A noise-resistant and robust modulated signal identification method and system

Abstract

The application discloses an anti-noise robust modulation signal identification method and system, and specifically comprises the following steps: preprocessing an IQ signal to generate an AP feature; inputting the normalized IQ signal and the AP feature into a GRU guided feature fusion module to generate a fusion feature; based on the fusion feature, performing channel calibration through a first SEBlock channel attention module, performing feature extraction and down-sampling through a convolutional network, and combining a position coding parameter to obtain a coding feature; inputting the coding feature into a multi-attention collaborative enhancement module for feature enhancement, and outputting a deep time sequence feature; inputting the deep time sequence feature into an LSTM network for sequence modeling, and outputting a prediction result of a modulation mode of the IQ signal through a classification head. The application realizes high-precision identification of modulation signals in a complex noise environment, and effectively improves the anti-noise robustness and generalization ability of modulation signal identification.

Description

Technical Field

[0001] This invention relates to the fields of communication signal processing and deep learning technology, and in particular to a noise-resistant and robust modulation signal recognition method and system. Background Technology

[0002] Modulation signal recognition is a core technology in fields such as communication reconnaissance and cognitive radio. Its core objective is to accurately identify the modulation scheme of received signals in complex noisy environments. With the development of communication technology, the types of modulation signals are becoming increasingly diverse, and real-world communication scenarios involve multi-scale noise interference (such as environments with different signal-to-noise ratios (SNR)). Traditional modulation recognition methods rely on manual feature extraction, which suffers from limited data representation capabilities and high labor costs. In contrast, deep learning-based modulation recognition methods eliminate the need for manual processing. Relying on their powerful automatic feature extraction capabilities, they are widely used in modulation classification tasks, improving the efficiency and accuracy of modulation recognition.

[0003] In the context of wireless signal modulation identification, the goal of this invention is to classify signal modulation under different signal-to-noise ratio (SNR) environments. Traditional modulation identification methods mainly rely on prior signal knowledge and complex preprocessing. The commonly used identification algorithms can be divided into two categories: likelihood-based (LB) and feature-based (FB). The former identifies the most probable modulation scheme by comparing the received signal with the likelihood functions of different modulation schemes, but it is not suitable for real-time applications. The latter utilizes features extracted from the signal and classifier training, enabling effective identification of different modulation schemes and improving adaptability and accuracy in practical communication. However, its accuracy is low for identifying high-order modulation schemes.

[0004] Deep learning-based methods have proven highly usable in various fields, including information security, the Internet of Things, vehicle prediction, and data recovery. Similarly, deep learning continues to demonstrate reliable performance in wireless signal modulation recognition. By automatically extracting features through deep neural networks, the cost of manual feature extraction is saved, and the training difficulty of the model is reduced.

[0005] Most existing deep learning-based modulation recognition methods employ classic convolutional neural networks (CNNs). While CNNs excel at local feature extraction, their limited receptive field prevents them from capturing global relational features. Furthermore, they fail to consider the complementarity between different signal representation dimensions, resulting in low feature utilization. Currently mainstream modulation recognition models include schemes combining CNNs and classic ResNet models, CNN-LSTM models that combine LSTM to learn signal data features, and Transformer models with a single attention mechanism. However, these models still suffer from insufficient capture of the sequence positional relationships of signals and inadequate utilization of representation dimensions. Therefore, current methods still cannot guarantee high recognition performance. Summary of the Invention

[0006] The purpose of this invention is to provide a noise-resistant and robust modulation signal recognition method and system. By using multi-feature intelligent fusion, multi-attention collaborative enhancement and LSTM gating mechanism, it achieves high-precision recognition of modulation signals in complex noise environments, effectively improving the noise resistance and generalization ability of modulation signal recognition, thereby solving at least one of the above-mentioned problems in the prior art.

[0007] In a first aspect, the present invention provides a noise-resistant and robust modulation signal identification method, the method specifically comprising:

[0008] The input IQ signal is preprocessed, and AP features are generated through amplitude-phase conversion. The IQ signal is then normalized.

[0009] The normalized IQ signal and AP features are input into the GRU-guided feature fusion module for adaptive weight learning and weighted calibration to generate fused features.

[0010] Based on the fusion features, channel calibration is performed through the first-level SEBlock channel attention module, feature extraction and downsampling are performed through a convolutional network, and the encoded features are obtained by combining the positional encoding parameters.

[0011] The encoded features are input into a multi-attention collaborative enhancement module that includes convolutional enhanced attention and global self-attention for feature enhancement, and secondary channel calibration is performed through a secondary SEBlock channel attention module to output deep temporal features.

[0012] Deep temporal features are input into an LSTM network for sequence modeling to capture long-term dependencies, and the prediction results of the modulation mode of the IQ signal are output through the classification head.

[0013] The model is trained using the cross-entropy loss function and an adaptive learning rate scheduler. The effectiveness of the method is verified through comparative and ablation experiments. The optimal model is saved and the recognition performance on the test set is evaluated.

[0014] Secondly, the present invention provides a noise-resistant and robust modulation signal identification system, the system specifically comprising:

[0015] The preprocessing module is used to preprocess the input IQ signal, generate AP features through amplitude-phase conversion, and normalize the IQ signal.

[0016] The feature fusion module is used to input the normalized IQ signal and AP features into the GRU-guided feature fusion module to perform adaptive weight learning and weighted calibration to generate fused features.

[0017] The feature encoding module is used to obtain encoded features based on fused features, through channel calibration by the first-level SEBlock channel attention module, feature extraction and downsampling by the convolutional network, and combined with position encoding parameters;

[0018] The feature enhancement module is used to input the encoded features into a multi-attention collaborative enhancement module containing convolutional enhanced attention and global self-attention for feature enhancement, and perform secondary channel calibration through a secondary SEBlock channel attention module to output deep temporal features.

[0019] The prediction output module is used to input deep temporal features into the LSTM network for sequence modeling to capture long-term dependencies, and output the prediction results of the modulation mode of the IQ signal through the classification head;

[0020] The model evaluation module is used to train the model using the cross-entropy loss function and an adaptive learning rate scheduler, verify the effectiveness of the method through comparative experiments and ablation experiments, save the optimal model, and evaluate the recognition performance on the test set.

[0021] Thirdly, the present invention provides a computer device, comprising: a memory and a processor, and a computer program stored in the memory, wherein when the computer program is executed on the processor, it implements a noise-resistant and robust modulation signal identification method as described in any of the above methods.

[0022] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a noise-resistant and robust modulation signal identification method as described in any of the above methods.

[0023] Compared with the prior art, the present invention has at least one of the following technical effects:

[0024] 1. This invention achieves high-precision identification of modulated signals in complex noise environments through multi-feature intelligent fusion, multi-attention collaborative enhancement and LSTM gating mechanism, effectively improving the noise robustness and generalization ability of modulated signal identification;

[0025] 2. This invention does not rely on complex manual feature design, but achieves high-precision modulation scheme discrimination in a wide signal-to-noise ratio (SNR) environment through an end-to-end deep learning architecture;

[0026] 3. This invention, by fusing multi-dimensional features and adaptively focusing on key information, can efficiently and accurately identify modulated signals over a wide SNR range;

[0027] 4. This invention generates AP features by performing amplitude-phase conversion on IQ signals and then normalizes them, providing a suitable data foundation for subsequent feature fusion and accurate identification;

[0028] 5. The GRU-guided feature fusion module of the present invention effectively fuses IQ signals and AP features through adaptive weight learning and weighted calibration, thereby improving feature quality;

[0029] 6. This invention obtains encoded features that are more conducive to subsequent processing by processing the fused features through a first-level SEBlock channel attention module, a convolutional network, and positional encoding parameters.

[0030] 7. The multi-attention collaborative enhancement module of the present invention combines convolutional enhanced attention and global self-attention, and performs secondary calibration through a two-level SEBlock channel attention module to output deep temporal features that can better characterize the signal;

[0031] 8. The convolutional enhanced attention submodule of the present invention effectively extracts and enhances local structural features through multi-scale depth-separable convolution and projection network transformation;

[0032] 9. This invention inputs deep temporal features into an LSTM network for sequence modeling and outputs prediction results through a classification head, which can capture long-term dependencies and accurately predict modulation patterns;

[0033] 10. The three-layer stacked LSTM network of the present invention updates the fused temporal information step by step through a gating mechanism, effectively extracting aggregated features for subsequent classification;

[0034] 11. This invention uses the cross-entropy loss function and an adaptive learning rate scheduler to train the model, and through experimental verification and evaluation, it can obtain an optimal model with excellent performance. Attached Figure Description

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

[0036] Figure 1This is a flowchart illustrating a noise-resistant and robust modulation signal identification method according to an embodiment of the present invention.

[0037] Figure 2 This is a diagram illustrating the overall architecture of a noise-resistant and robust modulation signal identification method according to an embodiment of the present invention.

[0038] Figure 3 This is a structural diagram of CAAttention, a convolutional enhanced attention mechanism provided in an embodiment of the present invention.

[0039] Figure 4 This is a structural diagram of an LSTM module provided in an embodiment of the present invention;

[0040] Figure 5 This is a comparison diagram between an embodiment of the present invention and existing mainstream modulation recognition network models;

[0041] Figure 6 This is a comparative ablation experiment diagram of each module provided in an embodiment of the present invention;

[0042] Figure 7 This is a graph showing the recognition accuracy of each modulated signal under different SNRs, provided by an embodiment of the present invention.

[0043] Figure 8 This is a confusion matrix diagram for identifying each modulation signal provided in an embodiment of the present invention;

[0044] Figure 9 This is a schematic diagram of the structure of a noise-resistant and robust modulation signal recognition system according to an embodiment of the present invention;

[0045] Figure 10 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0046] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0047] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0048] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0049] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0050] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0051] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0052] In this application embodiment, the entity executing the process includes a terminal device. This terminal device includes, but is not limited to, devices capable of executing the methods disclosed in this application, such as servers, computers, smartphones, and tablets. Figure 1 A flowchart illustrating a noise-resistant robust modulation signal identification method according to an embodiment of the present invention is shown below in detail:

[0053] S101 preprocesses the input IQ signal, generates AP features through amplitude-phase conversion, and normalizes the IQ signal.

[0054] In this embodiment, the input in-phase / quadrature (IQ) modulated signal is preprocessed, AP features are generated through amplitude-phase conversion, and sample-level normalization is performed on the IQ data.

[0055] S102 inputs the normalized IQ signal and AP features into the GRU-guided feature fusion module to perform adaptive weight learning and weighted calibration, generating fused features.

[0056] In this embodiment, the GRU-guided feature fusion module learns the adaptive weights of IQ and AP features, and generates highly recognizable fused features after weighted fusion.

[0057] The GRU-guided feature fusion module takes the superimposed signal of IQ and AP features as input, learns the collaborative weights of time steps and channels through the GRU network, and generates a weight matrix with the same dimension as the input features through the Sigmoid activation function. The IQ and AP features are calibrated by weighted multiplication, and then the calibrated dual features are concatenated to generate fused features, which fully explores the complementary information of different representation dimensions and improves feature recognition.

[0058] S103, based on fused features, performs channel calibration through a first-level SEBlock channel attention module, extracts and downsamples features through a convolutional network, and combines positional encoding parameters to obtain encoded features.

[0059] In this embodiment, a feature extraction network is constructed based on a convolutional neural network and a convolutional enhanced attention Transformer (CAMTransformer) to capture the local structure and long-term dependency features of the signal.

[0060] The Convolutional Enhanced Attention Transformer (CAMTransformer) integrates multi-scale convolutional attention and global self-attention. The convolutional attention uses 8 attention heads, corresponding to multi-scale convolutional kernels of 3, 5, 7...17, and extracts local structural features of the signal through depthwise separable convolution. The global self-attention captures long-term dependencies of the signal, and, together with residual connections and layer normalization, alleviates the gradient vanishing problem and enhances the ability to express deep features. A two-stage CAMTransformer cascaded structure is adopted to further enhance the feature extraction effect.

[0061] S104 inputs the encoded features into a multi-attention collaborative enhancement module that includes convolutional enhanced attention and global self-attention for feature enhancement, and performs secondary channel calibration through a secondary SEBlock channel attention module to output deep temporal features.

[0062] In this embodiment, a multi-level SEBlock channel attention mechanism and a convolutional attention mechanism are set in the model to enhance the expression of key features and suppress noise interference without significantly increasing the number of model parameters and complexity.

[0063] The multi-level SEBlock channel attention mechanism consists of a two-level structure. The first level has 4 input channels (corresponding to the feature channels after IQ and AP weighted fusion), and the second level has 64 input channels (corresponding to the channels after convolutional feature extraction). It compresses the temporal dimension features through adaptive average pooling, and generates channel weights through a link of "dimensionality reduction fully connected layer → ReLU activation function → dimension increase fully connected layer → Sigmoid activation function". This adaptively calibrates the channel feature distribution, suppresses redundant channel information, and focuses on key feature dimensions.

[0064] S105 inputs deep temporal features into an LSTM network for sequence modeling to capture long-term dependencies and outputs the prediction results of the modulation mode of the IQ signal through a classification head.

[0065] In this embodiment, the signal time series dependency is modeled using an LSTM network, and the modulation prediction result is output through a fully connected layer.

[0066] The LSTM network adopts a 3-layer stacked structure, with the input dimension and feature dimension being the same (64-dimensional). The hidden layer dimension is 64, and the output of the last time step is taken as the classification feature. The fully connected layer adopts the structure of "64→32→num_classes", combined with ReLU activation and Dropout regularization (dropout=0.1), and outputs the probability distribution of 11 modulation methods.

[0067] S106 uses the cross-entropy loss function and adaptive learning rate scheduler to train the model. The effectiveness of the method is verified through comparative experiments and ablation experiments. The optimal model is saved and the recognition performance on the test set is evaluated.

[0068] In this embodiment, a training dataset is prepared, which contains IQ signal samples of various modulation schemes under different signal-to-noise ratio environments. Each sample is labeled with its corresponding real modulation scheme label. The dataset is divided into a training set, a validation set, and a test set according to a certain ratio to ensure that the samples of various modulation schemes are evenly distributed in each dataset, thereby ensuring the stability of model training and the accuracy of evaluation.

[0069] During model training, the cross-entropy loss function is used as the objective function. The cross-entropy loss function measures the difference between the model's predictions and the true labels; the smaller the value, the closer the model's predictions are to reality. Specifically, for each IQ signal sample input to the model, the model outputs a predicted probability distribution, representing the probability that the sample belongs to each modulation scheme. The cross-entropy loss function calculates a loss value by comparing the difference between this predicted probability distribution and the one-hot encoding of the true labels. During training, the model parameters are continuously adjusted to gradually reduce this loss value, thereby improving the model's prediction accuracy.

[0070] Meanwhile, to accelerate the model's convergence process and improve training effectiveness, an adaptive learning rate scheduler is employed. The adaptive learning rate scheduler automatically adjusts the learning rate based on the model's performance during training. In the early stages of training, a larger learning rate is set so that the model can quickly learn the general features of the data; as training progresses and the model approaches convergence, the learning rate gradually decreases, allowing for more precise parameter adjustments and preventing the model from missing the optimal solution. For example, a cosine annealing learning rate scheduler can be used, where the learning rate gradually decreases according to a cosine function as the number of training epochs increases, effectively balancing training speed and model accuracy.

[0071] In each training round, IQ signal samples from the training set are sequentially input into the model. First, the input IQ signals are preprocessed, generating AP features through amplitude-phase conversion and normalizing the IQ signals. Next, the normalized IQ signals and AP features are input into a GRU-guided feature fusion module for adaptive weight learning and weighted calibration, generating fused features. Then, based on the fused features, channel calibration is performed through a first-level SEBlock channel attention module, followed by feature extraction and downsampling via a convolutional network, combined with positional encoding parameters to obtain encoded features. These encoded features are then input into a multi-attention collaborative enhancement module containing convolutional enhanced attention and global self-attention for feature enhancement, and a second-level SEBlock channel attention module performs secondary channel calibration, outputting deep temporal features. Finally, these deep temporal features are input into an LSTM network for sequence modeling to capture long-term dependencies, and the prediction result of the IQ signal modulation mode is output through a classification head. The cross-entropy loss value is calculated based on the prediction results and the true labels, and the loss value is backpropagated back into the model using a backpropagation algorithm to update the model parameters. After each round of training, the model's performance is evaluated using a validation set, and metrics such as accuracy and loss value on the validation set are recorded.

[0072] Continue training the model until its performance on the validation set no longer shows significant improvement or the preset number of training epochs is reached. During training, save multiple models from different training stages. After training, select the model with the best performance on the validation set from these saved models as the candidate optimal model. The specific selection criteria can be determined according to actual needs, such as selecting the model with the highest accuracy on the validation set, or a selection based on a combination of factors such as accuracy and loss value.

[0073] To verify the effectiveness of the proposed method, comparative and ablation experiments were conducted.

[0074] Comparative Experiments: Several existing mainstream modulation recognition models were selected as comparison models, such as the scheme combining CNN and the classic ResNet model, the CNN-LSTM network model combining LSTM to learn signal data features, and the Transformer network model with a single attention mechanism. These comparison models were trained and evaluated on the same training, validation, and test sets using the same training parameters and methods. The performance of the optimal model of the proposed method was compared with that of these comparison models on the test set, with comparison metrics including accuracy, recall, and F1 score. The comparative experiments clearly demonstrate the advantages and improvements of the proposed method compared to existing methods.

[0075] Ablation experiments: To verify the effectiveness of each module in the proposed method, ablation experiments were conducted. These ablation models were trained and evaluated on the same training, validation, and test sets using the same training parameters and methods. The performance of the ablation model on the test set was compared with that of the complete model on the test set to analyze the contribution of each module to the model's performance. Ablation experiments provide a deeper understanding of the role and importance of each module in the proposed method, further validating the rationality of the method design.

[0076] After verifying the effectiveness of the proposed method through comparative and ablation experiments, the optimal model was determined. A test set was then used to perform a final performance evaluation of this optimal model. The test set contains data that the model has never encountered during training and validation, thus accurately reflecting its performance in real-world applications. Evaluation metrics included accuracy, recall, and F1 score. Evaluating the performance on the test set allows for a precise understanding of the proposed method's performance in real-world scenarios, providing a basis for subsequent applications and improvements. Furthermore, comparing the evaluation results of the optimal model with those of existing methods further highlights the advantages and innovations of the proposed method.

[0077] As an overview of the above steps S101 to S105, refer to Figure 2The overall architecture of the noise-resistant robust modulation signal recognition method of the present invention includes the following modules: (1) data preprocessing module; (2) GRU-guided feature fusion module; (3) multi-attention enhancement module; (4) sequence modeling module and classification module. The first part, the data preprocessing module, firstly preprocesses the original IQ signal (representing a complex signal containing in-phase component I and quadrature component Q) to generate AP features (the core features extracted after preprocessing, corresponding to the two key physical properties of the IQ signal: A (Amplitude) and P (Phase)), and at the same time normalizes the IQ signal. The second part, the GRU-guided feature fusion module, uses GRU to weightedly fuse IQ and AP features, and the weighted fused IQ and AP features effectively enhance their respective features. The third part, the convolutional enhanced attention Transformer and SE channel attention work together to extract deep features. The fourth part, the sequence modeling module, uses LSTM to model sequence dependencies and complete classification, ensuring the robustness of the model over a wide SNR range throughout the process.

[0078] First, amplitude-phase transformation is performed on the input modulation signal IQ data (dimension (N, 2, 128), where N is the number of samples, 2 corresponds to the IQ component, and 128 is the time step) to generate AP features (dimension consistent with IQ). Then, sample-level normalization is performed on the IQ data, and the maximum value of each sample is calculated to standardize the IQ data to the [-1, 1] interval to avoid numerical drift affecting training stability.

[0079] After the input data is constructed, the superimposed signal of IQ and AP features is used as input, and the collaborative weights of time step and channel are learned through GRU network. The result obtained by GRU is then used to generate an adaptive weight matrix through Sigmoid activation function, which is used to weight and enhance IQ and AP features respectively to achieve feature calibration. The weighted IQ features and AP features are concatenated to generate fused features, which fully explores the complementarity of the two features.

[0080] After feature fusion, the SEBlock module is connected, which compresses the time dimension features through adaptive average pooling, generates channel weights through two fully connected layers and a Sigmoid activation function, and performs channel importance verification on the 4-dimensional fused features to suppress redundant information. Then, the multi-head local convolution in the CAMTransformer module extracts local structural features and captures global dependencies through standard multi-head self-attention, and combines them with a feedforward network to achieve nonlinear feature enhancement.

[0081] The features output by the Transformer are input into a bidirectional LSTM network to capture the long-term time series dependence of the signal. The output features of the last time step of the LSTM are taken, and after passing through a fully connected layer, ReLU activation function and Dropout regularization, the probability distribution of various modulation methods is output.

[0082] In some embodiments, step S101 above, which involves preprocessing the input IQ signal to generate AP features through amplitude-phase conversion and normalizing the IQ signal, specifically includes:

[0083] Obtain raw IQ sample data, which includes in-phase component I and quadrature component Q;

[0084] Amplitude-phase transformation is performed on the IQ sample data, and amplitude and phase features are calculated based on the in-phase component I and quadrature component Q to form the AP feature matrix;

[0085] The IQ sample data are normalized to the maximum value at the sample level, and the in-phase component I and quadrature component Q of each sample are scaled to the interval [-1,1].

[0086] In this embodiment, the RadioML2016.10a public dataset is used, containing 11 modulation schemes and 21 SNR levels (-20dB to 18dB, with a 2dB interval). Each "modulation scheme + SNR" combination corresponds to 4000 samples, and each sample is IQ data (2×128 dimension) with 128 time steps, where I represents the in-phase (I) component of the signal data and Q represents the quadrature (Q) component of the signal data. In the data preprocessing stage, this invention normalizes the IQ data to the [-1, 1] interval using a sample-level maximum value normalization strategy to avoid numerical overflow. Simultaneously, amplitude-phase transformation is performed on the IQ data to generate an AP feature matrix, forming a dual-feature input with the IQ data. The specific calculation formula for converting the IQ signal into amplitude and phase features through mathematical transformation is as follows:

[0087] ;

[0088] ;

[0089] Where amp represents amplitude characteristics, phase represents phase characteristics, arctan2 represents the arctangent function in the four quadrants, I represents the in-phase component of the IQ signal, and Q represents the quadrature component of the IQ signal.

[0090] In some embodiments, step S102 above, which involves inputting the normalized IQ signal and AP features into the GRU-guided feature fusion module for adaptive weight learning and weighted calibration to generate fused features, specifically includes:

[0091] The normalized IQ signal and AP feature are superimposed element-wise to generate superimposed features;

[0092] The superimposed features are transposed to generate a feature sequence that conforms to the input format of the GRU network.

[0093] The feature sequence is input into the GRU network, which learns the cooperative relationship between time steps and channels, and outputs the hidden state sequence corresponding to the time step.

[0094] The hidden state sequence is transposed and processed by the Sigmoid activation function to generate an adaptive weight matrix corresponding to the input feature dimension.

[0095] The normalized IQ signal and AP feature are weighted element-wise using an adaptive weight matrix to obtain the weighted and calibrated IQ signal and AP feature.

[0096] The weighted and calibrated IQ signal and AP features are concatenated along the channel dimension to output the fused features.

[0097] In this embodiment, traditional feature fusion methods (such as direct concatenation and simple addition) do not consider the correlation and redundancy between features, resulting in poor fusion performance. This invention designs a method that adaptively learns the weights of IQ and AP features using a gated recurrent unit (GRU) to achieve dynamic calibration and efficient fusion.

[0098] First, the IQ and AP features are element-wise superimposed to obtain a fused input matrix. Then, the matrix is ​​adapted to the input format of the GRU network by dimensional transposition (the GRU input dimension is "batch × time step × number of features"). The specific calculation is as follows:

[0099] ;

[0100] ;

[0101] Formula (3) achieves element-wise superposition of IQ and AP features to obtain Enhance feature correlation. Let T(1,2) represent the intermediate feature tensor after the fusion of IQ and AP features; in formula (4), T(1,2) represents the first and second dimensions of the exchanged features. Indicates to The tensor after transposing the first and second dimensions, The dimension is The matrix ( (where 128 represents the batch size, 2 represents the feature dimension, and 2 represents the feature branches). The original "2×128" channel-time step dimension is converted to a "128×2" time step-feature number dimension, resulting in a GRU network input that meets the requirements. , This represents the input feature tensor of the GRU (Gated Recurrent Unit). The superimposed features are processed by the GRU network for time-step dependency learning, outputting a feature matrix with the same time-step length as the input. A weight matrix is ​​then generated using the Sigmoid activation function. Since GRU and LSTM have similar structures, but GRU has one less gate structure, resulting in lower computational complexity, and because this part of the invention only involves time-step weighting, GRU is used to achieve this effect while reducing complexity. The specific calculation is as follows:

[0102] ;

[0103] ;

[0104] In formula (5), the GRU network adaptively captures the dependencies between time steps through update gates and reset gates. Indicates a gated loop unit. This represents the output feature matrix of the GRU. Indicates to The tensor after transposing the first and second dimensions, The formula (6) represents the sigmoid activation function. Formula (6) restores the GRU output to a "2×128" channel-time step dimension through dimension transpose, and then generates a weight matrix W through sigmoid activation. A larger weight value indicates a more important feature at the corresponding location. Next, the learned weight matrix is ​​used to perform element-wise weighted calibration on the IQ and AP features respectively, suppressing redundant information and strengthening key features. Finally, the features are concatenated along the channel dimension to generate the final fused feature. .

[0105] In some embodiments, in step S103 above, the step of obtaining encoded features based on fused features, performing channel calibration through a first-level SEBlock channel attention module, extracting and downsampling features through a convolutional network, and combining positional encoding parameters, specifically includes:

[0106] The fused features are input into the first-level SEBlock channel attention module, and the time dimension is compressed through adaptive average pooling to obtain global channel features;

[0107] The global channel features are processed sequentially through a dimension-reducing fully connected layer, a ReLU activation function, a dimension-upgrading fully connected layer, and a Sigmoid activation function to generate channel weight vectors.

[0108] The channel weight vector is multiplied element-wise with the fused feature to complete the channel calibration, and the calibrated feature is output.

[0109] The calibrated features are input into a convolutional network containing two one-dimensional convolutional layers. Convolution, batch normalization and ReLU activation operations are performed sequentially to achieve feature extraction and downsampling, resulting in the first convolutional feature.

[0110] The first convolutional feature is transposed in dimension and added to the learnable positional encoding parameters to generate encoded features.

[0111] In this embodiment, the 4-channel features after GRU fusion are targeted. There is a channel redundancy issue. Before convolutional feature extraction, SEBlock adaptively learns channel weights to strengthen important channels and suppress redundant channels, providing more accurate input features for subsequent convolutional extraction. Specifically, adaptive average pooling is first performed to compress the 128 time-step features of each channel into one global feature, achieving global aggregation of channel information. Then, a channel activation network is constructed through two fully connected layers to learn the dependencies between channels and generate accurate channel weights, as shown in the following formula:

[0112] ;

[0113] ;

[0114] in, This represents the feature vector after global average pooling, used to process the fused 4-channel features. Take a global average along the sequence length dimension. This represents a 1D adaptive average pooling layer. The target length of the pooling is indicated and is used to specify that the length dimension of the input sequence is compressed to 1. Linear1 indicates a dimension-reducing fully connected layer, and Linear2 indicates a dimension-upgrading fully connected layer. ReLU indicates the ReLU activation function, which introduces non-linearity, enabling the network to learn complex dependencies between channels. This represents the Sigmoid activation function, which restricts the weight output to the range of 0 to 1, thereby quantifying the channel importance. This represents a tensor shape transformation, used to transform... The dimensions were adjusted from the original shape to The 2D tensor. Next, the channel weights are expanded to the dimension of fused feature matching, and adaptive calibration of each channel is achieved through element-wise multiplication, as shown in the following formula:

[0115] ;

[0116] in, This represents the channel-weighted feature tensor output by the SE module. This represents element-wise multiplication. This represents the channel attention weights generated by the SE module. Indicates to A tensor dimension expansion operation is performed to add a dimension in the second dimension (time step dimension), expanding the weight matrix from B×4 to B×4×1. Then, element-wise multiplication is performed through a broadcast mechanism. After calibration, important channel features are enhanced, and redundant channel features are suppressed, significantly improving the quality of the fused features.

[0117] The fused features after SEBlock calibration need to be processed by a convolutional network to extract deep abstract features, and positional encoding is used to supplement temporal step positional information, providing high-quality input for the subsequent attention mechanism. Specifically, this is achieved by constructing a concatenated convolutional network to process the calibrated features. After downsampling and feature abstraction, the network structure consists of two Conv1d convolutions, as shown in the following formula:

[0118] ;

[0119] ;

[0120] in, This represents the output feature tensor of the first convolutional block. ReLU represents the output feature tensor of the second convolutional block, and ReLU represents the ReLU activation function. This represents a one-dimensional batch normalization layer. Representing a one-dimensional convolutional layer, in In this example, setting the kernel size to 4 and the stride to 2 halves the time step length after each convolutional layer, increasing the number of channels from 4 to 64, thus achieving spatial compression and dimensionality enhancement of features. Standardizing input features accelerates model convergence, while Dropout suppresses overfitting by randomly deactivating some neurons. Since SEBlock pre-calibrates channel features, the convolutional network efficiently extracts deep abstract features. Because the subsequent Transformer module lacks time-step position awareness, positional encoding needs to be manually added to supplement the temporal order information. Therefore, learnable positional encoding parameters need to be constructed, aligned with the time-step dimension of the convolutional features, and then added together. The specific formula is as follows:

[0121] ;

[0122] in, This represents the input feature tensor of the Transformer module. This represents the tensor transpose operation, used to swap the first and second dimensions of a tensor. PosEmb represents the output feature tensor after the dimension transpose operation of the second-layer one-dimensional convolutional block. It is a learnable positional encoding that matches the time step and number of channels of the convolutional features. By adding it to the positional encoding, the model can distinguish features at different time steps and capture the temporal dependencies of the signal.

[0123] In some embodiments, in step S104 above, the step of inputting the encoded features into a multi-attention collaborative enhancement module that includes convolutional enhanced attention and global self-attention for feature enhancement, and performing secondary channel calibration through a secondary SEBlock channel attention module to output deep temporal features specifically includes:

[0124] The encoded features are input into the convolutional enhanced attention submodule, which extracts local structural features through multi-scale depthwise separable convolution, and after transformation by a projection network, they are weighted with the value features to output local enhanced features.

[0125] The local enhancement features and the encoded features are residually connected and layer normalized to obtain the first residual feature;

[0126] The first residual feature is simultaneously used as a query, key and value input to the global self-attention submodule. The global dependency relationship is captured through the multi-head attention mechanism, and the global dependency feature is output.

[0127] The global dependency feature and the first residual feature are joined by residual connection and layer normalization to obtain the second residual feature;

[0128] The second residual feature is nonlinearly transformed through a feedforward network to output a nonlinear enhanced feature.

[0129] The nonlinear enhancement features are input into the second-level SEBlock channel attention submodule. Channel weights are generated through adaptive average pooling, fully connected layers, and Sigmoid activation to calibrate the feature channels and output deep temporal features.

[0130] Furthermore, the step of inputting the encoded features into the convolutional enhancement attention submodule, extracting local structural features through multi-scale depthwise separable convolution, and then weighting them with the value features after transformation by a projection network to output local enhancement features, specifically includes:

[0131] Linear projection is performed on the encoded features to generate value feature matrix and spatial feature matrix respectively;

[0132] The spatial feature matrix is ​​split and its dimensions are transformed according to the preset number of attention heads to obtain multiple parallel multi-head feature maps;

[0133] Each multi-head feature map is assigned a convolutional kernel of a different scale, and depthwise separable convolution is applied to process them to obtain the corresponding convolutional output features.

[0134] All convolutional output features are concatenated along the channel dimension to obtain a second convolutional feature that integrates multi-scale information;

[0135] The second convolutional feature is input into the projection network, and then passes through the first linear layer, Gaussian error linear unit activation function, layer normalization, the second linear layer and nonlinear transformation in sequence to output the projection feature;

[0136] The projected features are multiplied element-wise with the value feature matrix to complete the weighted enhancement of local structural features and output the local enhanced features.

[0137] In this embodiment, to address the problem of key features being interfered with in complex noisy environments, a three-level attention mechanism of "convolutional enhanced attention + global self-attention + secondary channel attention" is designed to enhance feature representation from three dimensions: local structure, global dependency, and channel optimization. The specific implementation is as follows:

[0138] First, there's convolution-enhanced attention. The structure diagram for convolution-enhanced attention is as follows: Figure 3 As shown. To address the local structural features of the modulated signal (such as abrupt amplitude changes and phase transitions at specific time steps), a multi-scale convolutional attention enhancement mechanism is designed, extracting local features through multi-head, multi-scale convolution. First, the position-encoded features ( Linear projection is performed to generate the Value matrix and Spatial matrix, which are used for subsequent feature weighting and local convolution enhancement, respectively. Next, the Spatial matrix is ​​split into 8 parallel sub-feature maps based on the number of attention heads (num_heads=8), with each sub-feature map having 8 channels. Different sized convolution kernels are assigned to each attention, and multi-scale local features are extracted through depthwise separable convolution, as shown in the following formula:

[0139] ;

[0140] ;

[0141] Formula (13) will perform S and through The dimensional transformation ultimately resulted in the splitting into 8 multi-head feature maps. S represents the feature tensor output from the preceding module. This means rearranging S to a dimension of The four-dimensional tensor (B represents the batch size, 32 represents the number of feature groups, and 8,8 represents the sub-feature dimension). This indicates a dimension permutation operation, used to rearrange the dimensions of the reshaped tensor from [B,32,8,8] to [8,B,8,32]. This represents a one-dimensional convolutional layer. This represents the permuted four-dimensional tensor. express The i-th feature group slice, Represents the output tensor of the convolution; Formula (14) enables depthwise separable convolution, that is, each feature map is convolved with different convolution kernels to obtain the corresponding tensor. This reduces the number of parameters while improving the targeting of local feature extraction. Next, the outputs of the eight multi-head convolutions are concatenated along the channel dimension and then subjected to a non-linear transformation through a feature projection network (Proj), as shown in the following formula:

[0142] ;

[0143] ;

[0144] in, This represents the one-dimensional convolutional output features of groups 0 to 7, and cat indicates channel-dimensional concatenation, merging the 8 multi-head feature maps; Proj represents the output features of the projection layer. This represents the input feature tensor of the projection layer. This indicates the Dropout layer. Indicates a linear layer. Presentation layer normalization, used for normalization operations. The network achieves non-threaded enhancement and dimension alignment of local features through the process of formula (16). represent The activation function is the Gaussian error linear unit. Finally, the projected local features are compared with the value features. This refers to element-wise multiplication, which achieves weighted enhancement of local structural features. The formula is as follows:

[0145] ;

[0146] in, This represents the feature tensor after channel attention weighting. This represents the input feature tensor.

[0147] Next is global self-attention (SA) and residual fusion: For the long-term time-step dependency of the modulated signal, standard multi-head attention is used to capture the global dependency, combined with residual connections and layer normalization to alleviate gradient vanishing. Specifically, the CAAttention output features ( Simultaneously serving as a query, key, and value, the global attention weights are calculated and summed using the following formula:

[0148] ;

[0149] ;

[0150] in, For each dimension of attention head, This means splitting Q, K, and V into 8 multi-head parallel attention computations, and then concatenating them to output global self-attention features. This represents the output feature tensor of a single-head self-attention system after weighted fusion of V through the interaction of query (Q), key (K), and value (V). This represents the query-key relevance score matrix. Indicates the scaling factor. This represents the Softmax activation function. This represents the attention features after weighted fusion. The output feature tensor representing multi-head attention. This indicates the number of attention heads. To avoid gradient vanishing during deep network training, residual connections and layer normalization are added after the attention module. Finally, a feedforward network is used to further enhance the nonlinear expressive power of the features. Due to the use of a two-stage CAMTransformer cascade, the local and global information of the features are fully integrated, resulting in a better final output. For output features SEBlock channel attention is applied again for secondary channel calibration to adapt to the 64-channel dimension of the convolutional features. For the specific process, please refer to the first-level SEBlock. Output features .

[0151] In some embodiments, step S105 above, which involves inputting deep temporal features into an LSTM network for sequence modeling to capture long-term dependencies and outputting the prediction result of the modulation mode of the IQ signal through a classification head, specifically includes:

[0152] The deep temporal features are transposed to obtain a temporal sequence that conforms to the input format of the LSTM network.

[0153] The time series is input into a three-layer stacked LSTM network. Through the forget gate, input gate, output gate and cell state gating mechanism, the time series information is updated and fused step by step, and the hidden state of the last time step is output as the aggregated feature.

[0154] The aggregated features are input into the classification head, and the features are subjected to dimensionality reduction and non-linear mapping. The classification head includes a first fully connected layer, a ReLU activation function, a Dropout layer and a second fully connected layer connected in sequence.

[0155] The second fully connected layer outputs a probability distribution corresponding to a preset modulation scheme type, and the category with the highest probability is used as the prediction result of the modulation scheme of the IQ signal.

[0156] Furthermore, the step of inputting the time-series sequence into a three-layer stacked LSTM network, updating and fusing the time-series information step by step through forget gates, input gates, output gates, and cell state gating mechanisms, and outputting the hidden state of the last time step as the aggregated feature, specifically includes:

[0157] At each time step, the input features of the current time step are concatenated with the hidden state of the previous time step to form concatenated features.

[0158] The concatenated features are input into the forget gate, and after linear transformation and Sigmoid activation function processing, the forget gate output vector is generated.

[0159] The spliced ​​features are simultaneously input into the input gate and the candidate cell state calculation module. The input gate output vector is obtained by linear transformation and Sigmoid activation function, and the candidate cell state at the current time step is obtained by linear transformation and Tanh activation function.

[0160] Based on the forget gate output vector, the cell state of the previous time step, the input gate output vector, and the candidate cell states, the cell state of the current time step is updated by weighted fusion.

[0161] The concatenated features are input into the output gate, and after linear transformation and Sigmoid activation function processing, the output vector of the output gate is generated.

[0162] The output vector of the output gate is multiplied element-wise with the result of applying Tanh activation to the cell state at the current time step to generate the hidden state at the current time step.

[0163] All time steps are processed sequentially, and the hidden state generated by the third-layer LSTM network in the last time step is used as the aggregated feature.

[0164] In this embodiment, the modulated signal is essentially time-series data. A Long Short-Term Memory (LSTM) network is used to capture long-term dependencies between time steps, and then a fully connected layer completes the modulation classification. Specifically, this is achieved by processing the features output by the second-level SEBlock. The LSTM network structure diagram is shown below, after dimensionality transpose to adapt the input features to the LSTM. Figure 4As shown. LSTM mainly achieves temporal memory through three gates and one cell state update. The three gates are the forget gate, input gate, and output gate. The forget gate determines how much of the cell state from the previous time step is retained in the current time step, the input gate determines how much of the current network's input is saved to the cell state, and the output gate controls how much of the cell state is output to the LSTM's current output value. The specific process is shown in the formulas, and all formulas are executed sequentially at time step t:

[0165] ;

[0166] This formula preserves long-term historical memories through the forgetting gate, in which... This represents the concatenation of the previous hidden state with the current input features. This represents the forget gate weight matrix. This is the output vector of the forget gate, with each dimension corresponding to a feature channel of the cell state, responsible for determining whether to retain the "temporal pattern memory" of the previous time step. This represents the forget gate bias coefficient. This represents the Sigmoid activation function;

[0167] ;

[0168] This formula controls whether to store the new feature at the current time step via an input gate. Its main function is to filter whether the key features at the current time step should be stored in long-term memory. Indicates the input gate weights. This represents the input gate bias coefficient. This represents the Sigmoid activation function. This represents the feature value after input gate filtering. This represents the concatenation of the previous hidden state with the current input features;

[0169] ;

[0170] This formula is responsible for refining new memories from candidate cell states. It can process the input features at the current time step. The most discriminative temporal information is refined and extracted, serving as candidate memories to be stored in the cellular state. Represents the candidate cell state weights. This represents the concatenation of the previous hidden state with the current input features. This represents the hyperbolic tangent activation function. Indicates the bias coefficient of the linear transformation layer;

[0171] ;

[0172] This formula updates the cell state, which is the core of LSTM. This indicates the result of the cell state update. This represents the output vector of the forget gate. This indicates the information of the old cells from the previous time step. This represents the feature value after input gate filtering. This represents the candidate cell state. In modulation recognition, It will continuously retain the long-term memory of "smooth frequency changes", and at the same time, it will store the "amplitude fluctuation characteristics" of the current time step through the input gate, ultimately forming a unique temporal memory for the modulation category;

[0173] ;

[0174] This formula controls whether the output gate outputs long-term memory to the hidden state, where... This represents the concatenation of the previous hidden state with the current input features. Indicates the output gate weights. This represents the output gate bias coefficient. This represents the characteristic state of the output. This represents the Sigmoid activation function. Its main function is to filter which information from long-term memory is used for the output of the current time step. For example, in the classification stage, if the last phase transition feature of QPSK is needed, then the last time step's... It will approach 1, and the phase transition memory stored in the cell state will be output to the hidden state;

[0175] ;

[0176] This formula represents the hidden state output, i.e., the short-term memory at the current time step. This represents the output at time step t. This represents the hyperbolic tangent activation function, at each time step. They are all fusions of local temporal features and long-term dependencies, with 32 time steps. Together, they constitute the complete temporal feature sequence of the modulated signal. Finally, the hidden state of the last time step of the third-layer LSTM is taken as the input of the classification head, and the probability distribution of 11 modulation modes is output.

[0177] In some embodiments, step S106 above, which involves training the model using a cross-entropy loss function and an adaptive learning rate scheduler, verifying the effectiveness of the method through comparative and ablation experiments, saving the optimal model, and evaluating the recognition performance on the test set, specifically includes:

[0178] The original dataset is divided into a training set, a validation set, and a test set;

[0179] The cross-entropy loss function is set as the objective function for model training, and the objective function is used to calculate the difference between the predicted result and the true label;

[0180] Configure an optimizer and use an adaptive learning rate scheduler to dynamically adjust the learning rate during training;

[0181] During training, an early stopping strategy is used to save the parameters of the optimal model based on the validation set loss;

[0182] The optimal model is evaluated using a test set to obtain the overall recognition accuracy and performance metrics at various signal-to-noise ratios.

[0183] In this embodiment, the model-related training configuration of the present invention is shown in Table 1. The training process initializes model parameters by fixing a random seed and moving the model to the target device (GPU). Training is conducted through epoch rounds, each round including a training phase and a validation phase. During the training phase, the model is set to training mode and gradient calculation is enabled; the training set data loader is traversed, the network training process is executed, the training loss and accuracy of each batch are recorded, and finally the average training loss and accuracy of each epoch are calculated. During the validation phase, the model is set to evaluation mode and gradient calculation is disabled; the validation set data loader is traversed, only forward propagation is performed, and the validation loss and accuracy are calculated; an early stopping strategy is adopted, i.e., if the validation loss does not decrease for 10 consecutive epochs, the process stops early, and the checkpoint and the best model are saved according to the saving strategy. The inference process loads the trained best model weights, sets the model to evaluation mode, disables gradient calculation, and performs performance evaluation on the test set to obtain the overall accuracy, the accuracy at each signal-to-noise ratio, and the confusion matrix for each modulation method.

[0184] Analysis of experimental results:

[0185] The method of this invention was compared with six other deep learning-based modulation recognition methods. These include ResNet, MCLDNN, MCformer, LSTM2, LSTM-DAE, and FEAT. Specifically, the inputs of ResNet, MCLDNN, and MCformer are the original I / Q signals, the inputs of LSTM2 and LSTM-DAE are the A / P signals, and the input of FEAT is the embedding tag generated by frame extraction of the I / Q signal through a sliding window. Table 1 shows the parameter count, complexity, and accuracy of all models in RadioML2016.01a. This invention defines -6dB as a low signal-to-noise ratio.

[0186] Table 1. Performance comparison of all methods;

[0187]

[0188] As shown in Table 1 and Figure 5As demonstrated, the model architecture proposed in this invention outperforms existing methods at all signal-to-noise ratio (SNR) levels in RadioML 2016.01a, achieving an average accuracy of up to 63.48%. Particularly noteworthy is the excellent performance of this model in the low to medium SNR range, where its performance is comparable to MCformer and significantly superior to other methods. Under low SNR conditions of -6 dB, the accuracy of this model reaches 60.95%. In high SNR scenarios, the model achieves a maximum accuracy of 93.95% at 14 dB, while MCformer only maintains a maximum accuracy of 89.3%. This demonstrates the unique advantages of the proposed model in noise handling, exhibiting excellent performance not only in low SNR environments but also maintaining good generalization ability across all noise levels.

[0189] In the complexity analysis, by comparing the number of parameters and floating-point operations per second of different models, it can be found that although the model of this invention has a higher number of parameters than ResNet, LSTM-DAE, and MCformer, it achieves improved accuracy. Furthermore, the model of this invention has fewer parameters than other models, indicating that it does not require a larger memory space. In terms of computational speed, the model of this invention significantly reduces computational cost compared to LSTM2 and MCLDNN.

[0190] In addition, ablation experiments were conducted in this invention. To verify the accuracy of each module in this invention, ablation experiments were performed to compare network architectures without the CATransformer module, network architectures without the GRU module, and network architectures without the LSTM module. See Table 2 and... Figure 6 The results shown are comparative ablation experiments of the present invention.

[0191] Table 2 Comparison of ablation test performance;

[0192]

[0193] Therefore, it can be seen that all the modules used in this invention play a role, especially the GRU fusion module, which, with only a 0.04k increase in parameters, achieved a 0.5% improvement in accuracy, demonstrating that this module effectively enhances the features. To more comprehensively demonstrate the indicators of this invention, such as... Figure 7 , Figure 8 The accuracy of this invention at each signal-to-noise ratio for 11 modulation schemes in the dataset and the final confusion matrix of this invention are shown respectively. The equipment configuration used for testing is as follows: 13th Gen Intel(R) Core(TM) i5-13400F CPU @ 2.50GHz and NVIDIA GeForce RTX 4070.

[0194] In summary, this invention proposes a noise-resistant and robust modulation signal recognition method and system. Addressing the issues of easily interfered temporal features of modulation signals and low efficiency in multi-feature fusion under complex noise environments, it achieves high-precision and efficient modulation scheme recognition. Specifically, the model first uses a GRU-guided intelligent multi-feature fusion module to adaptively learn the weights of IQ and AP features, dynamically calibrating and fusing the dual features to mine complementary information in the temporal and physical properties of the signal. Before convolutional feature extraction, a first-level SEBlock channel attention is introduced to strengthen important channels and suppress redundant channels, improving the efficiency of convolutional feature extraction. Subsequently, convolutional enhanced attention (CAAttention) combined with global self-attention and second-level SEBlock channel attention strengthens feature expression from three dimensions: local structure, global dependency, and channel optimization, accurately capturing the temporal patterns of the modulation signal. Finally, a three-layer LSTM is used to capture long-term temporal dependencies of the signal, further aggregating key features. This model optimizes recognition performance under different signal-to-noise ratio environments without significantly increasing the number of parameters, while also balancing training and inference efficiency through reasonable network structure design. Experimental results show that the model of the present invention is superior to similar modulation signal recognition models. In particular, the proposed model has excellent noise robustness while maintaining high recognition accuracy, and can be adapted to the modulation signal recognition needs under different signal-to-noise ratio scenarios.

[0195] Reference Figure 9 An embodiment of the present invention provides a noise-resistant robust modulation signal identification system 9 for performing a noise-resistant robust modulation signal identification method as described in any of the above methods. The system 9 specifically includes:

[0196] The preprocessing module 901 is used to preprocess the input IQ signal, generate AP features through amplitude-phase conversion, and normalize the IQ signal.

[0197] The feature fusion module 902 is used to input the normalized IQ signal and AP feature into the GRU-guided feature fusion module to perform adaptive weight learning and weighted calibration to generate fused features.

[0198] The feature encoding module 903 is used to obtain encoded features based on fused features, through channel calibration by the first-level SEBlock channel attention module, feature extraction and downsampling by the convolutional network, and combined with position encoding parameters;

[0199] The feature enhancement module 904 is used to input the encoded features into a multi-attention collaborative enhancement module containing convolutional enhanced attention and global self-attention for feature enhancement, and perform secondary channel calibration through a secondary SEBlock channel attention module to output deep temporal features.

[0200] The prediction output module 905 is used to input deep temporal features into the LSTM network for sequence modeling to capture long-term dependencies, and output the prediction results of the modulation mode of the IQ signal through the classification head.

[0201] The model evaluation module 906 is used to train the model using the cross-entropy loss function and the adaptive learning rate scheduler, verify the effectiveness of the method through comparative experiments and ablation experiments, save the optimal model, and evaluate the recognition performance on the test set.

[0202] It is understandable that, such as Figure 1 The content of the noise-resistant robust modulation signal recognition method embodiments shown are all applicable to the noise-resistant robust modulation signal recognition system embodiments. The specific functions implemented by the noise-resistant robust modulation signal recognition system embodiments are the same as those shown in the examples. Figure 1 The noise-resistant robust modulation signal identification method shown in the embodiment is the same, and the beneficial effects achieved are the same as those described above. Figure 1 The beneficial effects achieved by the noise-resistant and robust modulation signal identification method embodiment shown are also the same.

[0203] It should be noted that the information interaction and execution process between the above systems are based on the same concept as the method embodiments of the present invention. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.

[0204] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0205] Reference Figure 10 The present invention also provides a computer device 10, including a memory 1002 and a processor 1001, and a computer program 1003 stored in the memory 1002. When the computer program 1003 is executed on the processor 1001, it implements the noise-resistant and robust modulation signal recognition method as described in any of the above methods.

[0206] The computer device 10 may be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device 10 may include, but is not limited to, a processor 1001 and a memory 1002. Those skilled in the art will understand that... Figure 10 The computer device 10 is merely an example and does not constitute a limitation on the computer device 10. It may include more or fewer components than shown, or combine certain components, or different components, such as input / output devices, network access devices, etc.

[0207] The processor 1001 may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0208] In some embodiments, the memory 1002 may be an internal storage unit of the computer device 10, such as a hard disk or memory of the computer device 10. In other embodiments, the memory 1002 may be an external storage device of the computer device 10, such as a plug-in hard disk, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the computer device 10. Further, the memory 1002 may include both internal and external storage units of the computer device 10. The memory 1002 is used to store the operating system, applications, boot loader, data, and other programs, such as the program code of the computer program. The memory 1002 can also be used to temporarily store data that has been output or will be output.

[0209] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the noise-resistant and robust modulation signal identification method as described in any of the above methods.

[0210] In this embodiment, if the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0211] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0212] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0213] In the embodiments disclosed in this application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling or direct coupling or communication connection may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0214] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

Claims

1. A method of anti-noise robust modulation signal recognition, characterized in that, The method specifically comprises: The input IQ signal is preprocessed, AP features are generated through amplitude-phase conversion, and the IQ signal is normalized; The normalized IQ signal and the AP features are input into a GRU guided feature fusion module to perform adaptive weight learning and weighted calibration, and to generate fusion features; Based on the fusion features, channel calibration is performed through a first SEBlock channel attention module, feature extraction and down-sampling are performed through a convolutional network, and coding features are obtained by combining position encoding parameters; The coding features are input into a multi-attention collaborative enhancement module containing convolutional enhanced attention and global self-attention for feature enhancement, and secondary channel calibration is performed through a second SEBlock channel attention module to output deep temporal features; The deep temporal features are input into an LSTM network for sequence modeling to capture long-time dependencies, and the prediction results of the modulation mode of the IQ signal are output through a classification head; The model is trained using a cross-entropy loss function and an adaptive learning rate scheduler, the effectiveness of the method is verified through comparative experiments and ablation experiments, the optimal model is saved, and the test set recognition performance is evaluated; The coding features are input into a convolutional enhanced attention submodule, local structure features are extracted through multi-scale deep separable convolution, and after transformation through a projection network, the local structure features are weighted with value features to output local enhanced features; The local enhanced features and the coding features are connected in residual and normalized by layers to obtain first residual features; The first residual features are input into a global self-attention submodule as queries, keys, and values, global dependency features are output through a multi-head attention mechanism; The global dependency features and the first residual features are connected in residual and normalized by layers to obtain second residual features; The second residual features are nonlinearly transformed through a feedforward network to output nonlinear enhanced features; The nonlinear enhanced features are input into a second SEBlock channel attention submodule, channel weights are generated through adaptive average pooling, a fully connected layer, and a Sigmoid activation to calibrate the feature channels, and deep temporal features are output. The input IQ signal is preprocessed, AP features are generated through amplitude-phase conversion, and the IQ signal is normalized; 2. The method of claim 1, wherein, Original IQ sample data is obtained, the IQ sample data including in-phase component I and quadrature component Q; Amplitude-phase conversion is performed on the IQ sample data, amplitude features and phase features are calculated based on the in-phase component I and the quadrature component Q to form an AP feature matrix; The IQ sample data is normalized by maximum value at the sample level, and the in-phase component I and the quadrature component Q of each sample are scaled to the [-1, 1] interval. The normalized IQ signal and the AP features are input into a GRU guided feature fusion module to perform adaptive weight learning and weighted calibration, and to generate fusion features, specifically comprising:

3. The method of claim 1, wherein, ​ Element-level superposition of the normalized IQ signal and the AP feature is performed to generate a superposition feature; Dimension transposition is performed on the superposition feature to generate a feature sequence conforming to the input format of the GRU network; The feature sequence is input into the GRU network to learn the cooperative relationship between the time step and the channel, and output a hidden state sequence corresponding to the time step; The hidden state sequence is transposed and processed through a Sigmoid activation function to generate an adaptive weight matrix corresponding to the input feature dimension; The normalized IQ signal and the AP feature are respectively weighted element by element using the adaptive weight matrix to obtain the weighted and calibrated IQ signal and AP feature; The weighted and calibrated IQ signal and AP feature are spliced along the channel dimension to output a fusion feature.

4. The method of claim 1, wherein, Based on the fusion feature, channel calibration is performed through a first SEBlock channel attention module, feature extraction and down-sampling are performed through a convolutional network, and an encoded feature is obtained in combination with a position encoding parameter, specifically including: The fusion feature is input into the first SEBlock channel attention module to compress the time dimension through adaptive average pooling to obtain a global channel feature; The global channel feature is processed through a dimension reduction fully connected layer, a ReLU activation function, a dimension increasing fully connected layer and a Sigmoid activation function in sequence to generate a channel weight vector; The channel weight vector is multiplied element by element with the fusion feature to complete channel calibration and output a calibrated feature; The calibrated feature is input into a convolutional network comprising two one-dimensional convolutional layers to perform convolution, batch normalization and ReLU activation in sequence to realize feature extraction and down-sampling and obtain a first convolutional feature; The first convolutional feature is transposed in dimension and added to a learnable position encoding parameter to generate an encoded feature.

5. The method of claim 1, wherein, The encoded feature is input into a convolutional enhanced attention sub-module to extract local structure features through multi-scale deep separable convolution and output local enhanced features after transformation through a projection network and weighting with value features, specifically including: Linear projection is performed on the encoded feature to generate a value feature matrix and a spatial feature matrix, respectively; The spatial feature matrix is split and dimension transformed according to a preset number of attention heads to obtain multiple parallel multi-head feature maps; Different scale convolution kernels are assigned to each multi-head feature map, and deep separable convolution is applied to each to obtain corresponding convolution output features; All convolution output features are spliced along the channel dimension to obtain a second convolutional feature fused with multi-scale information; The second convolutional feature is input into the projection network to sequentially pass through a first linear layer, a Gaussian error linear unit activation function, layer normalization, and a second linear layer and a nonlinear transformation to output a projection feature; The projection feature is multiplied element by element with the value feature matrix to complete the weighted enhancement of the local structure features and output local enhanced features.

6. The method of claim 1, wherein, The deep layer time sequence feature is input into an LSTM network for sequence modeling to capture long-time dependency relationships, and the prediction result of the modulation mode of the IQ signal is output through a classification head, specifically including: The deep layer time sequence feature is transposed in dimension to obtain a time sequence sequence conforming to the input format of the LSTM network; The time sequence is input into a three-layer stacked LSTM network, and the time sequence information is updated and fused step by step through the gating mechanism of the forgetting gate, the input gate, the output gate and the cell state, and the hidden state of the last time step is output as the aggregated feature; The aggregated feature is input into a classification head for dimension reduction and nonlinear mapping of the feature, and the classification head comprises a first fully connected layer, a ReLU activation function, a Dropout layer and a second fully connected layer connected in sequence; The probability distribution corresponding to the preset modulation mode category is output through the second fully connected layer, and the class with the highest probability is taken as the prediction result of the modulation mode of the IQ signal.

7. The method of claim 6, wherein, The time sequence is input into a three-layer stacked LSTM network, and the time sequence information is updated and fused step by step through the gating mechanism of the forgetting gate, the input gate, the output gate and the cell state, and the hidden state of the last time step is output as the aggregated feature, specifically comprising: At each time step, the input feature of the current time step is spliced with the hidden state of the previous time step to form a spliced feature; The spliced feature is input into the forgetting gate, and a forgetting gate output vector is generated through linear transformation and Sigmoid activation function processing; The spliced feature is simultaneously input into the input gate and the candidate cell state calculation module, and an input gate output vector is obtained through linear transformation and Sigmoid activation function processing, and a candidate cell state of the current time step is obtained through linear transformation and Tanh activation function processing; Based on the forgetting gate output vector, the cell state of the previous time step, the input gate output vector and the candidate cell state, the cell state of the current time step is updated through weighted fusion; The spliced feature is input into the output gate, and an output gate output vector is generated through linear transformation and Sigmoid activation function processing; The output gate output vector is multiplied element by element with the result of applying Tanh activation to the cell state of the current time step to generate the hidden state of the current time step; All time steps are processed in sequence, and the hidden state generated by the third layer LSTM network at the last time step is taken as the aggregated feature.

8. The method of claim 1, wherein, The model is trained by using the cross-entropy loss function and the adaptive learning rate scheduler, the effectiveness of the method is verified through comparative experiments and ablation experiments, the optimal model is saved and the test set recognition performance is evaluated, and specifically, the method comprises the following steps: The original data set is divided into a training set, a validation set and a test set; The cross-entropy loss function is set as the objective function of model training, and the objective function is used to calculate the difference between the predicted result and the true label; An optimizer is configured and an adaptive learning rate scheduler is used to dynamically adjust the learning rate during training; During the training process, the parameters of the optimal model are saved based on the validation set loss using an early stopping strategy; The optimal model is used to evaluate the recognition performance, and the overall recognition accuracy and performance indicators under each signal-to-noise ratio are obtained.

9. A noise-robust modulation signal identification system configured to perform a noise-robust modulation signal identification method according to any one of claims 1 to 8, characterized in that The system specifically comprises: A preprocessing module for preprocessing the input IQ signal, generating AP features through amplitude-phase conversion, and normalizing the IQ signal; The feature fusion module is configured to input the normalized IQ signal and the AP feature into a GRU guided feature fusion module to perform adaptive weight learning and weighted calibration, and generate a fusion feature; The feature encoding module is configured to perform channel calibration through a first SEBlock channel attention module, perform feature extraction and down-sampling through a convolution network, and obtain an encoded feature by combining a position encoding parameter based on the fusion feature; The feature enhancement module is configured to input the encoded feature into a multi-attention collaborative enhancement module including a convolution enhanced attention and a global self-attention to perform feature enhancement, and perform secondary channel calibration through a second SEBlock channel attention module to output a deep time sequence feature; The prediction output module is configured to input the deep time sequence feature into an LSTM network to perform sequence modeling to capture long-time dependency, and output a prediction result of a modulation mode of the IQ signal through a classification head; The model evaluation module is configured to train the model by using a cross-entropy loss function and an adaptive learning rate scheduler, verify the effectiveness of the method through comparative experiments and ablation experiments, save an optimal model, and evaluate the recognition performance of a test set. The feature enhancement module is configured to input the encoded feature into a multi-attention collaborative enhancement module including a convolution enhanced attention and a global self-attention to perform feature enhancement, and perform secondary channel calibration through a second SEBlock channel attention module to output a deep time sequence feature, and specifically includes: The encoded feature is input into a convolution enhanced attention submodule, local structure features are extracted through multi-scale deep separable convolution, and the transformed features are weighted with value features after projection network transformation, and local enhanced features are output; The local enhanced features and the encoded features are connected in residual and normalized by layers to obtain first residual features; The first residual features are input into a global self-attention submodule as queries, keys and values, global dependency features are captured through a multi-head attention mechanism, and the global dependency features are output; The global dependency features and the first residual features are connected in residual and normalized by layers to obtain second residual features; The second residual features are nonlinearly transformed through a feedforward network to output nonlinear enhanced features; The nonlinear enhanced features are input into a second SEBlock channel attention submodule, channel weights are generated through adaptive average pooling, a fully connected layer and a Sigmoid activation, the feature channels are calibrated, and deep time sequence features are output.

Images