Modulation identification and code element rate estimation method and device, storage medium and equipment

Abstract

The invention discloses a modulation identification and code element rate 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; the multi-task model comprises a convolution module, a pooling module, a DRA module, an SST module and a full connection module, and the DRA module is composed of a plurality of DA units and a jump connection structure with an ECA unit; the SST module is composed of a Patch Partition unit, a linear embedding unit and a Swin TransformerBlock unit, and the Swin TransformerBlock unit is composed of a Patch Partition unit, a linear embedding unit and a Swin TransformerBlock unit; and processing the two-dimensional matrix by using a multi-task model to obtain a modulation mode and a code element rate. According to the method, modulation identification and code rate estimation can be synchronously carried out, the performance is good, the complexity and parameter quantity are low, the overhead is small, the deduction speed is high, and the generalization ability and robustness are high.

Description

Technical Field

[0001] The present application relates to the field of communication technologies, and particularly to a modulation recognition and symbol rate estimation method, apparatus, storage medium, and device. Background Art

[0002] With the rapid development of global technology, radio communication technology has made remarkable progress in terms of transmission efficiency, reliability, and diverse applications. In modern communication systems, all communication signals (radios, televisions, mobile communications, etc.) need to go through a modulation process before transmission. However, in the context of non-cooperative radio communication, the interactivity between the two communication parties is significantly weakened, and information such as the type of the transmitted signal, communication parameters, and pilot header information is unknown. In this context, modulation methods and symbol rates, as key communication parameters, their accurate recognition is directly related to the success of signal demodulation and decoding. This problem is not only one of the core technical difficulties in the field of wireless communication but also an important technical support in electronic countermeasures and reconnaissance in the military field.

[0003] Although traditional methods have achieved certain results in the fields of modulation recognition and parameter estimation, the technical bottlenecks and deficiencies they face are gradually emerging, making it difficult to meet the increasingly complex wireless communication requirements. Most traditional algorithms rely on feature extraction means and prior knowledge designed by experts, which not only leads to a high dependence on specific scenarios but also significantly reduces their adaptability in dynamic and complex environments. In addition, the performance of traditional methods in dealing with high-dimensional complex signals is also relatively limited. With the development of communication technology towards higher frequency bands and broadband, the complexity of signals also poses higher requirements for the flexibility and scalability of algorithms. Therefore, it is particularly important to develop more flexible and efficient methods to overcome these limitations.

[0004] In recent years, the rapid development of artificial intelligence technology has triggered profound changes in multiple fields. As an important branch of artificial intelligence, deep learning has shown excellent performance in tasks such as image recognition and natural language processing due to its powerful data processing capabilities and the advantages of modeling complex non-linear relationships. In addition, thanks to the availability of large-scale datasets, the significant improvement in the performance of computing hardware, and the continuous innovation of advanced algorithms, the application of deep learning has rapidly expanded to multiple fields such as healthcare, finance, and autonomous driving. In these fields, deep learning demonstrates excellent performance, especially in solving complex problems that are difficult to handle with traditional methods, and has irreplaceable advantages. The innovation and development of deep learning technology have made it an important driving force for promoting progress in modern science and engineering fields. In modern communication systems, with the continuous growth of the demand for data transmission rate and spectrum utilization efficiency, signal processing technology is facing unprecedented challenges. Traditional methods are limited by complex feature engineering and dynamic adaptability, while deep learning improves task accuracy and efficiency by automatically extracting features, providing new solutions and methods for solving problems in modern communication systems, and demonstrating great potential in processing complex signal environments, optimizing system performance, and improving the level of intelligence. It performs outstandingly in key tasks such as modulation recognition and symbol rate estimation, and can especially adapt to diverse signal conditions and dynamic channel environments. Through parallel computing and optimization, deep learning further enhances the real-time performance and reliability of communication systems, promoting the intelligent development of the systems. Deep learning models can also adapt to different signal-to-noise ratios, spectrum conditions, and modulation schemes, and have strong generalization capabilities, which are crucial in modern communication systems. In the future, deep learning technology will unleash its potential in higher dimensions and complex scenarios, providing new impetus for the innovation of modern communication technology. Summary of the Invention

[0005] This application provides a method, device, storage medium, and equipment for modulation recognition and symbol rate estimation, which are used to solve the problems that traditional algorithms have a high dependence on specific scenarios, low adaptability in dynamic and complex environments, and poor performance in processing high-dimensional complex signals. The technical solutions are as follows:

[0006] According to the first aspect of this application, a method for modulation recognition and symbol rate estimation is provided, and the method includes:

[0007] Convert the radio frequency signal transmitted through the wireless channel to obtain a baseband modulation signal. The radio frequency signal is obtained by processing the transmission data with a modulation method and a symbol rate, and the wireless channel is a channel using Rayleigh fading superimposed with additive white Gaussian noise;

[0008] Separate the real part and the imaginary part of the baseband modulation signal, and splice the separated real part and imaginary part into a two-dimensional matrix;

[0009] Obtain a pre-trained multi-task model, where the multi-task model includes a convolutional module, a pooling module, a depthwise separable convolution DRA module with an attention mechanism and residual connections, a simplified version SST module of SwinTransformer, and a fully connected module. The DRA module consists of p depthwise separable convolution DA units with an attention mechanism and q skip connection structures with an efficient channel attention mechanism ECA units. The DA unit is a depthwise separable convolution DSC with a convolutional block attention CBAM mechanism, and p and q are positive integers. The SST module consists of a Patch Partition unit, a linear embedding unit, and two cascaded SwinTransformerBlock units.

[0010] Process the two-dimensional matrix using the multi-task model to obtain the modulation method and symbol rate of the transmission data.

[0011] In a possible implementation manner, the process of using the multi-task model to process the two-dimensional matrix to obtain the modulation method and symbol rate of the transmission data includes:

[0012] Perform convolution on the two-dimensional matrix using the convolutional module to obtain convolution features.

[0013] Perform downsampling on the convolution features using the pooling module to obtain first pooling features.

[0014] Perform local detail feature extraction on the first pooling features using the DRA module to obtain detail features.

[0015] Perform aggregation and significant feature selection on the detail features using the pooling module to obtain second pooling features.

[0016] Perform global correlation feature extraction on the second pooling features using the SST module to obtain temporal features.

[0017] Process the temporal features using the fully connected module to obtain the modulation method and symbol rate of the transmission data.

[0018] In a possible implementation manner, the DRA module includes ten cascaded DA units and three skip connection structures with ECA units. The first ECA unit skip connects the fourth DA unit and the sixth DA unit, the second ECA unit skip connects the second DA unit and the eighth DA unit, and the third ECA unit skip connects the input end and the output end of the DRA module. The process of using the DRA module to perform local detail feature extraction on the first pooling features to obtain detail features includes:

[0019] Process the first pooling feature using the first to sixth DA units connected in series;

[0020] Add the first DA feature output by the fourth DA unit and the second DA feature output by the sixth DA unit through the first ECA unit to obtain the third DA feature;

[0021] Process the third DA feature using the seventh and eighth DA units connected in series;

[0022] Add the fourth DA feature output by the second DA unit and the fifth DA feature output by the eighth DA unit through the second ECA unit to obtain the sixth DA feature;

[0023] Process the sixth DA feature using the ninth and tenth DA units connected in series;

[0024] Add the first pooling feature and the seventh DA feature output by the tenth DA unit through the third ECA unit to obtain the detail feature.

[0025] In a possible implementation, the using the SST module to extract global correlation features from the second pooling feature to obtain a timing feature includes:

[0026] Use the Patch Partition unit to partition the second pooling feature and flatten it in the channel direction to obtain a partitioned feature;

[0027] Use the linear embedding unit to perform a linear transformation on the partitioned feature to obtain a linear feature;

[0028] Use two Swin TransformerBlock units connected in series to extract time features from the linear feature to obtain the timing feature.

[0029] In a possible implementation, the two Swin TransformerBlock units include layer normalization LN, window-based multi-head self-attention mechanism W-MSA, LN, multi-layer perceptron MLP, LN, shifted window-based multi-head self-attention mechanism SW-MSA, LN, and MLP. The using two Swin TransformerBlock units connected in series to extract time features from the linear feature to obtain the timing feature includes:

[0030] Process the linear feature using the LN and W-MSA connected in series to obtain the first ST feature;

[0031] Add the linear feature and the first ST feature to obtain the second ST feature;

[0032] Process the second ST feature using a cascaded LN and MLP to obtain a third ST feature;

[0033] Add the second ST feature and the third ST feature to obtain a fourth ST feature;

[0034] Process the fourth ST feature using a cascaded LN and SW-MSA to obtain a fifth ST feature;

[0035] Add the fourth ST feature and the fifth ST feature to obtain a sixth ST feature;

[0036] Process the sixth ST feature using a cascaded LN and MLP to obtain a seventh ST feature;

[0037] Add the sixth ST feature and the seventh ST feature to obtain the timing feature.

[0038] In a possible implementation, the method further includes:

[0039] Create a training data set, where the training samples in the training data set include a two-dimensional matrix, a true modulation mode, and a true symbol rate. The two-dimensional matrix is obtained by processing simulation data using the true modulation mode and the true symbol rate, transmitting the processed radio frequency signal using an analog wireless channel, and separating the real and imaginary parts of the complex signal obtained after transmission. The analog wireless channel is an analog channel using Rayleigh fading superimposed with additive white Gaussian noise;

[0040] Create the multi-task model;

[0041] For each training sample, process the two-dimensional matrix in the training sample using the multi-task model to obtain a predicted modulation mode and a predicted symbol rate;

[0042] Calculate the loss value of the multi-task model according to the predicted modulation mode and the true modulation mode, the predicted symbol rate and the true symbol rate, and train the model parameters of the multi-task model according to the loss value.

[0043] In a possible implementation, the creating of the training data set includes:

[0044] Randomly select a true modulation mode and a true symbol rate using a signal modulator, modulate a randomly generated binary sequence using the selected true modulation mode and true symbol rate to obtain modulation data, where the modulation data includes N symbols;

[0045] Adjust the waveform of the modulated data using a square root raised cosine pulse filter, and convert the shaped data into a radio frequency signal after energy normalization;

[0046] Transmit the radio frequency signal using an analog wireless channel, where the analog wireless channel is an analog channel using Rayleigh fading superimposed with additive white Gaussian noise;

[0047] Convert the radio frequency signal transmitted through the wireless channel, separate the real and imaginary parts of the baseband modulated signal obtained after conversion, and splice the separated real and imaginary parts into a two-dimensional matrix;

[0048] Form a training sample by combining the two-dimensional matrix, the true modulation method, and the true symbol rate.

[0049] According to the second aspect of the present application, there is provided a modulation recognition and symbol rate estimation device, the device includes:

[0050] A conversion module for converting the radio frequency signal transmitted through the wireless channel to obtain a baseband modulated signal, where the radio frequency signal is obtained by processing transmission data with a modulation method and a symbol rate, and the wireless channel is a channel using Rayleigh fading superimposed with additive white Gaussian noise;

[0051] A splicing module for separating the real and imaginary parts of the baseband modulated signal and splicing the separated real and imaginary parts into a two-dimensional matrix;

[0052] An acquisition module for acquiring a pre-trained multi-task model, the multi-task model includes a convolution module, a pooling module, a depthwise separable convolution DRA module with an attention mechanism and a residual connection, a simplified version SST module of Swin Transformer, and a fully connected module. The DRA module consists of p depthwise separable convolution DA units with an attention mechanism and q skip connection structures with an efficient channel attention mechanism ECA unit. The DA unit is a depthwise separable convolution DSC with a convolutional block attention CBAM mechanism, and p and q are positive integers; the SST module consists of a Patch Partition unit, a linear embedding unit, and two cascaded Swin Transformer Block units;

[0053] A processing module for processing the two-dimensional matrix using the multi-task model to obtain the modulation method and symbol rate of the transmission data.

[0054] According to the third aspect of the present application, there is provided a computer-readable storage medium, in which at least one instruction is stored, and the at least one instruction is loaded and executed by a processor to implement the modulation recognition and symbol rate estimation method as described above.

[0055] According to the fourth aspect of the present application, a computer device is provided, and the computer device includes the above modulation recognition and symbol rate estimation device.

[0056] The beneficial effects of the technical solution provided by the present application at least include:

[0057] Aiming at the problem of performance degradation caused by separating the two dimensions of space and channel in DSC, we designed to add the CBAM mechanism after depth convolution and pointwise convolution respectively to learn the correlation in the spatial dimension and channel dimension, thereby reducing the complexity and the number of parameters of the multi-task model and improving the model performance. While improving DSC, we added a skip connection structure dominated by ECA to form a DRA module, which can alleviate the problem of insufficient deep weight update caused by gradient disappearance or explosion during backpropagation of the model, enrich the information flow of the shallow and deep layers of the model, so as to efficiently capture the interaction relationship between cross-channels and further improve the overall performance of the model.

[0058] By introducing the window-based multi-head attention mechanism and the sliding window-based multi-head attention mechanism of the SwinTransformerBlock unit to form the SST module, while ensuring that the model has the ability to extract global features, it maintains the characteristics of low complexity.

[0059] The multi-task model with the DRA module and the SST module can perform modulation recognition and symbol rate estimation synchronously. Compared with the mainstream models, this model has the advantages of good performance, low complexity and low number of parameters. And, compared with two corresponding single-task models, the system overhead of the multi-task model is basically the same as that of one of the single-task models, and the accuracy of classification and prediction is slightly improved, greatly reducing the deduction speed of the model. In addition, the multi-task model can share the feature extraction layer, effectively avoiding overfitting and parameter complication, and improving the generalization ability and robustness of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0061] Figure 1 is a schematic structural diagram of a multi-task model shown according to an embodiment of the present application;

[0062] Figure 2 is a schematic structural diagram of a DRA module shown according to an embodiment of the present application;

[0063] Figure 3 It is a schematic structural diagram of an SST module shown according to an embodiment of the present application;

[0064] Figure 4 It is a schematic structural diagram of a multi-task model shown according to an embodiment of the present application;

[0065] Figure 5 It is a training flow chart of a multi-task model shown according to an embodiment of the present application;

[0066] Figure 6 It is a schematic diagram of a communication system model shown according to an embodiment of the present application;

[0067] Figure 7 It is a flow chart of a modulation recognition and symbol rate estimation method shown in an embodiment of the present application;

[0068] Figure 8 It is a detailed flow chart of a modulation recognition and symbol rate estimation method shown in an embodiment of the present application;

[0069] Figure 9 It is a structural block diagram of a modulation recognition and symbol rate estimation device shown in an embodiment of the present application. Specific Embodiments

[0070] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the following will further describe the embodiments of the present application in detail with reference to the accompanying drawings.

[0071] Existing mainstream methods that apply deep learning to the communication field all model a single task. For example, a model that only classifies the types of debugging signals. However, the present application proposes a multi-task model DRASST-MTL for joint modulation recognition and symbol rate estimation. As Figure 1 shown, the multi-task model DRASST-MT includes a convolutional module, a pooling module, a DRA module, an SST module, and a fully connected module. Among them, the convolutional module, the pooling module, the DRA module, the pooling module, the SST module, and the fully connected module are connected in sequence.

[0072] The convolutional module can be Conv2d, that is, a two-dimensional convolutional module, which is used to perform a convolutional operation on the input data. The convolutional operation includes three basic units, namely two-dimensional convolution (Conv2d), two-dimensional normalization (BatchNorm2d), and an activation function (ReLU). Among them, the convolution kernel size of the two-dimensional convolution is 1*3, and 1 unit is filled on each side in the width direction of the input data, and the height direction remains unchanged. The number of channels after convolution is 32. If the input is data of 1*2*2048, after the convolutional operation, the data is converted into data of 32*2*2048.

[0073] The first pooling module can be MaxPool2d, that is, a two-dimensional maximum pooling module, which is used to perform pooling operations on the input data. The pooling operation uses max pooling (MaxPool2d), and the size of the convolutional kernel is 1*2. Its main function is to downsample the data to reduce the network complexity. If the input data is 32*2*2048, after the pooling operation, the data is converted into 32*2*1024 data.

[0074] DRA is a depthwise separable convolution with an attention mechanism and residual connection (Depthwise separable convolution with Residual connection and Attention mechanism). The DRA module consists of p DA units and q skip connection structures with ECA units. DA is a depthwise separable convolution with an attention mechanism (Depthwise separable convolution with Attention mechanism). The attention mechanism mentioned here is the convolutional block attention (Convolutional Block Attention Module, CBAM) mechanism, that is, the DA unit is a DSC with the CBAM mechanism, aiming to alleviate the problem of cross-channel information loss.

[0075] In the DA module, the input feature F is depth-convolved to obtain the feature Fd of each channel; the CBAM mechanism is added after each channel to obtain the feature F' with optimized channel dimension attention weights; the feature F' then passes through the spatial attention module to obtain the feature F” with double optimized attention weights in both the channel dimension and the spatial dimension, realizing the extraction of key information and obtaining the feature F” with enhanced features; the features F” of each channel are concatenated in the channel dimension to obtain Fc; pointwise convolution is performed on Fc to achieve channel fusion and change the number of output channels to obtain the feature Fp; then the CBAM mechanism is added and the output feature F”' is obtained through equations (2) and (3). The whole process can be summarized as:

[0076] Fd = DWConv(F, K, S, P) (1)

[0077]

[0078] Fp = PWConv(Fc, K) (4)

[0079]

[0080] Where: K is the convolutional kernel, S is the stride, P is the padding, Mc represents channel attention, Ms represents spatial attention, DWConv represents depth convolution, and PWConv represents pointwise convolution.

[0081] ECA is the Efficient Channel Attention mechanism. The skip connection structure with ECA units can enhance the feature flow between the shallow and deep layers of the network. In the DRA module, a skip connection is provided between every two DA units, and the features are fused through ECA in the middle, realizing efficient significant information interaction and extraction in both the spatial and channel dimensions and across channels to extract local detailed features.

[0082] Among them, p and q are positive integers. In one example, p = 10 and q = 3, that is, a DRA module includes ten DA units and three ECAs. As Figure 2 shown, the first ECA unit skip-connects the fourth DA unit and the sixth DA unit, the second ECA unit skip-connects the second DA unit and the eighth DA unit, and the third ECA unit skip-connects the input end and the output end of the DRA module. The basic structure of the DA unit is DSC, and their output channels are 32, 64, 64, 128, 128, 128, 64, 64, 32, 32 respectively, and the channel dimensions of the three ECAs are 32, 64, and 128 respectively.

[0083] The DRA module used by the multi-task model does not change the dimension and size of the data. Therefore, the data output by the DRA module still remains 32*2*1024 in size.

[0084] The second pooling module can be MaxPool2d, that is, a two-dimensional maximum pooling module. The second pooling module downsamples the data to reduce the network complexity. A maximum pooling operation is performed on the data output by the DRA module, and the output is data of 32*2*512.

[0085] SST is the Simplified Swin Transformer. As Figure 3As shown in the figure, the SST module consists of a PatchPartition unit, a LinearEmbedding unit, and two cascaded SwinTransformerBlock units. The two SwinTransformerBlock units include Layer Normalization (LN), Window-based Multi-head Self-attention (W-MSA), LN, Multi-Layer Perceptron (MLP), LN, Shifted Window-based Multi-head Self-attention (SW-MSA), LN, and MLP. Among them, the settings of SwinTransformer are as follows: the patch_size is (2, 4), the sliding window size is 7*7, the model depth is 4, one block contains 8 attention heads, and the number of classes is set to 8 according to the classification task.

[0086] The input data is first partitioned into blocks by the Patch Partition unit, then flattened in the channel direction, and then linearly transformed by the Linear Embedding unit for input into the two Swin Transformer Block units. Finally, the two SwinTransformerBlock units globally extract the time series features. In the SwinTransformerBlock unit, W-MSA significantly reduces the computational complexity but ignores the information interaction between windows. Therefore, to strengthen the connection of data features between windows, SW-MSA is introduced on the basis of W-MSA, and the adjacent and non-overlapping windows are connected to each other by means of cyclic shifting, thereby expanding the receptive field. The cooperation of W-MSA and SW-MSA can not only capture the time series features of the signal and the global features of the data, but also significantly reduce the computational complexity.

[0087] The fully connected module has two branches. One branch outputs an 8-classification structure, that is, the result corresponding to modulation recognition, and the other branch outputs 1 prediction result, that is, the result corresponding to symbol rate estimation. As Figure 4 shown, the input data passes through the DRA module, the SST module, and the fully connected module respectively, and then outputs the results of modulation recognition and code rate estimation.

[0088] As Figure 5 shown, the training process of the multi-task model DRASST-MTL is described below. The training process of the multi-task model DRASST-MTL includes:

[0089] Step 501: Create a training dataset. The training samples in the training dataset include a two-dimensional matrix, the true modulation mode, and the true symbol rate. The two-dimensional matrix is obtained by processing simulation data with the true modulation mode and the true symbol rate, transmitting the processed RF signal using an analog wireless channel, and separating the real and imaginary parts of the complex signal obtained after transmission. The analog wireless channel is an analog channel using Rayleigh fading superimposed with additive white Gaussian noise.

[0090] Specifically, creating the training dataset may include:

[0091] (1) Randomly select a true modulation mode and a true symbol rate using a signal modulator, and modulate the randomly generated binary sequence with the selected true modulation mode and true symbol rate to obtain modulated data, where the modulated data includes N symbols.

[0092] In this embodiment, a simulation dataset is generated through MATLAB software. The dataset has a total of 880,000 pieces of data, including 8 modulation modes, 10 symbol rates (randomly generated from 100K to 200K), and 11 signal-to-noise ratios (spaced at 2dB intervals from 0 to 20dB). The sampling frequency is 400KHz. For each symbol rate of each modulation mode at each signal-to-noise ratio, there are 1,000 pieces of data, and the size of each piece of data is 2×2048. The 8 modulation modes are Binary Phase Shift Keying (BPSK), Quadrature Phase Shift Keying (QPSK), Offset-QPSK, 8 Phase Shift Keying (8PSK), 16 Quadrature Amplitude Modulation (QAM), 64QAM, Minimum Shift Keying (MSK), and Gaussian Filtered Minimum Shift Keying (GMSK).

[0093] As Figure 6 shown, first generate a random binary sequence a(t), and generate modulated data s(t) through a signal modulator, which contains N symbols.

[0094] (2) Use a square root raised cosine pulse filter to adjust the waveform of the modulated data, and perform energy normalization on the shaped data and then convert it into an RF signal.

[0095] The pulse filter adopted by this model is the Square-Root Raised Cosine (SRRC) pulse filter \(g(t)\), and its expression is as follows:

[0096]

[0097] Where: \(g(t)\) is the SRRC filter, \(t\) is the signal duration, \(T\) is the symbol period, and \(\alpha\) is the roll-off factor.

[0098] The expression of the signal \(x'(t)\) after \(s(t)\) passes through the square-root raised cosine pulse filter \(g(t)\) can be written as:

[0099]

[0100] Where: \(x'(t)\) is the output after the data passes through the square-root raised cosine pulse filter, \(t\) is the signal duration, \(T\) is the symbol period; \(T_x\) is the sampling time interval. The oversampling multiple \(P\) of the square-root raised cosine pulse filter can be expressed as \(T / T_x\), that is, \(P = T / T_x\). Also, knowing that the sampling frequency \(F_s = 1 / T_x\), then the symbol rate \(R = 1 / T = F_s / P\) (this provides data labels for symbol rate estimation based on neural networks). Subsequently, the data \(x'(t)\) after waveform adjustment undergoes an energy normalization operation to obtain the data \(x(t)\), and the expression of \(x(t)\) is:

[0101]

[0102] Where: \(x(t)\) is the output after the data undergoes energy normalization.

[0103] The data after energy normalization is transformed into a radio frequency signal and then transmitted through a wireless channel.

[0104] (3) Use an analog wireless channel to transmit the radio frequency signal. The analog wireless channel is an analog channel that uses Rayleigh fading superimposed with additive white Gaussian noise.

[0105] The modeled channel environment is (Rayleigh Fading Channel), and through the channel environment \(w(t)\) of additive white Gaussian noise (AWGN), the expression of the baseband modulated signal \(y(t)\) can be generated as follows:

[0106] \(y(t)=hx(t)+w(t)\ (9)\)

[0107] Where: \(h\) is the Rayleigh fading parameter, and \(w(t)\) is the additive white Gaussian noise.

[0108] (4) Convert the radio frequency signal transmitted through the wireless channel, separate the real part and the imaginary part of the obtained baseband modulated signal, and splice the separated real part and imaginary part into a two-dimensional matrix.

[0109] Separate the real part r(t) and the imaginary part i(t) of the generated complex signal data y(t), and then splice them into a two-dimensional matrix A(t).

[0110] (5) Combine the two-dimensional matrix, the true modulation mode, and the true symbol rate to form a training sample.

[0111] For each two-dimensional matrix A(t) and the corresponding labeled tags (modulation mode and symbol rate), divide the training set, validation set, and test set according to the ratio of 7:1:2, that is, use 700 pieces of data for training, 100 pieces of data for validation, and finally 200 pieces of data for testing. Train the model with the training set data, verify whether the model state converges or overfits with the validation set data, and finally save the converged model and its parameters.

[0112] Step 502, create a multi-task model.

[0113] The multi-task model is Figure 1 the DRASST-MTL shown.

[0114] Step 503, for each training sample, use the multi-task model to process the two-dimensional matrix in the training sample to obtain the predicted modulation mode and the predicted symbol rate.

[0115] Specifically, receive the two-dimensional matrix as the input. First, perform data channel dimension transformation on this two-dimensional matrix through convolution operations. Then, use the DRA module to extract the local detailed features of the input data, and add max pooling before and after the DRA module for downsampling and feature dimensionality reduction. Then, extract the data temporal features through the SST module to establish global feature dependencies. Finally, map to the output layer through the fully connected module. The output layer has two sub-tasks: modulation recognition and symbol rate estimation, and can perform them simultaneously, reducing the computational resource consumption during training and improving the network's deduction speed.

[0116] Step 504, calculate the loss value of the multi-task model according to the predicted modulation mode and the true modulation mode, the predicted symbol rate and the true symbol rate, and train the model parameters of the multi-task model according to the loss value.

[0117] First, read in the encapsulated training set data for training. During the training process, the Adam optimizer is used with an initial learning rate of 0.001, and the dynamic learning rate method is used for training. In terms of network error, a composite loss function can be used to evaluate the error. Among them, the cross-entropy loss function (CrossEntropyLoss) is used to evaluate the error of the modulation recognition task, and the mean squared error loss function (MSELoss) is used to evaluate the error of the symbol rate estimation task. In addition, to ensure that the proportions of the two tasks in the network are the same, it is necessary to set their loss ratios. We use the loss value ratio of 4:1 when the individual tasks converge to set the loss ratio of the multi-task, so as to ensure that the loss weights of the two tasks are the same, and further enable the network to achieve the optimal effect for each subtask. After each round of training, we send the validation set data into the network for validation to monitor the network status in real time and prevent overfitting.

[0118] The test model module mainly uses the test set data to verify the accuracy of the model and judge the precision of the model. The main steps are as follows: First, take out the test set data, and then send it into the trained multi-task model. The multi-task model can automatically output the modulation mode and symbol rate of the input data. Then, by comparing with the true modulation mode and symbol rate, the accuracy and precision of the model can be judged.

[0119] After training, save the finally converged network model to obtain the trained multi-task model.

[0120] As Figure 7 shown, it shows the method flow chart of the modulation recognition and symbol rate estimation method provided by an embodiment of the present application. The modulation recognition and symbol rate estimation method can be applied to a computer device. The modulation recognition and symbol rate estimation method may include:

[0121] Step 701: Convert the radio frequency signal transmitted through the wireless channel to obtain a baseband modulation signal. The radio frequency signal is obtained by processing the transmission data with a modulation mode and a symbol rate, and the wireless channel is a channel using Rayleigh fading superimposed with additive white Gaussian noise.

[0122] Step 702: Separate the real part and the imaginary part of the baseband modulation signal, and splice the separated real part and imaginary part into a two-dimensional matrix.

[0123] Step 703: Obtain a pre-trained multi-task model.

[0124] Among them, the multi-task model is Figure 1 the multi-task model shown.

[0125] Step 704: Use the multi-task model to process the two-dimensional matrix to obtain the modulation mode and symbol rate of the transmission data.

[0126] Specifically, first, a two-dimensional matrix is subjected to dimensional transformation through a convolution module, then a pooling operation is used for downsampling, and then a DRA module is used to extract local detailed features of the signal. A pooling operation is performed on the output of the DRA module to further aggregate features and select significant features. Then, an SST module is used to capture the global correlation of the signal to extract the temporal features of the data. Finally, a fully connected module is used to output the modulation category and symbol rate of the signal.

[0127] In summary, the modulation recognition and symbol rate estimation method provided by the embodiment of the present application has a multi-task model with a DRA module and an SST module, which can perform modulation recognition and symbol rate estimation synchronously. Compared with the mainstream models, this model has the advantages of good performance, low complexity, and low number of parameters. Moreover, compared with two corresponding single-task models, the system overhead of the multi-task model is basically the same as that of one of the single-task models, and the accuracy of classification and prediction is slightly improved, greatly reducing the deduction speed of the model. In addition, the multi-task model can share the feature extraction layer, effectively avoiding overfitting and parameter complexity, and improving the generalization ability and robustness of the model.

[0128] As Figure 8 shown, it shows a flowchart of a modulation recognition and symbol rate estimation method provided by an embodiment of the present application. This modulation recognition and symbol rate estimation method can be applied to a computer device. The modulation recognition and symbol rate estimation method may include:

[0129] Step 801, convert the radio frequency signal transmitted through the wireless channel to obtain a baseband modulation signal. The radio frequency signal is obtained by processing the transmission data in a modulation mode and a symbol rate, and the wireless channel is a channel using Rayleigh fading superimposed with additive white Gaussian noise.

[0130] Step 802, separate the real part and the imaginary part of the baseband modulation signal, and splice the separated real part and imaginary part into a two-dimensional matrix.

[0131] Step 803, obtain a pre-trained multi-task model.

[0132] Among them, the multi-task model is Figure 1 the multi-task model shown.

[0133] Step 804, perform convolution on the two-dimensional matrix using a convolution module to obtain convolution features.

[0134] If the input data is 1*2*2048, after the convolution operation, the data is converted into 32*2*2048 data.

[0135] Step 805: Use the pooling module to downsample the convolutional features to obtain the first pooled features.

[0136] If the input data is 32 * 2 * 2048, after the pooling operation, the data is converted into 32 * 2 * 1024 data.

[0137] Step 806: Use the DRA module to extract local detailed features from the first pooled features to obtain detailed features.

[0138] As Figure 2 shown, the DRA module includes ten serially connected DA units and three skip connection structures with ECA units. Moreover, the first ECA unit makes a skip connection to the fourth DA unit and the sixth DA unit, the second ECA unit makes a skip connection to the second DA unit and the eighth DA unit, and the third ECA unit makes a skip connection to the input end and the output end of the DRA module. Using the DRA module to extract local detailed features from the first pooled features to obtain detailed features may include:

[0139] (1) Process the first pooled features using the first to sixth serially connected DA units.

[0140] (2) Add the first DA feature output by the fourth DA unit and the second DA feature output by the sixth DA unit through the first ECA unit to obtain the third DA feature.

[0141] (3) Process the third DA feature using the seventh and eighth serially connected DA units.

[0142] (4) Add the fourth DA feature output by the second DA unit and the fifth DA feature output by the eighth DA unit through the second ECA unit to obtain the sixth DA feature.

[0143] (5) Process the sixth DA feature using the ninth and tenth serially connected DA units.

[0144] (6) Add the first pooled features and the seventh DA feature output by the tenth DA unit through the third ECA unit to obtain the detailed features.

[0145] The DRA module used by the multi-task model does not change the dimension and size of the data. Therefore, the data output by the DRA module still maintains a size of 32 * 2 * 1024.

[0146] Step 807: Use the pooling module to aggregate the detailed features and select significant features to obtain the second pooled features.

[0147] Perform a max-pooling operation on the data output by the DRA module, and the output is data of 32 * 2 * 512.

[0148] Step 808: Use the SST module to extract global correlation features from the second pooled features to obtain temporal features.

[0149] Specifically, using the SST module to extract global correlation features from the second pooled features to obtain temporal features may include:

[0150] (1) Use the Patch Partition unit to partition the second pooled features and flatten them in the channel direction to obtain partitioned features.

[0151] (2) Use the linear embedding unit to perform a linear transformation on the partitioned features to obtain linear features.

[0152] (3) Use two cascaded SwinTransformer Block units to extract temporal features from the linear features to obtain temporal features.

[0153] As Figure 3 shown, the two Swin Transformer Block units include LN, W-MSA, LN, MLP, LN, SW-MSA, LN, and MLP. Using two cascaded SwinTransformer Block units to extract temporal features from the linear features to obtain temporal features may include: Processing the linear features with cascaded LN and W-MSA to obtain the first ST feature; Adding the linear features to the first ST feature to obtain the second ST feature; Processing the second ST feature with cascaded LN and MLP to obtain the third ST feature; Adding the second ST feature to the third ST feature to obtain the fourth ST feature; Processing the fourth ST feature with cascaded LN and SW-MSA to obtain the fifth ST feature; Adding the fourth ST feature to the fifth ST feature to obtain the sixth ST feature; Processing the sixth ST feature with cascaded LN and MLP to obtain the seventh ST feature; Adding the sixth ST feature to the seventh ST feature to obtain temporal features.

[0154] Step 809: Use the fully connected module to process the temporal features to obtain the modulation mode and symbol rate of the transmitted data.

[0155] One branch of the fully connected module outputs one of the 8 modulation modes, and the other branch outputs one symbol rate.

[0156] In summary, the modulation recognition and symbol rate estimation method provided by the embodiments of the present application, with a multi-task model having a DRA module and an SST module, can perform modulation recognition and symbol rate estimation synchronously. Compared with mainstream models, this model has the advantages of good performance, low complexity, and low number of parameters. Moreover, compared with two corresponding single-task models, the system overhead of the multi-task model is basically the same as that of one of the single-task models, and the accuracy of classification and prediction is slightly improved, greatly reducing the inference speed of the model. In addition, the multi-task model can share the feature extraction layer, effectively avoiding overfitting and parameter complexity, and improving the generalization ability and robustness of the model.

[0157] As Figure 9 shown, it shows a structural block diagram of a modulation recognition and symbol rate estimation device provided by an embodiment of the present application. The modulation recognition and symbol rate estimation device can be applied to a computer device. The modulation recognition and symbol rate estimation device may include:

[0158] A conversion module 910, configured to convert a radio frequency signal transmitted through a wireless channel to obtain a baseband modulation signal. The radio frequency signal is obtained by processing transmission data in a modulation method and a symbol rate, and the wireless channel is a channel using Rayleigh fading superimposed with additive white Gaussian noise;

[0159] A splicing module 920, configured to separate the real part and the imaginary part of the baseband modulation signal, and splice the separated real part and imaginary part into a two-dimensional matrix;

[0160] An acquisition module 930, configured to acquire a pre-trained multi-task model. The multi-task model includes a convolution module, a pooling module, a DRA module, an SST module, and a fully connected module. The DRA module is composed of p DA units and q skip connection structures with ECA units. The DA unit is a DSC of the CBAM mechanism, and p and q are positive integers; the SST module is composed of a Patch Partition unit, a linear embedding unit, and two cascaded Swin Transformer Block units;

[0161] A processing module 940, configured to process the two-dimensional matrix using the multi-task model to obtain the modulation method and symbol rate of the transmission data.

[0162] In an optional embodiment, the processing module 940 is further configured to:

[0163] Perform convolution on the two-dimensional matrix using the convolution module to obtain convolution features;

[0164] Perform downsampling on the convolution features using the pooling module to obtain first pooling features;

[0165] The DRA module is used to extract local detailed features from the first pooled feature to obtain detailed features;

[0166] The pooling module is used to aggregate the detailed features and select significant features to obtain the second pooled feature;

[0167] The SST module is used to extract global correlation features from the second pooled feature to obtain temporal features;

[0168] The fully connected module is used to process the temporal features to obtain the modulation mode and symbol rate of the transmitted data.

[0169] In an optional embodiment, the DRA module includes ten serially connected DA units and three skip connection structures with ECA units, and the first ECA unit skip connects the fourth DA unit and the sixth DA unit, the second ECA unit skip connects the second DA unit and the eighth DA unit, and the third ECA unit skip connects the input end and the output end of the DRA module. The processing module 940 is further configured to:

[0170] Process the first pooled feature using the first to sixth serially connected DA units;

[0171] Add the first DA feature output by the fourth DA unit and the second DA feature output by the sixth DA unit through the first ECA unit to obtain the third DA feature;

[0172] Process the third DA feature using the seventh and eighth serially connected DA units;

[0173] Add the fourth DA feature output by the second DA unit and the fifth DA feature output by the eighth DA unit through the second ECA unit to obtain the sixth DA feature;

[0174] Process the sixth DA feature using the ninth and tenth serially connected DA units;

[0175] Add the first pooled feature and the seventh DA feature output by the tenth DA unit through the third ECA unit to obtain detailed features.

[0176] In an optional embodiment, the processing module 940 is further configured to:

[0177] Use the Patch Partition unit to partition the second pooled feature and flatten it in the channel direction to obtain partitioned features;

[0178] Use the linear embedding unit to perform a linear transformation on the partitioned features to obtain linear features;

[0179] Use two cascaded Swin Transformer Block units to perform temporal feature extraction on linear features to obtain temporal features.

[0180] In an alternative embodiment, the two Swin Transformer Block units include LN, W-MSA, LN, MLP, LN, SW-MSA, LN, and MLP. The processing module 940 is further configured to:

[0181] Process the linear features using cascaded LN and W-MSA to obtain the first ST feature;

[0182] Add the linear features to the first ST feature to obtain the second ST feature;

[0183] Process the second ST feature using cascaded LN and MLP to obtain the third ST feature;

[0184] Add the second ST feature to the third ST feature to obtain the fourth ST feature;

[0185] Process the fourth ST feature using cascaded LN and SW-MSA to obtain the fifth ST feature;

[0186] Add the fourth ST feature to the fifth ST feature to obtain the sixth ST feature;

[0187] Process the sixth ST feature using cascaded LN and MLP to obtain the seventh ST feature;

[0188] Add the sixth ST feature to the seventh ST feature to obtain the temporal feature.

[0189] In an alternative embodiment, the apparatus further includes:

[0190] A creation module for creating a training data set. The training samples in the training data set include a two-dimensional matrix, a true modulation mode, and a true symbol rate. The two-dimensional matrix is obtained by processing simulation data in the true modulation mode and true symbol rate, transmitting the processed RF signal using an analog wireless channel, and separating the real and imaginary parts of the transmitted complex signal. The analog wireless channel is an analog channel using Rayleigh fading superimposed with additive white Gaussian noise;

[0191] The creation module is further configured to create a multi-task model;

[0192] The processing module 940 is further configured to, for each training sample, process the two-dimensional matrix in the training sample using the multi-task model to obtain a predicted modulation mode and a predicted symbol rate;

[0193] A training module, configured to calculate the loss value of the multi-task model according to the predicted modulation mode and the true modulation mode, and the predicted symbol rate and the true symbol rate, and train the model parameters of the multi-task model according to the loss value.

[0194] In an optional embodiment, the creation module is further configured to:

[0195] Randomly select a true modulation mode and a true symbol rate by using a signal modulator, and modulate the randomly generated binary sequence by using the selected true modulation mode and true symbol rate to obtain modulated data, where the modulated data includes N symbols;

[0196] Perform waveform adjustment on the modulated data by using a square root raised cosine pulse filter, and convert the shaped data into a radio frequency signal after energy normalization;

[0197] Transmit the radio frequency signal by using an analog wireless channel, where the analog wireless channel is an analog channel using Rayleigh fading superimposed with additive white Gaussian noise;

[0198] Convert the radio frequency signal transmitted by the wireless channel, separate the real part and the imaginary part of the baseband modulated signal obtained after conversion, and splice the separated real part and imaginary part into a two-dimensional matrix;

[0199] Form a training sample from the two-dimensional matrix, the true modulation mode, and the true symbol rate.

[0200] In summary, the modulation recognition and symbol rate estimation device provided by the embodiment of the present application has a multi-task model with a DRA module and an SST module, and can perform modulation recognition and symbol rate estimation synchronously. Compared with the mainstream models, this model has the advantages of good performance, low complexity, and low number of parameters. Moreover, compared with two corresponding single-task models, the system overhead of the multi-task model is basically the same as that of one of the single-task models, and the classification and prediction accuracy is slightly improved, greatly reducing the model deduction speed. In addition, the multi-task model can share the feature extraction layer, effectively avoiding overfitting and parameter complexity, and improving the generalization ability and robustness of the model.

[0201] An embodiment of the present application provides a computer-readable storage medium, in which at least one instruction is stored, and the at least one instruction is loaded and executed by a processor to implement the modulation recognition and symbol rate estimation method as described above.

[0202] An embodiment of the present application provides a computer device, and the computer device includes any of the above modulation recognition and symbol rate estimation devices.

[0203] It should be noted that when the modulation recognition and symbol rate estimation device provided in the above embodiments performs modulation recognition and symbol rate estimation, only the division of the above functional modules is used as an example for illustration. In actual applications, the above functions can be assigned to different functional modules according to needs, that is, the internal structure of the modulation recognition and symbol rate estimation device is divided into different functional modules to complete all or part of the functions described above. In addition, the modulation recognition and symbol rate estimation device provided in the above embodiments and the embodiments of the modulation recognition and symbol rate estimation method belong to the same concept. For the specific implementation process, please refer to the method embodiments and will not be elaborated here.

[0204] Those of ordinary skill in the art can understand that all or part of the steps to implement the above embodiments can be completed by hardware or by a program instructing relevant hardware. The program can be stored in a computer-readable storage medium. The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disc, etc.

[0205] The above does not intend to limit the embodiments of the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the embodiments of the present application shall be included in the protection scope of the embodiments of the present application.

Claims

1. A modulation identification and symbol rate estimation method, characterized in that: The method comprises: Converting a radio frequency signal transmitted through a wireless channel to obtain a baseband modulation signal, wherein the radio frequency signal is obtained by processing transmission data in a modulation mode and a symbol rate, and the wireless channel is a channel using Rayleigh attenuation superimposed with additive white Gaussian noise; Separating the real part and the imaginary part of the baseband modulation signal, and splicing the separated real part and the imaginary part into a two-dimensional matrix; Obtain a pre-trained multi-task model, wherein the multi-task model includes a convolution module, a pooling module, a deep separable convolution DRA module with an attention mechanism and a residual connection, a simplified version of the SwinTransformer SST module, and a fully connected module, wherein the DRA module is composed of p deep separable convolution DA units with an attention mechanism and q skip connection structures with efficient channel attention mechanisms ECA units, and the DA unit is a deep separable convolution DSC with a convolutional block attention CBAM mechanism, where p and q are positive integers; the SST module is composed of a Patch Partition unit, a linear embedding unit, and two serially connected SwinTransformerBlock units; The two-dimensional matrix is ​​processed using the multi-task model to obtain the modulation mode and symbol rate of the transmission data.

2. The modulation identification and symbol rate estimation method according to claim 1, characterized in that: The using the multi-task model to process the two-dimensional matrix to obtain the modulation mode and symbol rate of the transmission data includes: Convolving the two-dimensional matrix using the convolution module to obtain convolution features; Down-sampling the convolution feature using the pooling module to obtain a first pooling feature; Using the DRA module to extract local detail features from the first pooled features to obtain detail features; Using the pooling module to aggregate the detail features and select significant features to obtain a second pooling feature; Using the SST module to extract global correlation features from the second pooling features to obtain time series features; The timing characteristics are processed using the fully connected module to obtain the modulation mode and symbol rate of the transmission data.

3. The modulation identification and symbol rate estimation method according to claim 2, characterized in that: The DRA module includes ten DA units connected in series and three jump connection structures with ECA units, and the first ECA unit jumps to connect the fourth DA unit and the sixth DA unit, the second ECA unit jumps to connect the second DA unit and the eighth DA unit, and the third ECA unit jumps to connect the input end and the output end of the DRA module. The DRA module is used to extract local detail features from the first pooling feature to obtain detail features, including: Processing the first pooled feature using the first to sixth DA units connected in series; The first DA feature output by the fourth DA unit and the second DA feature output by the sixth DA unit are added by the first ECA unit to obtain a third DA feature; processing the third DA feature using a seventh and an eighth DA unit connected in series; adding the fourth DA feature output by the second DA unit and the fifth DA feature output by the eighth DA unit through the second ECA unit to obtain a sixth DA feature; processing the sixth DA feature using a ninth and a tenth DA unit connected in series; The first pooling feature is added to the seventh DA feature output by the tenth DA unit through the third ECA unit to obtain the detail feature.

4. The modulation identification and symbol rate estimation method according to claim 2, characterized in that: The using the SST module to extract global correlation features from the second pooling features to obtain time series features includes: The second pooling feature is divided into blocks using the Patch Partition unit, and flattened in the channel direction to obtain a block feature; Using the linear embedding unit to perform linear transformation on the block features to obtain linear features; Two serially connected Swin TransformerBlock units are used to extract time features from the linear features to obtain the time series features.

5. The modulation identification and symbol rate estimation method according to claim 4, characterized in that: The two Swin Transformer Block units include layer normalization LN, window-based multi-head self-attention mechanism W-MSA, LN, multi-layer perceptron MLP, LN, shift window-based multi-head self-attention mechanism SW-MSA, LN and MLP, and the two serially connected Swin Transformer Block units are used to extract time features from the linear features to obtain the time series features, including: The linear feature is processed by using LN and W-MSA in series to obtain a first ST feature; Adding the linear feature to the first ST feature to obtain a second ST feature; The second ST feature is processed by using LN and MLP in series to obtain a third ST feature; Adding the second ST feature to the third ST feature to obtain a fourth ST feature; Processing the fourth ST feature by using LN and SW-MSA connected in series to obtain a fifth ST feature; Adding the fourth ST feature to the fifth ST feature to obtain a sixth ST feature; Processing the sixth ST feature by using LN and MLP connected in series to obtain a seventh ST feature; The sixth ST feature and the seventh ST feature are added to obtain the timing feature.

6. The modulation identification and symbol rate estimation method according to any one of claims 1 to 5, characterized in that: The method further comprises: Creating a training data set, wherein the training samples in the training data set include a two-dimensional matrix, a real modulation mode, and a real symbol rate, wherein the two-dimensional matrix is ​​obtained by processing simulation data with the real modulation mode and the real symbol rate, transmitting the processed radio frequency signal using a simulated wireless channel, and separating the real part and the imaginary part of the complex signal obtained after the transmission, wherein the simulated wireless channel is a simulated channel using Rayleigh attenuation superimposed with additive white Gaussian noise; Creating the multi-task model; For each training sample, using the multi-task model to process the two-dimensional matrix in the training sample to obtain a predicted modulation mode and a predicted symbol rate; The loss value of the multi-task model is calculated according to the predicted modulation mode and the actual modulation mode, the predicted codeword rate and the actual codeword rate, and the model parameters of the multi-task model are trained according to the loss value.

7. The modulation identification and symbol rate estimation method according to claim 6, characterized in that: The creating of the training data set comprises: A signal modulator is used to randomly select a real modulation mode and a real symbol rate, and a randomly generated binary sequence is modulated using the selected real modulation mode and the real symbol rate to obtain modulated data, wherein the modulated data includes N symbols; The modulated data is subjected to waveform adjustment using a square root raised cosine pulse filter, and the shaped data is subjected to energy normalization and then converted into a radio frequency signal; The radio frequency signal is transmitted by using an analog wireless channel, wherein the analog wireless channel is an analog channel using Rayleigh attenuation superimposed with additive white Gaussian noise; Convert the radio frequency signal transmitted by the wireless channel, separate the real part and the imaginary part of the baseband modulated signal obtained after the conversion, and splice the separated real part and the imaginary part into a two-dimensional matrix; The two-dimensional matrix, the actual modulation mode and the actual symbol rate are combined into a training sample.

8. A modulation identification and symbol rate estimation device, characterized in that: The device comprises: A conversion module, used for converting a radio frequency signal transmitted by a wireless channel to obtain a baseband modulation signal, wherein the radio frequency signal is obtained by processing the transmission data in a modulation mode and a symbol rate, and the wireless channel is a channel using Rayleigh attenuation superimposed with additive white Gaussian noise; A splicing module, used for separating the real part and the imaginary part of the baseband modulated signal, and splicing the separated real part and the imaginary part into a two-dimensional matrix; An acquisition module is used to acquire a pre-trained multi-task model, wherein the multi-task model includes a convolution module, a pooling module, a deep separable convolution DRA module with an attention mechanism and a residual connection, a simplified version of the Swin Transformer SST module, and a fully connected module. The DRA module is composed of p deep separable convolution DA units with an attention mechanism and q skip connection structures with efficient channel attention mechanisms ECA units. The DA unit is a deep separable convolution DSC with a convolution block attention CBAM mechanism, where p and q are positive integers. The SST module is composed of a Patch Partition unit, a linear embedding unit, and two serially connected Swin TransformerBlock units. A processing module is used to process the two-dimensional matrix using the multi-task model to obtain the modulation mode and code element rate of the transmission data.

9. A computer-readable storage medium, characterized in that: The storage medium stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the modulation identification and codeword rate estimation method as described in any one of claims 1 to 7.

10. A computer device, characterized in that: The computer device comprises: the modulation identification and code element rate estimation device as described in claim 8.

Images