Modulation identification and signal-to-noise ratio estimation based on multi-task neural network

Abstract

The invention discloses a modulation identification and signal-to-noise ratio estimation method and device, a storage medium and equipment, and belongs to the technical field of communication. The method comprises the following steps: converting a radio frequency signal transmitted by a wireless channel, separating a real part and an imaginary part of a baseband modulation signal, and splicing the real part and the imaginary part into a two-dimensional matrix; obtaining a multi-task model, wherein the multi-task model comprises an input mapping module, a Mama module, an MSDCAM module, an MHAAP module and a full connection module; the Mmba module and the MSDCAM module jointly form a double-branch structure; the MSDCAM module is composed of four groups of multi-scale expansion convolution, channel attention and space attention; the MHAAP module is formed by connecting attention pooling, global average pooling and maximum pooling in parallel; and processing the two-dimensional matrix by using a multi-task model to obtain a modulation identification and signal-to-noise ratio estimation value. According to the method, modulation identification and signal-to-noise ratio estimation can be synchronously carried out, the performance is good, the complexity and parameter quantity are low, the overhead is small, and the generalization ability and robustness are high.

Description

Technical Field

[0001] This invention belongs to the field of communication technology, specifically the design of an algorithm for modulation recognition and signal-to-noise ratio estimation based on a multi-task neural network. Background Technology

[0002] In modern wireless communication systems, modulation identification and signal-to-noise ratio (SNR) estimation are two crucial signal processing techniques, especially important in non-cooperative communication environments. Non-cooperative communication refers to communication where the receiver communicates without prior negotiation with the transmitter; the receiver cannot know the modulation scheme, SNR, or other key parameters of the transmitted signal beforehand. This situation is prevalent in military communications, spectrum monitoring, cognitive radio, and electronic warfare. The development of modulation identification technology has gone through several stages. Early modulation identification methods were mainly based on statistical pattern recognition theory, classifying signals by extracting statistical features. These methods include those based on higher-order statistics, those based on cyclostationary properties, and those based on decision theory. Methods based on higher-order statistics use higher-order cumulants of the signal as feature vectors, classifying signals by calculating the differences in cumulants between different modulated signals. However, these methods experience a sharp performance drop in low SNR environments and have high computational complexity. Methods based on cyclostationary properties use the cyclic frequency characteristics of the modulated signal for identification, exhibiting some robustness to noise, but are sensitive to multipath fading channels. Decision-theoretic methods classify data by constructing decision trees or using machine learning algorithms such as support vector machines. While they perform well in specific scenarios, their generalization ability is limited.

[0003] Traditional modulation identification methods have several limitations. First, these methods heavily rely on manually designed features, requiring in-depth signal processing expertise and an accurate understanding of the characteristics of different modulation schemes. Second, hand-designed features often fail to adequately characterize signal properties in complex channel environments, leading to decreased identification performance. Third, traditional methods are highly sensitive to non-ideal factors in real-world communication environments, such as channel distortion, multipath fading, and frequency shift. Finally, with the continuous emergence of new modulation schemes, traditional methods struggle to adapt and expand quickly. Signal-to-noise ratio (SNR) estimation techniques face similar challenges. Traditional SNR estimation methods mainly include the method of moments (moments), maximum likelihood estimation (maximum likelihood estimation), and subspace methods. The method of moments estimates the SNR by calculating the second and fourth moments of the received signal; the algorithm is simple but has significant errors under low SNR conditions. The maximum likelihood estimation method can theoretically reach the Cramer-Rao bound, but its computational complexity is extremely high and requires prior information about the signal. The subspace method estimates based on the orthogonality of the signal and noise subspaces, but it places strict requirements on the assumptions of the signal model. These traditional methods generally suffer from problems such as strict assumptions about noise characteristics, poor adaptability to changes in the channel environment, and difficulty in balancing computational complexity and estimation accuracy.

[0004] In recent years, the rapid development of deep learning technology has brought new opportunities to the field of communication signal processing. Deep neural networks possess powerful nonlinear mapping capabilities and automatic feature learning capabilities, enabling them to automatically extract discriminative features from raw signals without the need for manually designed complex feature extraction algorithms. In modulation recognition, convolutional neural network-based methods have achieved good classification performance by learning the time-frequency features of signals. Recurrent neural network-based methods can capture the temporal dependencies of signals, making them particularly suitable for handling dynamically changing channel environments. Attention-based methods can adaptively focus on important information in the signal, improving recognition accuracy. Deep learning methods also demonstrate advantages in signal-to-noise ratio (SNR) estimation. By training deep networks to learn the complex mapping relationship between signal power and noise power, relatively accurate estimation can be achieved under various channel conditions. Especially when dealing with complex environments such as non-Gaussian noise and multipath fading, the robustness of deep learning methods is significantly better than that of traditional methods.

[0005] However, existing deep learning methods mainly employ a single-task learning strategy, i.e., training the modulation recognition network and the signal-to-noise ratio (SNR) estimation network separately. This approach has significant drawbacks: First, it consumes large computational resources, requiring the maintenance of two independent sets of network parameters, which consumes significant storage space and computational resources in practical deployments. Second, it fails to fully utilize the inherent correlation between the modulation recognition and SNR estimation tasks; different modulation schemes affect the power distribution characteristics of the signal, while the SNR level also affects the discernibility of modulation features. Third, separately trained networks are prone to overfitting, resulting in limited generalization ability. Fourth, it is difficult to deploy on resource-constrained mobile communication devices, embedded systems, or edge computing devices.

[0006] Multi-task learning, as an important branch of machine learning, improves learning efficiency and generalization performance by simultaneously optimizing multiple related tasks. In the field of communication signal processing, modulation recognition and signal-to-noise ratio estimation tasks are naturally related, both requiring the extraction of information from the same received signal, and the modulation and noise characteristics of the signal interact with each other. Therefore, employing a multi-task learning framework to jointly optimize these two tasks has significant technical value and practical implications.

[0007] Currently, research on multi-task learning for communication signals is relatively limited. Existing multi-task methods mainly employ simple parameter-sharing strategies, lacking in-depth consideration of the characteristics of different tasks, and require improvement in network structure design, feature fusion strategies, and loss function optimization. Especially when dealing with challenging scenarios such as complex modulation schemes, low signal-to-noise ratio (SNR) environments, and multipath fading channels, the performance of existing methods still has significant room for improvement. Therefore, developing an efficient multi-task neural network method capable of simultaneously performing modulation recognition and SNR estimation, while significantly reducing computational complexity and parameter size while maintaining recognition accuracy and estimation accuracy, is of great significance for promoting the development and practical application of communication signal processing technology. Summary of the Invention

[0008] This invention proposes a multi-task neural network model for joint modulation recognition and signal-to-noise ratio (SNR) estimation. It designs a dual-path processing architecture, combining MambaBlock for long-range temporal dependency modeling and a multi-scale dilated convolutional attention module (MSDCAM) for hierarchical feature extraction. Through a unified feature learning framework, the MTMCN neural network simultaneously performs modulation recognition and SNR estimation tasks. The MTMCN network can receive complex modulation signals generated by MATLAB, separating the real and imaginary parts of the complex signal to form a two-dimensional matrix, which is then fed into the network for training. After network convergence, the model parameters are saved, enabling the network to perform modulation recognition and SNR estimation simultaneously.

[0009] The system flowchart of the modulation recognition and signal-to-noise ratio estimation method based on multi-task neural networks is attached. Figure 1 As shown, the process includes generating a simulation dataset, preprocessing the dataset, building and training a multi-task network model, and testing and outputting the predicted modulation type and signal-to-noise ratio on a test set.

[0010] The modulation recognition and signal-to-noise ratio estimation method based on a multi-task neural network includes the following steps:

[0011] Step 1: Generate a dataset of modulated signals using MATLAB software simulation. (See attached image) Figure 2 As shown, a random binary sequence d(t) is first generated, which is then modulated by a signal modulator to produce modulated data s(t). The modulated data s(t) is then upsampled and subjected to a pulse shaping filter to adjust its waveform. The filter used in this model is a square-root raised cosine (SRRC) pulse shaping filter g(t), the expression of which is as follows:

[0012]

[0013] In the formula: t is the signal duration; Ts is the symbol period; α is the roll-off factor. The expression for the signal x′(t) after s(t) passes through the filter g(t) can be written as:

[0014]

[0015] In the formula: Tx is the sampling time interval. The waveform-adjusted x'(t) data is then power normalized to ensure signal power consistency across different modulation schemes. The normalization process first calculates the average power of the filtered signal, then scales the signal to achieve the required symbol energy. The resulting expression for the data x(t) is:

[0016]

[0017] Step 2: After the power-normalized data is converted into an radio frequency signal, it is transmitted through the wireless channel. The modeled channel environment is an additive white Gaussian noise (AWGN) channel environment w(t), with a fixed noise power. The signal-to-noise ratio (SNR) is controlled by adjusting the signal power to achieve the desired SNR level. After this process, the expression for the modulated signal y(t) can be generated as follows:

[0018] y(t)=x(t)+n(t) (5)

[0019] In the formula: x(t) is the power-scaled transmission signal, and n(t) is additive white Gaussian noise.

[0020] Step 3: Separate the real part r(t) and imaginary part i(t) of the generated complex signal data y(t), and then concatenate them into a two-dimensional matrix A(t). Then, feed A(t) and the corresponding labels of the two-dimensional array into the MTMCN multi-task neural network for training, dynamically adjust the learning rate to accelerate network convergence, and finally save the converged model parameters.

[0021] The multi-task network model MTMCN, which combines modulation recognition and signal-to-noise ratio estimation, is attached. Figure 3 As shown, it includes: input mapping operation, Mamba module, MSDCAM module, MHAAP module, and feature enhancement module.

[0022] The input mapping operation linearly transforms the input two-dimensional matrix, converting the input data feature space from 2 channels to 64 channels through convolution. This facilitates the subsequent extraction of local detail features from the modulated signal by the Mamba and MSDCAM modules.

[0023] The Mamba module is attached. Figure 4As shown, it contains two basic blocks consisting of an SSM block, a LayerNorm layer, and a Dropout layer. This branch combines a positional coding layer and a Mamba module to focus on long-term temporal dependencies and capture global signal information.

[0024] The MSDCAM module is attached. Figure 3 As shown in (d), the algorithm includes a multi-scale extended convolutional module, a CBAM attention mechanism, channel concatenation, and channel convolution fusion. The features after input mapping are first passed through four multi-scale average channels into corresponding dilated convolutional layers. Different dilation factors and kernel sizes are used to expand the receptive field and extract more features. Channel attention is added after each branch channel to focus on the features at each scale. Then, the four sets of extracted features are concatenated and fused before being input into spatial attention to extract the temporal region containing the most information, while suppressing noise-dominated parts. This branch is mainly used to capture local variations in the signal.

[0025] The feature stitching operation performs channel stitching on the data after passing through the Mamba and MSDCAM branches to create information containing comprehensive features through complementary combinations.

[0026] The MHAAP module is attached. Figure 5 As shown, this module implements a complex pooling mechanism that combines attention-based pooling with traditional statistical pooling methods (global average pooling and max pooling). By adaptively focusing on the signal segment with the most information while preserving global sequence statistics, it overcomes the limitations of traditional pooling operations.

[0027] After receiving information from MHAAP, the feature enhancement module outputs the modulation recognition result and the signal-to-noise ratio estimation result through two output mappings.

[0028] Step 4: Input the test set data into the trained multi-task neural network model, which will then automatically output the classification results of the data modulation method and the signal-to-noise ratio prediction results.

[0029] The advantages of this invention are:

[0030] 1. A dual-path processing architecture is designed, which combines the Mamba module for global temporal dependency modeling and is connected in parallel with the MSDCAM module for focusing on local sequence feature information. Both processing paths provide feature diversity and robustness. Their outputs are then concatenated to form a comprehensive feature representation, combining the complementary advantages of the two methods.

[0031] 2. An MSDCAM module was designed, which integrates multi-scale dilated convolution and CBAM attention mechanisms. It implements four parallel expanding convolutional paths, each progressively expanding the receptive field through different combinations of kernel size and dilation rate. This design enables the simultaneous capture of local feature information and global feature information across different temporal scales, enriching the information flow between deep and shallow layers of the network. This allows for efficient capture of cross-channel and spatial interactions, further improving the overall performance of the network.

[0032] 3. An MHAAP module was designed. This module generates global sequence information through a learnable attention mechanism weight, allowing the network to adaptively focus on the most informative part of the input signal while suppressing less relevant information. In addition, it is connected in parallel with average pooling and max pooling, which enables the network to pay more comprehensive attention to key feature information and improve the performance of the neural network.

[0033] 4. Building upon the previous work, a multi-task neural network model, MTMCN, was constructed, capable of simultaneously performing modulation recognition and signal-to-noise ratio estimation. Compared with mainstream network models, this model boasts superior performance while also possessing the advantages of low complexity and a small number of parameters.

[0034] 5. Compared with two corresponding single-task networks, the proposed multi-task network model has a system cost that is basically the same as that of a single task network, while slightly improving the accuracy of classification and prediction and significantly reducing the inference speed of the network. Furthermore, the multi-task network model can share feature extraction layers, effectively avoiding overfitting and parameter complexity, and improving the network's generalization ability and robustness. Attached Figure Description

[0035] Figure 1 This is a system flowchart for a modulation recognition and signal-to-noise ratio estimation method based on a multi-task neural network.

[0036] Figure 2 This is a system framework diagram for a modulation recognition and signal-to-noise ratio estimation method based on a multi-task neural network.

[0037] Figure 3 This is a model structure diagram of the multi-task neural network MTMCN constructed in this invention.

[0038] Figure 4 This is a model structure diagram of the Mamba module introduced in this invention.

[0039] Figure 3 This is a model structure diagram of the MSDCAM module designed in this invention.

[0040] Figure 5 This is a model structure diagram of the MHAAP module designed in this invention.

[0041] Figure 6 This is a diagram of the MTMCN structure constructed in this invention. Detailed Implementation

[0042] The following is in conjunction with the instruction manual appendix. Figures 1 to 6 The specific embodiments of the present invention will be described in further detail below.

[0043] Current mainstream methods for applying deep learning to the communications field model single tasks, such as neural networks that only classify the types of modulation signals. This invention proposes a multi-task neural network model, MTMCN, that jointly performs modulation recognition and signal-to-noise ratio (SNR) estimation. The network receives a two-dimensional matrix composed of the real and imaginary parts of the signal data as input. First, this matrix is ​​transformed to linear channel dimensions through a projection input layer. Then, the Mamba module extracts global features from the input data and is connected in parallel with the MSDCAM module for extracting local details to achieve complementarity. Next, the two branches are concatenated and the network extracts key temporal features through the MHAAP module, establishing global focal information relationships. Finally, the network maps these features to the output layer through a fully connected layer. The output layer has two sub-tasks: modulation recognition and SNR estimation, which can be performed simultaneously, reducing computational resource consumption during training and improving the network's inference speed.

[0044] The modulation recognition and signal-to-noise ratio estimation method based on multi-task neural networks is attached. Figure 1 As shown, it includes the following steps:

[0045] Step 1: Generate the simulation dataset using Matlab. The dataset used in this experiment contains 310,000 data points, including 10 modulation schemes (BPSK, QPSK, OQPSK, 8PSK, 16QAM, 64QAM, MSK, GMSK, 16APSK, and 32APSK) and 31 signal-to-noise ratios (-10-20dB intervals in 1dB increments). Each modulation scheme has 1000 data points per signal-to-noise ratio, and each data point is 2×128 pixels in size.

[0046] Step 2: Separate the real and imaginary parts of the dataset to form a two-dimensional matrix. Divide the dataset into training, validation, and test sets in a 7:1:2 ratio, i.e., use 700 data points for training, 100 data points for validation, and 200 data points for testing.

[0047] Step 3: Build a modulation recognition and signal-to-noise ratio estimation model based on a multi-task neural network. Input the training dataset and the validation dataset, dynamically adjust the learning rate to make the model converge, and finally save the converged model and parameters.

[0048] Step 4: Feed the test set data into the converged model, which will then output the modulation type and signal-to-noise ratio of the data.

[0049] As attached Figure 3 As shown, the Modulation Recognition and Signal-to-Noise Ratio Estimation Network (MTMCN) model based on a multi-task network includes an input mapping operation, a Mamba module, an MSDCAM module, an MHAAP module, and a feature enhancement module.

[0050] The input mapping operation comprises three basic units: a fully connected layer (Linear), layer normalization (LayerNorm), activation function (PReLU), and dropout. The first fully connected layer has 2 input channels and 32 output channels, and performs layer normalization and PReLU on the resulting features. The second fully connected layer outputs 64 channels. The input 1*128*2 data is transformed into batch_size*128*64 data after the input mapping operation.

[0051] The Mamba branch uses Mamba2Simple, where data is position-encoded before entering the Mamba module, such as... Figure 4 As shown, the state dimension is 64, the convolution kernel dimension is 4, and the expansion factor is 2. Its main function is to extract global feature information from the data. The input batch_size*128*64 data becomes batch_size*128*64 after this operation.

[0052] MSDCAM module, such as Figure 5 As shown, it consists of multi-scale dilated convolution (DCONV) and convolutional block attention (CBAM), designed in parallel to enrich the network's receptive field for extracting broader features and obtaining local feature information of the modulated signal. After receiving the output from the input projection layer, it divides the channels into four groups for operation to enhance the network's feature representation capability. The input feature F... C / 4 After dilated convolution, the feature F of each channel is obtained. i (i = 1, 2, 3, 4), a CA attention mechanism is added after each channel to obtain the feature F after channel-dimensional attention weight optimization. i ′(i=1,2,3,4),F i After feature concatenation, F″ is obtained, which is then fused with a kernel size of 1 to obtain F″′. F″′ enters the SA attention module to obtain the feature F″″ after double optimization of attention weights in both the channel dimension and the spatial dimension, thereby extracting key information and obtaining the feature-enhanced feature F″″. The entire process can be summarized as follows:

[0053] F i=Dilated_Conv(F i (5)

[0054] F′ i =M C (F i (6)

[0055]

[0056] F″′=Conv(F″,K) (8)

[0057] F″″=M s (F″′) (9)

[0058] In the formula: K is the convolution kernel, S is the stride, P is the padding, D is the dilation factor, and M is the expansion factor. C Represents channel attention; M S _ represents spatial attention, and _Dliated_Conv represents dilated convolution.

[0059] The MSDCAM module we use does not change the dimensions and size of the data, so the output feature size remains batch_size*128*64.

[0060] The data output from the two branches are concatenated to obtain the output feature size batch_size*128*128, which is then passed to the MHAAP module.

[0061] The MHAAP module consists of learnable multi-head attention pooling, average pooling, and max pooling, as shown in the appendix. Figure 6 As shown, the original input features are first divided into three paths and subjected to learnable multi-head attention pooling, average pooling, and max pooling respectively. The results from the three paths are then concatenated. The core objective is to maximize the multi-dimensional information value of the input features while taking into account the key focus, global representativeness, and local saliency of the features, ultimately improving the model's ability to express the features. Finally, the output of the MHAAP module is passed to the feature enhancement module.

[0062] The feature enhancement module outputs two branches after passing through multiple fully connected layers: one branch outputs a 10-class classification structure, and the other branch outputs a single prediction result. The former corresponds to the modulation recognition task, while the latter corresponds to the signal-to-noise ratio estimation task.

[0063] Example:

[0064] The modulation recognition and signal-to-noise ratio estimation method based on multi-task neural networks includes: a data generation module, a training model module, and a testing model module.

[0065] The data generation module primarily prepares the dataset needed for neural network training. This dataset was generated through MATLAB simulation and includes 10 modulation schemes: BPSK, QPSK, OQPSK, 8PSK, 16QAM, 64QAM, MSK, GMSK, 16APSK, and 32APSK. The signal-to-noise ratio ranges from -10dB to 20dB in 1dB increments, with an oversampling factor of 4 and a signal length of 128. The channel used is additive white Gaussian noise (AWGN). To facilitate neural network training, the real and imaginary parts of the data are separated into a 2*128 two-dimensional matrix. Each data point is labeled accordingly, and the dataset is divided into training, validation, and test sets in a 7:1:2 ratio.

[0066] The main process of the model training module is to train the model using the training set data, and then verify the model's convergence and overfitting using the validation set data. Finally, the converged model and its parameters are saved. Details are as follows:

[0067] First, the pre-packaged training set data is read in for training. The Adam optimizer is used during training with an initial learning rate of 0.001, employing a dynamic learning rate method. For network error, a composite loss function is used to evaluate the error. Cross-entropy loss is used to evaluate the error of the modulation recognition task, and smooth L1 loss is used to evaluate the signal-to-noise ratio estimation task error. Furthermore, to ensure that the two tasks have equal weight in the network, their loss ratios are set to ensure that the loss weights of the two tasks are the same, thus allowing the network to achieve optimal performance for each sub-task. After each training round, validation set data is fed into the network for validation to monitor the network state in real time and prevent overfitting. Finally, after training is complete, the converged network model is saved.

[0068] The test model module primarily uses test set data to verify the model's accuracy and determine its precision. The main steps are as follows: First, retrieve the test set data and then feed it into the trained multi-task neural network model. The model can automatically output the modulation scheme and signal-to-noise ratio (SNR) of the input data. By comparing this output with the actual modulation scheme and SNR, the model's accuracy and precision can be determined.

[0069] Finally, the data generation module, training model module, and testing model module are linked together in sequence.

[0070] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A modulation identification and signal-to-noise ratio estimation method, characterized in that, The method includes: The radio frequency signal transmitted through the wireless channel is converted to obtain a baseband modulated signal. The radio frequency signal is obtained by processing the transmitted data with a modulation method and a signal-to-noise ratio. The wireless channel is a channel using additive white Gaussian noise. The real and imaginary parts of the baseband modulation signal are separated, and the separated real and imaginary parts are concatenated into a two-dimensional matrix; A pre-trained multi-task model is obtained, which includes an input mapping operation, a Mamba module, a multi-scale grouped dilated convolution MSDCAM module with CBAM attention mechanism, a MHAAP module with three pooling methods in parallel, and a fully connected module. The MSDCAM module is composed of four groups of dilated convolutions with different kernel sizes and dilation rates connected in parallel with channel attention mechanisms and then fused by a spatial attention mechanism. The MHAAP module is composed of multi-head attention pooling, global average pooling, and max pooling in parallel. The modulation scheme and signal-to-noise ratio of the transmitted data are obtained by processing the two-dimensional matrix using the multi-task model.

2. The modulation identification and signal-to-noise ratio estimation method according to claim 1, characterized in that, The step of processing the two-dimensional matrix using the multi-task model to obtain the modulation scheme and signal-to-noise ratio of the transmitted data includes: The input mapping module is used to perform channel mapping on the two-dimensional matrix to obtain high-dimensional channel features; The Mamba module is used to perform position encoding and global correlation feature extraction on the high-dimensional channel features to obtain time-series features; The MSDCAM module is used to extract local detail features from the high-dimensional channel features to obtain detail features; The MHAAP module is used to aggregate the temporal features and detail features and extract salient features to obtain pooled features; The pooling features are processed using the fully connected module to obtain the modulation scheme and signal-to-noise ratio of the transmitted data.

3. The modulation identification and signal-to-noise ratio estimation method according to claim 2, characterized in that, The MSDCAM module includes four sets of dilated convolutional cascaded channel attention mechanisms with different kernel sizes and dilation rates. These four branches are then fused into a spatial attention mechanism. The MSDCAM module is used to extract local detail features from the high-dimensional channel features, yielding detailed features including: Channel segmentation is used to perform four-branch average division of high-dimensional channel features; The first DConv feature is obtained by using a dilated convolutional cascaded channel attention mechanism with a kernel size of three and a dilation rate of one. The second DConv feature is obtained by using a dilated convolutional cascaded channel attention mechanism with a kernel size of five and a dilation rate of two. The third DConv feature is obtained by using a dilated convolutional cascaded channel attention mechanism with a kernel size of seven and a dilation rate of four. The fourth DConv feature is obtained through a dilated convolutional cascaded channel attention mechanism with a kernel size of nine and an inflation rate of eight. Feature concatenation is used to perform feature concatenation processing on the first to fourth features; The fused features are obtained by performing convolution with a kernel size of one. The detailed features are obtained by processing the fused features using a channel attention mechanism.

4. The modulation identification and signal-to-noise ratio estimation method according to claim 2, characterized in that, The process of aggregating the detailed features and extracting salient features using the MHAAP module to obtain pooled features includes: The attention pooling method is used to adaptively and dynamically assign weights to the detailed features to focus on key features relevant to the task, thereby obtaining attention features. The global average pooling is used to significantly extract global information from the detailed features, resulting in global average features; The maximum feature is obtained by extracting the local most salient features from the detailed features using the max pooling method. The pooling features are obtained by combining the attention features, the global average features, and the maximum features.

5. The modulation identification and signal-to-noise ratio estimation method according to any one of claims 1 to 4, characterized in that, The method further includes: A training dataset is created, in which training samples include a two-dimensional matrix, a true modulation scheme, and a true signal-to-noise ratio (SNR). The two-dimensional matrix is ​​obtained by processing simulation data using the true modulation scheme and the true SNR. The processed radio frequency signal is transmitted using a simulated wireless channel. The resulting complex signal is obtained by separating the real and imaginary parts. The simulated wireless channel uses additive Gaussian noise to simulate a channel. Create the aforementioned multi-task model; For each training sample, the two-dimensional matrix in the training sample is processed using the multi-task model to obtain the predicted modulation scheme and the predicted signal-to-noise ratio; The loss value of the multi-task model is calculated based on the predicted modulation scheme and the actual modulation scheme, the predicted signal-to-noise ratio and the actual signal-to-noise ratio, and the model parameters of the multi-task model are trained based on the loss value.

6. The modulation identification and signal-to-noise ratio estimation method according to claim 5, characterized in that, The creation of the training dataset includes: A real modulation scheme and a real signal-to-noise ratio are randomly selected using a signal modulator. The randomly generated binary sequence is modulated using the selected real modulation scheme and the real signal-to-noise ratio to obtain modulated data. The waveform of the modulated data is adjusted using a square root raised cosine pulse filter, and the shaped data is then converted into a radio frequency signal after power normalization. The radio frequency signal is transmitted using an analog wireless channel, which is an analog channel using additive white Gaussian noise; The radio frequency signal transmitted through the wireless channel is converted, the real and imaginary parts of the baseband modulation signal obtained after conversion are separated, and the separated real and imaginary parts are concatenated into a two-dimensional matrix. The two-dimensional matrix, the actual modulation scheme, and the actual signal-to-noise ratio are combined to form a training sample.

7. A modulation identification and signal-to-noise ratio estimation device, characterized in that... The device includes: The conversion module is used to convert the radio frequency signal transmitted through the wireless channel to obtain a baseband modulated signal. The radio frequency signal is obtained by processing the transmitted data with a modulation method and a signal-to-noise ratio. The wireless channel is a channel using additive white Gaussian noise. The splicing module is used to separate the real and imaginary parts of the baseband modulation signal and splice the separated real and imaginary parts into a two-dimensional matrix. The acquisition module is used to acquire a pre-trained multi-task model. The multi-task model includes an input mapping module, a Mamba module, an MSDCAM module, an MHAAP module, and a fully connected module. The MSDCAM module is composed of four groups of dilated convolutions with different kernel sizes and dilation rates connected in parallel with channel attention mechanisms and then fused by a spatial attention mechanism. The MHAAP module is composed of attention pooling, global average pooling, and max pooling connected in parallel. The processing module is used to process the two-dimensional matrix using the multi-task model to obtain the modulation scheme and signal-to-noise ratio of the transmitted data.

8. A computer-readable storage medium, characterized in that... The storage medium stores at least one instruction, which is loaded and executed by a processor to implement the modulation identification and signal-to-noise ratio estimation method as described in any one of claims 1 to 6.

9. A computer device, characterized in that, The computer device includes: the modulation identification and signal-to-noise ratio estimation apparatus as described in claim 7.

Images