Modulation signal identification method based on convolution attention and multi-dimensional feature fusion

Abstract

The invention discloses a modulation signal identification method based on convolution attention and multi-dimensional feature fusion, and solves the problems that in the prior art, different signals in the same signal family always show similar features under the condition of a low signal-to-noise ratio, so that a model is difficult to distinguish effectively, the identification accuracy is reduced, and the identification accuracy is low. The modulation signal identification under a low signal-to-noise ratio is realized, the identification precision is higher, and the calculation complexity is lower; the method comprises the following steps: acquiring an original complex signal, normalizing the original complex signal to obtain an IQ sequence, and obtaining an AP sequence through the IQ sequence; then, a pre-trained signal identification network is used for identification; a preprocessing module converts the input IQ and AP sequences into IQ and AP feature maps; the feature extraction module extracts local and global features to form a path feature map; the WAFF module performs weighted fusion on the two paths of features; and finally, the classification head outputs a classification result according to the fusion features.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a modulation signal recognition method based on the fusion of convolutional attention and multi-dimensional features. Background Technology

[0002] Automatic modulation identification (AMR), a crucial step between signal detection and demodulation, provides modulation information for wireless signals, playing a key role in intelligent wireless communication. Traditional modulation identification techniques can be broadly categorized into two types: likelihood-based (LB) and feature-based (FB) methods. LB methods use Bayesian theory to calculate the matching degree between a signal and different modulation schemes. In contrast, FB methods manually extract features and combine them with traditional machine learning classifiers to distinguish modulation types. Deep learning, capable of automatically extracting features from received signals and classifying them, has made significant progress in AMR in recent years, with examples such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory Networks (LSTMs). CNNs are typically used to extract spatial features from data, while LSTMs excel at modeling time-series data. Self-attention mechanisms and Transformers, as novel methods in natural language processing and computer vision, have also shown great potential in radio signal classification. Transformers, through multi-head self-attention (MHSA) mechanisms, can extract global features containing spatiotemporal information.

[0003] Furthermore, most existing modulation recognition methods only use a single dimension of the signal as the model input, failing to fully capture the feature representation of the modulated signal across different dimensions. In contrast, multi-dimensional input signals exhibit better recognition capabilities and stronger robustness. Therefore, this invention designs a deep learning model based on a surrogate attention mechanism and multi-dimensional feature fusion. This model uses convolution plus attention to extract multi-dimensional features from the input signal, and then uses an attention fusion module to weightedly fuse information from different dimensions. Through weighted fusion, the model can flexibly adjust the influence of different input signal dimensions, thereby improving the richness and accuracy of feature representation under low signal-to-noise ratio conditions.

[0004] Traditional likelihood-based and feature-based modulation identification methods rely on a large amount of prior knowledge, making it difficult to achieve fast and accurate modulation identification in modern complex wireless communication environments.

[0005] Deep learning-based automatic modulation recognition techniques have several drawbacks. First, CNNs are typically used to extract spatial features from data but lack attention to the temporal features of the signal. LSTMs, while adept at modeling time-series data, cannot be parallelized and are computationally inefficient with large datasets. This limitation makes it difficult for these deep learning models to comprehensively capture the overall features of complex modulated signals, restricting their robustness and generalization ability in practical applications, thus affecting their recognition performance. Second, compared to CNNs and LSTMs, Transformers utilize self-attention mechanisms to capture global features but neglect local feature modeling. Furthermore, Transformers have high computational complexity and resource consumption, significantly limiting their deployment on resource-constrained edge devices. Finally, many methods rely solely on the IQ features of the signal itself, ignoring the potential information in other dimensions. This single-dimensional input limits the model's comprehensive modeling of signal features, resulting in poor recognition performance in complex signal environments. Especially under low signal-to-noise ratio conditions, different signals within the same signal family (QAM16, QAM64, 8PSK, BPSK) often exhibit similar characteristics, making it difficult for the model to effectively distinguish them and reducing the recognition accuracy. Summary of the Invention

[0006] This invention provides a modulation signal recognition method based on convolutional attention and multi-dimensional feature fusion. This method solves the problem in the prior art where different signals within the same signal family often exhibit similar features under low signal-to-noise ratio conditions, making it difficult for the model to effectively distinguish them and reducing the recognition accuracy. This invention achieves higher recognition accuracy for modulation signals under low signal-to-noise ratio conditions and has lower computational complexity.

[0007] This invention provides a modulation signal recognition method based on convolutional attention and multi-dimensional feature fusion, the method comprising: The original complex signal to be identified is acquired, the original complex signal is processed to obtain a normalized IQ sequence, and the AP sequence is obtained based on the normalized IQ sequence. The IQ sequence and the AP sequence are used for signal recognition based on a pre-trained signal recognition network to obtain recognition results. The signal recognition network includes a preprocessing module, a feature extraction module, a WAFF module, and a classification head. Specifically: the preprocessing module performs preliminary shape transformations on the IQ and AP sequences to obtain processed IQ and AP feature maps; the feature extraction module extracts local detail features and global features from the processed IQ and AP feature maps to obtain IQ and AP path feature maps; the WAFF module performs weighted fusion of the IQ and AP path feature maps to obtain fused features; and the classification head maps the fused features to class prediction to obtain classification results.

[0008] In one possible implementation, the step of processing the original complex signal to obtain a normalized IQ sequence, and obtaining the AP sequence based on the normalized IQ sequence, includes: The real part is extracted from the original complex signal to form the I-sequence signal, and the imaginary part is extracted to form the Q-sequence signal. The I-sequence signal and the Q-sequence signal are normalized respectively to obtain the normalized IQ sequence; The normalized IQ sequence is converted into an AP sequence.

[0009] In one possible implementation, the preprocessing module includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, and a channel splicing layer; The preprocessing module is used to perform preliminary shape transformations on the IQ sequence and the AP sequence to obtain processed IQ feature maps and AP feature maps, including: The normalized IQ sequence is input into the first convolutional layer, the second convolutional layer, and the third convolutional layer respectively for multi-scale feature extraction to obtain the first feature, the second feature, and the third feature; wherein, the convolutional kernels of the first convolutional layer and the third convolutional layer are different, and the second convolutional layer is a max pooling layer; The first feature, the second feature, and the third feature are fused together through the channel splicing layer to form an IQ feature map; The IQ feature map is transformed into an amplitude feature map and a phase feature map through polar coordinate transformation, and the two are then concatenated to obtain the AP feature map.

[0010] In one possible implementation, the feature extraction module includes a first stage module, a second stage module, a third stage module, and a fourth stage module connected in sequence. The feature extraction module is used to extract local detail features and global features from the processed IQ feature map and the AP feature map to obtain the IQ path feature map and the AP path feature map, including: The first stage module includes: a first downsampling block and a first depthwise separable convolution module; the first stage module performs local spatiotemporal feature extraction on the processed IQ feature map and the AP feature map to obtain the IQ path primary feature map and the AP path primary feature map; The second stage module includes: a second downsampling block and a second depthwise separable convolution module; the second depthwise separable unit performs multi-scale local feature extraction on the primary feature map of the IQ path and the primary feature map of the AP path to obtain intermediate feature maps of the IQ path and the AP path. The third stage module includes: a third downsampling block and a first proxy attention module; the third stage module introduces an attention mechanism to capture the global context dependency of the intermediate feature map of the IQ path and the intermediate feature map of the AP path, so as to obtain the IQ path-aware feature map and the AP path-aware feature map. The fourth stage module includes: a fourth downsampling block and a second proxy attention module; the fourth stage module further refines the global features of the IQ path-aware feature map and the AP path-aware feature map at the highest semantic level to obtain the IQ path high-level feature map and the AP path high-level feature map.

[0011] In one possible implementation, the first depthwise separable convolutional module includes: a first normalization layer, a first depthwise separable convolutional layer, a first multilayer perceptron layer, and a first residual connection layer. The first stage module performs local spatiotemporal feature extraction on the processed IQ feature map and the AP feature map to obtain a primary IQ path feature map and a primary AP path feature map, including: The initial downsampled features output by the first downsampled block are input into the first normalization layer for normalization processing; The normalized features are input into the first depthwise separable convolutional layer, and depthwise convolution and pointwise convolution are performed sequentially to extract local spatial features. The local spatial features are input into the first multilayer perceptron layer, and a nonlinear transformation of the channel dimension is performed to obtain the nonlinearly transformed features. The nonlinearly transformed features are added to the initial downsampled features through the first residual connection layer to obtain the IQ path primary feature map and the AP path primary feature map.

[0012] In one possible implementation, the first proxy attention module includes: a second normalization layer, a second depthwise separable convolutional layer, a second multilayer perceptron layer, and a second residual connection layer. The third-stage module introduces an attention mechanism to capture the global contextual dependencies of the intermediate feature maps of the IQ path and the AP path, resulting in IQ path-aware feature maps and AP path-aware feature maps, including: The high-level downsampling features output by the third downsampling block are input into the second normalization layer for normalization processing; The normalized features are input into the proxy attention calculation layer, and a global context-enhanced feature map is obtained by generating proxy vectors, proxy-key value attention calculation and query-proxy attention calculation; The global context-enhanced feature map is input into the second depthwise separable convolutional layer to perform local feature recovery and enhancement, thereby obtaining local enhanced features. The local enhancement features are added to the high-level downsampling features through the second residual connection layer to obtain the IQ path-aware feature map and the AP path-aware feature map.

[0013] In one possible implementation, the WAFF module includes: a local attention branch, a global attention branch, a weight generation layer, and a weighted fusion layer; The WAFF module performs a weighted fusion of the IQ path feature map and the AP path feature map to obtain fused features, including: The initial fused features are obtained by adding the IQ path feature map and the AP path feature map; The initial fusion features are simultaneously input into the local attention branch and the global attention branch, and the local attention weights and global attention weights are extracted respectively. The attention weight map is generated by adding the local attention weights and the global attention weights and then applying the sigmoid activation function. The weighted fusion layer adaptively weights and fuses the IQ path feature map and the AP path feature map according to the attention weight map to obtain the weighted fusion feature. The weighted fusion feature is added to the initial fusion feature through a residual connection to obtain the final fusion feature.

[0014] In one possible implementation, the classification head includes: a global average pooling layer and a fully connected classification layer; The classification head is used to map the fused features onto category prediction to obtain classification results, including: The fused features are input into the global average pooling layer to compress the spatial dimension and obtain the feature vector. The feature vector is input into the fully connected classification layer, and the output is the probability distribution corresponding to each modulation category; The category with the highest probability is output as the final recognition result.

[0015] In one possible implementation, the training process of the signal recognition network includes: Using a signal dataset that includes multiple modulation types and signal-to-noise ratios, the dataset is divided proportionally into training, validation, and test sets. The model was trained using the cross-entropy loss function and the AdamW optimizer. During training, a cosine annealing strategy is used to adjust the learning rate, and Dropout is set to prevent overfitting.

[0016] In one possible implementation, the signal dataset includes RML2016.10a, wherein the modulation type includes at least one of: 8PSK, BPSK, CPFSK, GFSK, PAM4, 16QAM, 64QAM, QPSK, AM-DSB, AM-SSB, and WBFM.

[0017] One or more technical solutions provided in this invention have at least the following technical effects or advantages: This invention acquires the original complex signal to be identified and processes it to obtain a normalized IQ sequence. Through standardization, amplitude deviations caused by factors such as transmission distance and channel gain are eliminated, providing a stable and unified input benchmark for subsequent deep learning models. This effectively improves the convergence speed and stability of model training and enhances the model's generalization ability to environments with different signal strengths, avoiding performance fluctuations caused by inconsistent input data scales. The signal is processed through a pre-trained signal recognition network. The core of this network lies in its four collaborative modules. First, the preprocessing module processes the normalized IQ sequence to generate IQ feature maps and AP feature maps. The beneficial effect of this step is the pioneering introduction of multi-dimensional feature input. It not only preserves the original instantaneous information of the signal in the complex domain I / Q but also generates amplitude / phase (A / P) features characterizing the signal's statistical properties through polar coordinate transformation. These two feature paths complement each other, providing the model with a richer and more robust feature foundation, fundamentally solving the problems of incomplete information and easy confusion between similar signals caused by single-dimensional input in traditional methods. Secondly, the feature extraction module processes the dual-path feature maps in parallel, extracting local and global features. This module strikes a balance between computational efficiency and feature extraction: the front-end uses depthwise separable convolutions for local feature extraction, significantly reducing the number of parameters and computational complexity while effectively capturing the spatiotemporal local details of the signal, making it suitable for deployment on resource-constrained edge devices. Efficient modeling of the global context: the rear-end introduces a surrogate attention mechanism, reducing the quadratic computational complexity of traditional self-attention to linear levels. This allows the model to capture the global long-range dependencies of the entire signal sequence at an affordable computational cost, understanding the macroscopic patterns of the signal and overcoming the bottlenecks of local CNN vision and serial computation of LSTM. Hierarchical feature learning: the cascaded structure of four stages achieves progressive abstraction and refinement of features from local to global and from low to high levels, constructing a multi-level feature representation with strong discriminative power. Then, the WAFF module adaptively weights and fuses the high-level features extracted from the dual paths. The beneficial effect of this step is that it achieves intelligent feature fusion. Unlike simple concatenation or averaging, the WAFF module dynamically learns and generates a weight map through parallel local and global attention branches. This weight map adaptively determines the relative weights of IQ path features and AP path features in the final fused features based on the specific content of the input signal. For example, at high signal-to-noise ratios (SNR), it may rely more on clear IQ features; at low SNRs, it may rely more on the more robust AP statistical features. This dynamic fusion mechanism maximizes the preservation of effective information from each path, suppresses noise or invalid features, and generates a more discriminative and robust unified feature representation. Finally, the classification head maps the fused features to class predictions and outputs the recognition result. Based on the high-quality, highly discriminative fused features extracted in the preceding steps, the classification head can make classification decisions more accurately and reliably.This achieves a significant and superior balance between overall recognition accuracy, recognition rate at low signal-to-noise ratios, and model computational complexity. Experimental data shows that the proposed method achieves leading average and maximum accuracy on publicly available datasets while maintaining computational overhead comparable to advanced lightweight models, validating the overall superiority of its technical solution. This method achieves the core objective of improving modulation recognition accuracy at low signal-to-noise ratios while maintaining model computational efficiency, demonstrating good practical value and application prospects. Attached Figure Description

[0018] Figure 1 A flowchart illustrating the steps of a modulation signal recognition method based on convolutional attention and multi-dimensional feature fusion provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a signal recognition network provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the preprocessing module structure provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the feature extraction module structure provided in an embodiment of the present invention; Figure 5 A schematic diagram of the structure of the first depthwise separable convolutional module and the second depthwise separable convolutional module provided in an embodiment of the present invention; Figure 6 A schematic diagram of the structure of the first depthwise separable convolutional module and the second depthwise separable convolutional module provided in an embodiment of the present invention; Figure 7 This is a computational structure diagram of the proxy attention mechanism provided in an embodiment of the present invention; Figure 8 A comparison of the accuracy of various recognition methods provided in the embodiments of the present invention under different signal-to-noise ratios on the RML2016.10A dataset; Figure 9 The confusion matrix of the present invention at a signal-to-noise ratio of -10dB is provided in the embodiments of the present invention; Figure 10 The confusion matrix of the present invention at a signal-to-noise ratio of 0dB is provided in the embodiments of the present invention; Figure 11 The confusion matrix of the present invention at a signal-to-noise ratio of 10dB is provided for an embodiment of the present invention; Figure 12 This is the confusion matrix of CV-TRN provided in this embodiment of the invention at a signal-to-noise ratio of -2dB; Figure 13 The confusion matrix of GAM-HRNN-GRU provided in this embodiment of the invention at a signal-to-noise ratio of -2dB; Figure 14 The confusion matrix provided in this embodiment of the invention has a signal-to-noise ratio of -2dB. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] This invention provides a modulation signal recognition method based on convolutional attention and multi-dimensional feature fusion, see [link to relevant documentation]. Figure 1 The method includes the following steps S101 to S102.

[0021] S101, acquire the original complex signal to be identified, perform data processing on the original complex signal to obtain the normalized IQ sequence, and obtain the AP sequence based on the normalized IQ sequence; Specifically, in step S101, the original complex signal is processed to obtain a normalized IQ sequence, and the AP sequence is obtained based on the normalized IQ sequence, including the following steps S1011 to S1013.

[0022] S1011, extract the real part from the original complex signal as the I sequence signal, and extract the imaginary part as the Q sequence signal; S1012, Normalize the I sequence signal and the Q sequence signal respectively to obtain the normalized IQ sequence; S1013 converts the normalized IQ sequence into an AP sequence.

[0023] For example, the I-sequence and Q-sequence signals are normalized, and then the complex domain signals are transformed into A / P (amplitude / phase) dimension data through polar coordinate transformation, enhancing the inherent separability of different signals. The transformation formula is as follows: .

[0024] S102, perform signal recognition on the IQ sequence and AP sequence according to the pre-trained signal recognition network to obtain the recognition result.

[0025] The signal recognition network includes a preprocessing module, a feature extraction module, a WAFF module, and a classification head.

[0026] The preprocessing module is used to normalize the I-sequence signal and the Q-sequence signal to obtain the processed IQ feature map and AP feature map; Specifically, the preprocessing module includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, and a channel splicing layer; The preprocessing module performs preliminary shape transformations on the IQ and AP sequences to obtain processed IQ and AP feature maps, including: (1) Input the normalized IQ sequence into the first convolutional layer, the second convolutional layer and the third convolutional layer respectively to extract multi-scale features, and obtain the first feature, the second feature and the third feature; wherein, the convolutional kernels of the first convolutional layer and the third convolutional layer are different, and the second convolutional layer is a max pooling layer; (2) The first feature, the second feature and the third feature are fused through the channel splicing layer to form an IQ feature map; (3) The IQ feature map is transformed into an amplitude feature map and a phase feature map through polar coordinate transformation, and the two are spliced ​​together to obtain the AP feature map.

[0027] Construct a network model based on agent attention mechanism and multi-dimensional feature fusion, the network structure is as follows: Figure 2 As shown, the overall structure consists of a preprocessing module, a feature extraction module, a WAFF module, and a classification head. The preprocessing module adjusts the input shape of the data sequence. The weighted attention feature fusion module (WAFF) weights and fuses information from different dimensions of the signal. Finally, the classification head maps the extracted features onto the category prediction, thus obtaining the classification result. Here, in Figure 2 In this system, the main network includes a preprocessing module and a feature extraction module.

[0028] For example, the preprocessing module is used to preprocess the IQ and AP sequences of the RML2016.10a dataset, such as... Figure 3 As shown, multi-scale convolution extracts features from a sequence through three convolutional layers of different sizes within the Inception architecture. These features are then concatenated using a concatenation operation.

[0029] The feature extraction module is used to extract local detail features and global features from the processed IQ feature map and AP feature map to obtain the IQ path feature map and AP path feature map; Specifically, the feature extraction module includes a first stage module, a second stage module, a third stage module, and a fourth stage module connected in sequence; The feature extraction module is used to extract local detail features and global features from the processed IQ feature map and AP feature map, resulting in IQ path feature map and AP path feature map, including: (1) The first stage module includes: a first downsampling block and a first depthwise separable convolution module; the first stage module performs local spatiotemporal feature extraction on the processed IQ feature map and AP feature map to obtain the IQ path primary feature map and the AP path primary feature map. Here, the first depthwise separable convolutional module includes: a first normalization layer, a first depthwise separable convolutional layer, a first multilayer perceptron layer, and a first residual connection layer; The first stage module performs local spatiotemporal feature extraction on the processed IQ and AP feature maps to obtain primary IQ path feature maps and primary AP path feature maps, including: (1.1) Input the initial downsampled features output by the first downsampled block into the first normalization layer for normalization processing; (1.2) Input the normalized features into the first depthwise separable convolutional layer, and perform depthwise convolution and pointwise convolution in sequence to extract local spatial features; (1.3) Input the local spatial features into the first multilayer perceptron layer and perform a nonlinear transformation of the channel dimension to obtain the nonlinearly transformed features; (1.4) The nonlinearly transformed features are added to the initial downsampled features through the first residual connection layer to obtain the primary feature map of the IQ path and the primary feature map of the AP path.

[0030] (2) The second stage module includes: a second downsampling block and a second depthwise separable convolution module; the second depthwise separable unit performs multi-scale local feature extraction on the primary feature map of the IQ path and the primary feature map of the AP path to obtain the intermediate feature map of the IQ path and the intermediate feature map of the AP path. It is important to note that the structure of the second depthwise separable convolutional module is the same as the hierarchical structure of the first depthwise separable convolutional module.

[0031] (3) The third stage module includes: the third downsampling block and the first proxy attention module; the third stage module introduces an attention mechanism to capture the global context dependency of the intermediate feature maps of the IQ path and the intermediate feature maps of the AP path, and obtains the IQ path-aware feature maps and the AP path-aware feature maps. Here, the first agent attention module includes: a second normalization layer, a second depthwise separable convolutional layer, a second multilayer perceptron layer, and a second residual connection layer; The third-stage module introduces an attention mechanism to capture the global contextual dependencies of intermediate feature maps in the IQ path and AP path, resulting in IQ path-aware feature maps and AP path-aware feature maps, including: (3.1) Input the high-level downsampling features output by the third downsampling block into the second normalization layer for normalization processing; (3.2) Input the normalized features into the proxy attention calculation layer, and obtain the global context-enhanced feature map by generating proxy vectors, proxy-key value attention calculation and query-proxy attention calculation; (3.3) Input the feature map of global context enhancement into the second deep separable convolutional layer to perform local feature recovery and enhancement to obtain local enhanced features; (3.4) The local enhancement features are added to the high-level downsampling features through the second residual connection layer to obtain the IQ path-aware feature map and the AP path-aware feature map.

[0032] (4) The fourth stage module includes: the fourth downsampling block and the second agent attention module; the fourth stage module further refines the global features of the IQ path-aware feature map and the AP path-aware feature map at the highest semantic level to obtain the IQ path high-level feature map and the AP path high-level feature map.

[0033] It is important to note that the first agent attention module and the second agent attention module have the same structure.

[0034] For example, see Figure 4 The feature extraction module consists of four stages. The first two stages consist of downsampling blocks and depthwise separable convolutional modules, while the last two stages consist of downsampling blocks and surrogate attention modules.

[0035] Construct a downsampling module. The downsampling layer consists of a layer normalization and a 2D convolution. The parameters of the convolution are set as follows: kernel_size=3, stride=2, padding=1. See [link to specific structure] for details. Figure 5 The input feature map is normalized by the layers and then downsampled by 2D convolution, reducing the height and width of the feature map to half of the original size. At the same time, the number of channels in the feature map is adjusted. The downsampling blocks of the four stages adjust the number of channels to 64, 128, 320 and 512 respectively.

[0036] Construct a depthwise separable convolution module. The depthwise separable convolution module consists of layer normalization, depthwise separable convolutions, an MLP (Multi-Level Processing), and residual connections. The specific structure can be found in [link to structural diagram]. Figure 6 .

[0037] The inputs to the two depthwise separable convolutional modules are the feature map data after the first two downsampling steps. The input to the first depthwise separable convolutional module is 64×64×64, and the input to the second depthwise separable convolutional module is 128×32×32. The input feature maps are first normalized by a normalization layer to smooth the data distribution and reduce training time. Then, the data enters the depthwise separable convolutional layers, where features are extracted. The computation process of depthwise separable convolution includes pointwise convolution and depthwise convolution operations.

[0038] Specifically, for the first depthwise separable convolution module, depthwise convolution first performs independent spatial convolution on each channel of the input feature map. The convolution parameters are set to kernel_size=3, stride=2, and padding=1, resulting in input and output dimensions of 64×64×64. Depthwise convolution can efficiently extract spatial features. Subsequently, pointwise convolution performs channel fusion on the output, combining and transforming the information from each channel. This significantly reduces the number of model parameters and computational cost while maintaining feature extraction capabilities. The kernel parameters for pointwise convolution are set to 1×1. After completing the depthwise separable convolution, the local spatial features are fed into a multilayer perceptron (MLP) module. MLPs typically consist of two fully connected layers with a ReLU activation function in between, which further performs nonlinear transformations along the channel dimension, enhancing the model's expressive power. To promote gradient flow and stabilize training, this module employs residual connections.

[0039] Specifically, the module's final output is the sum of two parts: first, the data processed by layer normalization, depthwise separable convolution, and MLP path; and second, the module's initial raw input. This residual structure effectively alleviates the vanishing gradient problem in deep networks, making the network easier to train and optimize. Finally, the feature map after this series of processing is passed as output to subsequent layers or other modules in the network for further processing. The entire depthwise separable convolution module design combines normalization, efficient convolution, nonlinear transformation, and residual learning, achieving lightweight and efficient computation while ensuring performance. The module's calculation formula is expressed as: ; in, These are the initial downsampling features. This is a primary feature map. For layer normalization operation, For depthwise separable convolution operations, addition represents residual connection operations; This represents the feature map after depthwise separable convolution and residual linking.

[0040] Construct a proxy attention module. The proxy attention module consists of layer normalization, a proxy attention mechanism, an MLP, and residual connections. The specific structure can be found in [link to specific structure]. Figure 7 .

[0041] The inputs to the two proxy attention modules are the feature map data after the last two downsampling steps, respectively. The input to the first proxy attention module is the high-level downsampling feature map with a size of 320×16×16, and the input to the second proxy attention module is the output of the fourth downsampling block with a size of 512×8×8. The input feature map is first normalized by a normalization layer to make the data distribution smoother. Then the data enters the proxy attention calculation, and the proxy attention calculation process is as follows: Figure 7 As shown. The normalized data is used for the query. ,key ,value Subsequently, the data enters the agent attention calculation: First, in order to significantly reduce traditional self-attention The computational complexity is determined by the query. .

[0042] Taking the first agent attention module as an example, its dimensions are reshaped to (256, 320), where, It is the number of spatial locations. The feature dimension is used to perform adaptive average pooling, compressing the spatial size from 16×16 to 1×1, thus obtaining a surrogate vector A with dimensions (1, 320), i.e.: ; in, This indicates an adaptive average pooling operation.

[0043] Next, the proxy vector Replace the original query , and key ,value Perform the first attention calculation, based on the formula: ; in, This indicates a self-attention computation operation; Indicates the feature dimension; This represents agent-based contextual features.

[0044] Obtain agent-based context features with dimensions (1, 320). The computational complexity of this step is reduced to: .

[0045] Then, the final proxy attention calculation is performed, which applies to the original query. With proxy vector Interact and with As a value, according to the formula: ; The output is obtained, and its dimensions are restored to (256, 20), which can be reshaped back to (320, 16, 16).

[0046] Since proxy attention may introduce feature loss, the module uses a depthwise separable convolution on the original features. Local feature recovery and enhancement are performed, and finally the output of the proxy attention is added to the features processed by deep convolution through residual connections to obtain the final output of the module. Its dimensions and input Maintain consistency.

[0047] Through this design, the proxy attention module captures global contextual information at a low computational cost, and effectively stabilizes the training process and integrates multi-scale features by combining local feature recovery and residual connections. The overall calculation formula is as follows: .

[0048] in, This indicates the proxy attention calculation operation; This indicates a depthwise separable convolution operation.

[0049] The calculation process for the second agent attention module with an input of 512×8×8 is similar, with only the dimension parameters changing accordingly.

[0050] A feature extraction model combining convolution and attention was constructed. Traditional modulation recognition methods rely on a large amount of prior knowledge, making them difficult to adapt to complex wireless environments. Mainstream deep learning models each have their shortcomings: CNNs ignore signal temporal features, LSTMs have poor parallel computing capabilities, and Transformers have insufficient local modeling and high computational cost. Starting with a preprocessing module, a cascaded model is constructed using a downsampling module, a depthwise separable convolution module, a surrogate attention module, a downsampling module, a surrogate attention module, and another surrogate attention module. The depthwise separable convolution module is responsible for lightweight local feature extraction. Through a split design of "depthwise convolution + pointwise convolution," it significantly reduces the computational and parameter requirements while efficiently capturing multi-scale local features of the signal. The surrogate attention module is responsible for global key feature enhancement. Through an adaptive weight allocation mechanism, it focuses on core features that significantly contribute to the recognition results and suppresses noise and redundant information. The downsampling module plays a role in feature scale compression and hierarchical connection between modules, gradually reducing the feature map size and achieving hierarchical representation of multi-scale features.

[0051] Introducing AP dimension features of the signal as a supplementary input precisely compensates for the shortcomings of existing methods that rely solely on IQ single-dimensional features. In existing technologies, modulation recognition models almost always use IQ features as the sole input. These features cannot fully capture the joint statistical characteristics of amplitude and phase. In low signal-to-noise ratio scenarios, the details of IQ features are easily obscured by noise, leading to high confusion of constellation diagram features between similar modulation signals such as QAM16 and QAM64, 8PSK and BPSK, and a sharp decline in model recognition accuracy. To address this, this invention innovatively introduces AP dimension features as a supplementary input, enabling the model to comprehensively utilize the complementary features between different dimensions, effectively overcoming the shortcomings of incomplete single-dimensional features. Especially under low signal-to-noise ratio conditions, this strategy can significantly enhance the model's ability to distinguish subtle differences between high-order modulated signals within the same signal family, such as QAM16 and QAM64, and approximate modulated signals, such as 8PSK and BPSK. Through cross-validation and feature enhancement between multiple dimensions, the recognition accuracy is significantly improved. At the same time, the inherent information redundancy of multi-dimensional inputs reduces the model's sensitivity to noise or distortion in specific dimensions. Combined with the adaptive weighting mechanism of the weight attention fusion module, the model's generalization ability and stability in different channel environments are further enhanced, thus achieving a significant improvement in recognition performance under complex electromagnetic environments.

[0052] The WAFF module performs weighted fusion of the IQ path feature map and the AP path feature map to obtain the fused feature; Specifically, the WAFF module includes: local attention branch, global attention branch, weight generation layer, and weighted fusion layer; The WAFF module performs a weighted fusion of the IQ path feature map and the AP path feature map to obtain fused features, including: (1) Add the IQ path feature map and the AP path feature map to obtain the initial fused features; (2) Input the initial fusion features into the local attention branch and the global attention branch at the same time, and extract the local attention weights and global attention weights respectively; (3) After adding the local attention weights and the global attention weights, an attention weight map is generated by using the Sigmoid activation function; (4) The weighted fusion layer is used to adaptively weight and fuse the IQ path feature map and the AP path feature map according to the attention weight map to obtain the weighted fusion feature; (5) The weighted fusion features and the initial fusion features are added together through residual connection to obtain the final fusion features.

[0053] For example, a weighted attention feature fusion module (WAFF) is constructed. The weighted attention feature fusion module (WAFF) is used to adaptively weight and fuse information from different dimensions of the signal. This module receives two input features. Figure X IQ and X AP The dimensions are (batch size, 512, 64), where the number of channels C=512 and the sequence length L=64. First, the preliminary fusion feature X = X is obtained by element-wise addition. IQ + X AP The dimensions are maintained at (B, 512, 64). Subsequently, the fused features are simultaneously fed into the local attention branch and the global attention branch for parallel processing. In the local attention branch, the input feature X is first compressed from 512 to 128 channels by a 1×1 convolution, then batch normalized and activated by ReLU, and finally restored from 128 to 512 channels by a second 1×1 convolution, resulting in a local attention weight with dimensions (B, 512, 64).

[0054] In the global attention branch, the input feature X is first compressed from 64 to 1 using adaptive average pooling, resulting in a global feature descriptor of dimension (B, 512, 1). It then undergoes a similar convolutional processing: first, a 1×1 convolution compresses the number of channels from 512 to 128; then, batch normalization and ReLU activation are applied; finally, a 1×1 convolution restores the number of channels from 128 to 512, outputting global attention weights of dimension (B, 512, 1). The outputs of the two branches are then combined element-wise, with the global attention weights expanded from (B, 512, 1) to (B, 512, 64) dimension via a broadcasting mechanism, aligning with the local attention weights (B, 512, 64). The combined features are then processed by a Sigmoid activation function to generate an attention weight map wei ranging from [0, 1], with dimension (B, 512, 64). Finally, this weight map is used for adaptive weighted fusion of the original input features: X IQ *wei+X AP *(1-wei) yields the weighted fusion feature with dimensions (B, 512, 64). This result is then added to the initial fusion feature X via a residual connection to obtain the module's final output Y, which has the same dimensions as the input (B, 512, 64).

[0055] This module addresses the inherent shortcomings of traditional feature fusion methods by achieving adaptive dynamic fusion of IQ and AP signal features. Traditional feature fusion often employs static methods such as simple concatenation or fixed weighting. These methods lack adaptability to dynamic changes in the signal environment. When the signal faces complex situations such as signal-to-noise ratio fluctuations and interference superposition, they cannot adjust the contribution weights of different dimensional features in real time, easily leading to underestimation of effective feature information and amplification of redundant features or noise interference, thus affecting subsequent feature processing and recognition performance. The WAFF module, through the collaborative design of local and global dual-branch attention mechanisms, can accurately capture fine-grained local correlation information and overall global distribution patterns of features. Based on the real-time characteristics of the signal, it dynamically generates an attention weight matrix, thereby achieving precise control of the fusion ratio of IQ and AP features, ensuring that features with stronger discriminative power receive higher weights in different signal scenarios. At the same time, the module incorporates a residual connection design, effectively preserving original feature information and reducing information loss during the feature fusion process. This module, together with the preceding convolutional-attention feature extraction architecture and multi-dimensional input strategy, forms a closely coordinated technical loop, significantly improving the discriminative power and anti-interference capability of the fused features. It provides key feature fusion support for high-precision identification of modulated signals in complex wireless communication environments, and further enhances the robustness and environmental adaptability of the overall model.

[0056] The classification head is used to map the fused features onto the category prediction to obtain the classification result.

[0057] Specifically, the classification head includes: a global average pooling layer and a fully connected classification layer; The classification head is used to map fused features onto class predictions to obtain classification results, including: (1) Input the fused features into the global average pooling layer to compress the spatial dimension and obtain the feature vector; (2) Input the feature vector into the fully connected classification layer and output the probability distribution corresponding to each modulation category; (3) Output the category with the highest probability as the final recognition result.

[0058] The training process of a signal recognition network includes: (1) Use signal datasets that include multiple modulation types and signal-to-noise ratios, and divide them into training set, validation set and test set according to the proportions; (2) The model was trained using the cross-entropy loss function and the AdamW optimizer; (3) During training, the learning rate is adjusted using a cosine annealing strategy, and Dropout is set to prevent overfitting.

[0059] The signal dataset includes the RML2016.10a dataset, in which the modulation type includes at least one of the following: 8PSK, BPSK, CPFSK, GFSK, PAM4, 16QAM, 64QAM, QPSK, AM-DSB, AM-SSB, and WBFM.

[0060] For example, a modulation signal dataset is constructed based on the publicly available RML2016.10a dataset. RML2016.10a contains 11 modulation signals: 8PSK, BPSK, CPFSK, GFSK, PAM4, 16QAM, 64QAM, QPSK, AM-DSB, AM-SSB, and WBFM. The data format is 2 (I and Q channels) × 128, the signal-to-noise ratio range is -20dB to 18dB, with 2dB intervals, and the total number of samples is 220,000.

[0061] The dataset is divided into training, validation, and test sets in a 6:2:2 ratio. The signals undergo multi-dimensional transformation: the original I / Q dimension data samples are normalized, and then polar coordinate transformation is used to convert the complex domain signals into A / P (amplitude / phase) dimension data, enhancing the inherent separability of different signals. The transformation formula is as follows: .

[0062] During training, the "adamw" optimizer was used to update the network parameters; the cross-entropy loss function was used; dropout was set to 0.6; the learning rate was set to 4e-6 and adjusted according to the cosine annealing function; the epoch was set to 100; and the batch size was set to 128.

[0063] In a specific simulation experiment provided by this invention, the trained model is used to classify and identify signals in the test set. To demonstrate the recognition effect of this invention, the following comparison and analysis of the embodiments of this invention with current advanced modulation signal recognition methods are based on three evaluation indicators: overall recognition accuracy, confusion matrix under low signal-to-noise ratio, and model computational complexity.

[0064] Comparison of recognition accuracy of various recognition methods under different signal-to-noise ratios, for example Figure 8 As shown in Table 1, the average accuracy and maximum accuracy of each model are presented. The confusion matrix of this invention at SNRs of 10, 0, and -10 dB is shown in Table 1. Figure 9 , Figure 10 and Figure 11 As shown; from Table 1 and Figure 8It can be seen that the present invention achieves the best average accuracy, maximum accuracy, and accuracy within the 2dB~18dB range. Compared to the CV-TRN model with the highest accuracy, the average accuracy and maximum accuracy of the present invention are improved by 0.46% and 0.73%, respectively. The accuracy of the present invention is improved across most signal-to-noise ratios. Figure 9 , Figure 10 and Figure 11 It is evident that at low signal-to-noise ratios (SNR=-10dB), noise severely interferes with the signal, making it difficult for the model to extract distinctive features, resulting in a significant decrease in the recognition accuracy of most signals. Most signals are misidentified as AM-SSB, but AM-SSB still maintains a 92% accuracy rate in this environment because AM-SSB signals exhibit a noise pattern in the time domain, and their characteristics remain relatively obvious even at low SNR. At medium SNR (SNR=0dB), signal features gradually become clearer, and the model's recognition performance significantly improves. Except for AM-SSB, AM-DSB, and WBFM analog modulation signals, the recognition accuracy of other signals generally exceeds 90%, with many even reaching over 97%. At high SNR (SNR=10dB), signal features are even more pronounced, with an overall accuracy of 95.43%. At this point, except for AM-DSB and WBFM, the recognition accuracy of other signal types is almost close to or reaches 100%, with virtually no confusion. However, the silent periods included in the dataset for AM-DSB and WBFM signals still cause confusion between the two at high signal-to-noise ratios, although the degree of confusion is reduced.

[0065] Table 1 shows the results of the recognition method on the experimental dataset.

[0066] To demonstrate the high discriminative power of this invention for the same family of signals at low signal-to-noise ratios, the confusion matrices of this invention at -2dB are compared with those of other well-performing models (CV-TRN, GAM-HRNN-GRU). Figure 12 , Figure 13 and Figure 14 As shown in the diagram, the confusion matrix reveals that when the SNR is -2 dB, the GAM-HRNN-GRU has difficulty distinguishing between QAM16 and QAM64 modulations, while the CV-TRN is able to classify both. Compared to these three models, this invention exhibits the highest performance in distinguishing between QAM16 and QAM64. Except for WBFM, the recognition accuracy of this invention is acceptable across all modulation types. For all models, many WBFM samples are misclassified as AM-DSB signals.

[0067] To demonstrate the good balance between recognition accuracy and computational complexity achieved by this invention, a comparison of computational complexity between this invention and existing lightweight models is presented in Table 2. Although the model of this invention has slightly more parameters than CV-TRN and GAM-HRNN-GRU, its testing time is not higher than these two models. Table 2 also shows that although the FLOPs of this invention are higher than CNN and GAM-HRNN-GRU, they are still on the same order of magnitude as CV-TRN. At the same time, this invention maintains the highest recognition accuracy and high resolution for similar signals at low signal-to-noise ratios. These results indicate that the model provided by this invention achieves a reasonable trade-off between recognition accuracy and computational complexity.

[0068] Table 2 Comparison of Complexity of Lightweight Models

[0069] Firstly, regarding the feature extraction architecture, traditional modulation recognition methods suffer from limitations such as insufficient temporal feature capture in CNNs, low efficiency due to serial computation of LSTMs, and difficulty adapting Transformers to edge devices due to missing local modeling and high computational complexity. The convolutional-attention cascaded feature extraction model constructed in this invention, through the "depth convolution + pointwise convolution" design of a depthwise separable convolution module, efficiently captures multi-scale local spatiotemporal features while significantly reducing the number of parameters and computational load. Combined with the adaptive weight allocation of the surrogate attention module to strengthen global key features, and the downsampling module to achieve orderly connection and hierarchical expression of feature scales, this architecture not only compensates for the inherent defects of single models but also improves inference efficiency through parallel computing, significantly enhancing the comprehensiveness and accuracy of feature extraction and providing a high-quality feature foundation for subsequent recognition. Secondly, regarding the input feature dimension, existing methods rely only on single-dimensional IQ features, making it difficult to capture the joint statistical characteristics of amplitude and phase, leading to easy confusion between similar modulation signals such as QAM16 and QAM64, and 8PSK and BPSK at low signal-to-noise ratios. This invention innovatively introduces AP dimension features as a supplement, utilizing the instantaneous constellation diagram information of IQ features to complement the statistical regularity of AP features. This multi-dimensional input strategy, through information cross-validation, effectively enhances the model's ability to distinguish subtle feature differences, especially in low signal-to-noise ratio scenarios, significantly reducing the probability of confusion between similar signals and laying a data foundation for improving recognition accuracy. Finally, in the feature fusion stage, traditional static fusion methods (such as simple splicing and fixed weighting) cannot adapt to dynamic changes in the signal environment, easily leading to the underestimation of effective features or the amplification of noise interference. The WAFF weighted fusion module designed in this invention dynamically generates a weight matrix through a local and global dual-branch attention mechanism, precisely controlling the fusion ratio of IQ and AP features, and reducing information loss by combining residual connections. This module enables the fusion process to have environmental adaptability, strengthening discriminative features and suppressing redundant noise even in complex scenarios such as signal-to-noise ratio fluctuations and interference superposition, further improving the robustness and discriminative power of the fused features.

[0070] The various embodiments described in this specification are presented in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. All or part of this invention can be used in numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, mobile communication terminals, multiprocessor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0071] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the present invention.

Claims

1. A modulation signal recognition method based on convolutional attention and multi-dimensional feature fusion, characterized in that, include: The original complex signal to be identified is acquired, the original complex signal is processed to obtain a normalized IQ sequence, and the AP sequence is obtained based on the normalized IQ sequence. The IQ sequence and the AP sequence are used for signal recognition based on a pre-trained signal recognition network to obtain recognition results. The signal recognition network includes a preprocessing module, a feature extraction module, a WAFF module, and a classification head. Specifically: the preprocessing module performs preliminary shape transformations on the IQ and AP sequences to obtain processed IQ and AP feature maps; the feature extraction module extracts local detail features and global features from the processed IQ and AP feature maps to obtain IQ and AP path feature maps; the WAFF module performs weighted fusion of the IQ and AP path feature maps to obtain fused features; and the classification head maps the fused features to class prediction to obtain classification results.

2. The modulation signal recognition method based on convolutional attention and multi-dimensional feature fusion according to claim 1, characterized in that, The step of processing the original complex signal to obtain a normalized IQ sequence, and then obtaining the AP sequence based on the normalized IQ sequence, includes: The real part is extracted from the original complex signal to form the I-sequence signal, and the imaginary part is extracted to form the Q-sequence signal. The I-sequence signal and the Q-sequence signal are normalized respectively to obtain the normalized IQ sequence; The normalized IQ sequence is converted into an AP sequence.

3. The modulation signal recognition method based on convolutional attention and multi-dimensional feature fusion according to claim 1, characterized in that, The preprocessing module includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, and a channel splicing layer; The preprocessing module is used to perform preliminary shape transformations on the IQ sequence and the AP sequence to obtain processed IQ feature maps and AP feature maps, including: The normalized IQ sequence is input into the first convolutional layer, the second convolutional layer, and the third convolutional layer respectively for multi-scale feature extraction to obtain the corresponding first feature, second feature, and third feature; wherein, the convolutional kernels of the first convolutional layer and the third convolutional layer are different, and the second convolutional layer is a max pooling layer; The first feature, the second feature, and the third feature are fused together through the channel splicing layer to form an IQ feature map; The IQ feature map is transformed into an amplitude feature map and a phase feature map through polar coordinate transformation, and the two are then concatenated to obtain the AP feature map.

4. The modulation signal recognition method based on convolutional attention and multi-dimensional feature fusion according to claim 1, characterized in that, The feature extraction module includes a first stage module, a second stage module, a third stage module, and a fourth stage module connected in sequence; The feature extraction module is used to extract local detail features and global features from the processed IQ feature map and the AP feature map to obtain the IQ path feature map and the AP path feature map, including: The first stage module includes: a first downsampling block and a first depthwise separable convolution module; the first stage module performs local spatiotemporal feature extraction on the processed IQ feature map and the AP feature map to obtain the IQ path primary feature map and the AP path primary feature map; The second stage module includes: a second downsampling block and a second depthwise separable convolution module; the second depthwise separable unit performs multi-scale local feature extraction on the primary feature map of the IQ path and the primary feature map of the AP path to obtain intermediate feature maps of the IQ path and the AP path. The third stage module includes: a third downsampling block and a first proxy attention module; the third stage module introduces an attention mechanism to capture the global context dependency of the intermediate feature map of the IQ path and the intermediate feature map of the AP path, so as to obtain the IQ path-aware feature map and the AP path-aware feature map. The fourth stage module includes: a fourth downsampling block and a second proxy attention module; the fourth stage module further refines the global features of the IQ path-aware feature map and the AP path-aware feature map at the highest semantic level to obtain the IQ path high-level feature map and the AP path high-level feature map.

5. The modulation signal recognition method based on convolutional attention and multi-dimensional feature fusion according to claim 4, characterized in that, The first depthwise separable convolutional module includes: a first normalization layer, a first depthwise separable convolutional layer, a first multilayer perceptron layer, and a first residual connection layer; The first stage module performs local spatiotemporal feature extraction on the processed IQ feature map and the AP feature map to obtain a primary IQ path feature map and a primary AP path feature map, including: The initial downsampled features output by the first downsampled block are input into the first normalization layer for normalization processing; The normalized features are input into the first depthwise separable convolutional layer, and depthwise convolution and pointwise convolution are performed sequentially to extract local spatial features. The local spatial features are input into the first multilayer perceptron layer, and a nonlinear transformation of the channel dimension is performed to obtain the nonlinearly transformed features. The nonlinearly transformed features are added to the initial downsampled features through the first residual connection layer to obtain the IQ path primary feature map and the AP path primary feature map.

6. The modulation signal recognition method based on convolutional attention and multi-dimensional feature fusion according to claim 4, characterized in that, The first proxy attention module includes: a second normalization layer, a second depthwise separable convolutional layer, a second multilayer perceptron layer, and a second residual connection layer; The third-stage module introduces an attention mechanism to capture the global contextual dependencies of the intermediate feature maps of the IQ path and the AP path, resulting in IQ path-aware feature maps and AP path-aware feature maps, including: The high-level downsampling features output by the third downsampling block are input into the second normalization layer for normalization processing; The normalized features are input into the proxy attention calculation layer, and a global context-enhanced feature map is obtained by generating proxy vectors, proxy-key value attention calculation and query-proxy attention calculation; The global context-enhanced feature map is input into the second depthwise separable convolutional layer to perform local feature recovery and enhancement, thereby obtaining local enhanced features. The local enhancement features are added to the high-level downsampling features through the second residual connection layer to obtain the IQ path-aware feature map and the AP path-aware feature map.

7. The modulation signal recognition method based on convolutional attention and multi-dimensional feature fusion according to claim 1, characterized in that, The WAFF module includes: a local attention branch, a global attention branch, a weight generation layer, and a weighted fusion layer; The WAFF module performs a weighted fusion of the IQ path feature map and the AP path feature map to obtain fused features, including: The initial fused features are obtained by adding the IQ path feature map and the AP path feature map; The initial fusion features are simultaneously input into the local attention branch and the global attention branch, and the local attention weights and global attention weights are extracted respectively. The attention weight map is generated by adding the local attention weights and the global attention weights and then applying the sigmoid activation function. The weighted fusion layer adaptively weights and fuses the IQ path feature map and the AP path feature map according to the attention weight map to obtain the weighted fusion feature. The weighted fusion feature is added to the initial fusion feature through a residual connection to obtain the final fusion feature.

8. The modulation signal recognition method based on convolutional attention and multi-dimensional feature fusion according to claim 1, characterized in that, The classification head includes: a global average pooling layer and a fully connected classification layer; The classification head is used to map the fused features onto category prediction to obtain classification results, including: The fused features are input into the global average pooling layer to compress the spatial dimension and obtain the feature vector. The feature vector is input into the fully connected classification layer, and the output is the probability distribution corresponding to each modulation category; The category with the highest probability is output as the final recognition result.

9. The modulation signal recognition method based on convolutional attention and multi-dimensional feature fusion according to claim 1, characterized in that, The training process of the signal recognition network includes: Using a signal dataset that includes multiple modulation types and signal-to-noise ratios, the dataset is divided proportionally into training, validation, and test sets. The model was trained using the cross-entropy loss function and the AdamW optimizer. During training, a cosine annealing strategy is used to adjust the learning rate, and Dropout is set to prevent overfitting.

10. The modulation signal recognition method based on convolutional attention and multi-dimensional feature fusion according to claim 9, characterized in that, The signal dataset includes the RML2016.10a dataset, wherein the modulation type includes at least one of the following: 8PSK, BPSK, CPFSK, GFSK, PAM4, 16QAM, 64QAM, QPSK, AM-DSB, AM-SSB, and WBFM.

Images