Signal modulation identification method and system based on feature pyramid and Mama hybrid model, computer equipment and readable storage medium

Abstract

The invention provides a signal modulation recognition method and system based on a feature pyramid and a Mama hybrid model, computer equipment and a readable storage medium, and solves the technical problems of low calculation efficiency and poor performance caused by bloated structure, high reasoning delay and insufficient multi-scale feature application of an existing signal modulation recognition algorithm. The method comprises the following steps: acquiring a signal data set, and dividing to obtain a training set, a verification set and a test set containing to-be-identified signals; constructing a signal modulation recognition model based on the feature pyramid and a Mama model, inputting the training set and the verification set into the signal modulation recognition model for training, and storing the trained signal modulation recognition model; and identifying the test set by adopting the trained signal modulation identification model to obtain an identification result. The method can be widely applied to the technical field of signal processing.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of signal processing, and more particularly relates to a signal modulation recognition method and system based on a feature pyramid and Mamba hybrid model, a computer device, and a readable storage medium. BACKGROUND

[0002] With the rapid development of wireless communication technology, the complex communication environment has put forward urgent demands for the automatic modulation recognition capability of the receiving end, and the development of a method with high recognition rate, strong robustness and easy engineering deployment has become the key. In recent years, automatic modulation recognition methods based on deep learning have attracted widespread attention due to their powerful end-to-end learning capabilities, without the need for complex manual feature extraction, and can automatically learn feature representations from raw signals. Existing automatic modulation recognition methods based on deep learning mainly include the following categories: multi-modal feature fusion-based methods, which transform I / Q signals into various forms of representation such as amplitude and phase, constellation diagrams, and time-frequency diagrams, and extract and fuse features from different representations of the signal; optimization methods for backbone networks, which introduce RNN variants and Transformers to address the shortcomings of CNNs in modeling long-distance temporal dependencies; and multi-model fusion methods, which embed the self-attention module of the Transformer model into existing CNN or RNN models to improve long-distance dependency modeling capabilities.

[0003] However, existing automatic modulation recognition methods based on deep learning have the following problems: First, multi-modal feature fusion and multi-model fusion methods can improve performance, but the fusion of multiple modalities and multiple models makes the model structure bulky and the parameter quantity large, resulting in high inference delay, which is not conducive to the engineering deployment of resource-constrained edge devices. The optimization method for backbone networks can extract local features, but the receptive field is limited, making it difficult to model long-distance signal temporal dependencies; RNN models can handle temporal relationships, but they suffer from gradient vanishing and low computational efficiency; and the Transformer model based on the self-attention mechanism has a computational and memory complexity that is proportional to the square of the sequence length, making it too costly to process long sequences of signals. In addition, most hierarchical model classification decisions rely only on the deepest features of the network, which can result in the dilution or loss of shallow information that is crucial for distinguishing certain modulation types during step-by-step downsampling, leading to insufficient application of multi-scale features, slow computation, and poor performance. SUMMARY

[0004] The purpose of the embodiments of the present application is to provide a signal modulation recognition method and system based on a feature pyramid and Mamba hybrid model, a computer device, and a readable storage medium to solve the technical problems of bulky structure, high inference delay, and insufficient application of multi-scale features in existing signal modulation recognition algorithms, resulting in slow computation and poor performance.

[0005] To achieve the above object, the first aspect of the embodiment of the application provides a signal modulation recognition method based on a feature pyramid and a Mamba mixed model, comprising the following steps: obtaining a signal data set, and dividing the signal data set to obtain a training set, a validation 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 validation set into the signal modulation recognition model for training, and saving the trained signal modulation recognition model; recognizing the test set by using the trained signal modulation recognition model 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 for converting the signal to be recognized into a feature sequence. The hierarchical feature extraction module is used for multi-stage deep feature extraction and transformation of the feature sequence, deepening and refining the feature representation while reducing the sequence resolution at each stage, and obtaining output feature sequences at different stages. The pyramid feature fusion classification module is used for aggregating the output feature sequences at different stages to obtain a feature vector, splicing to obtain a fusion feature vector, and mapping to obtain a modulation type probability distribution of the signal to be recognized. The output module is used for converting the modulation type probability distribution of the signal to be recognized into a probability distribution vector, finding an element with the maximum probability value in the probability distribution vector, determining the index corresponding to the element as the final predicted class index, mapping the predicted class index to a predefined modulation type label, and obtaining the recognition result.

[0006] Preferably, the hierarchical feature extraction module comprises a first feature extraction stage, a first downsampling module, a second feature extraction stage, a second downsampling module and a third feature extraction stage in sequence. The first feature extraction stage, the second feature extraction stage and the third feature extraction stage each comprise a core processing unit. The core processing unit comprises a depth separable convolution unit, a layer normalization layer, a Mamba unit and a residual connection, is used for local feature extraction of the feature sequence to obtain a convolution output sequence, pre-normalization processing to obtain a normalized feature sequence, bidirectional global temporal dependence modeling to obtain a modeled feature sequence, residual connection of the modeled feature sequence and the feature sequence to obtain an output feature sequence.

[0007] Preferably, the internal variation formula of the core processing unit is: ; In the formula, x represents the feature sequence, y represents the convolution output sequence, z represents the normalized feature sequence, and h represents the output feature sequence. an output feature sequence processed by the core processing unit, a feature sequence input to the core processing unit, a depth separable convolution unit, a Mamba unit in a bidirectional selection state space, a regularization technique in a deep neural network.

[0008] Preferably, 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, and is configured to receive the output feature sequences of different stages and unify the output feature sequences to obtain a feature vector; The fusion unit is configured to merge feature dimensions of the feature vector to obtain a fused feature vector; The classification head comprises a layer normalization layer, a Dropout layer and a fully connected layer, and is configured to receive the fused feature vector and map the fused feature vector to obtain a modulation type probability distribution.

[0009] Preferably, the signal embedding module comprises a dimension expansion unit, a parallel multi-scale processing unit and a fusion projection unit; The dimension expansion unit comprises a point-wise convolution layer, a batch normalization layer and a GELU activation function layer, and is configured to receive a to-be-recognized signal, expand a channel dimension of the to-be-recognized signal, 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, and is configured 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 reduced feature map; The fusion projection unit is configured to splice the reduced feature map in the channel dimension to obtain a fused feature map, project a channel dimension of the fused feature map to a preset output dimension through point-wise convolution, and obtain a feature sequence.

[0010] Preferably, a formula for obtaining a probability distribution vector is: ; In the formula, is a value of an i-th element in the probability distribution vector, representing a probability that the to-be-recognized signal belongs to an i-th modulation mode, is a natural constant, is an index variable of a summation term, is an output logic value of the pyramid feature fusion classification module for an i-th category, and N is a total number of categories.

[0011] ​​​The second aspect of the embodiment of the application provides a signal modulation recognition system based on a feature pyramid and a Mamba mixed model, comprising: The preprocessing module is used for acquiring a signal data set, performing division to obtain a training set, a verification set and a test set containing a signal to be recognized; The model construction and training module is used for 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; The recognition module is used for recognizing the test set by using the trained signal modulation recognition model to obtain a recognition result.

[0012] The third aspect of the embodiment of the application provides a signal modulation recognition computer device based on a feature pyramid and a Mamba mixed model, comprising: A memory and a processor, the memory stores a computer program, and the processor implements the steps of the above-mentioned signal modulation recognition method based on a feature pyramid and a Mamba mixed model when executing the computer program.

[0013] The fourth aspect of the embodiment of the application provides a readable storage medium for signal modulation recognition based on a feature pyramid and a Mamba mixed model, comprising: A computer program is stored thereon, and the computer program implements the steps of the above-mentioned signal modulation recognition method based on a feature pyramid and a Mamba mixed model when executed by a processor.

[0014] The application has the beneficial effects that the application provides a signal modulation recognition method, system, computer device and readable storage medium based on a feature pyramid and a Mamba mixed model, a signal modulation recognition model fusing a feature pyramid and a Mamba model is constructed to recognize a signal to be recognized and obtain a recognition result. The signal modulation recognition model constructed by the application combines the linear calculation complexity advantage of the Mamba model and the multi-scale feature utilization capability of the feature pyramid, maintains the lightweight of the model, significantly improves the recognition accuracy and robustness of the model in a complex scene, realizes the balance of high precision, strong robustness and high efficiency, and greatly improves the calculation efficiency and performance.

[0015] Specifically, the Mamba model adopted by the present application can efficiently process long time sequence dependence modeling with linear computational complexity, effectively solving the problems of limited receptive field of traditional CNN, easy occurrence of gradient vanishing of RNN, low computational efficiency, and high cost of processing long sequence of Transformer, greatly improving the modeling capability of long time sequence signal in wireless communication. At the same time, the Mamba unit simplifies the signal modulation recognition model parameters, so that the signal modulation recognition model is more lightweight while maintaining high efficiency, and can realize convenient engineering deployment in resource-limited edge devices and other scenarios, meeting the low latency demand of modern communication systems. Secondly, the feature pyramid structure is introduced in the signal modulation recognition model, and the multi-scale signal information extracted at different depths is fully utilized in the model training process, avoiding the defect that the traditional layered model only relies on deep features, causing the loss of shallow key information. Therefore, when the trained signal modulation recognition model identifies the test set, it can more comprehensively capture the subtle differences and essential features of the signal, thereby significantly improving the recognition accuracy of multiple modulation methods while maintaining the lightweight of the model. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0017] Figure 1 The overall block diagram of the signal modulation recognition model provided by an embodiment of the present application is shown in the figure. Figure 2 The working principle block diagram of the signal embedding module provided by an embodiment of the present application is shown in the figure. Figure 3 The working principle block diagram of the layered feature extraction module provided by an embodiment of the present application is shown in the figure. Figure 4 The working principle block diagram of the pyramid feature fusion classification module provided by an embodiment of the present application is shown in the figure. Figure 5 The comparison diagram of the recognition accuracy of the signal modulation recognition model of the present application and other classical modulation recognition models under different signal-to-noise ratios is shown in the figure. Figure 6 The recognition accuracy diagram of the signal modulation recognition model of the present application under different signal-to-noise ratios is shown in the figure. Figure 7 The comparison diagram of the ablation experiment provided by an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0018] In order to make the technical problems, technical solutions and beneficial effects of the present application clearer, the present application will be further described in detail below in conjunction with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0019] The first embodiment of the present application provides a signal modulation recognition method based on a feature pyramid and Mamba hybrid model, comprising: S1: Obtain a signal data set and divide it to obtain a training set, a validation set and a test set containing signals to be recognized.

[0020] Specifically, a signal data set containing multiple modulation types and signal-to-noise ratios is obtained, and a hierarchical sampling method is used to divide it into a training set, a validation set and a test set.

[0021] In an optional embodiment, the present application obtains an in-phase quadrature (I / Q) signal data set and divides it using a hierarchical sampling method to obtain a training set, a validation set and a test set. Among them, the selected in-phase quadrature (I / Q) signal data set is the RadioML2016.10a data set, which contains 220000 signal samples, covers 11 modulation modes, including 8 digital modulation modes BPSK, QPSK, 8PSK, 16QAM, 64QAM, CPFSK, GFSK, PAM4 and three analog modulation modes AM-DSB, AM-SSB, WBFM; The signal-to-noise ratio range covers -20dB to 18dB, with an interval of 2dB, a total of 20 signal-to-noise ratio conditions; The signal format is complex I / Q signal, stored as a two-dimensional array with a shape of (2, 128), the sampling rate is 2.048MHz, and the signal length is 128 sampling points. The selection of the signal data set by the present application is not limited, and can be set according to the actual situation.

[0022] S2: Construct a signal modulation recognition model based on a feature pyramid and a Mamba model.

[0023] Please refer to 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.

[0024] Please refer to Figure 2 , the working principle diagram of the signal embedding module, the signal embedding module includes a dimensionality expansion unit, a parallel multi-scale processing unit and a fusion projection unit, which is used to convert the signal to be recognized into a feature sequence.

[0025] The dimension expansion unit includes a point-wise convolution layer (Conv1d), a batch normalization layer (BatchNorm), and a GELU activation function layer, which are configured to receive the to-be-identified signal, expand the channel dimension of the to-be-identified signal, and obtain an expanded feature map. In this process, to ensure the flexibility of subsequent multi-scale grouping, the channel dimension of the expanded feature map can be dynamically calculated and rounded up to ensure that it can be evenly divided by the number of parallel processing branches, thereby achieving self-adaptation to different parameter configurations.

[0026] The parallel multi-scale processing unit includes a depthwise separable convolution unit (DWConv), a batch normalization layer, and a GELU activation function layer, which are configured to segment the expanded feature map in the channel dimension, obtain three feature groups, and input the three feature groups into three independent parallel processing branches for feature extraction and dimension reduction, thereby obtaining a reduced feature map.

[0027] The fusion projection unit is configured to splice the reduced feature maps output by all parallel processing branches in the channel dimension, obtain a fused feature map, and project the channel dimension of the fused feature map to an expected output dimension through point-wise convolution, thereby obtaining a feature sequence.

[0028] Referring to Figure 3 The hierarchical feature extraction module sequentially includes a first feature extraction stage (Stage1), a first downsampling module, a second feature extraction stage (Stage2), a second downsampling module, and a third feature extraction stage (Stage3), which are configured to perform multi-stage deep feature extraction and transformation on the feature sequence, deepen and refine the feature representation while reducing the sequence resolution at each stage, and obtain output feature sequences at different stages.

[0029] In this application, the channel dimensions of the first feature extraction stage (Stage1), the second feature extraction stage (Stage2), and the third feature extraction stage (Stage3) are all configured as 64. Each stage is formed by a plurality of core processing units (Mamba units) stacked in series, and the distribution numbers of the core processing units (Mamba units) in the stages are 2, 3, and 2, respectively. The first downsampling module and the second downsampling module each include a convolution layer (Conv) and a batch normalization layer (BatchNorm), which are configured to perform a halving operation on the length of the feature sequence of the previous layer.

[0030] The core processing unit (Mamba unit) includes a depthwise separable convolution unit, a layer normalization layer (LayerNorm), a Mamba unit, and a residual connection.

[0031] Specifically, the processing steps of the core processing unit (Mamba unit) include: Firstly, the input feature sequence is input into a depth separable convolution unit to reduce the parameter quantity and extract local features of the input feature sequence to obtain a convolution output sequence. The depth separable convolution unit includes a depth convolution layer, a batch normalization layer, a GELU activation function layer, a 1x1 pointwise convolution layer and a batch normalization layer, and is configured to reduce the parameter quantity and efficiently extract local features of the input feature sequence.

[0032] Then, the convolution output sequence is input into a layer normalization layer (LayerNorm) for pre-normalization to obtain a normalized feature sequence. The layer normalization layer is arranged between the depth separable convolution unit and the Mamba unit to improve the stability of model training.

[0033] After that, the normalized feature sequence is input into the Mamba unit for bidirectional global time sequence dependence modeling to obtain a modeled feature sequence. The Mamba unit is a selection state space model implemented by a hardware-aware parallel algorithm, and the ZigZag scanning mechanism is used to implement bidirectional information modeling by using only one Mamba unit.

[0034] Finally, the modeled feature sequence output by the Mamba unit is connected in residual with the feature sequence input into the depth separable convolution to obtain an output feature sequence of the core processing unit.

[0035] Further, the core processing unit (Mamba unit) is connected in residual throughout to ensure the stability of the depth model, and the internal change can be represented by the following formula: ; In the formula, is the output feature sequence processed by the core processing unit, is the feature sequence input into the core processing unit, is a depth separable convolution unit for enhancing local features, is a Mamba unit in a bidirectional selection state space using a ZigZag scanning mechanism for global time sequence dependence modeling, is a regularization technique in a deep neural network for improving the generalization performance and training stability of the model.

[0036] Please refer to Figure 4This 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.

[0037] 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.

[0038] 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.

[0039] 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: ; 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.

[0040] 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.

[0041] S3: Put the training set and validation set data into the signal modulation identification model for training, and save the trained signal modulation identification model.

[0042] Specifically, before training the signal modulation identification model, data augmentation is performed on the training set, including additive white Gaussian noise, random phase offset or random time masking. Among them, the additive white Gaussian noise is used to add white Gaussian noise to the signal to simulate channel interference, and the noise standard deviation is set to 0.05. Random phase offset is used to apply a random phase rotation within a range to the signal. Random time masking is used to randomly mask a small area on the time axis of the signal, and the maximum masking ratio is set to 0.1 of the signal length.

[0043] Among them, the model training uses label smoothing strategy when calculating loss, and the label smoothing value is 0.1. The AdamW optimizer is used to optimize the parameters in the iteration process, and the initial learning rate of the optimizer is set to 1e-3 and the weight decay coefficient is set to 1e-4. The model training adopts the cosine annealing strategy with restart to dynamically adjust the learning rate, which is used to realize stable and efficient convergence. Among them, the restart period is set to 20 training rounds, and the early stopping strategy is adopted, that is, when the performance index on the validation set does not improve for 10 consecutive rounds, the training is stopped.

[0044] The cross-entropy loss function is used to evaluate the difference between the model prediction result and the true label. In the model training process, by calculating the loss value and backpropagating the error, the optimizer (such as AdamW) is guided to gradually update the model weights, so that the prediction probability distribution output by the model is closer to the true label distribution. At the same time, by monitoring the performance index (such as average recognition accuracy or loss value) on the validation set, the optimal model weight is saved, and the trained signal modulation identification model is obtained. Its calculation formula is as follows: ; In the formula, is the loss value, is the one-hot encoding of the true label, is the prediction probability of the signal modulation identification model for the th class.

[0045] S4: Use the trained signal modulation identification model to identify the test set to obtain the identification result.

[0046] The test set containing the signal to be identified (I / Q signal) is input, and a 1x1 pointwise convolution layer (Conv1d) is used to expand the channel of the signal to be identified, and then a batch normalization layer (BatchNorm) and a GELU activation function layer are sequentially input to stabilize the data distribution, and an expanded feature map is obtained. In this process, the shape of the I / Q signal is changed from the original (B, 2, 128) to (B, 384, 128).

[0047] The extended feature map is evenly divided into 3 groups, each with a shape of (B, 128, 128). The 3 feature groups are input into depth separable convolution units with kernel sizes of 1x1, 1x3 and 1x5 in parallel to extract local features, and at the same time, the convolution units with a convolution stride of 2 are used to reduce the dimension of the sequence length to obtain the reduced feature map. The reduced feature map is sequentially input into a batch normalization layer (BatchNorm) and a GELU activation function layer to stabilize the data distribution, and 3 feature maps with a shape of (B, 128, 64) are obtained.

[0048] The feature maps output by the 3 parallel branches are spliced in the channel dimension to obtain a fused feature map with a shape of (B, 384, 64), which is then input into a 1x1 pointwise convolution layer for information fusion and dimension projection to obtain a feature sequence, which is sequentially input into a batch normalization layer (BatchNorm) and a GELU activation function layer to stabilize the data distribution. Finally, a feature sequence with rich local information and a shape of (B, 64, 64) is output. The feature sequence is input into a hierarchical feature extraction module for compression at different levels to obtain output feature sequences at different levels. The dimension changes of the feature sequence in the hierarchical feature extraction module are as follows: The input feature sequence has a shape of (B, 64, 64), and after the first feature extraction stage (Stage1) (i.e., passing through 2 core processing units Mamba units), a first feature map with a shape of (B, 64, 64) is output.

[0049] After inputting the first feature map into the first downsampling module (convolution with a stride of 2), the shape is (B, 64, 32), and after the second feature extraction stage (Stage2) (i.e., passing through 3 core processing units Mamba units), a second feature map with a shape of (B, 64, 32) is output.

[0050] After inputting the second feature map into the second downsampling module (convolution with a stride of 2), the shape is (B, 64, 16), and after the third feature extraction stage (Stage3) (i.e., passing through 2 core processing units Mamba units), a third feature map with a shape of (B, 64, 16) is output.

[0051] The sequence information of the first, second and third feature maps is aggregated by the 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, and then sequentially pass through a flattening layer (Flatten), a layer normalization layer (LN), a linear projection layer (Linear) and a GELU activation function layer to output a feature vector with a shape of (B, 128). All feature vectors processed by the feature adaptation unit and containing different hierarchical information are spliced into a fusion feature vector with a shape of (B, 384). The high-dimensional fusion feature vector is mapped to a preset category space to finally obtain a modulation type probability distribution of the signal to be identified with a shape of (B, 11).

[0052] The modulation type probability distribution of the signal to be identified is received and converted into a probability distribution vector by a SoftMax function. Finally, the element with the maximum probability value in the probability distribution vector is found by a maximum value selection operation, and the index corresponding to the element is determined as the final prediction class index. The prediction class index is mapped to a predefined modulation type label to obtain the recognition result.

[0053] The second embodiment of the present application provides a signal modulation recognition system based on a feature pyramid and Mamba hybrid model, comprising: The preprocessing module is used to obtain a signal data set and divide it to obtain a training set, a validation set and a test set containing signals to be identified; The model construction and training module is used to construct a signal modulation recognition model, input the training set and validation set data 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 by using the trained signal modulation recognition model to obtain a recognition result.

[0054] The third embodiment of the present application provides a signal modulation recognition computer device based on a feature pyramid and Mamba hybrid model, comprising a memory and a processor; The memory stores a computer program, and the processor executes the computer program to realize the steps of the signal modulation recognition method based on the feature pyramid and the Mamba hybrid model.

[0055] The fourth embodiment of the present application provides a signal modulation recognition computer readable storage medium based on a feature pyramid and Mamba hybrid model, which stores a computer program. The computer program is executed by the processor to realize the steps of the signal modulation recognition method based on the feature pyramid and the Mamba hybrid model.

[0056] Embodiment 1: Comparative experiment Please refer toFigure 5 Figure 6 is a comparison diagram of the recognition accuracy of the signal modulation recognition model of the present application and other signal modulation recognition models under different signal-to-noise ratios (SNR). It can be seen from the figure that the signal modulation recognition model of the present application (IQMambaFPN) exhibits the best performance in the entire signal-to-noise ratio range, and the accuracy curve (IQMambaFPN) is significantly higher than that of all the comparison models. In particular, in the high signal-to-noise ratio range of 0 dB to 18 dB, the peak accuracy of the signal modulation recognition model of the present application is as high as 94%. In the key low signal-to-noise ratio range of -10 dB to 0 dB, the performance of the signal modulation recognition model of the present application rises rapidly, showing stronger noise robustness.

[0057] Figure 7 is a comparison diagram of the recognition accuracy of the signal modulation recognition model of the present application and other signal modulation recognition models under different signal-to-noise ratios (SNR). Figure 6 Figure 8 is a comparison diagram of the recognition accuracy of the signal modulation recognition model of the present application under different signal-to-noise ratios (SNR) for each modulation mode.

[0058] Figure 8 is a comparison diagram of the recognition accuracy of the signal modulation recognition model of the present application under different signal-to-noise ratios (SNR) for each modulation mode.

[0059] Example 2: Ablation experiment In order to illustrate the influence of each module on the performance of the model, the following ablation experiments are performed: Remove the multi-scale convolution unit in the signal embedding module, denoted as NoMulConv; remove the feature pyramid mechanism in the hierarchical feature extraction module, denoted as NoFPN; remove both of the above, denoted as NoMulConv+NoFPN.

[0060] Figure 9 is a comparison diagram of the results of the ablation experiment. The performance comparison is clearly presented in the figure: complete model (IQMambaFPN) > model with removed multi-scale convolution (NoMulConv) > model with removed pyramid head (NoFPN) > baseline model (NoMulConv+NoFPN). Figure 7

[0061] ​By comparing the accuracy curves of the complete model and the model removing the pyramid head, it can be seen that there is a significant performance gap in the full signal-to-noise ratio interval. It shows that the pyramid feature fusion strategy adopted in the present application, which aggregates all hierarchical features, is the biggest contributor to the performance improvement of the model, effectively solving the performance bottleneck caused by the loss of shallow detail information in traditional models. By comparing the accuracy curves of the model removing the pyramid head and the baseline model, it can be seen that the multi-scale convolution front end also has stable and clear gains in model performance, proving its ability to capture diverse local information in the feature extraction stage.

[0062] The following Table 1 is a summary table of model performance indicators of the present application and other classical modulation recognition models.

[0063] Table 1: Summary table of model performance indicators of the present application and other classical modulation recognition models:

[0064] As shown in Table 1, the model (IQMambaFPN) proposed in the present application exhibits significant comprehensive advantages in various performance indicators. In terms of recognition performance, the average accuracy of the model in the present application reaches 63.7%, which is the best among all the compared models; in terms of model efficiency and inference speed, the model in the present application achieves the highest recognition accuracy while maintaining a moderate parameter amount (0.330M). Its average inference time per sample is only 0.0650ms, which is nearly three times faster than other models. Thus, the present application improves the inference efficiency, increases the recognition accuracy, and achieves significant improvement in computational efficiency and performance.

[0065] Those skilled in the art can appreciate that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Professionals can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0066] The above-described embodiments are only used to illustrate the technical solutions of the present application, but not limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present 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 the 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.

2. The signal modulation recognition method based on the feature pyramid and the Mamba hybrid model according to claim 1, wherein the hierarchical feature extraction module comprises a first feature extraction stage, a first downsampling module, a second feature extraction stage, a second downsampling module and a third feature extraction stage in sequence; the first feature extraction stage, the second feature extraction stage and the third feature extraction stage each comprise a core processing unit; the core processing unit comprises a depth separable convolution unit, a layer normalization layer, a Mamba unit and a residual connection, is used to perform local feature extraction on the feature sequence to obtain a convolution output sequence, perform pre-normalization processing to obtain a normalized feature sequence, perform bidirectional global temporal dependence modeling to obtain a modeled feature sequence, and perform the residual connection on the modeled feature sequence and the feature sequence to obtain the output feature sequence.

3. The signal modulation recognition method based on the feature pyramid and the Mamba hybrid model according to claim 2, wherein an internal variation 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 the feature pyramid and the Mamba hybrid model according to claim 1, wherein 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 perform uniform processing to obtain the feature vector. The fusion unit is configured to fuse feature dimensions of the feature vector to obtain the fusion feature vector. The classification head includes a layer normalization layer, a Dropout layer, and a fully connected layer, configured to receive the fusion feature vector and perform mapping to obtain the modulation type probability distribution. 5.The signal modulation identification method based on the feature pyramid and Mamba hybrid model according to claim 1, wherein The signal embedding module includes a dimension expansion unit, a parallel multi-scale processing unit, and a fusion projection unit. The dimension expansion unit includes a point-wise convolution layer, a batch normalization layer, and a GELU activation function layer, configured to receive the to-be-identified signal, expand the channel dimension of the to-be-identified signal, and obtain an expanded feature map. The parallel multi-scale processing unit includes a depth separable convolution unit, a batch normalization layer, and a GELU activation function layer, configured 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 reduced feature map. The fusion projection unit is configured to splice the reduced feature map in the channel dimension to obtain a fusion feature map, and project the channel dimension of the fusion feature map to a preset output dimension through point-wise convolution to obtain the feature sequence. 6.The signal modulation identification method based on the feature pyramid and Mamba hybrid model according to claim 1, wherein 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.

7. 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 1-6, characterized in that, including The preprocessing module is configured to obtain a signal dataset, divide the signal dataset to obtain a training set, a validation set, and a test set containing to-be-identified signals; The model construction and training module is configured to construct a signal modulation identification model based on a feature pyramid and a Mamba model, input the training set and the validation set into the signal modulation identification model for training, and save the trained signal modulation identification model; The identification module is configured to identify the test set by using the trained signal modulation identification model to obtain an identification result.

8. A signal modulation recognition computer device based on a feature pyramid and Mamba hybrid model, characterized in that, including 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 identification method based on the feature pyramid and Mamba hybrid model according to any one of claims 1-6. 9.A readable storage medium based on a feature pyramid and Mamba hybrid model for signal modulation identification, comprising: The readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the signal modulation identification method based on the feature pyramid and Mamba hybrid model according to any one of claims 1-6.

Images