Modulation identification and symbol rate estimation method, apparatus, storage medium, and device

Abstract

This application discloses a modulation identification and symbol rate estimation method, apparatus, storage medium, and device, belonging to the field of communication technology. It converts the radio frequency signal transmitted through a wireless channel, separates the real and imaginary parts of the baseband modulation signal, and concatenates the real and imaginary parts into a two-dimensional matrix; it obtains a multi-task model, which includes a convolution module, a pooling module, a DRA module, an SST module, and a fully connected module. The DRA module consists of multiple DA units and a jump-connected structure with ECA units; the SST module consists of Patch Partition units, linear embedding units, and Swing TransformerBlock units; the multi-task model is used to process the two-dimensional matrix to obtain the modulation scheme and symbol rate. This application can simultaneously perform modulation identification and symbol rate estimation, exhibiting good performance, low complexity and parameter quantity, low overhead, fast derivation speed, and high generalization ability and robustness.

Description

Technical Field

[0001] This application relates to the field of communication technology, and in particular to a modulation identification and symbol rate estimation method, apparatus, storage medium and device. Background Technology

[0002] With the rapid development of global science and technology, radio communication technology has made significant progress in transmission efficiency, reliability, and diversified applications. In modern communication systems, all communication signals (radio, television, mobile communication, etc.) must undergo modulation before transmission. However, in the context of non-cooperative radio communication, the interactivity between the communicating parties is greatly reduced, and information such as the type of transmitted signal, communication parameters, and pilot header information is unknown. In this context, the modulation method and symbol rate, as key communication parameters, are directly related to the success of signal demodulation and decoding. This problem is not only one of the core technical challenges in the field of wireless communication, but also an important technical support for electronic warfare and reconnaissance in the military field.

[0003] While traditional methods have achieved some success in modulation recognition and parameter estimation, their technical bottlenecks and shortcomings are becoming increasingly apparent, making it difficult to meet the increasingly complex demands of wireless communication. Traditional algorithms largely rely on expert-designed feature extraction methods and prior knowledge, which not only leads to high dependence on specific scenarios but also significantly reduces their adaptability to dynamic and complex environments. Furthermore, traditional methods also have limited performance when processing high-dimensional complex signals. As communication technologies develop towards higher frequencies and wider bandwidths, the complexity of signals places higher demands on the flexibility and scalability of algorithms. Therefore, developing more flexible and efficient methods to overcome these limitations is particularly important.

[0004] In recent years, the rapid development of artificial intelligence (AI) technology has triggered profound changes in multiple fields. As an important branch of AI, deep learning has demonstrated outstanding performance in tasks such as image recognition and natural language processing due to its powerful data processing capabilities and advantages in modeling complex nonlinear relationships. Furthermore, thanks to the availability of large-scale datasets, significant improvements in computing hardware performance, and continuous innovation in 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 has exhibited superior performance, especially in solving complex problems that traditional methods struggle with, possessing irreplaceable advantages. The innovation and development of deep learning technology has made it a crucial driving force for progress in modern science and engineering. In modern communication systems, with the continuous growth in data transmission rates and spectrum utilization demands, signal processing technology faces 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 addressing problems in modern communication systems. It demonstrates enormous potential for handling complex signal environments, optimizing system performance, and improving intelligence levels. It excels in key tasks such as modulation recognition and symbol rate estimation, and is particularly adaptable 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, driving the intelligent development of these systems. Deep learning models can also adapt to different signal-to-noise ratios, spectral conditions, and modulation schemes, exhibiting powerful generalization capabilities, which are crucial in modern communication systems. In the future, deep learning technology will unleash its potential in higher dimensions and more complex scenarios, providing new impetus for innovation in modern communication technologies. Summary of the Invention

[0005] This application provides a modulation recognition and symbol rate estimation method, apparatus, storage medium, and device to address the problems of traditional algorithms being highly dependent on specific scenarios, having low adaptability in dynamic and complex environments, and performing poorly when processing high-dimensional complex signals. The technical solution is as follows:

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

[0007] 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 symbol rate. The wireless channel is a channel using Rayleigh attenuation superimposed with additive white Gaussian noise.

[0008] 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;

[0009] A pre-trained multi-task model is obtained, comprising a convolutional module, a pooling module, a deep separable convolutional (DRA) module with attention mechanisms and residual connections, a simplified Swin Transformer (SST) module, and a fully connected module. The DRA module consists of a skip-connected structure of p deep separable convolutional (DA) units with attention mechanisms and q units with efficient channel attention (ECA) mechanisms. The DA units are deep separable convolutional (DSC) units with convolutional block attention (CBAM) mechanisms, where 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.

[0010] The modulation scheme and symbol rate of the transmitted data are obtained by processing the two-dimensional matrix using the multi-task model.

[0011] In one possible implementation, processing the two-dimensional matrix using the multi-task model to obtain the modulation scheme and symbol rate of the transmitted data includes:

[0012] The convolution module is used to convolve the two-dimensional matrix to obtain convolutional features;

[0013] The convolutional features are downsampled using the pooling module to obtain the first pooling feature;

[0014] The DRA module is used to extract local detail features from the first pooling feature to obtain detail features;

[0015] The pooling module is used to aggregate the detailed features and select salient features to obtain the second pooled feature.

[0016] The SST module is used to extract global correlation features from the second pooling feature to obtain temporal features;

[0017] The timing characteristics are processed using the fully connected module to obtain the modulation scheme and symbol rate of the transmitted data.

[0018] In one possible implementation, the DRA module includes ten cascaded DA units and three skip-connected structures with ECA units. The first ECA unit skips to the fourth and sixth DA units, the second ECA unit skips to the second and eighth DA units, and the third ECA unit skips to the input and output of the DRA module. The step of using the DRA module to extract local detail features from the first pooling feature to obtain detail features includes:

[0019] The first pooling feature is processed using the first to sixth DA units in series;

[0020] The first DA feature is obtained by adding the first DA feature output from the fourth DA unit to the second DA feature output from the sixth DA unit through the first ECA unit;

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

[0022] The fourth DA feature output by the second DA unit is added to the fifth DA feature output by the eighth DA unit through the second ECA unit to obtain the sixth DA feature;

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

[0024] 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.

[0025] In one possible implementation, the step of using the SST module to perform global correlation feature extraction on the second pooling feature to obtain time-series features includes:

[0026] The second pooling feature is divided into blocks using the Patch Partition unit and flattened in the channel direction to obtain the block feature;

[0027] The block features are linearly transformed using the linear embedding unit to obtain linear features;

[0028] The temporal features are obtained by extracting time features from the linear features using two cascaded Swing Transformer Block units.

[0029] In one possible implementation, the two Swin Transformer Block units include Layer Normalization (LN), Window-based Multi-Head Self-Attention (W-MSA) LN, Multilayer Perceptron (MLP) LN, and Shift-Window-based Multi-Head Self-Attention (SW-MSA) LN and MLP. The extraction of temporal features from the linear features using the two cascaded Swin Transformer Block units to obtain the temporal features includes:

[0030] The linear features are processed using cascaded LN and W-MSA to obtain the first ST feature;

[0031] The linear feature is added to the first ST feature to obtain the second ST feature;

[0032] The second ST feature is processed using cascaded LN and MLP to obtain the third ST feature;

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

[0034] The fourth ST feature is processed using cascaded LN and SW-MSA to obtain the fifth ST feature;

[0035] The fourth ST feature is added to the fifth ST feature to obtain the sixth ST feature;

[0036] The sixth ST feature is processed using cascaded LN and MLP to obtain the seventh ST feature;

[0037] The sixth ST feature is added to the seventh ST feature to obtain the time series feature.

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

[0039] A training dataset is created, in which training samples include a two-dimensional matrix, a real modulation scheme, and a real symbol rate. The two-dimensional matrix is ​​obtained by processing the simulation data with the real modulation scheme and the real symbol rate. The processed radio frequency signal is transmitted using a simulated wireless channel. The real and imaginary parts of the complex signal obtained after transmission are separated. The simulated wireless channel is a simulated channel using Rayleigh attenuation superimposed with additive white Gaussian noise.

[0040] Create the aforementioned multi-task model;

[0041] 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 symbol rate.

[0042] The loss value of the multi-task model is calculated based on the predicted modulation scheme and the actual modulation scheme, the predicted symbol rate and the actual symbol rate, and the model parameters of the multi-task model are trained based on the loss value.

[0043] In one possible implementation, creating the training dataset includes:

[0044] A signal modulator randomly selects a real modulation scheme and a real symbol rate, and modulates a randomly generated binary sequence using the selected real modulation scheme and the real symbol rate to obtain modulated data, which includes N symbols.

[0045] The waveform of the modulated data is adjusted using a square root raised cosine pulse filter, and the shaped data is converted into a radio frequency signal after energy normalization.

[0046] The radio frequency signal is transmitted using an analog wireless channel, which is an analog channel using Rayleigh attenuation superimposed with additive white Gaussian noise;

[0047] 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.

[0048] The two-dimensional matrix, the actual modulation scheme, and the actual symbol rate are combined to form a training sample.

[0049] According to a second aspect of this application, a modulation identification and symbol rate estimation apparatus is provided, the apparatus comprising:

[0050] 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 symbol rate. The wireless channel is a channel using Rayleigh attenuation superimposed with additive white Gaussian noise.

[0051] 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.

[0052] The acquisition module is used to acquire a pre-trained multi-task model, which includes a convolution module, a pooling module, a deep separable convolutional (DRA) module with attention mechanisms and residual connections, a simplified Swin Transformer (SST) module, and a fully connected module. The DRA module consists of a skip-connected structure of p deep separable convolutional (DA) units with attention mechanisms and q units with efficient channel attention (ECA) mechanisms. The DA units are deep separable convolutional (DSC) units with convolutional block attention (CBAM) mechanisms, where 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] The processing module is used to process the two-dimensional matrix using the multi-task model to obtain the modulation scheme and symbol rate of the transmitted data.

[0054] According to a third aspect of this application, a computer-readable storage medium is provided, wherein at least one instruction is stored therein, the at least one instruction being loaded and executed by a processor to implement the modulation identification and symbol rate estimation method as described above.

[0055] According to a fourth aspect of this application, a computer device is provided, the computer device including the above-described modulation identification and symbol rate estimation apparatus.

[0056] The beneficial effects of the technical solution provided in this application include at least the following:

[0057] To address the performance degradation caused by separating the spatial and channel dimensions in DSC, we designed a CBAM mechanism after depthwise convolution and pointwise convolution to learn the correlation between the spatial and channel dimensions. This reduces the complexity and number of parameters in the multi-task model and improves its performance. In addition to improving DSC, we added a skip connection structure led by ECA to form a DRA module. This alleviates the problem of insufficient deep weight updates caused by gradient vanishing or exploding during backpropagation, enriches the information flow between shallow and deep layers, and efficiently captures cross-channel interactions, further improving the overall performance of the model.

[0058] The SST module is constructed by introducing window-based multi-head attention mechanism and sliding window-based multi-head attention mechanism of the Swin Transformer Block unit, which ensures that the model has the ability to extract global features while maintaining low complexity.

[0059] This multi-task model, featuring DRA and SST modules, can simultaneously perform modulation recognition and symbol rate estimation. Compared to mainstream models, it offers superior performance while maintaining low complexity and a small number of parameters. Furthermore, compared to two corresponding single-task models, the multi-task model's system overhead is essentially the same as that of a single-task model, with slightly improved classification and prediction accuracy, significantly reducing the model's inference speed. In addition, the multi-task model can share feature extraction layers, effectively avoiding overfitting and parameter complexity, thus improving the model's generalization ability and robustness. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a schematic diagram of a multi-task model according to an embodiment of this application;

[0062] Figure 2 This is a schematic diagram of the structure of a DRA module according to an embodiment of this application;

[0063] Figure 3 This is a schematic diagram of the structure of an SST module according to an embodiment of this application;

[0064] Figure 4 This is a schematic diagram of a multi-task model according to an embodiment of this application;

[0065] Figure 5 This is a flowchart illustrating the training process of a multi-task model according to an embodiment of this application;

[0066] Figure 6 This is a model diagram of a communication system according to an embodiment of this application;

[0067] Figure 7 This is a flowchart illustrating the modulation identification and symbol rate estimation method in the embodiments of this application;

[0068] Figure 8 This is a detailed flowchart illustrating the modulation identification and symbol rate estimation method in the embodiments of this application;

[0069] Figure 9 This is a structural block diagram of the modulation identification and symbol rate estimation device illustrated in the embodiments of this application. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0071] Current mainstream methods for applying deep learning to the communications field model single tasks, such as models that only classify the types of modulation signals. This application proposes a multi-task model, DRASST-MTL, that jointly performs modulation recognition and symbol rate estimation. Figure 1 As shown, the multi-task model DRAST-MT includes a convolution module, a pooling module, a DRA module, an SST module, and a fully connected module, wherein the convolution module, pooling module, DRA module, pooling module, SST module, and fully connected module are connected in sequence.

[0072] The convolution module can be Conv2d, i.e., a two-dimensional convolution module, used to perform convolution operations on the input data. The convolution operation consists of three basic units: two-dimensional convolution (Conv2d), two-dimensional normalization (BatchNorm2d), and an activation function (ReLU). The two-dimensional convolution kernel is 1*3 in size, with one unit padded on each side of the input data's width, while the height remains unchanged. The number of channels after convolution is 32. If the input data is 1*2*2048, after the convolution operation, the data is transformed into 32*2*2048 data.

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

[0074] DRA stands for Depthwise Separable Convolution with Residual Connection and Attention Mechanism. A DRA module consists of p DA units and q skip-connected structures with ECA units. DA stands for Depthwise Separable Convolution with Attention Mechanism, specifically Convolutional Block Attention Module (CBAM). In other words, the DA unit is a DSC with CBAM, designed to mitigate cross-channel information loss.

[0075] In the DA module, the input feature F is processed by depthwise convolution to obtain the feature Fd of each channel; a CBAM mechanism is added after each channel to obtain the feature F' after channel-dimensional attention weight optimization; the feature F' is then processed by the spatial attention module to obtain the feature F” after dual optimization of attention weights in both channel and spatial dimensions, realizing the extraction of key information and obtaining the feature F” after feature enhancement; 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 follows:

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

[0077]

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

[0079]

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

[0081] ECA stands for Efficient Channel Attention. Skip connections with ECA units enhance the fluidity of features between shallow and deep layers of the network. In the DRA module, skip connections are set between every two DA units, and features are fused through ECA in between. This enables efficient interaction and extraction of salient information in both spatial and channel dimensions, as well as across channels, to extract local detail features.

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

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

[0084] The second pooling module can be MaxPool2d, i.e., a two-dimensional max pooling module. This second pooling module downsamples the data to reduce network complexity. Max pooling is performed on the data output from the DRA module, resulting in a 32*2*512 data set.

[0085] SST is a simplified version of the Swing Transformer. For example... Figure 3As shown, the SST module consists of a PatchPartition unit, a LinearEmbedding unit, and two cascaded SwinTransformer Block units. The two Swin Transformer Block units include LayerNormalization (LN), Windows Multi-head SelfAttention (W-MSA), LN, Multi-Layer Perceptron (MLP), LN, and Windows Multi-head Self Attention (SW-MSA), LN, and MLP. The Swin Transformer settings are as follows: patch_size is (2, 4), sliding window size is 7*7, 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 divided into blocks by the Patch Partition unit, then flattened along the channel direction, and then linearly transformed by the Linear Embedding unit before being input into two Swing Transformer Block units. Finally, the time series features are extracted globally by the two Swing Transformer Block units. While W-MSA significantly reduces computational complexity within the Swing Transformer Block units, it neglects information exchange between windows. Therefore, to strengthen the connection between data features across windows, Swing Transformer Block (SW-MSA) is introduced on top of W-MSA. This cyclic shift method establishes connections between adjacent and non-overlapping windows, thereby expanding the receptive field. The collaboration between W-MSA and SW-MSA not only captures both the time series features of the signal and the global features of the data but also significantly reduces computational complexity.

[0087] The fully connected module has two branches: one branch outputs an 8-class classification structure, corresponding to the modulation recognition result, and the other branch outputs a single prediction result, corresponding to the symbol rate estimation result. For example... Figure 4 As shown, the input data passes through the DRA module, SST module, and fully connected module, respectively, and outputs the modulation recognition and code rate estimation results.

[0088] like Figure 5 As shown below, the training process of the multi-task model DRASST-MTL is described. 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, a real modulation scheme, and a real symbol rate. The two-dimensional matrix is ​​obtained by processing the simulation data with the real modulation scheme and the real symbol rate. The processed radio frequency signal is transmitted using a simulated wireless channel. The real and imaginary parts of the complex signal obtained after transmission are separated. The simulated wireless channel is a simulated channel using Rayleigh attenuation superimposed with additive white Gaussian noise.

[0090] Specifically, creating a training dataset can include:

[0091] (1) Use a signal modulator to randomly select a real modulation method and a real symbol rate, and use the selected real modulation method and real symbol rate to modulate the randomly generated binary sequence to obtain modulated data. The modulated data includes N symbols.

[0092] In this embodiment, a simulation dataset was generated using MATLAB software. The dataset contains 880,000 data points, including 8 modulation schemes, 10 symbol rates (randomly generated from 100K to 200K), and 11 signal-to-noise ratios (SNRs) (0-20dB intervals of 2dB). The sampling frequency is 400kHz. Each SNR for each modulation scheme has 1000 data points, and each data point is 2×2048 in size. The 8 modulation schemes are Binary Phase Shift Keying (BPSK), Quadrature Phase Shift Keying (QPSK), Offset-QPSK, 8-Phase Shift Keying (8PSK), Quadrature Amplitude Modulation (QAM), 64QAM, Minimum Shift Keying (MSK), and Gaussian Filtered Minimum Shift Keying (MSK). ShiftKeying (GMSK).

[0093] like Figure 6 As shown, a random binary sequence a(t) is first generated, and then modulated data s(t) containing N symbols is generated by the signal modulator.

[0094] (2) The waveform of the modulated data is adjusted by using a square root raised cosine pulse filter, and the data obtained after shaping is converted into radio frequency signals after energy normalization.

[0095] The pulse filter used in this model is a square-root raised cosine (SRRC) pulse filter g(t), and its expression is as follows:

[0096]

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

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

[0099]

[0100] In the formula: x'(t) is the output of the data after passing through the square root raised cosine pulse filter, t is the signal duration, T is the symbol period, and Tx is the sampling time interval. The oversampling factor P of the square root raised cosine pulse filter can be expressed as T / Tx, i.e., P = T / Tx. Given that the sampling frequency Fs = 1 / Tx, the symbol rate R = 1 / T = Fs / P can be derived (this provides a data label for symbol rate estimation based on neural networks). Subsequently, the waveform-adjusted x'(t) data is normalized to obtain the data x(t), and the expression for x(t) is:

[0101]

[0102] In the formula: x(t) is the output of the data after energy normalization.

[0103] The energy-normalized data is converted into a radio frequency signal and then transmitted through a wireless channel.

[0104] (3) Radio frequency signals are transmitted using an analog wireless channel, which is an analog channel using Rayleigh attenuation superimposed with additive white Gaussian noise.

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

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

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

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

[0109] The generated complex signal data y(t) is separated into real part r(t) and imaginary part i(t), and then concatenated into a two-dimensional matrix A(t).

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

[0111] Each two-dimensional matrix A(t) and its corresponding labels (modulation scheme and symbol rate) are divided into training, validation, and test sets in a 7:1:2 ratio; that is, 700 data points are used for training, 100 for validation, and 200 for testing. The model is trained using the training set data, and the model's convergence and overfitting are verified using the validation set data. Finally, the converged model and its parameters are saved.

[0112] Step 502: Create a multi-task model.

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

[0114] Step 503: For each training sample, the two-dimensional matrix in the training sample is processed using a multi-task model to obtain the predicted modulation scheme and the predicted symbol rate.

[0115] Specifically, the system receives a two-dimensional matrix as input. First, this matrix is ​​transformed to reduce the data channel dimension through convolution. Then, the DRA module extracts local detail features from the input data, and max pooling is applied before and after the DRA module for downsampling and feature dimensionality reduction. Next, the SST module extracts temporal features from the data and establishes global feature dependencies. Finally, a fully connected module maps the data to the output layer. The output layer has two subtasks: modulation recognition and symbol rate estimation, which can be performed simultaneously, reducing computational resource consumption during training and improving the network's inference speed.

[0116] Step 504: Calculate the loss value of the multi-task model based on the predicted modulation scheme and the actual modulation scheme, the predicted symbol rate and the actual symbol rate, and train the model parameters of the multi-task model based on the loss value.

[0117] 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. Specifically, the cross-entropy loss function 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. Furthermore, to ensure that the two tasks have equal weight in the network, their loss ratios need to be set. A 4:1 ratio of the loss values ​​at convergence of individual tasks is used to set the loss ratio for multiple tasks, ensuring that the loss weights of the two tasks are equal, thus allowing the network to achieve optimal performance for each sub-task. After each round of training, validation set data is fed into the network for validation, allowing real-time monitoring of the network's state and preventing overfitting.

[0118] 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 model. The multi-task model can automatically output the modulation scheme and symbol rate of the input data. By comparing this with the actual modulation scheme and symbol rate, the model's accuracy and precision can be determined.

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

[0120] like Figure 7 The diagram illustrates a flowchart of a modulation identification and symbol rate estimation method according to an embodiment of this application. This modulation identification and symbol rate estimation method can be applied to a computer device. The modulation identification and symbol rate estimation method may include:

[0121] Step 701: 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 symbol rate. The wireless channel is a channel using Rayleigh attenuation superimposed with additive white Gaussian noise.

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

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

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

[0125] Step 704: Process the two-dimensional matrix using a multi-task model to obtain the modulation scheme and symbol rate of the transmitted data.

[0126] Specifically, a convolution module is first used to transform the dimension of the two-dimensional matrix, followed by pooling for downsampling. Then, the DRA module is used to extract local detail features of the signal. The output of the DRA module is pooled to further aggregate and select salient features. Next, the 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 identification and symbol rate estimation method provided in this application has a multi-task model with DRA and SST modules, enabling simultaneous modulation identification and symbol rate estimation. Compared with mainstream models, this model offers better performance while also boasting low complexity and a small number of parameters. Furthermore, compared to two corresponding single-task models, the system overhead of the multi-task model is essentially the same as that of a single-task model, with slightly improved classification and prediction accuracy, significantly reducing the model's inference speed. In addition, the multi-task model can share feature extraction layers, effectively avoiding overfitting and parameter complexity, thus improving the model's generalization ability and robustness.

[0128] like Figure 8 The diagram illustrates a flowchart of a modulation identification and symbol rate estimation method according to an embodiment of this application, which can be applied to a computer device. The modulation identification and symbol rate estimation method may include:

[0129] Step 801: 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 symbol rate. The wireless channel is a channel using Rayleigh attenuation superimposed with additive white Gaussian noise.

[0130] Step 802: Separate the real and imaginary parts of the baseband modulation signal, and then concatenate the separated real and imaginary parts into a two-dimensional matrix.

[0131] Step 803: Obtain the pre-trained multi-task model.

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

[0133] Step 804: Use the convolution module to convolve the two-dimensional matrix to obtain convolution features.

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

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

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

[0137] Step 806: Use the DRA module to extract local detail features from the first pooling feature to obtain detail features.

[0138] like Figure 2 As shown, the DRA module includes ten cascaded DA units and three skip connections with ECA units. The first ECA unit skips to the fourth and sixth DA units, the second ECA unit skips to the second and eighth DA units, and the third ECA unit skips to the input and output of the DRA module. The DRA module is used to extract local detail features from the first pooling features, yielding detail features that may include:

[0139] (1) The first pooling feature is processed using the first to sixth DA units in series.

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

[0141] (3) The third DA feature is processed using the seventh and eighth DA units in series.

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

[0143] (5) The sixth DA feature is processed using the ninth and tenth DA units connected in series.

[0144] (6) 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.

[0145] The DRA module used in the multi-task model does not change the dimensions and size of the data, so the data output by the DRA module remains at 32*2*1024.

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

[0147] Max pooling is performed on the data output by the DRA module, resulting in a data set of 32*2*512.

[0148] Step 808: Use the SST module to extract global correlation features from the second pooling features to obtain time-series features.

[0149] Specifically, the SST module is used to extract global correlation features from the second pooling features to obtain time-series features, which may include:

[0150] (1) The second pooling feature is divided into blocks using the Patch Partition unit and flattened in the channel direction to obtain the block feature.

[0151] (2) Linear features are obtained by using linear embedding units to perform linear transformation on the block features.

[0152] (3) Use two cascaded Swing Transformer Block units to extract time features from linear features to obtain time-series features.

[0153] like Figure 3 As shown, the two Swin Transformer Block units include LN, W-MSA, LN, MLP, LN, SW-MSA, LN, and MLP. Using the two cascaded Swin Transformer Block units to extract time features from linear features to obtain time-series features can include: processing the linear features using cascaded LN and W-MSA to obtain a first ST feature; adding the linear feature to the first ST feature to obtain a second ST feature; processing the second ST feature using cascaded LN and MLP 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 using cascaded LN and SW-MSA 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 using cascaded LN and MLP to obtain a seventh ST feature; and adding the sixth ST feature to the seventh ST feature to obtain the time-series feature.

[0154] Step 809: Use the fully connected module to process the timing characteristics to obtain the modulation scheme and symbol rate of the transmitted data.

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

[0156] In summary, the modulation identification and symbol rate estimation method provided in this application has a multi-task model with DRA and SST modules, enabling simultaneous modulation identification and symbol rate estimation. Compared with mainstream models, this model offers better performance while also boasting low complexity and a small number of parameters. Furthermore, compared to two corresponding single-task models, the system overhead of the multi-task model is essentially the same as that of a single-task model, with slightly improved classification and prediction accuracy, significantly reducing the model's inference speed. In addition, the multi-task model can share feature extraction layers, effectively avoiding overfitting and parameter complexity, thus improving the model's generalization ability and robustness.

[0157] like Figure 9 The diagram illustrates a structural block diagram of a modulation identification and symbol rate estimation apparatus according to an embodiment of this application. This modulation identification and symbol rate estimation apparatus can be applied in a computer device. The modulation identification and symbol rate estimation apparatus may include:

[0158] The conversion module 910 is used to convert the radio frequency signal transmitted through the wireless channel to obtain the baseband modulated signal. The radio frequency signal is obtained by processing the transmitted data with a modulation method and a symbol rate. The wireless channel is a channel using Rayleigh attenuation superimposed with additive white Gaussian noise.

[0159] The splicing module 920 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.

[0160] The acquisition module 930 is used to acquire a pre-trained multi-task model. The multi-task model includes a convolutional module, a pooling module, a DRA module, an SST module, and a fully connected module. The DRA module consists of p DA units and q skip-connected structures with ECA units. The DA units are DSCs with the CBAM mechanism, where p and q are positive integers. The SST module consists of a Patch Partition unit, a linear embedding unit, and two cascaded Swing Transformer Block units.

[0161] The processing module 940 is used to process the two-dimensional matrix using a multi-task model to obtain the modulation scheme and symbol rate of the transmitted data.

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

[0163] The convolution module is used to convolve a two-dimensional matrix to obtain convolutional features;

[0164] The convolutional features are downsampled using the pooling module to obtain the first pooled features;

[0165] The DRA module is used to extract local detail features from the first pooling features to obtain detail features.

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

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

[0168] The timing characteristics are processed using a fully connected module to obtain the modulation scheme and symbol rate of the transmitted data.

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

[0170] The first pooling feature is processed using the first to sixth DA units in a series.

[0171] The first DA feature is obtained by adding the first DA feature output from the fourth DA unit to the second DA feature output from the sixth DA unit through the first ECA unit;

[0172] The third DA feature is processed using the seventh and eighth DA units connected in series;

[0173] The fourth DA feature output by the second DA unit is added to the fifth DA feature output by the eighth DA unit through the second ECA unit to obtain the sixth DA feature;

[0174] The sixth DA feature is processed using the ninth and tenth DA units connected in series;

[0175] 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.

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

[0177] The second pooling feature is divided into blocks using the Patch Partition unit and flattened in the channel direction to obtain the block features;

[0178] Linear features are obtained by using linear embedding units to perform linear transformation on the block features;

[0179] Temporal features are obtained by extracting temporal features from linear features using two cascaded Swing Transformer Block units.

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

[0181] The first ST feature is obtained by processing the linear features using cascaded LN and W-MSA.

[0182] The linear feature is added to the first ST feature to obtain the second ST feature;

[0183] The second ST feature is processed 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] The fourth ST feature is processed 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] The sixth ST feature is processed 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 time series feature.

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

[0190] The creation module is used to create a training dataset. The training samples in the training dataset include a two-dimensional matrix, a real modulation scheme, and a real symbol rate. The two-dimensional matrix is ​​a simulation data processed with the real modulation scheme and the real symbol rate. The processed radio frequency signal is transmitted using a simulated wireless channel. The real and imaginary parts of the complex signal obtained after transmission are separated. The simulated wireless channel is a simulated channel using Rayleigh attenuation superimposed with additive white Gaussian noise.

[0191] The module is also used to create multi-task models;

[0192] The processing module 940 is also used to process the two-dimensional matrix in the training sample using a multi-task model for each training sample to obtain the predicted modulation scheme and the predicted symbol rate.

[0193] The training module is used to calculate the loss value of the multi-task model based on the predicted modulation scheme and the actual modulation scheme, the predicted symbol rate and the actual symbol rate, and to train the model parameters of the multi-task model based on the loss value.

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

[0195] Using a signal modulator, a real modulation scheme and a real symbol rate are randomly selected. The selected real modulation scheme and real symbol rate are used to modulate the randomly generated binary sequence to obtain modulated data, which includes N symbols.

[0196] The waveform of the modulated data is adjusted using a square root raised cosine pulse filter, and the shaped data is then converted into an radio frequency signal after energy normalization.

[0197] Radio frequency signals are transmitted using an analog wireless channel, which is an analog channel that uses Rayleigh attenuation superimposed with additive white Gaussian noise;

[0198] 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.

[0199] A training sample is composed of a two-dimensional matrix, the actual modulation scheme, and the actual symbol rate.

[0200] In summary, the modulation identification and symbol rate estimation device provided in this application has a multi-task model with DRA and SST modules, enabling simultaneous modulation identification and symbol rate estimation. Compared with mainstream models, this model offers better performance while also boasting advantages in low complexity and low parameter count. Furthermore, compared to two corresponding single-task models, the system overhead of the multi-task model is essentially the same as that of a single-task model, with slightly improved classification and prediction accuracy, significantly reducing the model's inference speed. In addition, the multi-task model can share feature extraction layers, effectively avoiding overfitting and parameter complexity, and improving the model's generalization ability and robustness.

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

[0202] One embodiment of this application provides a computer device, which includes the above-described arbitrary modulation identification and symbol rate estimation device.

[0203] It should be noted that the modulation identification and symbol rate estimation device provided in the above embodiments is only illustrated by the division of the above functional modules when performing modulation identification and symbol rate estimation. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the modulation identification and symbol rate estimation device can be divided into different functional modules to complete all or part of the functions described above. In addition, the modulation identification and symbol rate estimation device provided in the above embodiments and the modulation identification and symbol rate estimation method embodiments belong to the same concept, and the specific implementation process is detailed in the method embodiments, which will not be repeated here.

[0204] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0205] The above description is not intended to limit the embodiments of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of the embodiments of this application.

Claims

1. A modulation identification and symbol rate 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 symbol rate. The wireless channel is a channel using Rayleigh attenuation superimposed with 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, comprising a convolutional module, a pooling module, a deep separable convolutional (DRA) module with attention mechanisms and residual connections, a simplified Swin Transformer (SST) module, and a fully connected module. The DRA module consists of a skip-connected structure of p deep separable convolutional (DA) units with attention mechanisms and q units with efficient channel attention (ECA) mechanisms. The DA units are deep separable convolutional (DSC) units with convolutional block attention (CBAM) mechanisms, where 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. The modulation scheme and symbol rate of the transmitted data are obtained by processing the two-dimensional matrix using the multi-task model.

2. The modulation identification and symbol rate 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 symbol rate of the transmitted data includes: The convolution module is used to convolve the two-dimensional matrix to obtain convolutional features; The convolutional features are downsampled using the pooling module to obtain the first pooling feature; The DRA module is used to extract local detail features from the first pooling feature to obtain detail features; The pooling module is used to aggregate the detailed features and select salient features to obtain the second pooled feature. The SST module is used to extract global correlation features from the second pooling feature to obtain temporal features; The timing characteristics are processed using the fully connected module to obtain the modulation scheme and symbol rate of the transmitted data.

3. The modulation identification and symbol rate estimation method according to claim 2, characterized in that, The DRA module includes ten cascaded DA units and three skip-connected structures with ECA units. The first ECA unit skips to the fourth and sixth DA units, the second ECA unit skips to the second and eighth DA units, and the third ECA unit skips to the input and output of the DRA module. The DRA module is used to extract local detail features from the first pooling feature to obtain detail features, including: The first pooling feature is processed using the first to sixth DA units in series; The first DA feature is obtained by adding the first DA feature output from the fourth DA unit to the second DA feature output from the sixth DA unit through the first ECA unit; The third DA feature is processed using the seventh and eighth DA units connected in series; The fourth DA feature output by the second DA unit is added to the fifth DA feature output by the eighth DA unit through the second ECA unit to obtain the sixth DA feature; The sixth DA feature is processed using the ninth and tenth DA units 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 step of using the SST module to extract global correlation features from the second pooling feature 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 the block feature; The block features are linearly transformed using the linear embedding unit to obtain linear features; The temporal features are obtained by extracting time features from the linear features using two cascaded Swing Transformer Block units.

5. The modulation identification and symbol rate estimation method according to claim 4, characterized in that, The two Swing Transformer Block units include Layer Normalization (LN), Window-based Multi-Head Self-Attention (W-MSA) LN, Multilayer Perceptron (MLP) LN, and Shift-Window-based Multi-Head Self-Attention (SW-MSA) LN and MLP. The temporal feature extraction of the linear features using the two cascaded Swing Transformer Block units to obtain the temporal features includes: The linear features are processed using cascaded LN and W-MSA to obtain the first ST feature; The linear feature is added to the first ST feature to obtain the second ST feature; The second ST feature is processed using cascaded LN and MLP to obtain the third ST feature; Add the second ST feature to the third ST feature to obtain the fourth ST feature; The fourth ST feature is processed using cascaded LN and SW-MSA to obtain the fifth ST feature; The fourth ST feature is added to the fifth ST feature to obtain the sixth ST feature; The sixth ST feature is processed using cascaded LN and MLP to obtain the seventh ST feature; The sixth ST feature is added to the seventh ST feature to obtain the time series 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 includes: A training dataset is created, in which training samples include a two-dimensional matrix, a real modulation scheme, and a real symbol rate. The two-dimensional matrix is ​​obtained by processing the simulation data with the real modulation scheme and the real symbol rate. The processed radio frequency signal is transmitted using a simulated wireless channel. The real and imaginary parts of the complex signal obtained after transmission are separated. The simulated wireless channel is a simulated channel using Rayleigh attenuation superimposed with additive white Gaussian noise. 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 symbol rate. The loss value of the multi-task model is calculated based on the predicted modulation scheme and the actual modulation scheme, the predicted symbol rate and the actual symbol rate, and the model parameters of the multi-task model are trained based on the loss value.

7. The modulation identification and symbol rate estimation method according to claim 6, characterized in that, The creation of the training dataset includes: A signal modulator randomly selects a real modulation scheme and a real symbol rate, and modulates a randomly generated binary sequence using the selected real modulation scheme and the real symbol rate to obtain modulated data, which includes N symbols. The waveform of the modulated data is adjusted using a square root raised cosine pulse filter, and the shaped data is converted into a radio frequency signal after energy normalization. The radio frequency signal is transmitted using an analog wireless channel, which is an analog channel using Rayleigh attenuation superimposed with 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 symbol rate are combined to form a training sample.

8. A modulation identification and symbol rate 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 symbol rate. The wireless channel is a channel using Rayleigh attenuation superimposed with 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, which includes a convolution module, a pooling module, a deep separable convolutional (DRA) module with attention mechanisms and residual connections, a simplified Swin Transformer (SST) module, and a fully connected module. The DRA module consists of a skip-connected structure of p deep separable convolutional (DA) units with attention mechanisms and q units with efficient channel attention (ECA) mechanisms. The DA units are deep separable convolutional (DSC) units with convolutional block attention (CBAM) mechanisms, where 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. The processing module is used to process the two-dimensional matrix using the multi-task model to obtain the modulation scheme and symbol rate of the transmitted data.

9. 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 symbol rate estimation method as described in any one of claims 1 to 7.

10. A computer device, characterized in that, The computer device includes: the modulation identification and symbol rate estimation apparatus as described in claim 8.

Images