A signal modulation recognition method, system, computer device and readable storage medium based on a feature pyramid and Mamba hybrid model

Abstract

This application provides a signal modulation recognition method, system, computer device, and readable storage medium based on a hybrid feature pyramid and Mamba model. It solves the technical problems of existing signal modulation recognition algorithms, such as bloated structure, high inference latency, and insufficient application of multi-scale features, leading to slow computational efficiency and poor performance. The method includes acquiring a signal dataset and dividing it into a training set, a validation set, and a test set containing the signal to be recognized; constructing a signal modulation recognition model based on the feature pyramid and Mamba model; inputting the training and validation sets into the signal modulation recognition model for training; and saving the trained signal modulation recognition model; and using the trained signal modulation recognition model to recognize the test set to obtain the recognition result. This application can be widely applied in the field of signal processing technology.

Description

Technical Field

[0001] This application belongs to the field of signal processing technology, and more specifically, relates to a signal modulation recognition method, system, computer device, and readable storage medium based on a hybrid model of feature pyramid and Mamba. Background Technology

[0002] With the rapid development of wireless communication technology, complex communication environments have created an urgent need for the receiver's automatic modulation and identification capabilities. Developing methods with high recognition rates, strong robustness, and easy engineering deployment has become crucial.

[0003] In recent years, deep learning-based automatic modulation recognition methods have attracted widespread attention due to their powerful end-to-end learning capabilities, enabling them to automatically learn feature representations from raw signals without complex manual feature extraction. Existing deep learning-based automatic modulation recognition methods mainly fall into the following categories: methods based on multimodal feature fusion, which transform I / Q signals into various representations such as amplitude-phase, constellation diagrams, and time-frequency diagrams, and then extract and fuse features from these different representations; optimization methods for the backbone network, which introduce RNN variants and Transformers to address the shortcomings of CNNs in long-range temporal dependency modeling; and multi-model fusion methods, which embed the self-attention module of the Transformer model into existing CNN or RNN models to improve long-range dependency modeling capabilities.

[0004] However, existing deep learning-based automatic modulation recognition methods suffer from the following problems: First, while methods based on multimodal feature fusion and multi-model fusion can improve performance, the fusion of multimodal and multi-model approaches results in bloated model structures and a large number of parameters, leading to high inference latency and hindering the engineering deployment of resource-constrained edge devices. Regarding backbone network optimization, although CNN models can extract local features, their limited receptive field makes it difficult to model long-distance signal temporal dependencies; RNN models, while capable of handling temporal relationships, suffer from gradient vanishing and computational inefficiency; and Transformer models based on self-attention mechanisms have computational and memory complexity proportional to the square of the sequence length, making them too costly to process long-sequence signals. Furthermore, most hierarchical models rely solely on the deepest network features for classification decisions, potentially leading to the dilution or loss of shallow information crucial for distinguishing certain modulation types during stepwise downsampling, resulting in insufficient application of multi-scale features, slow computational efficiency, and poor performance. Summary of the Invention

[0005] The purpose of this application is to provide a signal modulation recognition method, system, computer device, and readable storage medium based on a hybrid model of feature pyramid and Mamba, so as to solve the technical problems of bloated structure, high inference latency, and insufficient application of multi-scale features in the existing signal modulation recognition algorithm, which leads to slow computation efficiency and poor performance.

[0006] To achieve the above objectives, a first aspect of this application provides a signal modulation recognition method based on a hybrid model of feature pyramid and Mamba, comprising the following steps:

[0007] Obtain the signal dataset and divide it into a training set, a validation set, and a test set containing the signal to be identified;

[0008] A signal modulation recognition model is constructed based on the feature pyramid and the Mamba model. The training set and the validation set are input into the signal modulation recognition model for training, and the trained signal modulation recognition model is saved.

[0009] The trained signal modulation recognition model is used to identify the test set, and the recognition results are obtained.

[0010] The signal modulation recognition model includes a signal embedding module, a hierarchical feature extraction module, a pyramid feature fusion classification module, and an output module;

[0011] The signal embedding module is used to convert the signal to be identified into a feature sequence;

[0012] The hierarchical feature extraction module is used to perform multi-stage deep feature extraction and transformation on the feature sequence. While reducing the sequence resolution stage by stage, it deepens and refines the feature representation to obtain the output feature sequence at different stages.

[0013] The pyramid feature fusion classification module is used to aggregate the output feature sequences from different stages to obtain feature vectors, concatenate them to obtain fused feature vectors, and map them to obtain the modulation type probability distribution of the signal to be identified.

[0014] The output module is used to convert the modulation type probability distribution of the signal to be identified into a probability distribution vector, find the element with the maximum probability value in the probability distribution vector, determine the index corresponding to the element as the final prediction category index, and map the prediction category index to the predefined modulation type label to obtain the recognition result.

[0015] Preferably, the hierarchical feature extraction module includes, in sequence, a first feature extraction stage, a first downsampling module, a second feature extraction stage, a second downsampling module, and a third feature extraction stage;

[0016] The first feature extraction stage, the second feature extraction stage, and the third feature extraction stage all include core processing units;

[0017] The core processing unit includes a depthwise separable convolutional unit, a layer normalization layer, a Mamba unit, and residual connections. It is used to extract local features from the feature sequence to obtain the convolutional output sequence, perform pre-normalization to obtain the normalized feature sequence, perform bidirectional global temporal dependency modeling to obtain the modeled feature sequence, and perform residual connections between the modeled feature sequence and the feature sequence to obtain the output feature sequence.

[0018] Preferably, the internal change formula of the core processing unit is:

[0019] ;

[0020] In the formula, The output feature sequence after processing by the core processing unit. The feature sequence is input to the core processing unit. It is a depthwise separable convolutional unit. For bidirectional selection of Mamba cells in the state space, This refers to regularization techniques used in deep neural networks.

[0021] Preferably, the pyramid feature fusion classification module includes a feature adaptation unit, a fusion unit, and a classification head;

[0022] The feature adaptation unit includes a global average pooling layer, a flattening layer, a layer normalization layer, a linear projection layer, and a GELU activation function layer, which are used to receive the output feature sequences from different stages and unify them to obtain feature vectors.

[0023] The fusion unit is used to merge the feature dimensions of the feature vectors to obtain the fused feature vector;

[0024] The classification head includes a layer normalization layer, a Dropout layer, and a fully connected layer, which are used to receive the fused feature vector and map it to obtain the modulation type probability distribution.

[0025] Preferably, the signal embedding module includes a dimension-up expansion unit, a parallel multi-scale processing unit, and a fusion projection unit;

[0026] The dimension expansion unit includes a pointwise convolutional layer, a batch normalization layer, and a GELU activation function layer, which are used to receive the signal to be identified and expand the channel dimension of the signal to be identified to obtain an expanded feature map.

[0027] The parallel multi-scale processing unit includes a depthwise separable convolutional unit, a batch normalization layer, and a GELU activation function layer, which are used to segment the extended feature map in the channel dimension to obtain n feature groups. The n feature groups are then input into n independent parallel processing branches for feature extraction and dimensionality reduction to obtain the dimensionality-reduced feature map.

[0028] The fusion projection unit is used to stitch together the dimensionality-reduced feature maps along the channel dimension to obtain a fused feature map. The channel dimension of the fused feature map is then projected onto the preset output dimension through pointwise convolution to obtain a feature sequence.

[0029] Preferably, the formula for obtaining the probability distribution vector is:

[0030] ;

[0031] In the formula, The probability distribution vector is the first... The value of the nth element represents the signal to be identified belonging to the nth element. The probability of a modulation scheme. It is a natural constant. For the index variable of the summation term, The pyramid feature fusion classification module is for the first The output logical value for each category, where N is the total number of categories.

[0032] A second aspect of this application provides a signal modulation recognition system based on a hybrid feature pyramid and Mamba model, comprising:

[0033] The preprocessing module is used to acquire the signal dataset and divide it into training set, validation set and test set containing the signal to be identified;

[0034] The model building and training module is used to build a signal modulation recognition model based on feature pyramids and Mamba models, feed the training set and validation set into the signal modulation recognition model for training, and save the trained signal modulation recognition model.

[0035] The recognition module is used to identify the test set using a trained signal modulation recognition model and obtain the recognition results.

[0036] A third aspect of this application provides a signal modulation recognition computer device based on a hybrid feature pyramid and Mamba model, comprising:

[0037] The system includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the signal modulation recognition method based on a feature pyramid and Mamba hybrid model.

[0038] A fourth aspect of this application provides a signal modulation identification readable storage medium based on a hybrid model of feature pyramids and Mamba, comprising:

[0039] It stores a computer program, which, when executed by a processor, implements the steps of the signal modulation recognition method based on a hybrid model of feature pyramid and Mamba.

[0040] The beneficial effects of this application are as follows: This application provides a signal modulation recognition method, system, computer device, and readable storage medium based on a hybrid model of feature pyramid and Mamba. By constructing a signal modulation recognition model that integrates feature pyramid and Mamba models, the signal to be recognized is identified, and the recognition result is obtained. The signal modulation recognition model constructed in this application combines the linear computational complexity advantage of the Mamba model with the multi-scale feature utilization capability of the feature pyramid. While maintaining the model's lightweight nature, it significantly improves the recognition accuracy and robustness in complex scenarios, achieving a balance between high precision, strong robustness, and high efficiency, and resulting in a substantial improvement in computational efficiency and performance.

[0041] Specifically, the Mamba model used in this application can efficiently handle long-term sequence dependency modeling with linear computational complexity, effectively solving the problems of limited receptive field of traditional CNNs, gradient vanishing in RNNs, low computational efficiency, and excessive cost of Transformer for long sequences, thus significantly improving the modeling capability for long-term signals in wireless communication. Simultaneously, the Mamba unit simplifies the parameters of the signal modulation recognition model, making it more lightweight while maintaining high performance. This allows for convenient engineering deployment in resource-constrained edge devices and other scenarios, meeting the low-latency requirements of modern communication systems. Secondly, this application introduces a feature pyramid structure into the signal modulation recognition model, fully utilizing multi-scale signal information extracted at different depths during model training. This avoids the defect of traditional hierarchical models relying solely on deep features, leading to the loss of key shallow information. This allows the trained signal modulation recognition model to more comprehensively capture subtle differences and essential features of signals when recognizing test sets, thereby significantly improving the recognition accuracy for various modulation methods while maintaining model lightweightness. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art 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.

[0043] Figure 1 This is an overall block diagram of a signal modulation recognition model provided in an embodiment of this application;

[0044] Figure 2 This is a block diagram illustrating the working principle of a signal embedding module provided in one embodiment of this application.

[0045] Figure 3A block diagram illustrating the working principle of a hierarchical feature extraction module provided in an embodiment of this application;

[0046] Figure 4 A block diagram illustrating the working principle of a pyramid feature fusion classification module provided in an embodiment of this application;

[0047] Figure 5 A comparison chart of the recognition accuracy of the signal modulation recognition model of this application with other classic modulation recognition models under different signal-to-noise ratios is provided for one embodiment of this application.

[0048] Figure 6 A schematic diagram illustrating the recognition accuracy of the signal modulation recognition model of this application under different signal-to-noise ratios, provided as an embodiment of this application;

[0049] Figure 7 This is a comparative ablation experiment diagram provided in one embodiment of this application. Detailed Implementation

[0050] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0051] The first embodiment of this application provides a signal modulation recognition method based on a hybrid model of feature pyramid and Mamba, including:

[0052] S1: Obtain the signal dataset and divide it into a training set, a validation set, and a test set containing the signal to be identified.

[0053] Specifically, a signal dataset containing multiple modulation types and signal-to-noise ratios is obtained, and it is divided into a training set, a validation set, and a test set using a stratified sampling method.

[0054] In an optional embodiment, this application obtains an in-phase quadrature (I / Q) signal dataset and divides it using a stratified sampling method to obtain a training set, a validation set, and a test set. The selected in-phase quadrature (I / Q) signal dataset is the RadioML2016.10a dataset, which contains 220,000 signal samples covering 11 modulation schemes, including eight digital modulation schemes (BPSK, QPSK, 8PSK, 16QAM, 64QAM, CPFSK, GFSK, PAM4) and three analog modulation schemes (AM-DSB, AM-SSB, WBFM). The signal-to-noise ratio (SNR) ranges from -20dB to 18dB, in 2dB increments, for a total of 20 SNR conditions. The signal format is a complex I / Q signal, stored as a two-dimensional array of shape (2, 128), with a sampling rate of 2.048MHz and a signal length of 128 sampling points. This application does not limit the selection of the signal dataset and can set it according to actual conditions.

[0055] S2: Construct a signal modulation recognition model based on feature pyramids and the Mamba model.

[0056] Please see Figure 1 The signal modulation recognition model includes a signal embedding module, a hierarchical feature extraction module, a pyramid feature fusion classification module, and an output module.

[0057] Please see Figure 2 This is a block diagram illustrating the working principle of the signal embedding module. The signal embedding module includes a dimension expansion unit, a parallel multi-scale processing unit, and a fusion projection unit, which are used to convert the signal to be identified into a feature sequence.

[0058] The dimension expansion unit comprises a pointwise convolutional layer (Conv1d), a batch normalization layer (BatchNorm), and a GELU activation function layer. It receives the signal to be identified and expands its channel dimensions to obtain an expanded feature map. During this process, to ensure flexibility in subsequent multi-scale grouping, the channel dimensions of the expanded feature map can be dynamically calculated and rounded up to ensure it is divisible by the number of parallel processing branches, thus achieving adaptability to different parameter configurations.

[0059] The parallel multi-scale processing unit includes a depthwise separable convolutional unit (DWConv), a batch normalization layer, and a GELU activation function layer, which are used to segment the extended feature map in the channel dimension to obtain three feature groups. These three feature groups are then input into three independent parallel processing branches for feature extraction and dimensionality reduction to obtain the dimensionality-reduced feature map.

[0060] The fusion projection unit is used to concatenate the dimensionality-reduced feature maps output from all parallel processing branches along the channel dimension to obtain a fused feature map. Then, the channel dimension of the fused feature map is projected onto the expected output dimension through pointwise convolution to obtain the feature sequence.

[0061] Please see Figure 3 This is a block diagram illustrating the working principle of the hierarchical feature extraction module. The hierarchical feature extraction module sequentially includes a first feature extraction stage (Stage 1), a first downsampling module, a second feature extraction stage (Stage 2), a second downsampling module, and a third feature extraction stage (Stage 3). It is used to perform multi-stage deep feature extraction and transformation on the feature sequence, deepening and refining the feature representation while gradually reducing the sequence resolution, and obtaining the output feature sequences at different stages.

[0062] In this application, the channel dimension of the first feature extraction stage (Stage 1), the second feature extraction stage (Stage 2), and the third feature extraction stage (Stage 3) is configured to be 64. Each stage is composed of multiple core processing units (Mamba units) stacked in series, with the number of core processing units (Mamba units) distributed in the order of 2, 3, and 2. The first downsampling module and the second downsampling module both include a convolutional layer (Conv) and a batch normalization layer (BatchNorm), which are used to halve the length of the feature sequence of the previous layer.

[0063] The core processing unit (Mamba unit) includes depthwise separable convolutional units, layer normalization layers (LayerNorm), Mamba units, and residual connections.

[0064] Specifically, the processing steps of the core processing unit (Mamba unit) include:

[0065] First, the input feature sequence is passed through a depthwise separable convolutional unit to reduce the number of parameters. Simultaneously, local feature extraction is performed on the input feature sequence to obtain the convolutional output sequence. The depthwise separable convolutional unit includes a depthwise convolutional layer, a batch normalization layer, a GELU activation function layer, a 1x1 pointwise convolutional layer, and a batch normalization layer, aiming to reduce the number of parameters while efficiently extracting local features from the input feature sequence.

[0066] Next, the input layer normalization layer (LayerNorm) of the convolutional output sequence is pre-normalized to obtain the normalized feature sequence. Here, a layer normalization layer is set between the depthwise separable convolutional units and the Mamba units to improve the stability of model training.

[0067] Subsequently, the normalized feature sequence is input into a Mamba cell for bidirectional global temporal dependency modeling, resulting in the modeled feature sequence. The Mamba cell is a selected state-space model implemented using a hardware-aware parallel algorithm, and a ZigZag scanning mechanism is employed to achieve bidirectional information modeling using only one Mamba cell.

[0068] Finally, the modeled feature sequence output by the Mamba unit is residually concatenated with the feature sequence of the previous input depth-separable convolution to obtain the output feature sequence of the core processing unit.

[0069] Furthermore, the core processing unit (Mamba unit) ensures the training stability of the deep model through a persistent residual connection, and its internal variations can be represented by the following equation:

[0070] ;

[0071] In the formula, The output feature sequence after processing by the core processing unit. The feature sequence is input to the core processing unit. These are depthwise separable convolutional units used to enhance local features. This refers to Mamba cells in the bidirectional selection state space using the ZigZag scanning mechanism, used for global temporal dependency modeling. Regularization is a technique used in deep neural networks to improve the generalization performance and training stability of models.

[0072] Please see Figure 4 This is a block diagram illustrating the working principle of the pyramid feature fusion classification module. The pyramid feature fusion classification module includes multiple parallel feature adaptation units, a fusion unit, and a classification head. It aggregates the output feature sequences extracted by the hierarchical feature extraction module at different stages to obtain feature vectors, concatenates them to obtain a fused feature vector that integrates information from all stages, and maps it to obtain the modulation type probability distribution of the signal to be identified. Each feature adaptation unit includes a Global Average Pooling (GAP) layer, a Flatten layer, a Layer Normalization (LN) layer, a Linear Projection layer, and a GELU activation function layer. These units receive the output feature maps from the hierarchical feature extraction module at different stages and unify their output dimensions to obtain a feature vector.

[0073] The fusion unit merges all the feature vectors processed by the feature adaptation unit along the feature dimension through a splicing operation, forming a fused feature vector that contains information from all stages.

[0074] The classification head includes a layer normalization layer (LN), a dropout layer (with a dropout rate set to 0.2), and a fully connected layer (FC), which are used to receive the fused feature vector, map the high-dimensional fused feature vector to a preset category space, and finally obtain the modulation type probability distribution of the signal to be identified.

[0075] The output module receives the modulation type probability distribution of the signal to be identified. Using the SoftMax function, it converts the modulation type probability distribution of the signal to be identified into a probability distribution vector. The calculation formula is as follows:

[0076] ;

[0077] In the formula, The probability distribution vector is the first... The value of the nth element represents the signal to be identified belonging to the nth element. The probability of a modulation scheme. It is a natural constant. For the index variable of the summation term, For the pyramid feature fusion classification module, the first The output logical value for each category, where N is the total number of categories, N=11.

[0078] Finally, through a maximum value selection operation, the element with the highest probability value is found in the probability distribution vector, and the index corresponding to this element is determined as the final predicted category index. The predicted category index is then mapped to a predefined modulation type label to obtain the recognition result. Specifically, this application implements the maximum value selection operation by calling the torch.argmax function in the PyTorch library to find the element with the highest probability value in the probability distribution vector.

[0079] S3: Input the training set and validation set data into the signal modulation recognition model for training, and save the trained signal modulation recognition model.

[0080] Specifically, before training the signal modulation recognition model, data augmentation is performed on the training set, including additive white Gaussian noise, random phase shift, or random temporal masking. Additive white Gaussian noise is used to add Gaussian white noise to the signal to simulate channel interference, with a noise standard deviation set to 0.05. Random phase shift applies a random phase rotation to the signal within a certain range. Random temporal masking randomly masks a small segment of the signal's time axis, with a maximum masking ratio set to 0.1 times the signal length.

[0081] In the model training, a label smoothing strategy with a smoothing value of 0.1 is used when calculating the loss. The AdamW optimizer is employed to optimize parameters during iteration, with an initial learning rate of 1e-3 and a weight decay coefficient of 1e-4. The model training uses a cosine annealing strategy with restart to dynamically adjust the learning rate for stable and efficient convergence. The restart period is set to 20 training epochs, employing an early stopping strategy: training stops when the performance metrics on the validation set fail to improve for 10 consecutive epochs.

[0082] This application employs the cross-entropy loss function to evaluate the difference between the model's predictions and the true labels. During model training, the loss value is calculated and the error is backpropagated to guide the optimizer (such as AdamW) to gradually update the model weights, making the predicted probability distribution of the model output closer to the true label distribution. Simultaneously, the optimal model weights are saved by monitoring performance metrics (such as average recognition accuracy or loss value) on the validation set, resulting in a well-trained signal modulation recognition model. The calculation formula is as follows:

[0083] ;

[0084] In the formula, The loss value. One-hot encoding for the real label. For the signal modulation recognition model, the first The predicted probability of each category.

[0085] S4: Use the trained signal modulation recognition model to identify the test set and obtain the recognition results.

[0086] The input consists of a test set containing the signals to be identified (I / Q signals). A 1x1 pointwise convolutional layer (Conv1d) is used to expand the channels of the signals. Then, a batch normalization layer (BatchNorm) and a GELU activation function layer are sequentially applied to stabilize the data distribution, resulting in an expanded feature map. During this process, the shape of the I / Q signals is transformed from (B, 2, 128) to (B, 384, 128).

[0087] The extended feature map was divided into three equal groups, each with a shape of (B, 128, 128). These three feature groups were then fed in parallel into depthwise separable convolutional units with kernel sizes of 1x1, 1x3, and 1x5 for local feature extraction. Simultaneously, the sequence length was reduced using convolutional units with a stride of 2, resulting in dimensionality-reduced feature maps. These dimensionality-reduced feature maps were then sequentially fed into a batch normalization layer and a GELU activation function layer to stabilize the data distribution, resulting in three feature maps with shapes of (B, 128, 64).

[0088] The feature maps output from the three parallel branches are concatenated along the channel dimension to obtain a fused feature map of shape (B, 384, 64). This fused feature map is then fed into a 1x1 pointwise convolutional layer for information fusion and dimensional projection, resulting in a feature sequence. This sequence is then fed into a batch normalization layer (BatchNorm) and a GELU activation function layer to stabilize the data distribution. The final output is a feature sequence of shape (B, 64, 64) containing rich local information.

[0089] The feature sequence is input into the hierarchical feature extraction module for compression at different levels, resulting in output feature sequences at different levels. The dimensionality changes of the feature sequence within the hierarchical feature extraction module are as follows:

[0090] The input feature sequence has a shape of (B, 64, 64). After the first feature extraction stage (Stage 1) (i.e., after two core processing units, the Mamba unit), the output feature map has a shape of (B, 64, 64).

[0091] After the first feature map is input into the first downsampling module (convolution with a stride of 2), the shape is (B, 64, 32). After passing through the second feature extraction stage (Stage 2) (i.e., after passing through 3 core processing units, Mamba units), the output is a second feature map with a shape of (B, 64, 32).

[0092] After the second feature map is input into the second downsampling module (convolution with stride of 2), the shape is (B, 64, 16). After passing through the third feature extraction stage (Stage 3) (i.e., after passing through 2 core processing units, Mamba units), the output is a third feature map with shape (B, 64, 16).

[0093] The first, second, and third feature maps are aggregated by a feature adaptation unit to obtain a feature vector. Specifically, the first, second, and third feature maps are input into a global average pooling layer (GAP) to aggregate the sequence information of the first, second, and third feature maps. Then, the sequence information is passed through a flattening layer, a layer normalization layer (LN), a linear projection layer, and a GELU activation function layer in sequence to output a feature vector with shape (B, 128).

[0094] All feature vectors, which contain information at different levels and have been processed by the feature adaptation unit, are concatenated into a fused feature vector of shape (B, 384). This high-dimensional fused feature vector is then mapped to a preset category space to obtain the modulation type probability distribution of the signal to be identified, which has shape (B, 11).

[0095] The modulation type probability distribution of the received signal to be identified is converted into a probability distribution vector using the SoftMax function. Finally, a maximum value selection operation is performed to find the element with the highest probability value in the probability distribution vector, and the index corresponding to this element is determined as the final predicted category index. This predicted category index is then mapped to a predefined modulation type label to obtain the identification result.

[0096] The second embodiment of this application provides a signal modulation recognition system based on a hybrid model of feature pyramid and Mamba, including:

[0097] The preprocessing module is used to acquire the signal dataset and divide it into training set, validation set and test set containing the signal to be identified;

[0098] The model building and training module is used to build a signal modulation recognition model, feed training set and validation set data into the signal modulation recognition model for training, and save the trained signal modulation recognition model.

[0099] The recognition module is used to identify the test set using a trained signal modulation recognition model and obtain the recognition results.

[0100] The third embodiment of this application provides a signal modulation recognition computer device based on a hybrid model of feature pyramid and Mamba, including a memory and a processor;

[0101] The memory stores a computer program, and when the processor executes the computer program, it implements the steps of a signal modulation recognition method based on a hybrid model of feature pyramids and Mamba.

[0102] The fourth embodiment of this application provides a computer-readable storage medium for signal modulation recognition based on a hybrid model of feature pyramid and Mamba, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of a signal modulation recognition method based on a hybrid model of feature pyramid and Mamba.

[0103] Example 1: Comparative Experiment

[0104] Please see Figure 5 The figure shows a comparison of the recognition accuracy of the signal modulation recognition model of this application with other signal modulation recognition models at different signal-to-noise ratios (SNRs). As can be seen from the figure, the signal modulation recognition model of this application (IQMambaFPN) exhibits the best performance across the entire SNR range, with its accuracy curve (IQMambaFPN) significantly higher than all the comparison models. Particularly in the high SNR range of 0dB to 18dB, the peak accuracy of the signal modulation recognition model of this application reaches 94%. In the critical low SNR range of -10dB to 0dB, the performance of the signal modulation recognition model of this application increases rapidly, demonstrating stronger noise robustness.

[0105] Please see Figure 6 This is a schematic diagram showing the recognition accuracy of the signal modulation recognition model of this application for each modulation method under different signal-to-noise ratios (SNK).

[0106] The figure illustrates the fine-grained recognition capability of the signal modulation recognition model of this application for 11 different modulation types. As can be seen from the figure, even in low signal-to-noise ratio (SNR) environments (e.g., -8dB to -4dB), the model still achieves a considerably high recognition accuracy for various signals such as PAM4, AM-SSB, and QAM64. Notably, for high-order modulation signals such as QAM16 and QAM64, which are easily confused due to their similar constellation diagrams, the model still achieves a recognition accuracy of over 90% even at high SNR, demonstrating its powerful fine-grained feature resolution capability. Furthermore, the model also exhibits stable recognition performance for the extremely difficult-to-recognize WBFM signal. Therefore, the signal modulation recognition model of this application demonstrates stronger noise robustness and more stable recognition performance.

[0107] Example 2: Ablation Experiment

[0108] To illustrate the impact of each module on model performance, the following ablation experiments were conducted:

[0109] Remove the multi-scale convolutional unit from the signal embedding module, denoted as NoMulConv; remove the feature pyramid mechanism from the hierarchical feature extraction module, denoted as NoFPN; remove both of the above simultaneously, denoted as NoMulConv+NoFPN.

[0110] Please see Figure 7 The figure shows a comparison of the ablation experiment results. The performance comparison is clearly presented in the figure: full model (IQMambaFPN) > model with multi-scale convolutions removed (NoMulConv) > model with pyramid heads removed (NoFPN) > baseline model (NoMulConv+NoFPN).

[0111] By comparing the accuracy curves of the complete model and the model with the pyramid head removed, a significant performance gap can be observed across the entire signal-to-noise ratio range. This indicates that the pyramid feature fusion strategy adopted in this application, which aggregates features from all levels, is the biggest contributor to the improved model performance, effectively solving the performance bottleneck caused by the loss of shallow detail information in traditional models. Comparing the accuracy curves of the model with the pyramid head removed and the baseline model shows that the multi-scale convolutional front-end also provides a stable and clear gain to the model performance, demonstrating its ability to capture diverse local information during the feature extraction stage.

[0112] Table 1 below is a summary table of the model performance indicators of this application and other classic modulation recognition models.

[0113] Table 1. Summary of model performance metrics of this application and other classic modulation recognition models:

[0114]

[0115] As shown in Table 1, the proposed model (IQMambaFPN) exhibits significant comprehensive advantages across various performance metrics. In terms of recognition performance, the model achieves an average accuracy of 63.7%, the best among all compared models. Regarding model efficiency and inference speed, the model achieves the highest recognition accuracy while maintaining a moderate number of parameters (0.330M). Its average inference time per sample is only 0.0650ms, representing nearly three times the inference efficiency compared to other models. Therefore, this application improves inference efficiency and recognition accuracy, achieving a significant improvement in computational efficiency and performance.

[0116] 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.

[0117] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for signal modulation recognition based on a feature pyramid and Mamba hybrid model, characterized in that, The method comprises the following steps: obtaining a signal data set, and dividing the signal data set to obtain a training set, a verification set and a test set containing signals to be recognized; constructing a signal modulation recognition model based on a feature pyramid and a Mamba model, inputting the training set and the verification set into the signal modulation recognition model for training, and saving the trained signal modulation recognition model; using the trained signal modulation recognition model to recognize the test set to obtain a recognition result; the signal modulation recognition model comprises a signal embedding module, a hierarchical feature extraction module, a pyramid feature fusion classification module and an output module; the signal embedding module is used to convert the signal to be recognized into a feature sequence; the hierarchical feature extraction module is used to perform multi-stage deep feature extraction and transformation on the feature sequence, deepens and refines the feature representation while reducing the sequence resolution at each stage, and obtains output feature sequences at different stages; the pyramid feature fusion classification module is used to aggregate the output feature sequences at different stages to obtain a feature vector, splice the feature vector to obtain a fusion feature vector, and map the fusion feature vector to obtain a modulation type probability distribution of the signal to be recognized; the output module is used to convert the modulation type probability distribution of the signal to be recognized into a probability distribution vector, find an element with a maximum probability value in the probability distribution vector, determine an index corresponding to the element as a final prediction class index, map the prediction class index to a predefined modulation type label, and obtain the recognition result; the pyramid feature fusion classification module comprises a feature adaptation unit, a fusion unit and a classification head; the feature adaptation unit comprises a global average pooling layer, a flattening layer, a layer normalization layer, a linear projection layer and a GELU activation function layer, is used to receive the output feature sequences at different stages, and unify the output feature sequences at different stages to obtain the feature vector; the fusion unit is used to merge the feature dimensions of the feature vector to obtain the fusion feature vector; the classification head comprises a layer normalization layer, a Dropout layer and a fully connected layer, is used to receive the fusion feature vector, and map the fusion feature vector to obtain the modulation type probability distribution; the signal embedding module comprises a dimension lifting expansion unit, a parallel multi-scale processing unit and a fusion projection unit; the dimension lifting expansion unit comprises a point-wise convolution layer, a batch normalization layer and a GELU activation function layer, is used to receive the signal to be recognized, expand the channel dimension of the signal to be recognized, and obtain an expanded feature map; the parallel multi-scale processing unit comprises a depth separable convolution unit, a batch normalization layer and a GELU activation function layer, is used to split the expanded feature map in the channel dimension to obtain n feature groups, input the n feature groups into n independent parallel processing branches respectively for feature extraction and dimension reduction, and obtain a dimension-reduced feature map; the fusion projection unit is used to splice the dimension-reduced feature map in the channel dimension to obtain a fusion feature map, project the channel dimension of the fusion feature map to a preset output dimension through point-wise convolution, and obtain the feature sequence.

2. The signal modulation recognition method based on a hybrid model of feature pyramid and Mamba as described in claim 1, characterized in that, The hierarchical feature extraction module includes, in sequence, a first feature extraction stage, a first downsampling module, a second feature extraction stage, a second downsampling module, and a third feature extraction stage; The first feature extraction stage, the second feature extraction stage, and the third feature extraction stage all include core processing units; The core processing unit includes a depthwise separable convolutional unit, a layer normalization layer, a Mamba unit, and a residual connection. It is used to extract local features from the feature sequence to obtain a convolutional output sequence, perform pre-normalization to obtain a normalized feature sequence, perform bidirectional global temporal dependency modeling to obtain a modeled feature sequence, and perform the residual connection between the modeled feature sequence and the feature sequence to obtain the output feature sequence.

3. The signal modulation recognition method based on a hybrid model of feature pyramid and Mamba as described in claim 2, characterized in that, The internal change formula of the core processing unit is: ； wherein, is the output feature sequence processed by the core processing unit, is the input feature sequence to the core processing unit, is a depth separable convolution unit, is a Mamba unit in a bidirectional selection state space, is a regularization technique in a deep neural network.

4. The signal modulation recognition method based on a hybrid model of feature pyramid and Mamba as described in claim 1, characterized in that, The formula for obtaining the probability distribution vector is: ； In the formula, The probability distribution vector is the first... The value of the nth element represents the signal to be identified belonging to the nth element. The probability of a modulation scheme. It is a natural constant. For the index variable of the summation term, For the pyramid feature fusion classification module, the first The output logical value for each category, where N is the total number of categories.

5. A signal modulation recognition system based on a feature pyramid and Mamba hybrid model, applied to the signal modulation recognition method based on a feature pyramid and Mamba hybrid model of any one of claims 1-4, characterized in that, include: The preprocessing module is used to acquire the signal dataset and divide it into training set, validation set and test set containing the signal to be identified; The model building and training module is used to build a signal modulation recognition model based on feature pyramids and Mamba model, feed the training set and validation set into the signal modulation recognition model for training, and save the trained signal modulation recognition model. The recognition module is used to recognize the test set using the trained signal modulation recognition model to obtain the recognition result.

6. A signal modulation recognition computer device based on a feature pyramid and Mamba hybrid model, characterized in that, Including memory and processor; The memory stores a computer program, and the processor executes the computer program to implement the steps of the signal modulation recognition method based on the feature pyramid and Mamba hybrid model as described in any one of claims 1-4.

7. A readable storage medium for signal modulation recognition based on a hybrid model of feature pyramid and Mamba, characterized in that, It stores a computer program, which, when executed by a processor, implements the steps of the signal modulation recognition method based on a feature pyramid and Mamba hybrid model as described in any one of claims 1-4.

Images