fusion. Residual connections are used to alleviate the gradient vanishing problem.

[0033] Figure 8 This is a schematic diagram of the model architecture of the multi-scale fusion module provided in an embodiment of the present invention. For example... Figure 8As shown, the multi-scale fusion module includes: a multi-scale convolutional unit, a channel attention unit, and a feature fusion unit. The multi-scale convolutional unit, used for parallel extraction of features from different receptive fields, includes: a first 1×1 convolutional layer, a fourth 3×3 convolutional layer, a first 3×3 dilated convolutional layer, a first global average pooling layer, a second interpolation layer, a first linear rectified layer, a second linear rectified layer, a third linear rectified layer, and a fourth linear rectified layer. The first 1×1 convolutional layer and the first linear rectified layer constitute the second branch; the fourth 3×3 convolutional layer and the second linear rectified layer constitute the second branch; and the first 3×3 dilated convolutional layer and the third linear rectified layer constitute the third branch. The system consists of a fourth branch, which is composed of a first global average pooling layer, a second interpolation layer, and a fourth linear rectified layer. A channel attention unit, comprising a second connection layer and a channel attention mechanism subunit, is used. The second connection layer concatenates the output features of the four branches along the channel dimension to obtain concatenated features. The channel attention mechanism subunit processes the concatenated features to generate channel weights. A feature fusion unit, comprising a second product unit and a second 1×1 convolutional layer, is used to perform element-wise multiplication of the channel weights and concatenated features using the second product unit to achieve weighting, and to integrate the channels using the second 1×1 convolutional layer, outputting the fused features as the second feature.

[0034] Furthermore, the channel attention mechanism subunit includes: a second global average pooling layer, a fourth fully connected layer, a fifth linear rectified layer, a fifth fully connected layer, and a Sigmoid activation layer connected in sequence; the second global average pooling layer is used to compress the spatial information of the spliced ​​features to generate channel descriptors; the fourth fully connected layer is used to perform dimensionality reduction transformation on the channel descriptors to obtain dimensionality-reduced features; the fifth linear rectified layer is used to perform nonlinear transformation on the dimensionality-reduced features; the fifth fully connected layer is used to restore the feature dimension after nonlinear transformation to obtain interactive data; and the Sigmoid activation layer is used to map the interactive data to channel weights in the range of 0 to 1.

[0035] For example, this module contains four branches, each extracting feature information at different scales. Branch 1 is a 1×1 convolution, mainly used to preserve the original features; Branch 2 is a 3×3 convolution, extracting local features within a smaller range; Branch 3 is a 3×3 dilated convolution with a dilation rate of 4, extracting local features within a larger range; Branch 4 processes global features by performing global average pooling on the input feature map, then mapping it to the same number of channels as the other branches using a 1×1 convolution, and finally upsampling to the spatial size of the original feature map using bilinear interpolation. In the channel attention mechanism subunit, the input is first compressed by global average pooling. After pooling, each channel in each batch produces a single value, with an output shape of (B, C), where B is the batch size and C is the number of channels. Then, a fully connected layer reduces the number of features per channel from C to C / r, where r controls the model complexity and the number of parameters. Next, the ReLU activation function is used to increase non-linear processing capability, which can be expressed as: Then, another fully connected layer increases the number of features back from C / r to C. Finally, the output is compressed to the range [0, 1] using the Sigmoid function, which can be expressed as: This process yields the importance weights for each channel. The outputs of all branches of MSConv are concatenated along the channel dimension. The concatenated output is then processed through the weighted portion of the SE attention mechanism. The generated weights are then multiplied element-wise with the original concatenated output to apply the learned channel weights. Finally, a 1×1 convolution maps the multi-channel features back to the number of output channels.

[0036] Here, the classification network consists of a fifth 1×1 convolutional layer, a global average pooling layer, and a fully connected layer connected in sequence.

[0037] For example, firstly, the fifth 1×1 convolutional layer adjusts the number of channels of the input feature to a preset dimension. Then, a global average pooling layer compresses the spatial dimension of the feature map, transforming it into a global feature vector for each channel, thus preserving the global contextual information of the feature. Finally, a fully connected layer maps the dimension of this feature vector to the number of target classes, generating the final classification probability distribution.

[0038] The above describes the data processing process of the modulated signal recognition network model to be recognized.

[0039] In one possible implementation, the modulation signal recognition network model is trained in the following way: S1: Obtain a dataset of modulated signals containing multiple modulation types and divide it into training set, validation set and test set according to a preset ratio; S2: Construct the modulation signal recognition network model as described in claim 1, and initialize the model parameters; S3: Iteratively train the model using the training set, with 100 training rounds, a batch size of 128, and the use of the cross-entropy loss function and AdamW optimizer, along with a phased learning rate scheduling strategy for optimization. S4: After each round of training, use the validation set to evaluate the model performance and save the optimal model parameters based on the validation results; S5: Use the test set to perform the final performance test on the trained model to obtain the trained modulation signal recognition network model.

[0040] It should be noted that a warm-up learning rate scheduling strategy is used for the first 30 epochs, causing the learning rate to increase linearly from zero to a set initial value; after the warm-up period, the learning rate decays according to a cosine annealing function to a preset minimum value of 4 × 10⁻⁶. -6 Furthermore, the initial learning rate varies for different datasets. For example, using RML2016.10a or RML2018.01a as the modulation signal dataset, the dataset is divided into training, validation, and test sets in a 6:2:2 ratio. The RML2016.10a dataset contains 8 digital modulation signals and 3 analog modulation signals, while the RML2018.01a dataset contains 19 digital modulation signals and 5 analog modulation signals. Therefore, the learning rate is set to 0.01 for the RML2016.10a dataset and 0.004 for the RML2018.10a dataset.

[0041] To evaluate the recognition performance of this invention for different modulated signals under low signal-to-noise ratio conditions, the modulation signal recognition results in the RML2016.10a and RML2018.01a datasets were compared and analyzed under -4dB conditions.

[0042] Figure 9 These are simulation graphs showing the performance of various modulation signal recognition methods provided in the embodiments of this invention and the method proposed in this invention on the RML2016.10a dataset. Figure 10These are simulation graphs showing the performance of various modulation signal recognition methods provided in this invention and the method proposed in this invention on the RML2018.01a dataset. The existing methods involved include: Residual Network (ResNet), Long Short-Term Memory (LSTM), a hybrid neural network architecture specifically designed for sequence signal processing that connects convolutional neural networks, recurrent neural networks, and deep fully connected networks (Convolutional, LSTM, Deep Neural Network; CLDNN), Multi-channel Long Short-Term Memory Deep Neural Network (MCLDNN), CNN-GRU Deep Neural Network based on phase estimation transform (PET-CGDNN), and a general visual foundation model (InternImage) with deformable convolution as its core and incorporating Transformer architecture ideas. The method proposed in this invention is represented as GM-DMR.

[0043] To facilitate accurate quantitative evaluation, Table 1 further summarizes the detailed experimental results of each method on key metrics when processing the RML2016.10a dataset, and Table 2 summarizes the detailed experimental results of each method on key metrics when processing the RML2018.01a dataset.

[0044] Table 1

[0045] Table 2

[0046] like Figure 9 and 10As shown in the charts, the data reveals that on the RML2016.10a dataset, when the signal-to-noise ratio (SNR) is below -10dB, there is no significant difference in recognition accuracy between this invention and other methods; when the SNR is between -10dB and 0dB, the recognition accuracy of this invention is higher than other methods; when the SNR is greater than 0dB, the recognition accuracy of all methods tends to stabilize, but the recognition accuracy of this invention remains higher than other methods. On the RML2018.01a dataset, when the SNR is below 10dB, the recognition accuracy of this invention is higher than other methods; when the SNR is above 10dB, the recognition accuracy of all methods tends to stabilize. Tables 1 and 2 show that PET-CGDNN has the fewest parameters and computational cost, but its overall performance on both datasets is only average. In contrast, although this invention has a slightly increased number of parameters and computational cost, its performance is significantly improved. Furthermore, while CLDNN performed reasonably well on the RML2016.10a dataset, its accuracy dropped significantly on the RML2018.01a dataset, reflecting its insufficient generalization ability in complex environments. In contrast, the performance of this invention remained stable on both datasets. Moreover, this invention significantly outperformed other methods at low signal-to-noise ratios (-10 to 0 dB), indicating its strong robustness under low signal-to-noise ratio conditions.

[0047] To evaluate the recognition performance of this invention for different modulated signals under low signal-to-noise ratio conditions, we compared and analyzed the modulation signal recognition results in the RML2016.10a and RML2018.01a datasets under -4dB conditions. Figure 11 and 12 The performance of each method on different modulation schemes is visually illustrated using radar charts. On the RML2016.10a dataset, all methods achieved classification accuracy exceeding 95% for GFSK, CPFSK, and PAM4 signals, demonstrating stable performance. However, CLDNN and LSTM showed lower classification performance for QAM16 and QAM64 modulation signals. Although PET-CGDNN and MCLDNN performed well in classifying various modulation signals, their performance in recognizing complex signals such as AM-DSB was still lower than the results of this invention. On the RML2018.01a dataset, all methods achieved classification accuracy exceeding 95% for BPSK, GFSK, and OOK signals. All methods showed decreased classification accuracy for higher-order modulation signals such as 256QAM and 128QAM, as well as complex signals such as AM-DSB-WC and AM-SSB-WC. The performance of this invention was superior to other methods, further validating its ability to recognize complex signals.

[0048] To specifically evaluate the contribution of each module to the performance of this invention, we designed the following ablation experiments, as shown in Table 3. In the ablation experiments, we denoted the performance of this invention on two datasets as Model 1, compared the effects of each ablation model, and removed each module one by one to evaluate its independent contribution. In Model 2, we replaced the convolutional linear gating unit with a multilayer perceptron to evaluate its specific contribution to feature fusion capability. In Model 3, we removed the spatial and channel redundancy suppression module, leaving other modules unchanged, to analyze the impact of reducing spatial and channel redundancy on the overall model performance; in Model 4, we removed the multi-scale convolution module based on the channel attention mechanism and discussed the role of multi-scale convolution and channel attention mechanisms in improving model performance. As shown in Table 3, the model performs best when all modules are present. Model 2 shows a significant decrease in recognition accuracy, particularly at low SNR, indicating that the feature fusion capability of the convolutional linear gating unit plays a crucial role in model performance. Model 3 exhibits a decrease in accuracy and a significant increase in complexity, demonstrating that suppressing spatial and channel redundancy is critical for reducing model complexity and improving performance. Model 4 shows a significant decrease in recognition accuracy at low SNR, suggesting that its design combining channel attention with multi-scale convolution is significant for capturing multi-scale features and thus improving model performance.

[0049] Table 3

[0050] To address the existing problems, this invention provides a signal recognition method based on variable convolution and convolution-gated linear unit modulation. This method systematically solves the inherent defects of existing methods in feature fusion, computational efficiency, and multi-scale perception. Specifically: (1) At the feature extraction and fusion level, the deformable convolutional layers in the backbone network can dynamically adjust the receptive field according to the input signal content, adaptively capturing complex signal structure features in the time-frequency domain, overcoming the limitations of fixed receptive field convolution. At the same time, the convolutional gated linear unit, through its embedded deep convolutional gated branch, realizes an adaptive feature selection mechanism based on local spatial context, which can actively enhance key information and suppress the interference of noise and irrelevant features in the feature stream, significantly improving the selectivity of feature fusion and the model's anti-interference ability; (2) In terms of computational efficiency and model optimization, the spatial and channel redundancy suppression unit integrated in the stem network of the model effectively suppresses the spatial and channel dimension redundancy caused by the inherent characteristics of convolution operation by evaluating the importance of feature maps, separating and reorganizing them. This reduces the computational complexity and number of parameters of the model while ensuring or even improving the recognition performance. (3) At the level of multi-scale information perception and robustness, the multi-scale fusion module with fusion channel attention introduced at the end of the backbone network explicitly extracts context features of different scales through parallel multi-branch structure and uses channel attention mechanism for adaptive weighted fusion, so that the model can capture the transient details and long-term steady-state mode of the signal at the same time, thereby greatly enhancing the recognition adaptability and overall robustness of multi-scale, non-stationary modulation signals in complex and variable electromagnetic environments. In summary, the technical solution provided by this invention achieves more accurate, efficient, and robust end-to-end identification of modulated signals through the synergistic effect of the above modules, and shows significant advantages, especially under harsh communication conditions such as low signal-to-noise ratio.

[0051] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for identifying modulated signals based on variable convolution and convolution-gated linear units, characterized in that, include: Acquire the modulation signal to be identified; The modulation signal to be identified is input into a trained modulation signal recognition network model, which includes a stem network, a trunk network, and a classification network. The stem network is used to perform data preprocessing, preliminary feature extraction, and redundancy suppression on the modulated signal to be identified, to obtain a first feature. The backbone network is used to perform deep feature extraction and adaptive filtering on the first feature to obtain the second feature; wherein, the backbone network includes several basic modules connected in sequence and a multi-scale fusion module; Each basic module includes a first normalized layer, a deformable convolutional layer, a second normalized layer, and a convolutionally gated linear unit connected in sequence; wherein, the deformable convolutional layer is used to dynamically adjust the sampling point position according to the input content to adaptively extract the structural features of the signal in the time-frequency domain; the convolutionally gated linear unit is used to generate spatially aware adaptive weights through the gated branches composed of convolutional layers to filter features; The multi-scale fusion module is used to receive the output features of the last basic module, extract features under different receptive fields through parallel multi-scale convolutional layers, and adaptively weight and fuse features of different scales using a channel attention mechanism to obtain the second feature. The classification network is used to perform dimensionality compression and classification decisions on the second feature, and outputs the probability distribution of the recognition result.

2. The signal recognition method based on variable convolution and convolution-gated linear unit modulation according to claim 1, characterized in that, The deformable convolutional layer in each basic module includes: a first 3×3 convolutional layer, a first Softmax layer, a first connection layer, a first interpolation layer, and a second 3×3 convolutional layer; The first 3×3 convolutional layer receives the output features of the first normalization layer and outputs the sampling point offset Δx, the sampling point offset Δy, and the weight parameter Δm. In the first Softmax layer, the weight parameter Δm is normalized to obtain normalized weights; The first connection layer dynamically adjusts the sampling point coordinates of the output features of the first normalization layer based on the offsets Δx and Δy to obtain the offset sampling coordinates. The first interpolation layer uses bilinear interpolation to calculate the features at the offset sampling coordinates to obtain the interpolated features; The second 3×3 convolutional layer performs a weighted summation of the normalized weights and the interpolated features to obtain the output features of the deformable convolutional layer.

3. The signal recognition method based on variable convolution and convolution-gated linear unit modulation according to claim 1, characterized in that, The convolutional gated linear unit in each basic module includes: a first fully connected layer, a second fully connected layer, a first depthwise convolutional layer, a Gaussian error linear unit, a first product unit, and a third fully connected layer; The first fully connected layer and the second fully connected layer process the output features of the second normalized layer in parallel to generate the main information flow features and the gated intermediate features, respectively. The first deep convolutional layer performs spatial convolution processing on the gated intermediate features to incorporate local contextual information, thereby obtaining the first convolutional features; The Gaussian error linear unit performs nonlinear activation on the first convolutional feature to generate gated weights; The first product unit performs element-wise multiplication of the gate weights and the main information flow features to achieve feature filtering and obtain the filtered features. The third fully connected layer performs channel dimension transformation on the filtered features to obtain the output features of the convolutional gated linear unit.

4. The signal recognition method based on variable convolution and convolution-gated linear unit modulation according to claim 1, characterized in that, The multi-scale fusion module includes: a multi-scale convolutional unit, a channel attention unit, and a feature fusion unit; The multi-scale convolutional unit is used to extract features under different receptive fields in parallel, including: a first 1×1 convolutional layer, a fourth 3×3 convolutional layer, a first 3×3 dilated convolutional layer, a first global average pooling layer, a second interpolation layer, a first linear rectified layer, a second linear rectified layer, a third linear rectified layer, and a fourth linear rectified layer. The first 1×1 convolutional layer and the first linear rectified layer constitute the second branch, the fourth 3×3 convolutional layer and the second linear rectified layer constitute the third branch, and the first global average pooling layer, the second interpolation layer and the fourth linear rectified layer constitute the fourth branch. The channel attention unit includes a second connection layer and a channel attention mechanism subunit. The second connection layer concatenates the output features of the four branches along the channel dimension to obtain concatenated features. The channel attention mechanism subunit processes the concatenated features to generate channel weights. The feature fusion unit includes a second product unit and a second 1×1 convolutional layer. The second product unit is used to multiply the channel weights and the concatenated features element by element to achieve weighting, and the second 1×1 convolutional layer is used to integrate the channels and output the fused features as the second feature.

5. The signal recognition method based on variable convolution and convolution-gated linear unit modulation according to claim 4, characterized in that, The channel attention mechanism subunit includes: a second global average pooling layer, a fourth fully connected layer, a fifth linear rectifier layer, a fifth fully connected layer, and a Sigmoid activation layer connected in sequence; The second global average pooling layer is used to compress the spatial information of the spliced ​​features and generate channel descriptors; The fourth fully connected layer is used to perform dimensionality reduction transformation on the channel descriptor to obtain dimensionality reduction transformation features; The fifth linear rectifier layer is used to perform nonlinear transformation on the dimensionality reduction transformation features; The fifth fully connected layer is used to restore the feature dimensions after nonlinear transformation to obtain interactive data; A Sigmoid activation layer is used to map the interactive data to channel weights in the range of 0 to 1.

6. The signal recognition method based on variable convolution and convolution-gated linear unit modulation according to claim 1, characterized in that, The stem network includes: a preprocessing module, a spatial redundancy suppression unit, and a channel redundancy suppression unit; The preprocessing module is used to perform dimensional unification and multi-scale initial feature extraction on the input modulated signal to be identified, and output a two-dimensional feature map with a unified spatial size. The spatial redundancy suppression unit and the channel redundancy suppression unit are used to suppress redundant information in the spatial and channel dimensions of the input feature map in sequence.

7. The signal recognition method based on variable convolution and convolution-gated linear unit modulation according to claim 6, characterized in that, The spatial redundancy suppression unit includes: a group normalization layer, a third product unit, a Sigmoid activation layer, a first adder, a second adder, and a third connection layer; The group normalization layer is used to perform channel grouping normalization on the input weighted feature map to obtain a normalized feature map. The third product unit and the sigmoid activation layer constitute a weight generation module, which is used to calculate the importance weight of each spatial location using the normalized feature map. The first adder and the second adder constitute a feature separation module. The feature separation module is used to divide the importance weight into importance weight and suppression weight according to a preset threshold, and multiply them by the original input feature map respectively to obtain a first feature subset and a second feature subset; the information richness of the first feature subset is greater than that of the second feature subset. The third connection layer concatenates the first feature subset and the second feature subset in the channel dimension to obtain a feature map after spatial redundancy suppression.

8. The signal recognition method based on variable convolution and convolution-gated linear unit modulation according to claim 7, characterized in that, The channel redundancy suppression unit includes: a third 1×1 convolutional layer, a fourth 1×1 convolutional layer, a first group of convolutional layers, a first point convolutional layer, a second point convolutional layer, a third adder, a fourth connection layer, a third global average pooling layer, and a second Softmax layer. The third 1×1 convolutional layer and the fourth 1×1 convolutional layer simultaneously perform channel dimensionality reduction on the feature map after spatial redundancy suppression to obtain the first channel dimensionality reduction feature map and the second channel dimensionality reduction feature map. The first group of convolutional layers and the first point convolutional layer simultaneously perform group convolution and point-by-point convolution on the first channel dimensionality reduction feature map to extract spatial features and restore inter-channel information exchange, respectively, to obtain group convolutional features and first point convolutional features; then, the third adder is used to sum them to obtain the channel information feature map; The second point convolutional layer performs point-by-point convolution on the second channel's dimensionality-reduced feature map to generate the second point convolutional feature; The fourth connection layer concatenates the second point convolutional feature and the feature map after spatial redundancy suppression in the channel dimension to obtain the output feature map of the fourth connection layer. The third global average pooling layer performs global average pooling on the channel information feature map and the output feature map of the fourth connection layer to obtain the first global spatial information and the second global spatial information. The second Softmax layer performs Softmax normalization on the first global spatial information and the second global spatial information to generate corresponding channel weights; the corresponding channel weights are then weighted and fused with the channel information feature map and the output feature map of the fourth connection layer to obtain a feature map after channel redundancy suppression.

9. The method for identifying modulated signals based on variable convolution and convolution-gated linear units according to claim 1, characterized in that, The classification network comprises a fifth 1×1 convolutional layer, a global average pooling layer, and a fully connected layer connected in sequence.

Images