Neural network-based signal parameter estimation method, device, medium and apparatus

Abstract

The application discloses a neural network-based signal parameter estimation method, device, medium and equipment, and belongs to the technical field of communication. The method comprises the following steps: acquiring a pre-trained neural network model and a complex modulation signal to be estimated, wherein the neural network model at least comprises a convolution module, a convolution attention module, a time convolution network module and a full connection module; the real part and the imaginary part of the complex modulation signal are combined to form a two-dimensional matrix; the two-dimensional matrix is processed by using the channel attention mechanism and the spatial attention mechanism in the convolution module and the convolution attention module, and the inflation causal convolution layer in the time convolution network module, so that a feature matrix is obtained; the feature matrix is processed by using a first branch in the full connection module, so that a modulation type parameter is obtained; and the feature matrix is processed by using a second branch in the full connection module, so that a symbol rate parameter is obtained. The application can simultaneously perform modulation type estimation and symbol rate estimation, saves system overhead, and improves network performance.

Description

Technical Field

[0001] This application relates to the field of communication technology, and in particular to a signal parameter estimation method, apparatus, medium and device based on neural networks. Background Technology

[0002] With the rapid development of global technology, radio communication technology has become increasingly mature. In current communication technologies, all communication signals (radio, television, mobile phones, etc.) must be modulated before transmission. In non-cooperative radio communication, the communicating parties cannot know in advance the type of signal, communication parameters, and pilot information transmitted by the other party. Among these, the modulation type and symbol rate are particularly important. Only by accurately estimating the modulation type and symbol rate of the signal transmitted by the other party can subsequent demodulation and decoding operations be ensured to proceed smoothly.

[0003] Currently, traditional modulation type estimation algorithms and symbol rate estimation algorithms have certain limitations. For example, they require prior knowledge, suffer from high latency, are sensitive to noise, have high computational complexity, limited applicability, require high signal sampling rates, are susceptible to errors due to channel variations and multipath channels, and require more computational resources. In general, traditional methods typically require prior information, making end-to-end systems difficult to implement, which is infeasible for unknown signals. Furthermore, traditional methods cannot achieve synchronous parallel computation, thus doubling the resource consumption. In addition, traditional methods often perform poorly with complex nonlinear signals; therefore, more flexible methods are needed to address these problems. Summary of the Invention

[0004] This application provides a signal parameter estimation method, apparatus, medium, and device based on neural networks to solve problems in related technologies. The technical solution is as follows:

[0005] According to a first aspect of this application, a signal parameter estimation method based on a neural network is provided, the method comprising:

[0006] Obtain a pre-trained neural network model and a complex modulation signal to be estimated, wherein the neural network model includes at least a convolutional module, a convolutional attention module, a temporal convolutional network module, and a fully connected module;

[0007] The real and imaginary parts of the complex modulated signal are combined to form a two-dimensional matrix;

[0008] The two-dimensional matrix is ​​processed using the convolution module, the channel attention mechanism and spatial attention mechanism in the convolutional attention module, and the dilated causal convolutional layer in the temporal convolutional network module to obtain the feature matrix;

[0009] The feature matrix is ​​processed using the first branch in the fully connected module to obtain the modulation type parameters;

[0010] The feature matrix is ​​processed using the second branch in the fully connected module to obtain the symbol rate parameter.

[0011] In one possible implementation, the process of processing the two-dimensional matrix using the convolution module, the channel attention mechanism and the spatial attention mechanism in the convolutional attention module, and the dilated causal convolutional layer in the temporal convolutional network module to obtain the feature matrix includes:

[0012] The first convolutional module is used to perform channel transformation on the two-dimensional matrix to obtain the first feature matrix;

[0013] The first feature matrix is ​​enhanced by using the channel attention mechanism and spatial attention mechanism in the first convolutional attention module to obtain the second feature matrix;

[0014] The second convolutional module is used to extract spatial features from the second feature matrix to obtain the third feature matrix;

[0015] The third feature matrix is ​​enhanced using the channel attention and spatial attention mechanisms in the second convolutional attention module to obtain the fourth feature matrix.

[0016] The fourth feature matrix is ​​obtained by extracting temporal features using the dilated causal convolutional layer in the temporal convolutional network module.

[0017] In one possible implementation, the dilated causal convolutional layer includes a convolutional kernel and a dilation rate. Then, the temporal feature extraction of the fourth feature matrix using the dilated causal convolutional layer in the temporal convolutional network module, to obtain the feature matrix, includes:

[0018] For any element in the fourth feature matrix at any time, dilated convolution operation is performed on the element using the convolution kernel and the dilation rate to obtain a feature matrix with temporal characteristics.

[0019] In one possible implementation, the neural network model further includes a pooling module, then the method further includes:

[0020] The second feature matrix is ​​downsampled using the first pooling module;

[0021] The fourth feature matrix is ​​downsampled using the second pooling module.

[0022] In one possible implementation, the method further includes:

[0023] Generate a dataset, wherein the data samples in the dataset include a two-dimensional matrix of complex modulated signals and labeled modulation type parameters and symbol rate parameters;

[0024] The neural network model is trained using the dataset.

[0025] In one possible implementation, the generation of the dataset includes:

[0026] Generate a binary random number sequence based on the sampling frequency and symbol rate parameters;

[0027] The binary random number sequence is modulated according to the modulation type parameter to obtain the first data sequence;

[0028] The waveform of the first data sequence is adjusted to obtain the second data sequence;

[0029] The second data sequence is energy normalized to obtain the third data sequence;

[0030] The third data sequence is passed through an additive white Gaussian noise channel environment to obtain a complex modulated signal;

[0031] Extract the real and imaginary parts from the complex modulated signal, and concatenate the real and imaginary parts into a two-dimensional matrix;

[0032] The two-dimensional matrix, the modulation type parameter, and the symbol rate parameter are used to form a data sample.

[0033] In one possible implementation, training the neural network model using the dataset includes:

[0034] Create a cross-entropy loss function for the first branch and a mean squared error loss function for the second branch;

[0035] When training the neural network model using the dataset, the error of the first branch is calculated using the cross-entropy loss function, and the error of the second branch is calculated using the mean squared error loss function.

[0036] The parameters of the neural network model are adjusted according to the error until the neural network model meets the preset conditions, at which point training stops.

[0037] On the one hand, a signal parameter estimation device based on a neural network is provided, the device comprising:

[0038] The acquisition module is used to acquire a pre-trained neural network model and a complex modulation signal to be estimated. The neural network model includes at least a convolution module, a convolutional attention module, a temporal convolutional network module, and a fully connected module.

[0039] A generation module is used to form a two-dimensional matrix from the real and imaginary parts of the complex modulated signal;

[0040] The processing module is used to process the two-dimensional matrix using the convolution module, the channel attention mechanism and spatial attention mechanism in the convolutional attention module, and the dilated causal convolutional layer in the temporal convolutional network module to obtain the feature matrix;

[0041] An estimation module is used to process the feature matrix using the first branch in the fully connected module to obtain modulation type parameters;

[0042] The estimation module is further configured to process the feature matrix using the second branch in the fully connected module to obtain the symbol rate parameter.

[0043] 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 neural network-based signal parameter estimation method as described above.

[0044] According to a fourth aspect of this application, a computer device is provided, the computer device including a neural network-based signal parameter estimation device.

[0045] The beneficial effects of the technical solution provided in this application include at least the following:

[0046] By introducing the Convolutional Block Attention Module (CBAM) mechanism, which combines channel attention and spatial attention mechanisms, into the neural network model, the overall performance of the neural network model can be further improved by helping the neural network model focus on relevant features and suppress noise.

[0047] Temporal Convolutional Network (TCN) modules in neural network models have fewer parameters and are better suited for processing long sequence data. They can reduce the complexity and number of parameters of neural network models and improve network performance.

[0048] The neural network model is a multi-task neural network model that can simultaneously perform modulation type estimation and symbol rate estimation. Compared with two corresponding single-task neural network models, the system overhead of the multi-task neural network model is basically the same as that of a single-task neural network model, while the classification and prediction accuracy is slightly improved, and the inference speed of the network is significantly reduced. In addition, the multi-task neural network model can share feature extraction layers, effectively avoiding overfitting and parameter complexity, and improving the network's generalization ability and robustness. Compared with mainstream network models, it has the advantages of good performance, low complexity, and low parameter count. Attached Figure Description

[0049] 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.

[0050] Figure 1 This is a flowchart of a signal parameter estimation method based on a neural network provided in one embodiment of this application;

[0051] Figure 2 This is a schematic diagram illustrating the generation of a dataset provided in one embodiment of this application;

[0052] Figure 3 This is a detailed structural diagram of a neural network model provided in one embodiment of this application;

[0053] Figure 4 This is a schematic diagram of the macroscopic structure of a neural network model provided in one embodiment of this application;

[0054] Figure 5 This is a flowchart of a signal parameter estimation method based on a neural network provided in one embodiment of this application;

[0055] Figure 6 This is a structural block diagram of a signal parameter estimation device based on a neural network provided in one embodiment of this application. Detailed Implementation

[0056] 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.

[0057] In recent years, with the development of artificial intelligence, deep learning has emerged as a crucial technology in the field, possessing numerous unique advantages. Inspired by the neural networks of the human brain, it can handle complex nonlinear relationships, thus excelling in tasks such as image recognition and natural language processing. Deep learning models, trained on large-scale datasets, can automatically learn features, reducing reliance on manual feature engineering and making them widely applicable. Furthermore, supported by large-scale computer clusters and graphics processing units (GPUs), deep learning models possess outstanding computational power, capable of handling massive amounts of data, increasing model complexity, and supporting transfer learning, enabling the transfer of knowledge gained in one domain to another, improving the model's generalization ability. Most importantly, deep learning models are constantly evolving, with new architectures and algorithms emerging, providing powerful tools for solving various complex problems. Therefore, deep learning has demonstrated enormous potential in multiple fields such as healthcare, autonomous driving, and finance, bringing limitless possibilities for future technological innovation and social progress.

[0058] In the field of communications, deep learning models also offer significant advantages. They can be used for modulation type estimation and symbol rate estimation, automatically extracting signal features and reducing reliance on complex signal processing algorithms, thereby improving estimation accuracy. Deep learning models can also adapt to different signal-to-noise ratios, spectral conditions, and modulation schemes, exhibiting strong generalization capabilities, which are crucial in modern communication systems, especially in dynamic wireless environments. Furthermore, deep learning models can be used for communication channel modeling and bit error rate prediction, contributing to the optimization of communication system performance. With continuous improvements in hardware and computing power, the real-time performance and efficiency of deep learning models in the communications field are constantly improving, providing broader possibilities for practical applications. Therefore, deep learning models provide a powerful tool for improving the intelligence and efficiency of signal processing in the communications field, and are expected to drive further development of communication technologies in the future.

[0059] Current mainstream methods for applying deep learning to the field of communication model a single task, such as neural network models that only estimate debugging type or neural network models that only estimate symbol rate.

[0060] This invention proposes a multi-task neural network model for joint modulation type estimation and symbol rate estimation, and designs an improved multi-task neural network CTDMTL-Net based on the composite neural network CLDNN (CNN+LSTM+DNN). This neural network model receives a two-dimensional matrix composed of the real and imaginary parts of a complex modulation signal. This two-dimensional matrix is ​​transformed in terms of data channel dimension through convolution. Then, a convolution module is used to extract the spatial features of the input data, and convolutional attention modules are added before and after the convolution module to reasonably optimize the feature weights. During this process, pooling operations can also be used to downsample the feature mapping. Then, a temporal convolutional network module is used to extract the temporal features of the input data. Finally, a fully connected module simultaneously performs modulation type estimation and symbol rate estimation, reducing the computational resource consumption during training and improving the network's inference speed.

[0061] like Figure 1 The diagram illustrates a flowchart of a neural network-based signal parameter estimation method according to an embodiment of this application. This neural network-based signal parameter estimation method can be applied to computer devices. The neural network-based signal parameter estimation method may include:

[0062] Step 101: Obtain a pre-trained neural network model and a complex modulation signal to be estimated. The neural network model includes at least a convolutional module, a convolutional attention module, a temporal convolutional network module, and a fully connected module.

[0063] Computer devices can train neural network models to obtain trained neural network models, or computer devices can obtain trained neural network models from other devices or networks. In this embodiment, the source of the neural network model is not limited.

[0064] Complex modulated signals are modulated signals. Computer equipment needs to obtain their modulation type parameters and symbol rate parameters to facilitate subsequent demodulation and decoding operations. Modulation types can include: BPSK (Binary Phase Shift Keying), QPSK (Quadrature Phase Shift Keying), OQPSK (Offset Quadrature Phase Shift Keying), 8PSK (8 Phase Shift Keying), 16QAM (Quadrature Amplitude Modulation), 64QAM, MSK (Minimum Shift Keying), and GMSK (Gaussian Filtered Minimum Shift Keying). There are 10 symbol rates, randomly generated between 100K and 200K. This embodiment uses 8 modulation types and 10 symbol rates as examples. In actual use, it may include fewer or more modulation types and symbol rates.

[0065] The convolution module is used to perform convolution operations on the input data.

[0066] The convolutional attention module is composed of channel attention and spatial attention mechanisms connected together, so as to consider the correlation between channel dimension and spatial dimension while processing data.

[0067] The temporal convolutional network module includes four dilated causal convolutional layers for extracting temporal features of the data.

[0068] The fully connected module includes two branches: the first branch is used for modulation type estimation, and the second branch is used for symbol rate estimation.

[0069] Step 102: Combine the real and imaginary parts of the complex modulated signal into a two-dimensional matrix.

[0070] The computer device separates the real part r[n] and the imaginary part i[n] of the complex signal data y[n], then concatenates them into a two-dimensional matrix A[r][i], and then inputs the two-dimensional matrix into the neural network model for processing.

[0071] Step 103: The two-dimensional matrix is ​​processed using the convolution module, the channel attention mechanism and spatial attention mechanism in the convolutional attention module, and the dilated causal convolutional layer in the temporal convolutional network module to obtain the feature matrix.

[0072] Step 104: Process the feature matrix using the first branch in the fully connected module to obtain the modulation type parameters.

[0073] The first branch in the fully connected module outputs the probabilities of 8 modulation types, and the modulation type with the highest probability is determined as the modulation type parameter.

[0074] Step 105: Process the feature matrix using the second branch in the fully connected module to obtain the symbol rate parameters.

[0075] The second branch in the fully connected module outputs a prediction result, which is then used as the symbol rate parameter.

[0076] In summary, the signal parameter estimation method based on neural networks provided in this application introduces a convolutional attention mechanism that combines channel attention and spatial attention mechanisms into the neural network model. This helps the neural network model focus its attention on relevant features and suppress noise, thereby further improving the overall performance of the neural network model.

[0077] Temporal convolutional network modules in neural network models have fewer parameters and are more suitable for processing long sequence data. They can reduce the complexity and number of parameters of neural network models and improve network performance.

[0078] The neural network model is a multi-task neural network model that can simultaneously perform modulation type estimation and symbol rate estimation. Compared with two corresponding single-task neural network models, the system overhead of the multi-task neural network model is basically the same as that of a single-task neural network model, while the classification and prediction accuracy is slightly improved, and the inference speed of the network is significantly reduced. In addition, the multi-task neural network model can share feature extraction layers, effectively avoiding overfitting and parameter complexity, and improving the network's generalization ability and robustness. Compared with mainstream network models, it has the advantages of good performance, low complexity, and low parameter count.

[0079] The training process for a neural network model is described below. The training process includes the following steps:

[0080] 1. Generate a dataset. The data samples in the dataset include a two-dimensional matrix of complex modulated signals, as well as labeled modulation type parameters and symbol rate parameters.

[0081] In this embodiment, a simulation dataset can be generated using MATLAB. The dataset used in this invention contains 880,000 data points, including 8 modulation schemes (BPSK, QPSK, OQPSK, 8PSK, 16QAM, 64QAM, MSK, and GMSK), 10 symbol rates (randomly generated from 100K to 200K), and 11 signal-to-noise ratios (0-20dB intervals in 2dB increments). Each signal-to-noise ratio for each modulation scheme at each symbol rate has 1000 data points, and each data point is 2×1024 in size.

[0082] Specifically, such as Figure 2 As shown, generating a dataset can include:

[0083] (1) Generate a binary random number sequence based on the sampling frequency and symbol rate parameters.

[0084] MATLAB generates a random binary number sequence s[n] based on the sampling frequency and symbol rate parameters.

[0085] (2) Modulate the binary random number sequence according to the modulation type parameter to obtain the first data sequence.

[0086] The signal modulator can select a modulation type parameter to modulate the binary random number sequence s[n] to obtain a first data sequence a[n] of length N.

[0087] (3) The waveform of the first data sequence is adjusted to obtain the second data sequence.

[0088] The pulse shaping filter adjusts the waveform of the first data sequence a[n] to obtain the second data sequence x'(n).

[0089] The model uses a square root raised cosine (SRRC) pulse shaping filter g[n], the expression of which is as follows:

[0090]

[0091] in, Let g[n] be an SRRC filter; t be the symbol duration; T be the symbol period; and α be the roll-off factor. Then, the expression for x'(n) after the first data sequence a[n] passes through filter g[n] can be written as:

[0092]

[0093] Where x'(n) is the output of the data after passing through the filter; t is the symbol duration; T is the symbol period; and Tx is the sample duration. The oversampling factor P of the filter can be expressed as T / Tx, i.e., P = T / Tx. Given that the sampling frequency Fs = 1 / Tx, the symbol rate parameter R = 1 / T = Fs / P can be obtained.

[0094] (4) Normalize the energy of the second data sequence to obtain the third data sequence.

[0095] After the waveform-adjusted x'(n) data is subjected to energy normalization, the third data sequence x[n] is obtained. The expression for x[n] is:

[0096]

[0097] Where x(t) is the output of the data after energy normalization.

[0098] (5) Pass the third data sequence through a channel environment of additive white Gaussian noise to obtain a complex modulated signal.

[0099] By passing the energy-normalized third data sequence x[n] through an additive white Gaussian noise (awgn) channel environment w[n], we can generate the complex modulation signal y[n] we need. The expression for y[n] is:

[0100] y(n)=x(n)+w(n) (4)

[0101] (6) Extract the real and imaginary parts from the complex modulated signal and concatenate the real and imaginary parts into a two-dimensional matrix.

[0102] The generated complex modulated signal y[n] is separated into real part r[n] and imaginary part i[n], and then concatenated into a two-dimensional matrix A[r][i].

[0103] (7) The two-dimensional matrix, modulation type parameters and symbol rate parameters are combined to form a data sample.

[0104] Each two-dimensional matrix, along with its corresponding modulation type parameters and symbol rate parameters, is labeled and used to form data samples, ultimately resulting in a dataset.

[0105] 2. Train the neural network model using the dataset.

[0106] like Figure 3 and 4 As shown, the multi-task neural network model CTDMTL includes: a convolutional module, a convolutional attention module, a pooling module, a temporal convolutional network module, and a fully connected module. The first convolutional module corresponds to... Figure 3The convolutional operation in the first layer is followed by the first CBAM module and the first pooling module. The second convolutional module is a CNN module, followed by the second CBAM module and the second pooling module. The temporal convolutional network module is a TCN module, followed by a fully connected module and the output layer.

[0107] (1) The first convolution module uses two-dimensional convolution to extract features from the input two-dimensional matrix, so that the input data is convolved from 1 channel to 32 channels, which facilitates the channel attention mechanism in the subsequent CABM attention mechanism to optimize the channel features.

[0108] Specifically, the first convolution module contains three basic units: two-dimensional convolution (Conv2d), two-dimensional normalization (BatchNorm2d), and activation function (PReLU). The kernel size of the two-dimensional convolution is 1*3, and the number of channels after convolution is 32. In this way, the input two-dimensional matrix of size 1*2*1024 is transformed into data of size 32*2*1022 after the convolution operation.

[0109] (2) The CBAM module includes channel attention mechanism and spatial attention mechanism. Using it before and after the convolution operation can take into account the correlation between channel dimension and spatial dimension, and better capture important information in the data.

[0110] Specifically, the input feature F passes through the channel attention module to obtain the channel-dimensional attention weight-optimized feature F'. F' then passes through the spatial attention module to obtain the feature F'', which is optimized in both the channel and spatial dimensions. This process extracts key information and yields the enhanced output feature. The entire process can be summarized as follows:

[0111]

[0112]

[0113] Here, Mc represents channel attention and Ms represents spatial attention.

[0114] It should be noted that the CBAM attention mechanism does not change the dimensions and size of the data. Therefore, the data remains 32*2*1022 in size after passing through the CBAM module.

[0115] (3) Pooling operations downsample the data after convolution to reduce network complexity.

[0116] The pooling operation uses max pooling (MaxPool2d), with a kernel size of 1*2. Its main function is to downsample the data to reduce complexity.

[0117] The input data of size 32*2*1022 is transformed into data of size 32*2*511 after pooling.

[0118] (4) The CNN module uses four consecutive two-dimensional convolutional layers to extract spatial features of the input data and passes the extracted feature information to the TCN module.

[0119] The CNN module is primarily used to extract spatial features from the data. It consists of four consecutive stacked 2D convolutions (Conv2d), each followed by a 2D normalization layer (BatchNorm2d) and an activation function layer (PReLU). The four convolutions use 1*3 linear kernels with output channels of 64, 128, 64, and 32 respectively. Therefore, the convolution module has a symmetrical structure, allowing it to transform the 32*2*511 feature map after the previous pooling operation into a 32*2*503 feature map with extracted spatial features. This maintains channel consistency before and after the CNN module, which is beneficial for feature optimization by the CBAM module. The feature map after CBAM attention mechanism optimization of feature weights retains its original size. It is then transformed into a 32*2*251 feature map after another pooling operation and fed into the TCN module.

[0120] (5) The TCN module uses four consecutive dilated causal convolutional layers to extract temporal features and introduces dilated causal convolution to expand the receptive field. Finally, the extracted feature information is passed to the fully connected module.

[0121] The TCN module is primarily used to extract temporal features from the data. It consists of four consecutive stacked dilated causal convolutional layers (DCCs), each followed by an activation function layer (PReLU) and a dropout layer. The four DCCs use a kernel size of 3, and the dropout layer has a parameter setting of 0.5. The output channels of the four DCCs are 128, 256, 128, and 64, respectively. The TCN module receives a feature map of size 32*2*251 from the previous module. Since the core DCCs used in the TCN module are one-dimensional, the first two dimensions of the received data need to be compressed to a size of 64*251 before being fed into the TCN module. After processing by the four DCCs, the output feature map size remains 64*251. In addition, a dilation rate was introduced in the TCN module to expand the receptive field. We introduced dilation rates of [1, 2, 4, 8] in the four dilated causal convolutional layers respectively. The introduction of the dilation rate does not change the output size of the data. Therefore, the data processed by the TCN module still maintains a size of 64*251 and is passed to the fully connected module.

[0122] (6) After receiving the feature information from the TCN, the fully connected module outputs the modulation type parameter and symbol rate parameter through two output mappings respectively.

[0123] The first branch of the fully connected module outputs the probabilities of eight modulation types, and the modulation type with the highest probability is determined as the modulation type parameter. The second branch of the fully connected module outputs a prediction result, which is determined as the symbol rate parameter.

[0124] Computer equipment can divide the dataset into training, validation, and test sets in a 7:1:2 ratio, that is, using 700 data points for training, 100 data points for validation, and finally 200 data points for testing.

[0125] Computer equipment can build a multi-task-based neural network model. The training set is input into the neural network model for training. A validation set is used to verify whether the neural network model has converged or is overfitting. The learning rate is dynamically adjusted to ensure convergence. Finally, the converged neural network model and its parameters are saved. The test set is then input into the converged neural network model, which can then output the modulation type parameters and symbol rate parameters of the complex modulation signal.

[0126] First, the pre-packaged training set is loaded 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. A cross-entropy loss function is created for the first branch to evaluate the error in the modulation type estimation task; a mean squared error loss function (MSELoss) is created for the second branch to evaluate the error in the bit rate estimation task. Furthermore, to ensure that the two tasks have equal weight in the network, their loss ratios are set. A ratio of 22:1 is used to set the loss ratio for multiple tasks, ensuring that the loss weights for both tasks are equal, thus allowing the neural network model to achieve optimal performance for each sub-task. After each training round, validation data is fed into the neural network model for validation, monitoring the network state in real time and preventing overfitting. Finally, the converged neural network model is saved after training.

[0127] Then, the accuracy and precision of the neural network model are verified using a test set. The main steps are as follows: First, the test set is retrieved and then fed into the trained neural network model. The neural network model can automatically output the modulation type parameters and symbol rate parameters of the input data. Then, the output modulation type parameters are compared with the actual modulation type, and the output symbol rate parameters are compared with the actual symbol rate. This allows us to determine the accuracy and precision of the neural network model.

[0128] like Figure 5 The diagram illustrates a flowchart of a neural network-based signal parameter estimation method according to an embodiment of this application. This neural network-based signal parameter estimation method can be applied to computer devices. The neural network-based signal parameter estimation method may include:

[0129] Step 501: Obtain a pre-trained neural network model and a complex modulation signal to be estimated. The neural network model includes at least a convolutional module, a convolutional attention module, a temporal convolutional network module, and a fully connected module.

[0130] The training process of the neural network model is shown above and will not be repeated here.

[0131] Complex modulated signals are modulated signals. Computer equipment needs to obtain their modulation type parameters and symbol rate parameters to facilitate subsequent demodulation, decoding, and other operations. Modulation types can include BPSK, QPSK, OQPSK, 8PSK, 16QAM, 64QAM, MSK, and GMSK. There are 10 symbol rates, randomly generated between 100K and 200K. This embodiment uses 8 modulation types and 10 symbol rates as an example; in actual use, fewer or more modulation types and symbol rates may be included.

[0132] Step 502: Combine the real and imaginary parts of the complex modulated signal into a two-dimensional matrix.

[0133] The computer device separates the real part r[n] and the imaginary part i[n] of the complex signal data y[n], then concatenates them into a two-dimensional matrix A[r][i], and then inputs the two-dimensional matrix into the neural network model for processing.

[0134] Step 503: Use the first convolution module to perform channel transformation on the two-dimensional matrix to obtain the first feature matrix.

[0135] The first convolutional module uses two-dimensional convolution to extract features from the input two-dimensional matrix, turning the input data from 1-channel convolution to 32-channel convolution, which facilitates the channel attention mechanism in the subsequent CABM attention mechanism to optimize channel features.

[0136] Step 504: Use the channel attention mechanism and spatial attention mechanism in the first convolutional attention module to enhance the features of the first feature matrix to obtain the second feature matrix.

[0137] The first convolutional attention module includes channel attention and spatial attention mechanisms. Using it before and after convolution operations can take into account the correlation between channel and spatial dimensions, thus better capturing important information in the data.

[0138] Step 505: Use the first pooling module to downsample the second feature matrix.

[0139] The main function of the first pooling module is to downsample the data to reduce complexity.

[0140] Step 506: Use the second convolution module to extract spatial features from the second feature matrix to obtain the third feature matrix.

[0141] The second convolutional module uses four consecutive two-dimensional convolutional layers to extract spatial features from the input data.

[0142] Step 507: Use the channel attention mechanism and spatial attention mechanism in the second convolutional attention module to enhance the features of the third feature matrix to obtain the fourth feature matrix.

[0143] The second convolutional attention module includes channel attention and spatial attention mechanisms. Using it before and after convolution operations can take into account the correlation between channel and spatial dimensions, thus better capturing important information in the data.

[0144] Step 508: Use the second pooling module to downsample the fourth feature matrix.

[0145] The main function of the second pooling module is to downsample the data to reduce complexity.

[0146] Step 509: Use the dilated causal convolutional layer in the temporal convolutional network module to extract temporal features from the fourth feature matrix to obtain the feature matrix.

[0147] The temporal convolutional network module uses four consecutive dilated causal convolutional layers to extract temporal features and introduces dilated causal convolution to expand the receptive field, ultimately extracting the feature matrix.

[0148] The dilated causal convolutional layer includes a convolution kernel and a dilation rate. The dilated causal convolutional layer in the temporal convolutional network module is used to extract temporal features from the fourth feature matrix to obtain a feature matrix. This includes: for any element in the fourth feature matrix at any time, dilation convolution is performed on the element using the convolution kernel and the dilation rate to obtain a feature matrix with temporal features.

[0149] Dilated causal convolution assumes the input sequence of the model is X, and the size of the convolution kernel f is k. The dilated convolution operation on any element at time t in the sequence can be expressed as:

[0150]

[0151] Where d is the expansion rate; X t-di This represents the value at time td·i in the input sequence.

[0152] Step 510: Process the feature matrix using the first branch in the fully connected module to obtain the modulation type parameters.

[0153] The first branch in the fully connected module outputs the probabilities of 8 modulation types, and the modulation type with the highest probability is determined as the modulation type parameter.

[0154] Step 511: The feature matrix is ​​processed using the second branch in the fully connected module to obtain the symbol rate parameters.

[0155] The second branch in the fully connected module outputs a prediction result, which is then used as the symbol rate parameter.

[0156] In summary, the signal parameter estimation method based on neural networks provided in this application introduces a convolutional attention mechanism that combines channel attention and spatial attention mechanisms into the neural network model. This helps the neural network model focus its attention on relevant features and suppress noise, thereby further improving the overall performance of the neural network model.

[0157] Temporal convolutional network modules in neural network models have fewer parameters and are more suitable for processing long sequence data. They can reduce the complexity and number of parameters of neural network models and improve network performance.

[0158] The neural network model is a multi-task neural network model that can simultaneously perform modulation type estimation and symbol rate estimation. Compared with two corresponding single-task neural network models, the system overhead of the multi-task neural network model is basically the same as that of a single-task neural network model, while the classification and prediction accuracy is slightly improved, and the inference speed of the network is significantly reduced. In addition, the multi-task neural network model can share feature extraction layers, effectively avoiding overfitting and parameter complexity, and improving the network's generalization ability and robustness. Compared with mainstream network models, it has the advantages of good performance, low complexity, and low parameter count.

[0159] like Figure 6The diagram illustrates a structural block diagram of a neural network-based signal parameter estimation device according to an embodiment of this application. This neural network-based signal parameter estimation device can be applied to computer devices. The neural network-based signal parameter estimation device may include:

[0160] The acquisition module 610 is used to acquire a pre-trained neural network model and a complex modulation signal to be estimated. The neural network model includes at least a convolutional module, a convolutional attention module, a temporal convolutional network module, and a fully connected module.

[0161] The generation module 620 is used to form a two-dimensional matrix from the real and imaginary parts of the complex modulated signal;

[0162] The processing module 630 is used to process the two-dimensional matrix using the convolution module, the channel attention mechanism and spatial attention mechanism in the convolutional attention module, and the dilated causal convolutional layer in the temporal convolutional network module to obtain the feature matrix;

[0163] The estimation module 640 is used to process the feature matrix using the first branch in the fully connected module to obtain the modulation type parameters;

[0164] The estimation module 640 is also used to process the feature matrix using the second branch in the fully connected module to obtain the symbol rate parameters.

[0165] In an optional embodiment, the processing module 630 is further configured to:

[0166] The first convolutional module is used to perform channel transformation on the two-dimensional matrix to obtain the first feature matrix;

[0167] The first feature matrix is ​​enhanced by using the channel attention mechanism and the spatial attention mechanism in the first convolutional attention module to obtain the second feature matrix;

[0168] The second convolutional module is used to extract spatial features from the second feature matrix to obtain the third feature matrix;

[0169] The third feature matrix is ​​enhanced by using the channel attention and spatial attention mechanisms in the second convolutional attention module to obtain the fourth feature matrix.

[0170] Temporal features are extracted from the fourth feature matrix using the dilated causal convolutional layer in the temporal convolutional network module to obtain the feature matrix.

[0171] In an optional embodiment, if the dilated causal convolutional layer includes a convolutional kernel and a dilation rate, then the processing module 630 is further configured to:

[0172] For any element in the fourth feature matrix at any time step, dilation convolution is performed on the element using a convolution kernel and dilation rate to obtain a feature matrix with temporal characteristics.

[0173] In an optional embodiment, the neural network model further includes a pooling module, then the processing module 630 is further configured to:

[0174] The second feature matrix is ​​downsampled using the first pooling module;

[0175] The fourth feature matrix is ​​downsampled using the second pooling module.

[0176] In an optional embodiment, the device further includes a training module for:

[0177] Generate a dataset containing data samples including a two-dimensional matrix of complex modulated signals and labeled modulation type parameters and symbol rate parameters;

[0178] The neural network model is trained using the dataset.

[0179] In an optional embodiment, the training module is further configured to:

[0180] Generate a binary random number sequence based on the sampling frequency and symbol rate parameters;

[0181] The binary random number sequence is modulated according to the modulation type parameter to obtain the first data sequence;

[0182] The waveform of the first data sequence is adjusted to obtain the second data sequence;

[0183] The energy of the second data sequence is normalized to obtain the third data sequence;

[0184] The third data sequence is passed through a channel environment with additive white Gaussian noise to obtain a complex modulated signal;

[0185] Extract the real and imaginary parts from the complex modulated signal, and concatenate the real and imaginary parts into a two-dimensional matrix;

[0186] The data sample is composed of a two-dimensional matrix, modulation type parameters, and symbol rate parameters.

[0187] In an optional embodiment, the training module is further configured to:

[0188] Create a cross-entropy loss function for the first branch and a mean squared error loss function for the second branch;

[0189] When training a neural network model using a dataset, the error of the first branch is calculated using the cross-entropy loss function, and the error of the second branch is calculated using the mean squared error loss function.

[0190] The parameters of the neural network model are adjusted based on the error until the neural network model meets the preset conditions, at which point training stops.

[0191] In summary, the signal parameter estimation device based on neural networks provided in this application improves the overall performance of the neural network model by introducing a convolutional attention mechanism that combines channel attention and spatial attention mechanisms into the neural network model.

[0192] Temporal convolutional network modules in neural network models have fewer parameters and are more suitable for processing long sequence data. They can reduce the complexity and number of parameters of neural network models and improve network performance.

[0193] The neural network model is a multi-task neural network model that can simultaneously perform modulation type estimation and symbol rate estimation. Compared with two corresponding single-task neural network models, the system overhead of the multi-task neural network model is basically the same as that of a single-task neural network model, while the classification and prediction accuracy is slightly improved, and the inference speed of the network is significantly reduced. In addition, the multi-task neural network model can share feature extraction layers, effectively avoiding overfitting and parameter complexity, and improving the network's generalization ability and robustness. Compared with mainstream network models, it has the advantages of good performance, low complexity, and low parameter count.

[0194] 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 neural network-based signal parameter estimation method described above.

[0195] One embodiment of this application provides a computer device, which includes any of the above-described neural network-based signal parameter estimation devices.

[0196] It should be noted that the signal parameter estimation device based on neural networks provided in the above embodiments is only illustrated by the division of the functional modules described above when performing signal parameter estimation based on neural networks. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the signal parameter estimation device based on neural networks can be divided into different functional modules to complete all or part of the functions described above. In addition, the signal parameter estimation device based on neural networks provided in the above embodiments and the signal parameter estimation method embodiments based on neural networks belong to the same concept, and the specific implementation process is detailed in the method embodiments, which will not be repeated here.

[0197] 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.

[0198] 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 signal parameter estimation method based on neural networks, characterized in that, The method includes: Obtain a pre-trained neural network model and a complex modulation signal to be estimated, wherein the neural network model includes at least a convolutional module, a convolutional attention module, a temporal convolutional network module, and a fully connected module; The real and imaginary parts of the complex modulated signal are combined to form a two-dimensional matrix; The two-dimensional matrix is ​​processed using the convolution module, the channel attention mechanism and spatial attention mechanism in the convolutional attention module, and the dilated causal convolutional layer in the temporal convolutional network module to obtain the feature matrix; The feature matrix is ​​processed using the first branch in the fully connected module to obtain the modulation type parameters; The feature matrix is ​​processed using the second branch in the fully connected module to obtain the symbol rate parameter; The process of processing the two-dimensional matrix to obtain a feature matrix using the convolution module, the channel attention mechanism and spatial attention mechanism in the convolutional attention module, and the dilated causal convolutional layer in the temporal convolutional network module includes: performing channel transformation on the two-dimensional matrix using the first convolution module to obtain a first feature matrix; performing feature enhancement on the first feature matrix using the channel attention mechanism and spatial attention mechanism in the first convolutional attention module to obtain a second feature matrix; performing spatial feature extraction on the second feature matrix using the second convolution module to obtain a third feature matrix; performing feature enhancement on the third feature matrix using the channel attention mechanism and spatial attention mechanism in the second convolutional attention module to obtain a fourth feature matrix; and performing temporal feature extraction on the fourth feature matrix using the dilated causal convolutional layer in the temporal convolutional network module to obtain the feature matrix. The channel attention mechanism and the spatial attention mechanism are used before and after the convolution operation to simultaneously consider the correlation between the channel dimension and the spatial dimension, thereby capturing important information in the data.

2. The signal parameter estimation method based on neural networks according to claim 1, characterized in that, The dilated causal convolutional layer includes a convolutional kernel and a dilation rate. The temporal feature extraction of the fourth feature matrix using the dilated causal convolutional layer in the temporal convolutional network module yields the feature matrix, which includes: For any element in the fourth feature matrix at any time, dilated convolution operation is performed on the element using the convolution kernel and the dilation rate to obtain a feature matrix with temporal characteristics.

3. The signal parameter estimation method based on neural networks according to claim 1, characterized in that, If the neural network model further includes a pooling module, then the method further includes: The second feature matrix is ​​downsampled using the first pooling module; The fourth feature matrix is ​​downsampled using the second pooling module.

4. The signal parameter estimation method based on neural networks according to any one of claims 1 to 3, characterized in that, The method further includes: Generate a dataset, wherein the data samples in the dataset include a two-dimensional matrix of complex modulated signals and labeled modulation type parameters and symbol rate parameters; The neural network model is trained using the dataset.

5. The signal parameter estimation method based on neural networks according to claim 4, characterized in that, The generated dataset includes: Generate a binary random number sequence based on the sampling frequency and symbol rate parameters; The binary random number sequence is modulated according to the modulation type parameter to obtain the first data sequence; The waveform of the first data sequence is adjusted to obtain the second data sequence; The second data sequence is energy normalized to obtain the third data sequence; The third data sequence is passed through an additive white Gaussian noise channel environment to obtain a complex modulated signal; Extract the real and imaginary parts from the complex modulated signal, and concatenate the real and imaginary parts into a two-dimensional matrix; The two-dimensional matrix, the modulation type parameter, and the symbol rate parameter are used to form a data sample.

6. The signal parameter estimation method based on neural networks according to claim 4, characterized in that, The step of training the neural network model using the dataset includes: Create a cross-entropy loss function for the first branch and a mean squared error loss function for the second branch; When training the neural network model using the dataset, the error of the first branch is calculated using the cross-entropy loss function, and the error of the second branch is calculated using the mean squared error loss function. The parameters of the neural network model are adjusted according to the error until the neural network model meets the preset conditions, at which point training stops.

7. A signal parameter estimation device based on a neural network, characterized in that, The device includes: The acquisition module is used to acquire a pre-trained neural network model and a complex modulation signal to be estimated. The neural network model includes at least a convolution module, a convolutional attention module, a temporal convolutional network module, and a fully connected module. A generation module is used to form a two-dimensional matrix from the real and imaginary parts of the complex modulated signal; The processing module is used to process the two-dimensional matrix using the convolution module, the channel attention mechanism and spatial attention mechanism in the convolutional attention module, and the dilated causal convolutional layer in the temporal convolutional network module to obtain the feature matrix; An estimation module is used to process the feature matrix using the first branch in the fully connected module to obtain modulation type parameters; The estimation module is further configured to process the feature matrix using the second branch in the fully connected module to obtain the symbol rate parameter; The processing module is further configured to: perform channel transformation on the two-dimensional matrix using a first convolution module to obtain a first feature matrix; enhance the features of the first feature matrix using the channel attention mechanism and spatial attention mechanism in the first convolution attention module to obtain a second feature matrix; extract spatial features from the second feature matrix using a second convolution module to obtain a third feature matrix; enhance the features of the third feature matrix using the channel attention mechanism and spatial attention mechanism in the second convolution attention module to obtain a fourth feature matrix; and extract temporal features from the fourth feature matrix using the dilated causal convolution layer in the temporal convolutional network module to obtain the feature matrix; wherein the channel attention mechanism and the spatial attention mechanism are used before and after the convolution operation, which can simultaneously consider the correlation between the channel dimension and the spatial dimension to capture important information in the data.

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

9. A computer device, characterized in that, The computer device includes: the signal parameter estimation device based on a neural network as described in claim 7.

Images