Attention-based communication signal receiving method, apparatus, device, and medium

Abstract

This application relates to a communication signal receiving method, apparatus, device, and medium based on an attention mechanism. The method includes: constructing and training a multi-source coupled signal separation network consisting of an end-to-end encoder-separation module-decoder. In the network, the encoder first maps short-time overlapping segments in a mixed signal into high-dimensional feature representations, which are then concatenated along the time axis to form a mixed signal feature vector. Next, based on the attention mechanism process modeled by the separation module, a mask of the target signal is extracted from the mixed signal feature vector, and the target signal feature representation is separated using the mask. Finally, the decoder receives the target signal feature representation, reconstructs it segment by segment, and outputs the complete target signal by overlapping and superimposing the reconstructed signal segments. This application enables efficient and accurate reconstruction of target signals in complex electromagnetic environments, achieving accurate reception of communication signals and enhancing the robustness of the communication system.

Description

Technical Field

[0001] This application relates to the field of communication signal reception technology, and in particular to a communication signal reception method, apparatus, device and medium based on an attention mechanism. Background Technology

[0002] Separating target signals from complex electromagnetic environments filled with interference in modern times and accurately and effectively receiving target information is a key challenge facing the field of wireless communication. Blind source separation technology is currently the main means to solve the problem of multi-source signal aliasing. However, traditional blind source separation algorithms extract signal features based on assumptions such as signal independence and linear mixing models. These assumptions cannot meet the requirements of nonlinear distortion, multipath delay, and dynamic interference in complex electromagnetic environments, resulting in poor separation accuracy of complex signals. This poses a severe challenge to modern communication receiving equipment. Summary of the Invention

[0003] Therefore, it is necessary to provide a communication signal receiving method, apparatus, device, and medium based on an attention mechanism to address the aforementioned technical problems.

[0004] A communication signal receiving method based on an attention mechanism, the method comprising:

[0005] Construct a mixed signal dataset for communication scenarios and divide it into training, validation, and test sets;

[0006] Construct a multi-source coupled signal separation network comprising an encoder, a separation module, and a decoder;

[0007] The training and validation sets are input into the multi-source coupled signal separation network. In the network, the encoder first segments the long mixed signal into short-time overlapping segments, maps these segments to high-dimensional feature representations, and then concatenates them along the time axis to form a mixed signal feature vector. Next, based on the attention mechanism modeled by the separation module, a mask for the target signal is extracted from the mixed signal feature vector, and the target signal feature representation is separated from the mixed signal feature vector using the mask. Finally, the decoder receives the target signal feature representation and reconstructs it segment by segment. By overlapping and superimposing the reconstructed signal segments, the complete target signal is output. The network training loss function is calculated by comparing the target signal with the reference signal in the training set. The network parameters are iteratively adjusted to minimize the loss function, and the signal separation performance is evaluated using the validation set until a well-trained multi-source coupled signal separation network is obtained.

[0008] The mixed signals from the test set are input into the trained multi-source coupled signal separation network. The target signal mask obtained by the network training is used to separate and output the complete target signal, thus completing the reception of the target signal.

[0009] In one embodiment, constructing a hybrid signal dataset for a communication scenario includes:

[0010] For modulated signals in communication scenarios, based on open-source modulation datasets, multi-source interference scenarios in complex electromagnetic environments are simulated through signal extension, time-domain aliasing, and noise injection, and a mixed modulation signal dataset containing at least two types of modulation signals is constructed.

[0011] For electromagnetic signals emitted by transmitters in communication scenarios, based on the open-source radio frequency fingerprint dataset, a mixed electromagnetic signal dataset containing electromagnetic signals emitted by at least two transmitters is constructed by simulating multi-source interference scenarios in complex electromagnetic environments through signal extension, time-domain aliasing, and noise injection.

[0012] In one embodiment, the encoder segments the long-segment mixed signal into short-time overlapping segments, maps the short-time overlapping segments into high-dimensional feature representations, and then concatenates them along the time axis to form a mixed signal feature vector, including:

[0013] A mixed signal containing N independent target signals is input into the encoder; the waveform of the mixed signal is represented as follows:

[0014] ;

[0015] in, T is the number of time-domain sampling points of the signal. For the real number field; For the first One target signal, and ;

[0016] In the encoder, the input mixed signal is first divided into several short-time overlapping segments of length L. and ;in, Indicates the index of each short-time overlapping segment and , Indicates the total number of segments;

[0017] Then, through a convolution operation in a 1-D convolutional block of the encoder, each short-time overlapping segment is mapped to an N-dimensional feature space, resulting in a high-dimensional feature representation of each short-time overlapping segment, as follows:

[0018] ;

[0019] in, High-dimensional feature representation for each short-time overlapping segment; Let L be the basis function matrix of the encoder, containing N basis functions of length L. for Transpose of; For activation functions;

[0020] Finally, the high-dimensional feature representations of all short-time overlapping segments in the mixed signal are concatenated along the time axis to form the feature vector of the mixed signal. and the feature vector of the mixed signal Input to the separation module.

[0021] In one embodiment, based on the attention mechanism process modeled by the separation module, a mask for the target signal is extracted from the mixed signal feature vector, and the target signal feature representation in the mixed signal feature vector is separated using the mask, including:

[0022] The separation module consists of a layer normalization module, a first 1×1 convolutional block, repeatedly stacked 1-D convolutional blocks, a PReLU activation function, a second 1×1 convolutional block, and a Sigmoid activation function connected in sequence.

[0023] The feature vector of the mixed signal output by the encoder The input separation module, based on the attention mechanism modeled by the separation module, first processes the mixed signal feature vector through the layer normalization module in the separation module and the first 1×1 convolutional block. Layer normalization and channel compression are performed sequentially to obtain H-dimensional feature vectors with H feature channels, as follows:

[0024] ;

[0025] in, This represents the learnable weight parameters of the first 1×1 convolutional block, used for feature channel compression;

[0026] Then, Feature extraction of the target signal is performed on repeatedly stacked 1-D convolutional blocks; wherein, in the repeatedly stacked 1-D convolutional blocks, the dilation factor is... M 1-D convolutional blocks are stacked repeatedly R times, the inflation factor of each 1-D convolutional block increases exponentially, each 1-D convolutional block ensures that the input is the same through zero-padding, and each 1-D convolutional block is connected based on residual connections and skip connections.

[0027] The target signal features are then extracted from the repeatedly stacked 1-D convolutional blocks. The PReLU activation function, the second 1×1 convolutional block, and the Sigmoid activation function are processed sequentially to output a mixed signal feature vector. The mask for the target signal is represented as:

[0028] ;

[0029] in, Indicates the first in the mixed signal The mask corresponding to each target signal. Indicates the first Learnable weight parameters of a 1-D convolutional block corresponding to a target signal;

[0030] Finally, the feature vector of the mixed signal is separated using the mask of the target signal. The target signal features in the data are represented as follows:

[0031] ;

[0032] in, Indicates the separated first Characteristic representation of a target signal.

[0033] In one embodiment, in repeatedly stacked 1-D convolutional blocks, for the input of the m-th layer 1-D convolutional block... First, the signal is processed by the first 1×1 convolutional block, the first PReLU activation function, and the first global layer normalization module, which are sequentially connected in the m-th 1-D convolutional block. Then, the target signal is extracted by the depthwise separable convolutional block in the m-th 1-D convolutional block, resulting in:

[0034] ;

[0035] in, For the m-th layer 1-D convolutional block, the depth-separable convolutional block is... Characteristics of the output target signal; Here, P represents the depthwise convolution kernel, and P represents the kernel size. For the real number field;

[0036] Then, the second PReLU activation function and the second global layer normalization module, which are sequentially connected in the m-th layer 1-D convolutional block, are used to... After processing, we get:

[0037] ;

[0038] ;

[0039] in, The target signal features are those output by the normalization module of the second global layer in the m-th layer 1-D convolutional block. This indicates global level normalization processing; For any variable, It is a slope parameter;

[0040] for On the one hand, based on the residual connection method, After performing a convolution operation on the second 1×1 convolutional block in the m-th layer 1-D convolutional block, the result is compared with the input of the m-th layer 1-D convolutional block. By concatenating the components, we obtain the input to the (m+1)th layer 1-D convolutional block. , represented as:

[0041] ;

[0042] in, This represents the learnable weight parameters of the 1-D convolutional block in the residual connection;

[0043] On the other hand, based on the skip connection method, After performing a convolution operation on the third 1×1 convolutional block in the m-th layer 1-D convolutional block, the target signal features output by the third 1×1 convolutional block in other repeatedly stacked 1-D convolutional blocks are accumulated to obtain the final target signal features output by the repeatedly stacked 1-D convolutional blocks. , represented as:

[0044] ;

[0045] in, This represents the learnable weight parameters of a 1-D convolutional block in a skip connection.

[0046] In one embodiment, the decoder receives the target signal feature representation, reconstructs it segment by segment, and outputs the complete target signal by overlapping and superimposing the reconstructed signal segments, including:

[0047] For the input decoder's... Feature representation of a target signal The first segment is reconstructed by transposing the 1-D convolutional blocks in the decoder, resulting in the second segment. The reconstructed signal segment corresponding to each target signal , represented as:

[0048] ;

[0049] in, , For the real number field; It is the basis function matrix of the decoder. Each row in the table corresponds to a basis function of length L; The number of target signals from mutually independent sources in the mixed signal;

[0050] For the The target signal is obtained by overlapping and superimposing all adjacent reconstructed signal segments corresponding to each target signal. Specifically, for any two adjacent reconstructed signal segments... and The overlapping process is represented as follows:

[0051] ;

[0052] in, This is the superimposed signal segment. for and The length of the overlapping portion, and It is the weight of the overlapping portion of adjacent reconstructed signal segments obtained through training, and .

[0053] In one embodiment, when the mixed signal includes both I-branch mixed signal and Q-branch mixed signal, the processing procedure of the multi-source coupled signal separation network during the training phase includes:

[0054] In the multi-source coupled signal separation network, the encoder first receives the mixed signal of the I branch and the mixed signal of the Q branch respectively and segments them into short-time overlapping segments. The short-time overlapping segments in the mixed signal of each branch are mapped into high-dimensional features and spliced ​​along the time axis to form the feature vector of the mixed signal of the corresponding branch.

[0055] Then, in the separation module, which consists of a layer normalization module, a first 1×1 convolutional block, a repeatedly stacked 1-D convolutional block, a PReLU activation function, a second 1×1 convolutional block, and a Sigmoid activation function connected in sequence, the target signal features in the I-branch mixed signal feature vector and the Q-branch mixed signal feature vector are extracted in parallel by sharing network parameters, based on the feature extraction link including the layer normalization module, the first 1×1 convolutional block, and the repeatedly stacked 1-D convolutional block. After the target signal features of the two branches are concatenated and combined, they are input into the PReLU activation function, the second 1×1 convolutional block, and the Sigmoid activation function connected in sequence for processing, and a mask of the target signal is jointly trained and output. The target signal feature representation in the mixed signal feature vector of each branch is separated using this mask.

[0056] Finally, the decoder receives the target signal feature representations of the I-branch and the target signal feature representations of the Q-branch from the separation module, and reconstructs them segment by segment. By overlapping and superimposing the reconstructed signal segments of each branch, the complete target signals of the I-branch and Q-branch are output in parallel.

[0057] A communication signal receiving device based on an attention mechanism, the device comprising:

[0058] The dataset construction unit is used to construct a mixed signal dataset in a communication scenario and divide it into a training set, a validation set, and a test set.

[0059] Network building unit, used to construct a multi-source coupled signal separation network including encoder, separation module and decoder;

[0060] The network training unit is used to input the training set and validation set into the multi-source coupled signal separation network. In the network, the encoder first segments the long mixed signal into short-time overlapping segments, and then maps the short-time overlapping segments into high-dimensional feature representations, which are then concatenated along the time axis to form a mixed signal feature vector. Next, according to the attention mechanism modeled by the separation module, the mask of the target signal is extracted from the mixed signal feature vector, and the target signal feature representation in the mixed signal feature vector is separated using the mask. Finally, the decoder receives the target signal feature representation and reconstructs it segment by segment. By overlapping and superimposing the reconstructed signal segments, the complete target signal is output. The network training loss function is calculated by comparing the target signal with the reference signal in the training set. The network parameters are iteratively adjusted step by step with the goal of minimizing the loss function. The signal separation effect of the network is evaluated using the validation set until a well-trained multi-source coupled signal separation network is obtained.

[0061] The network test unit is used to input the mixed signals from the test set into the trained multi-source coupled signal separation network, and use the target signal mask obtained by network training to separate and output the complete target signal, thus completing the reception of the target signal.

[0062] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps:

[0063] Construct a mixed signal dataset for communication scenarios and divide it into training, validation, and test sets;

[0064] Construct a multi-source coupled signal separation network comprising an encoder, a separation module, and a decoder;

[0065] The training and validation sets are input into the multi-source coupled signal separation network. In the network, the encoder first segments the long mixed signal into short-time overlapping segments, maps these segments to high-dimensional feature representations, and then concatenates them along the time axis to form a mixed signal feature vector. Next, based on the attention mechanism modeled by the separation module, a mask for the target signal is extracted from the mixed signal feature vector, and the target signal feature representation is separated from the mixed signal feature vector using the mask. Finally, the decoder receives the target signal feature representation and reconstructs it segment by segment. By overlapping and superimposing the reconstructed signal segments, the complete target signal is output. The network training loss function is calculated by comparing the target signal with the reference signal in the training set. The network parameters are iteratively adjusted to minimize the loss function, and the signal separation performance is evaluated using the validation set until a well-trained multi-source coupled signal separation network is obtained.

[0066] The mixed signals from the test set are input into the trained multi-source coupled signal separation network. The target signal mask obtained by the network training is used to separate and output the complete target signal, thus completing the reception of the target signal.

[0067] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0068] Construct a mixed signal dataset for communication scenarios and divide it into training, validation, and test sets;

[0069] Construct a multi-source coupled signal separation network comprising an encoder, a separation module, and a decoder;

[0070] The training and validation sets are input into the multi-source coupled signal separation network. In the network, the encoder first segments the long mixed signal into short-time overlapping segments, maps these segments to high-dimensional feature representations, and then concatenates them along the time axis to form a mixed signal feature vector. Next, based on the attention mechanism modeled by the separation module, a mask for the target signal is extracted from the mixed signal feature vector, and the target signal feature representation is separated from the mixed signal feature vector using the mask. Finally, the decoder receives the target signal feature representation and reconstructs it segment by segment. By overlapping and superimposing the reconstructed signal segments, the complete target signal is output. The network training loss function is calculated by comparing the target signal with the reference signal in the training set. The network parameters are iteratively adjusted to minimize the loss function, and the signal separation performance is evaluated using the validation set until a well-trained multi-source coupled signal separation network is obtained.

[0071] The mixed signals from the test set are input into the trained multi-source coupled signal separation network. The target signal mask obtained by the network training is used to separate and output the complete target signal, thus completing the reception of the target signal.

[0072] The aforementioned communication signal receiving method, apparatus, device, and medium based on the attention mechanism constructs and trains a multi-source coupled signal separation network consisting of an end-to-end encoder-separation module-decoder. In this network, the encoder maps short-time overlapping segments in the mixed signal into high-dimensional feature representations, which are then concatenated along the time axis to form a mixed signal feature vector. This allows direct manipulation of the time-domain waveform within the encoder, more completely preserving the time-domain characteristics of the input signal and providing optimized feature representations for subsequent separation modules. Furthermore, based on the attention mechanism process modeled by the separation module, even when the separation module itself does not contain an attention mechanism module, the signal separation process is referenced through end-to-end training. The weight mapping of the computational neural network enables the separation module to automatically learn prior information in complex aliasing scenarios and learn a mask from the mixed signal to recover the target signal. After the network training is completed, the mask can be directly used to separate the corresponding target signal, breaking through the dependence of traditional methods on the assumption of signal independence and improving the accuracy and efficiency of target signal separation. Furthermore, by reconstructing and superimposing the segmented feature representation of the target signal output by the separation module received by the decoder, the time-domain waveform of the target signal can be restored, and the reconstructed target signal is continuous and smooth in the time domain, restoring the original features of the target signal to the greatest extent and further improving the accuracy of target signal reception. Attached Figure Description

[0073] Figure 1 This is a flowchart illustrating a communication signal receiving method based on an attention mechanism in one embodiment;

[0074] Figure 2 This is a schematic diagram of the structure of the separation module in one embodiment;

[0075] Figure 3 This is a schematic diagram of the structure of a 1-D convolutional block in a separation module in one embodiment;

[0076] Figure 4 This is a schematic diagram of the data processing process of a multi-source coupled signal separation network when the mixed signal includes I-branch mixed signal and Q-branch mixed signal in one embodiment;

[0077] Figure 5 This is a schematic diagram of the 4ASK-BPSK mixed signal separation result at 16dB in one embodiment;

[0078] Figure 6 This is a schematic diagram of the BPSK-GMSK mixed signal separation result at SNR=20dB in one embodiment;

[0079] Figure 7 This is a schematic diagram of the BPSK-GMSK mixed signal separation result at SNR=16dB in one embodiment;

[0080] Figure 8 This is a schematic diagram of the BPSK-GMSK mixed signal separation result at SNR=12dB in one embodiment;

[0081] Figure 9 This is a schematic diagram of the BPSK-GMSK mixed signal separation result at SNR=8dB in one embodiment;

[0082] Figure 10 This is a schematic diagram of the mixed signal separation result of 3123D7B-3123D7D in one embodiment;

[0083] Figure 11 This is a schematic diagram of the mixed signal separation results of 3123D54-3123D64-3123D58 in one embodiment;

[0084] Figure 12 This is a schematic diagram of the mixed signal separation results of 3123D7B-3124E4A at 8 feet in one embodiment;

[0085] Figure 13 This is a schematic diagram of the mixed signal separation results of 3123D7B-3124E4A at 26 feet in one embodiment;

[0086] Figure 14 This is a schematic diagram of the mixed signal separation results of 3123D7B-3124E4A at 50 feet in one embodiment;

[0087] Figure 15 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0088] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0089] In one embodiment, such as Figure 1 As shown, a communication signal receiving method based on an attention mechanism is provided, including the following steps:

[0090] Step S1: Construct a mixed signal dataset for the communication scenario and divide it into a training set, a validation set, and a test set.

[0091] It should be understood that a mixed signal dataset in a communication scenario contains target signals from at least two independent sources.

[0092] Step S2: Construct a multi-source coupled signal separation network including an encoder, a separation module, and a decoder.

[0093] It should be understood that the multi-source coupled signal separation network mainly adopts an end-to-end three-level architecture of "encoder-separation module-decoder" to realize the feature extraction and signal reconstruction of the target signal. The network uses the mask extracted by the separation module as the parameter of the separation network. In essence, it realizes the extraction of the target component in the mixed signal through a learnable weight allocation mechanism. Through a data-driven method, it learns the feature representation of the target signal from the mixed signal, thereby realizing the reconstruction of the target signal in a complex electromagnetic environment, completing the reception of communication signals, solving the problem of communication signal reception in a complex electromagnetic environment, and enhancing the robustness of the communication system.

[0094] Step S3: Input the training set and validation set into the multi-source coupled signal separation network. In the network, the encoder first segments the long mixed signal into short-time overlapping segments, and then maps the short-time overlapping segments into high-dimensional feature representations, which are then concatenated along the time axis to form a mixed signal feature vector. Next, according to the attention mechanism modeled by the separation module, the mask of the target signal is extracted from the mixed signal feature vector, and the target signal feature representation in the mixed signal feature vector is separated using the mask. Finally, the decoder receives the target signal feature representation and reconstructs it segment by segment. By overlapping and superimposing the reconstructed signal segments, the complete target signal is output. The network training loss function is calculated by comparing the target signal with the reference signal in the training set. With the goal of minimizing the loss function, the network parameters are iteratively adjusted step by step, and the signal separation effect of the network is evaluated using the validation set until a well-trained multi-source coupled signal separation network is obtained.

[0095] Step S4: Input the mixed signals from the test set into the trained multi-source coupled signal separation network, use the target signal mask obtained by the network training to separate and output the complete target signal, and complete the reception of the target signal.

[0096] The aforementioned communication signal receiving method based on the attention mechanism maps short-time overlapping segments in the mixed signal into high-dimensional feature representations by the encoder and splices them along the time axis to form a mixed signal feature vector. This allows direct manipulation of the time-domain waveform within the encoder, more completely preserving the time-domain characteristics of the input signal and providing optimized feature representations for the subsequent separation module. Based on the attention mechanism modeling process of the separation module, even when the separation module itself does not contain an attention mechanism module, end-to-end training parameterizes the separation process into weight mappings of a computational neural network. This enables the separation module to automatically learn prior information in complex aliasing scenarios and learn a mask from the mixed signal that can recover the target signal. After network training, the mask can be directly used to separate the corresponding target signal, breaking through the dependence of traditional methods on the assumption of signal independence and improving the accuracy and efficiency of target signal separation. Furthermore, by receiving the target signal feature representation output by the separation module from the decoder and reconstructing and overlapping segments, the time-domain waveform of the target signal can be restored, ensuring that the reconstructed target signal is continuous and smooth in the time domain, maximizing the restoration of the original features of the target signal and further improving the accuracy of target signal reception.

[0097] In one embodiment, constructing a hybrid signal dataset for a communication scenario includes:

[0098] For modulated signals in communication scenarios, based on open-source modulation datasets, a mixed modulation signal dataset containing at least two types of modulation signals is constructed by simulating multi-source interference scenarios in complex electromagnetic environments through signal extension, time-domain aliasing, and noise injection.

[0099] For electromagnetic signals emitted by transmitters in communication scenarios, based on the open-source radio frequency fingerprint dataset, a mixed electromagnetic signal dataset containing electromagnetic signals emitted by at least two transmitters is constructed by simulating multi-source interference scenarios in complex electromagnetic environments through signal extension, time-domain aliasing, and noise injection.

[0100] In one embodiment, the core function of the encoder is to map short-time overlapping segments of the mixed signal into a high-dimensional feature representation, which replaces the traditional short-time Fourier transform (STFT). Through data-driven convolution operations, the time-domain waveform of the mixed signal is mapped into a high-dimensional feature space, providing an optimized feature representation for the subsequent separation module.

[0101] The encoder's data processing procedure includes: assuming there are N independent target signals, The encoder receives a mixed signal containing N independent target signals; the waveform of the mixed signal is represented as follows:

[0102] ;

[0103] in, T is the number of time-domain sampling points of the signal. For the real number field; For the first One target signal, and .

[0104] To effectively extract features from the mixed signal, the encoder first divides the input mixed signal into several short-time overlapping segments of length L. and This is to ensure the continuity of local features in the mixed signal and avoid the loss of key information due to segmented processing. Indicates the index of each short-time overlapping segment and , This indicates the total number of segments. For example, the overlap rate of short-term overlapping segments is fixed at 50%.

[0105] Then, through the convolution operation of a 1-D convolutional block (one-dimensional convolutional block) in the encoder, each short-term overlapping segment is mapped to an N-dimensional feature space, obtaining a high-dimensional feature representation of each short-term overlapping segment. Here, the encoder operation is expressed as matrix multiplication, specifically as follows:

[0106] ;

[0107] in, High-dimensional feature representation for each short-time overlapping segment; Let be the basis function matrix of the encoder. It contains N basis functions of length L. for The transpose of the basis function matrix End-to-end training and automatic optimization can enhance feature representation capabilities; The activation function ensures the non-negativity of the result. It should be understood that in actual network model implementations, convolutional layers can better handle overlap between short-term overlapping segments, thus achieving a faster training process and better convergence.

[0108] For example, a common activation function is the ReLU function, represented as:

[0109] .

[0110] Finally, the high-dimensional feature representations of all short-time overlapping segments in the mixed signal are concatenated along the time axis to form the feature vector of the mixed signal. and the feature vector of the mixed signal Input to the separation module.

[0111] It should be understood that traditional short-time Fourier transforms suffer from amplitude-phase decoupling during time-frequency domain conversion, potentially leading to signal feature loss and impacting subsequent signal processing. The encoder design in this application avoids this problem by directly manipulating the time-domain waveform, thus preserving the signal's time-domain characteristics more completely. Furthermore, the encoder's basis functions are data-driven optimized, automatically adjusting filter parameters based on the characteristics of the input data to better adapt to complex electromagnetic environments.

[0112] In one embodiment, the separation module aims to extract the mask of the target signal from the mixed signal, a concept similar to estimating the source signal separation matrix in blind source separation. Traditional blind source separation methods primarily approximate the separation matrix through iterative optimization algorithms. However, the separation module in this application models an attention mechanism, parameterizing the separation process as a weight mapping of a computational neural network through end-to-end training. The separation module can automatically learn prior information in complex aliasing scenarios from the input data, overcoming the reliance of traditional methods on the assumption of signal independence.

[0113] like Figure 2 As shown, the separation module consists of a layer normalization module, a first 1×1 convolutional block, repeatedly stacked 1-D convolutional blocks, a PReLU activation function, a second 1×1 convolutional block, and a Sigmoid activation function connected in sequence.

[0114] The data processing procedure of the separation module is as follows: the feature vector of the mixed signal output by the encoder is... The input separation module, based on the attention mechanism modeled by the separation module, first processes the mixed signal feature vector through the layer normalization module in the separation module and the first 1×1 convolutional block. Layer normalization and channel compression are performed sequentially to obtain H-dimensional feature vectors with H feature channels, as follows:

[0115] ;

[0116] in, This represents the learnable weight parameters of the first 1×1 convolutional block, used for feature channel compression. It should be understood that channel compression reduces subsequent computational complexity.

[0117] Then, Feature extraction of the target signal is performed on repeatedly stacked 1-D convolutional blocks; wherein, in the repeatedly stacked 1-D convolutional blocks, the dilation factor is... M 1-D convolutional blocks are repeatedly stacked R times, with the dilation factor of each 1-D convolutional block increasing exponentially to ensure a sufficiently large temporal context window and to ensure long-term dependency of the communication signal. Each 1-D convolutional block ensures that the input is the same through zero-padding, and each 1-D convolutional block is connected based on residual connections and skip connections.

[0118] The target signal features are then extracted from the repeatedly stacked 1-D convolutional blocks. The PReLU activation function, the second 1×1 convolutional block, and the Sigmoid activation function are processed sequentially to output a mixed signal feature vector. The mask for the target signal is represented as:

[0119] ;

[0120] in, Indicates the first in the mixed signal The mask corresponding to each target signal. Indicates the first Learnable weight parameters of a 1-D convolutional block corresponding to a target signal.

[0121] Finally, the feature vector of the mixed signal is separated using the mask of the target signal. The target signal features in the data are represented as follows:

[0122] ;

[0123] in, Indicates the separated first Characteristic representation of a target signal.

[0124] Furthermore, in repeatedly stacked 1-D convolutional blocks, such as Figure 3 As shown, for the input of the m-th layer 1-D convolutional block First, the signal is processed by the first 1×1 convolutional block, the first PReLU activation function, and the first global layer normalization module, which are sequentially connected in the m-th 1-D convolutional block. Then, the target signal is extracted by the depthwise separable convolutional block in the m-th 1-D convolutional block, resulting in:

[0125] ;

[0126] in, For the m-th layer 1-D convolutional block, the depth-separable convolutional block is... Characteristics of the output target signal; Here, P represents the depthwise convolution kernel, and P represents the kernel size. For the real number field. For example, the depthwise convolution kernel size in a 1-D convolution block is all set to... This reduces the number of parameters and the computational load of the model.

[0127] Then, the second PReLU activation function and the second global layer normalization module, which are sequentially connected in the m-th layer 1-D convolutional block, are used to... After processing, we get:

[0128] ;

[0129] ;

[0130] in, The target signal features are those output by the normalization module of the second global layer in the m-th layer 1-D convolutional block. This indicates global level normalization processing; For any variable, Here is a slope parameter. It should be understood that using the PReLU activation function to learn non-linear features can improve the model's fitting ability, and using the gLN method for global layer normalization can improve training stability.

[0131] for On the one hand, based on the residual connection method, After performing a convolution operation on the second 1×1 convolutional block in the m-th layer 1-D convolutional block, the result is compared with the input of the m-th layer 1-D convolutional block. By concatenating the components, we obtain the input to the (m+1)th layer 1-D convolutional block. , represented as:

[0132] ;

[0133] in, This represents the learnable weight parameters of the 1-D convolutional block in the residual connection;

[0134] On the other hand, based on the skip connection method, After performing a convolution operation on the third 1×1 convolutional block in the m-th layer 1-D convolutional block, the target signal features output by the third 1×1 convolutional block in other repeatedly stacked 1-D convolutional blocks are accumulated to obtain the final target signal features output by the repeatedly stacked 1-D convolutional blocks. , represented as:

[0135] ;

[0136] in, This represents the learnable weight parameters of a 1-D convolutional block in a skip connection.

[0137] Figure 3In the left-hand connection, B represents the number of channels in the 1×1 convolution in the residual connection and the second 1×1 convolution block, and Sc represents the number of channels in the 1×1 convolution in the skip connection.

[0138] It should be understood that the design of the 1-D convolutional blocks in the separation module incorporates the concept of depthwise separable convolution, while also utilizing residual connections and skip connections to significantly improve separation performance. Residual connections specifically refer to using the remaining paths of one block as input to the next, preventing training failures in deep networks. Skip connections specifically refer to summing the skip connection paths of all blocks as the final output, aggregating local and global temporal information and enhancing the robustness of mask estimation. Furthermore, replacing the standard convolutions in each convolutional block with depthwise separable convolutions further reduces the number of model parameters.

[0139] It should be understood that the design of the separation module significantly reduces processing latency by employing inter-frame parallel computation. The use of depthwise separable convolution reduces the number of parameters and improves computational efficiency. Furthermore, by combining dilation factor convolution to model long-term context dependencies, signal features are effectively captured. Finally, the introduction of residual connections and skip connections suppresses error accumulation in deep networks, enhances the stability of feature propagation, and improves separation accuracy, thereby achieving efficient and robust target signal extraction in complex electromagnetic environments.

[0140] In one embodiment, the main function of the decoder module is to inversely transform the target signal feature table output by the separation module into a time-domain waveform, thereby reconstructing the target signal. When designing the decoder, it is necessary to ensure the complete reconstruction of the target signal, achieving an end-to-end optimization process through cooperation with the encoder.

[0141] The decoder's data processing procedure includes: for the input decoder's first... Feature representation of a target signal The first segment is reconstructed by performing a transpose convolution operation (i.e., deconvolution) on the 1-D convolutional blocks in the decoder, resulting in the second segment. The reconstructed signal segment corresponding to each target signal , represented as:

[0142] ;

[0143] in, , For the real number field; It is the basis function matrix of the decoder. Each row in the table corresponds to a basis function of length L; This represents the number of target signals from mutually independent sources in the mixed signal.

[0144] Because the input mixed signal is segmented into overlapping short-time overlapping segments (overlap rate 50%) during the encoder stage, the decoder needs to superimpose the overlapping parts after segment-by-segment reconstruction to generate the final complete signal waveform, ensuring the continuity of the time-domain signal. For the first... The target signal is obtained by overlapping and superimposing all adjacent reconstructed signal segments corresponding to each target signal. Specifically, for any two adjacent reconstructed signal segments... and The overlapping process is represented as follows:

[0145] ;

[0146] in, This is the superimposed signal segment. for and The length of the overlapping portion, and It is the weight of the overlapping portion of adjacent reconstructed signal segments obtained through training, and By superimposing signals based on this formula, discontinuities at signal splicing points can be avoided, ensuring that the reconstructed target signal is continuous and smooth in the time domain, and restoring the original characteristics of the target signal to the greatest extent.

[0147] In one embodiment, such as Figure 4 As shown, when the mixed signal includes both I-branch mixed signal and Q-branch mixed signal, the processing procedure of the multi-source coupled signal separation network during the training phase includes:

[0148] In the multi-source coupled signal separation network, the encoder first receives the mixed signal of the I branch and the mixed signal of the Q branch respectively and segments them into short-time overlapping segments. The short-time overlapping segments in the mixed signal of each branch are mapped into high-dimensional features and spliced ​​along the time axis to form the mixed signal feature vector of the corresponding branch.

[0149] Then, in the separation module, which consists of a layer normalization module, a first 1×1 convolutional block, repeatedly stacked 1-D convolutional blocks, a PReLU activation function, a second 1×1 convolutional block, and a Sigmoid activation function connected in sequence, the target signal features in the I-branch mixed signal feature vector and the Q-branch mixed signal feature vector are extracted in parallel by sharing network parameters, based on the feature extraction link including the layer normalization module, the first 1×1 convolutional block, and the repeatedly stacked 1-D convolutional block. After the target signal features of the two branches are concatenated and combined, they are input into the PReLU activation function, the second 1×1 convolutional block, and the Sigmoid activation function connected in sequence for processing, and jointly trained to output a mask of the target signal. The target signal feature representation in the mixed signal feature vector of each branch is then separated using this mask.

[0150] Finally, the decoder receives the target signal feature representations of the I-branch and the target signal feature representations of the Q-branch from the separation module, and reconstructs them segment by segment. By overlapping and superimposing the reconstructed signal segments of each branch, the complete target signals of the I-branch and Q-branch are output in parallel.

[0151] Furthermore, to verify the separation effect of this application on different types of modulated signals aliased in the same frequency band, a hybrid modulated signal dataset was constructed based on the open-source modulation dataset RML2018.01A. Multi-source interference scenarios in complex electromagnetic environments were simulated through signal spreading, time-domain aliasing, and noise injection. Based on this hybrid modulated signal dataset, the separation effect of this application on mixed signals of two modulated signals under high signal-to-noise ratio (SNR), mixed signals of two modulated signals under different SNR, and mixed signals of multiple modulated signals under high SNR was verified.

[0152] First, to verify the signal separation effect of two mixed modulation signals under high signal-to-noise ratio (SNR), two different types of modulation signals with a SNR of 16 dB were selected and mixed to generate seven mixed modulation signal datasets: 4ASK-BPSK (quadruple amplitude shift keying-binary phase shift keying), 4ASK-16APSK (quadruple amplitude shift keying-hexadecimal amplitude and phase joint keying), 4ASK-GMSK (quadruple amplitude shift keying-Gaussian minimum frequency shift keying), BPSK-16APSK (binary phase shift keying-hexadecimal amplitude and phase joint keying), BPSK-GMSK (binary phase shift keying-Gaussian minimum frequency shift keying), 16APSK-GMSK (hexadecimal amplitude and phase joint keying-Gaussian minimum frequency shift keying), and 16QAM-GMSK (hexadecimal quadrature amplitude modulation-Gaussian minimum frequency shift keying). The multi-source coupled signal separation network constructed in this application was trained using the training set of these seven mixed modulation signal datasets. The test set was input into the trained multi-source coupled signal separation network for separation experiments, and compared with traditional blind source separation algorithms ICA (Independent Component Analysis) and NMF (Non-negative Matrix Factorization). The average SI-SNR (Scale-Invariant Signal-to-Noise Ratio) results on the test set are shown in Table 1. The unit of SI-SNR is dB.

[0153] Table 1. Separation results of different types of mixed modulation signals

[0154]

[0155] The experimental results shown in Table 1 demonstrate that this application exhibits strong separation capabilities for mixed signals of any two modulation signals. Compared to traditional blind source separation algorithms, this application shows a significant improvement in the separation performance for mixed communication modulation signals. The visualization of the separation performance after mixing 4ASK and BPSK modulation signals is shown below. Figure 5As shown. By Figure 5 It can be seen that after the 4ASK and BPSK signals with a signal-to-noise ratio of 16dB are mixed, the 4ASK and BPSK signals are successfully separated after processing by this application, and they basically overlap with the clean signal.

[0156] Secondly, to verify the signal separation effect of the two modulated signals under different signal-to-noise ratios, BPSK and GMSK signals with signal-to-noise ratios ranging from 0dB to 20dB were selected and mixed to generate a total of 11 mixed modulated signal datasets. After training, the mixed signals were separated, and the experimental results are shown in Table 2 below.

[0157] Table 2. Separation results of BPSK-GMSK mixed signals under different signal-to-noise ratios

[0158]

[0159] The experimental results shown in Table 2 indicate that the signal-to-noise ratio of the mixed signal is [value missing]. At that time, the SI-SNR of this application for separating mixed signals remained stable above 10 dB, indicating good separation performance and strong adaptability to noise. However, when the signal-to-noise ratio... Subsequently, noise dominated the signal features, causing the model's attention weight allocation for the target signal to become ambiguous. At that time, SI-SNR is still positive, indicating that this application can still partially preserve the target signal waveform under extreme conditions and has a certain noise resistance capability. Figures 6 to 9 A visualization of BPSK and GMSK mixed signals under different signal-to-noise ratios.

[0160] Finally, considering the more complex electromagnetic environment and the greater variety of modulation signals in real-world communication scenarios, to verify the signal separation effect of mixed signals with multiple modulation signals under high signal-to-noise ratio conditions, the following was conducted: Nine types of modulation signals—16QAM (hexadecimal quadrature amplitude modulation), 128QAM (hexadecimal quadrature amplitude modulation), BPSK (binary phase shift keying), QPSK (quadratic phase shift keying), 8PSK (octal phase shift keying), 16PSK (hexadecimal phase shift keying), 8ASK (octal amplitude shift keying), GMSK (Gaussian minimum frequency shift keying), and 16APSK (hexadecimal amplitude and phase combined keying)—were used as target signals. Five modulation signals of different types from the target signals were randomly selected and mixed with the target signals to generate nine datasets, creating a more complex electromagnetic environment. The separation effect of this application on these nine mixed signals is shown in Table 3.

[0161] Table 3. Single-target separation results under various mixed modulation signal conditions

[0162]

[0163] As shown in Table 3, the experimental results demonstrate that this application can still focus on the characteristics of the target signal and successfully separate it even in scenarios with mixed modulated signals. The separation effect is best for GMSK and BPSK signals because these two signals have the most significant characteristics in the time domain. The multi-source coupled signal separation network constructed in this application is better able to focus on the characteristics of these two signals, thus achieving efficient separation of them.

[0164] Furthermore, to verify the separation effect of this application on electromagnetic signals emitted by the same type of transmitters in the same frequency band with aliasing, a mixed electromagnetic signal dataset was constructed based on the open-source ORACLE RF fingerprint dataset. Signal spreading, time-domain aliasing, and noise injection were used to simulate multi-source interference scenarios in complex electromagnetic environments. Based on this dataset, the separation effect of this application on mixed signals emitted by two identical transmitters at close range, mixed signals emitted by multiple identical transmitters at close range, and mixed signals emitted by two identical transmitters at different distances was verified.

[0165] First, to verify the separation effect of mixed signals emitted by two transmitters of the same type at close range, relatively clean signals emitted by the 3123D7B transmitter at a distance of 2 feet from six other different transmitters (3123D7D, 3123D7E, 3123D52, 3123D54, 3123D58, and 3123D64) were combined to form six mixed electromagnetic signal datasets, which were then used to train the multi-source coupled signal separation network. The separation performance of the multi-source coupled signal separation network under different datasets is shown in Table 4.

[0166] Table 4. Separation performance under different mixed datasets

[0167]

[0168] As shown in Table 4, this application can successfully separate mixed signals transmitted by two transmitters of the same type. The separation effect is particularly good when the signals from transmitters 3123D7B and 3123D7E, 3123D52, and 3123D54 are mixed pairwise, with average SI-SNR values ​​of 15.61 dB, 15.54 dB, and 16.25 dB, respectively. Analysis indicates that the better separation effect after mixing with these transmitters is due to the more significant difference in RF fingerprint characteristics between 3123D7B and these three transmitters, thus allowing the multi-source coupled signal separation network to better detect these differences. A visualization of the separation effect of the mixed signals transmitted by transmitters 3123D7B and 3123D7D is shown below. Figure 10 As shown.

[0169] Secondly, to further verify the target signal separation capability of this application under mixed signal conditions emitted by multiple transmitters of the same type, this application uses the signal emitted by transmitter labeled 3123D54 as the target signal and mixes it with the signals emitted by transmitters 3123D58 and 3123D64 to construct a mixed electromagnetic signal dataset, simulating a more complex communication environment. After training, the multi-source coupled signal separation network successfully separated the 3123D54 signal from the mixed signal. The average value of the separated signal compared to the reference signal is [data missing]. The separation results are as follows Figure 11 As shown. By Figure 11 The separation results shown demonstrate that the proposed method can still focus solely on the radiation source fingerprint features of the target signal in complex scenarios where signals from multiple transmitters overlap, thereby achieving efficient separation of the target signal.

[0170] To further investigate the separation effect of the proposed method on mixed signals transmitted from transmitters at different distances, an experiment was conducted, mixing signals transmitted from transmitters 3123D7B and 3124E4A at distances ranging from 2 feet to 50 feet. The separation effect of the multi-source coupled signal separation network on mixed signals at different transmission distances is shown in Table 5.

[0171] Table 5. Results of mixed signal separation of 3123D7B-3124E4A at different transmission distances

[0172]

[0173] As shown in Table 5, the separation results demonstrate that this application exhibits good robustness in separating mixed signals transmitted by the same type of transmitter, and is almost unaffected by changes in transmitter distance. After separating the mixed signals transmitted by the 3123D7B and 3124E4A transmitters, the scale-invariant signal-to-noise ratios of the signals are all above 8dB, exhibiting stable separation performance.

[0174] A visualization of the mixed signal separation effect between the 3123D7B and 3124E4A at transmission distances of 8 feet, 26 feet, and 50 feet, as shown below. Figures 12-14 As shown.

[0175] In one embodiment, an attention-based communication signal receiving device is provided, comprising:

[0176] The dataset construction unit is used to construct a mixed signal dataset in a communication scenario and divide it into a training set, a validation set, and a test set.

[0177] Network building unit, used to construct a multi-source coupled signal separation network including encoder, separation module and decoder;

[0178] The network training unit is used to input the training set and validation set into the multi-source coupled signal separation network. In the network, the encoder first segments the long mixed signal into short-time overlapping segments, and then maps the short-time overlapping segments into high-dimensional feature representations, which are then concatenated along the time axis to form a mixed signal feature vector. Next, according to the attention mechanism modeled by the separation module, the mask of the target signal is extracted from the mixed signal feature vector, and the target signal feature representation in the mixed signal feature vector is separated using the mask. Finally, the decoder receives the target signal feature representation and reconstructs it segment by segment. By overlapping and superimposing the reconstructed signal segments, the complete target signal is output. The network training loss function is calculated by comparing the target signal with the reference signal in the training set. The network parameters are iteratively adjusted step by step with the goal of minimizing the loss function. The signal separation effect of the network is evaluated using the validation set until a well-trained multi-source coupled signal separation network is obtained.

[0179] The network test unit is used to input the mixed signals from the test set into the trained multi-source coupled signal separation network, and use the target signal mask obtained by network training to separate and output the complete target signal, thus completing the reception of the target signal.

[0180] Specific limitations regarding the attention-based communication signal receiving device can be found in the limitations of the attention-based communication signal receiving method described above, and will not be repeated here. Each module in the aforementioned attention-based communication signal receiving device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independently of the processor in a computer device, or stored in software in the memory of a computer device, so that the processor can call and execute the corresponding operations of each module.

[0181] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 15As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a communication signal reception method based on an attention mechanism. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0182] Those skilled in the art will understand that Figure 15 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0183] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to perform the following steps:

[0184] Construct a mixed signal dataset for communication scenarios and divide it into training, validation, and test sets;

[0185] Construct a multi-source coupled signal separation network comprising an encoder, a separation module, and a decoder;

[0186] The training and validation sets are input into the multi-source coupled signal separation network. In the network, the encoder first segments the long mixed signal into short-time overlapping segments, maps these segments to high-dimensional feature representations, and then concatenates them along the time axis to form a mixed signal feature vector. Next, based on the attention mechanism modeled by the separation module, a mask for the target signal is extracted from the mixed signal feature vector, and the target signal feature representation is separated from the mixed signal feature vector using the mask. Finally, the decoder receives the target signal feature representation and reconstructs it segment by segment. By overlapping and superimposing the reconstructed signal segments, the complete target signal is output. The network training loss function is calculated by comparing the target signal with the reference signal in the training set. The network parameters are iteratively adjusted to minimize the loss function, and the signal separation performance is evaluated using the validation set until a well-trained multi-source coupled signal separation network is obtained.

[0187] The mixed signals from the test set are input into the trained multi-source coupled signal separation network. The target signal mask obtained by the network training is used to separate and output the complete target signal, thus completing the reception of the target signal.

[0188] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0189] Construct a mixed signal dataset for communication scenarios and divide it into training, validation, and test sets;

[0190] Construct a multi-source coupled signal separation network comprising an encoder, a separation module, and a decoder;

[0191] The training and validation sets are input into the multi-source coupled signal separation network. In the network, the encoder first segments the long mixed signal into short-time overlapping segments, maps these segments to high-dimensional feature representations, and then concatenates them along the time axis to form a mixed signal feature vector. Next, based on the attention mechanism modeled by the separation module, a mask for the target signal is extracted from the mixed signal feature vector, and the target signal feature representation is separated from the mixed signal feature vector using the mask. Finally, the decoder receives the target signal feature representation and reconstructs it segment by segment. By overlapping and superimposing the reconstructed signal segments, the complete target signal is output. The network training loss function is calculated by comparing the target signal with the reference signal in the training set. The network parameters are iteratively adjusted to minimize the loss function, and the signal separation performance is evaluated using the validation set until a well-trained multi-source coupled signal separation network is obtained.

[0192] The mixed signals from the test set are input into the trained multi-source coupled signal separation network. The target signal mask obtained by the network training is used to separate and output the complete target signal, thus completing the reception of the target signal.

[0193] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0194] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0195] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application.

Claims

1. A communication signal receiving method based on an attention mechanism, characterized in that, The method includes: Construct a mixed signal dataset for communication scenarios and divide it into training, validation, and test sets; Construct a multi-source coupled signal separation network comprising an encoder, a separation module, and a decoder; The training and validation sets are input into the multi-source coupled signal separation network. In the network, the encoder first segments the long-segment mixed signal into short-time overlapping segments, maps these segments to high-dimensional feature representations, and then concatenates them along the time axis to form a mixed signal feature vector. Next, based on the attention mechanism modeled by the separation module, a mask for the target signal is extracted from the mixed signal feature vector, and the target signal feature representation is separated from the mixed signal feature vector using the mask. Finally, the decoder receives the target signal feature representation and reconstructs it segment by segment. By overlapping and superimposing the reconstructed signal segments, a complete target signal is output. The target signal is compared with the reference signal in the training set to calculate the network training loss function. The network parameters are iteratively adjusted to minimize the loss function, and the signal separation performance is evaluated using the validation set until a well-trained multi-source coupled signal separation network is obtained. The mixed signals from the test set are input into the trained multi-source coupled signal separation network. The target signal mask obtained by the network training is used to separate and output the complete target signal, thus completing the reception of the target signal. Specifically, the encoder segments the long-segment mixed signal into short-time overlapping segments, maps these segments to high-dimensional feature representations, and then concatenates them along the time axis to form a mixed signal feature vector, including: A mixed signal containing N mutually independent target signals is input to the encoder; wherein the waveform of the mixed signal is represented as: ； in, T is the number of time-domain sampling points of the signal. For the real number field; For the first One target signal, and ; In the encoder, the input mixed signal is first divided into several short-time overlapping segments of length L. and ;in, Indicates the index of each short-time overlapping segment and , Indicates the total number of segments; Then, through a convolution operation in a 1-D convolutional block of the encoder, each short-time overlapping segment is mapped to an N-dimensional feature space, resulting in a high-dimensional feature representation of each short-time overlapping segment, as follows: ； in, High-dimensional feature representation for each short-time overlapping segment; Let L be the basis function matrix of the encoder, containing N basis functions of length L. for Transpose of; For activation functions; Finally, the high-dimensional feature representations of all short-time overlapping segments in the mixed signal are concatenated along the time axis to form a feature vector of the mixed signal. and the mixed signal feature vector Input to the separation module; Based on the attention mechanism process modeled by the separation module, a mask for the target signal is extracted from the mixed signal feature vector, and the target signal feature representation in the mixed signal feature vector is separated using the mask, including: The separation module consists of a layer normalization module, a first 1×1 convolutional block, repeatedly stacked 1-D convolutional blocks, a PReLU activation function, a second 1×1 convolutional block, and a Sigmoid activation function connected in sequence. The feature vector of the mixed signal output by the encoder The input separation module, based on the attention mechanism process modeled by the separation module, first processes the mixed signal feature vector through the layer normalization module and the first 1×1 convolutional block in the separation module. Layer normalization and channel compression are performed sequentially to obtain H-dimensional feature vectors with H feature channels, as follows: ； in, This represents the learnable weight parameters of the first 1×1 convolutional block, used for feature channel compression; Then, Feature extraction of the target signal is performed on the repeatedly stacked 1-D convolutional blocks; wherein, in the repeatedly stacked 1-D convolutional blocks, the dilation factor is... M 1-D convolutional blocks are stacked repeatedly R times, the inflation factor of each 1-D convolutional block increases exponentially, each 1-D convolutional block ensures that the input is the same through zero-padding, and each 1-D convolutional block is connected based on residual connections and skip connections. The target signal features are then finally output from the repeatedly stacked 1-D convolutional blocks. The PReLU activation function, the second 1×1 convolutional block, and the Sigmoid activation function are sequentially input for processing, and the mixed signal feature vector is output. The mask for the target signal is represented as: ； in, Indicates the first in the mixed signal The mask corresponding to each target signal. Indicates the first Learnable weight parameters of a 1-D convolutional block corresponding to a target signal; Finally, the feature vector of the mixed signal is separated using a mask of the target signal. The target signal features in the data are represented as follows: ； in, Indicates the separated first Characteristic representation of a target signal; In the repeatedly stacked 1-D convolutional blocks, for the input of the m-th layer 1-D convolutional block... First, the signal is processed by the first 1×1 convolutional block, the first PReLU activation function, and the first global layer normalization module, which are sequentially connected in the m-th 1-D convolutional block. Then, the target signal is extracted by the depthwise separable convolutional block in the m-th 1-D convolutional block, resulting in: ； in, For the m-th layer 1-D convolutional block, the depth-separable convolutional block is... Characteristics of the output target signal; Here, P represents the depthwise convolution kernel, and P represents the kernel size. For the real number field; Then, the second PReLU activation function and the second global layer normalization module, which are sequentially connected in the m-th layer 1-D convolutional block, are used to perform the operation on the m-th layer 1-D convolutional block. After processing, we get: ； in, The target signal features are those output by the normalization module of the second global layer in the m-th layer 1-D convolutional block. This indicates global level normalization processing; For any variable, It is a slope parameter; for On the one hand, based on the residual connection method, the aforementioned After performing a convolution operation on the second 1×1 convolutional block in the m-th layer 1-D convolutional block, the result is compared with the input of the m-th layer 1-D convolutional block. By concatenating the components, we obtain the input to the (m+1)th layer 1-D convolutional block. , represented as: ； in, This represents the learnable weight parameters of the 1-D convolutional block in the residual connection; On the other hand, based on the skip connection method, the... After performing a convolution operation on the third 1×1 convolutional block in the m-th layer 1-D convolutional block, the result is accumulated with the target signal features output by the third 1×1 convolutional blocks of other 1-D convolutional blocks in the repeatedly stacked 1-D convolutional blocks to obtain the final target signal features output by the repeatedly stacked 1-D convolutional blocks. , represented as: ； in, This represents the learnable weight parameters of a 1-D convolutional block in a skip connection.

2. The method according to claim 1, characterized in that, Constructing a mixed-signal dataset for communication scenarios, including: For modulated signals in communication scenarios, based on open-source modulation datasets, multi-source interference scenarios in complex electromagnetic environments are simulated through signal extension, time-domain aliasing, and noise injection, and a mixed modulation signal dataset containing at least two types of modulation signals is constructed. For electromagnetic signals emitted by transmitters in communication scenarios, based on the open-source radio frequency fingerprint dataset, a mixed electromagnetic signal dataset containing electromagnetic signals emitted by at least two transmitters is constructed by simulating multi-source interference scenarios in complex electromagnetic environments through signal extension, time-domain aliasing, and noise injection.

3. The method according to claim 1, characterized in that, The decoder receives the target signal feature representation, reconstructs it segment by segment, and outputs the complete target signal by overlapping and superimposing the reconstructed signal segments, including: For the input decoder's... Feature representation of a target signal The first segment is reconstructed by transposing the 1-D convolutional blocks in the decoder, resulting in the second segment. The reconstructed signal segment corresponding to each target signal , represented as: ； in, , For the real number field; It is the basis function matrix of the decoder. Each row in the table corresponds to a basis function of length L; The number of target signals from mutually independent sources in the mixed signal; For the The target signal is obtained by overlapping and superimposing all adjacent reconstructed signal segments corresponding to each target signal. Specifically, for any two adjacent reconstructed signal segments... and The overlapping process is represented as follows: ； in, This is the superimposed signal segment. for and The length of the overlapping portion, and It is the weight of the overlapping portion of adjacent reconstructed signal segments obtained through training, and .

4. The method according to claim 1, characterized in that, When the mixed signal includes both I-branch mixed signal and Q-branch mixed signal, the processing procedure of the multi-source coupled signal separation network during the training phase includes: In the multi-source coupled signal separation network, the encoder first receives the mixed signal of the I branch and the mixed signal of the Q branch respectively and segments them into short-time overlapping segments. The short-time overlapping segments in the mixed signal of each branch are mapped into high-dimensional features and spliced ​​along the time axis to form the feature vector of the mixed signal of the corresponding branch. Then, in the separation module, which consists of a layer normalization module, a first 1×1 convolutional block, repeatedly stacked 1-D convolutional blocks, a PReLU activation function, a second 1×1 convolutional block, and a Sigmoid activation function connected in sequence, the target signal features in the I-branch mixed signal feature vector and the Q-branch mixed signal feature vector are extracted in parallel by sharing network parameters, based on the feature extraction link including the layer normalization module, the first 1×1 convolutional block, and the repeatedly stacked 1-D convolutional block. After the target signal features of the two branches are concatenated and combined, they are input into the PReLU activation function, the second 1×1 convolutional block, and the Sigmoid activation function connected in sequence for processing, and the target signal mask is trained and output. The target signal feature representation in the mixed signal feature vector of each branch is separated using the mask. Finally, the decoder receives the target signal feature representations of the I-branch and the target signal feature representations of the Q-branch output by the separation module, and reconstructs them segment by segment. By overlapping and superimposing the reconstructed signal segments of each branch, the complete target signal of the I-branch and the target signal of the Q-branch are output in parallel.

5. A communication signal receiving device based on an attention mechanism, characterized in that, The device is implemented based on the attention mechanism-based communication signal receiving method according to any one of claims 1-4, comprising: The dataset construction unit is used to construct a mixed signal dataset in a communication scenario and divide it into a training set, a validation set, and a test set. Network building unit, used to construct a multi-source coupled signal separation network including encoder, separation module and decoder; The network training unit is used to input the training set and validation set into the multi-source coupled signal separation network. In the network, the encoder first segments the long mixed signal into short-time overlapping segments, and then maps the short-time overlapping segments into high-dimensional feature representations, which are then concatenated along the time axis to form a mixed signal feature vector. Then, according to the attention mechanism process modeled by the separation module, the mask of the target signal is extracted from the mixed signal feature vector, and the target signal feature representation in the mixed signal feature vector is separated using the mask. Finally, the decoder receives the target signal feature representation and reconstructs it segment by segment. By overlapping and superimposing the reconstructed signal segments, the complete target signal is output. The network training loss function is calculated by comparing the target signal with the reference signal in the training set. The network parameters are iteratively adjusted step by step with the goal of minimizing the loss function. The signal separation effect of the network is evaluated using the validation set until a well-trained multi-source coupled signal separation network is obtained. The network testing unit is used to input the mixed signals from the test set into the trained multi-source coupled signal separation network, and use the mask of the target signal obtained by network training to separate and output the complete target signal, thereby completing the reception of the target signal.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

Images