A wireless signal separation method, device, computer equipment and storage medium

Abstract

The application relates to a wireless signal separation method and device, computer equipment and a storage medium, and belongs to the technical field of wireless communication. The method comprises the following steps: obtaining an aliasing signal in a co-frequency interference scene; constructing an orthogonal decoupling network model combined with an information bottleneck constraint; adopting a multi-task learning strategy during model training; and jointly optimizing an encoder, an orthogonal decoupling module and a parallel decoder connected in sequence in the model under the information bottleneck constraint; the training total loss function comprises a scale invariant signal-to-interference-and-noise ratio loss, a variational information bottleneck loss and an orthogonality loss; the aliasing signal is input into the trained model, latent feature representation containing a target signal and redundant interference noise is obtained through the encoder, the orthogonal decoupling module is used for explicit decomposition into target signal related features and orthogonal redundant features, and finally wireless signal separation is completed through the parallel decoder. The method can realize stable target signal separation in a complex electromagnetic environment.

Description

Technical Field

[0001] This application relates to the field of wireless communication technology, and in particular to a wireless signal separation method, apparatus, computer device, and storage medium. Background Technology

[0002] With the explosive growth in demand for wireless communication services and the increasing scarcity of spectrum resources, co-channel interference (CCI) has become one of the main bottlenecks restricting the reliability and capacity of modern communication systems. In complex electromagnetic environments, the target signal (SOI) at the receiver is often severely interfered with. Therefore, how to accurately separate the SOI from aliased signals has become a core technical problem that urgently needs to be solved in the field of wireless communication.

[0003] Existing Target Signal Separation (TSS) methods are mainly divided into two categories: traditional signal processing methods and deep learning methods. Traditional methods are mostly based on filtering, subspace projection, or blind source separation techniques. For example, interference is suppressed by designing frequency domain or spatial domain filters, or blind separation is performed using the assumption of signal statistical independence. However, these methods usually have high computational complexity, and their separation performance deteriorates significantly in real-world scenarios such as severe overlap between the target signal and interference spectra, time-varying channels, or non-independent interference sources, making it difficult to meet the robustness requirements in complex electromagnetic environments.

[0004] In recent years, deep learning technology has made significant progress in the field of signal separation due to its powerful data-driven modeling capabilities. Existing work typically employs end-to-end supervised learning frameworks, achieving source signal separation by constructing encoder-decoder structures or time-frequency domain segmentation networks. Some studies have drawn on successful experiences in speech separation, designing complex-valued neural networks, recurrent neural networks, or convolutional recurrent network structures suitable for communication signals. Although these methods improve separation performance to some extent, existing network models lack selective preservation of SOI-related information and explicit suppression mechanisms for redundant interference during feature learning. This results in insufficient generalization ability of the models when facing dynamically changing interference conditions, thus limiting their separation performance. Summary of the Invention

[0005] Therefore, it is necessary to provide a wireless signal separation method, apparatus, computer equipment, and storage medium to address the above-mentioned technical problems, which can achieve robust TSS in complex electromagnetic environments and enhance the model's robustness and adaptability to interference by optimizing the information structure and decoupling capability of feature representation.

[0006] A method for separating wireless signals, the method comprising:

[0007] Acquire aliased signals in co-channel interference scenarios; An orthogonal decoupling network model combining information bottleneck constraints is constructed. During model training, a multi-task learning strategy is adopted to jointly optimize the encoder, orthogonal decoupling module and parallel decoder connected sequentially within the model under the information bottleneck constraint. The total loss function of model training includes scale-invariant signal-to-interference-plus-noise ratio loss, variational information bottleneck loss and orthogonality loss. The aliased signal is input into the trained model. First, the encoder maps the aliased signal to a high-dimensional feature space to obtain a latent feature representation containing the target signal and redundant interference noise. Then, the orthogonal decoupling module performs an orthogonal mapping operation on the latent feature representation to explicitly decompose it into target signal related features and orthogonal redundant features. Finally, the parallel decoder reconstructs the time-domain waveform of the target signal based on the target signal related features to complete the wireless signal separation.

[0008] In one embodiment, aliasing signals in a co-channel interference scenario Represented as: ; in, and The first The transmitter and the first The transmitter-to-receiver channel attenuation coefficient; For the first The source signal from each transmitter is set as the target signal to be separated. SOI ; For the first The source signal of each transmitter, for the target signal SOI Constitutes co-channel interference CCI ; , The total number of transmitters; It is Gaussian white noise. Indicates a Gaussian distribution. Let be the variance of the Gaussian distribution.

[0009] In one embodiment, the encoder employs a network structure that combines multi-scale parallel feature extraction branches, and the processing logic is as follows: The encoder input is of size The aliased signal is first processed by a convolution kernel with a size of 11 and an output channel number of... A one-dimensional convolutional layer yields a size of... The feature map; then set three parallel branches with kernel sizes of 3, 5 and 7 respectively, each branch consisting of... It consists of several deep residual shrinkage modules. The first deep residual shrinkage module in each branch reduces the number of channels from... Expand to The following The number of output channels for each depth residual shrinkage module remains at [value]. Size is The feature maps are processed by three parallel branches to obtain three values ​​of size 1. The parallel outputs are stacked along the output channel dimension to obtain a size of [size missing]. The fused feature map is passed sequentially through two one-dimensional convolutional layers, each with a kernel size of 3, and the number of output channels is set to... and The final latent feature representation is obtained. As input for subsequent modules; among which, For the set of real numbers, The batch size for encoder input.

[0010] In one embodiment, the processing logic of the orthogonal decoupling module is as follows: Introducing the projection matrix Represent the latent features output by the encoder The decomposition is divided into two complementary orthogonal subspaces, one of which preserves the correlation features of the target signal. The other subspace corresponds to redundant interference and noise components, represented by orthogonal redundancy features. ;in, It is an identity matrix, and .

[0011] In one embodiment, the decoder includes two independent decoding branches: a target signal decoding branch for reconstructing the target signal and an interference reconstruction branch for reconstructing residual interference and noise components. The target signal decoding branch takes the relevant features of the target signal output from the orthogonal decoupling module as input, and first uses an output channel with a number of... A one-dimensional convolutional layer with a kernel size of 5 performs initial decoding of the input features, followed by a convolutional layer consisting of... Multiple output heads composed of parallel fully connected layers reconstruct the signal during the model training phase. Each output head undergoes synchronous end-to-end training, and the loss function for the target signal decoding branch training is the average of the losses from each output head; during the inference phase, a random selection strategy is adopted, starting from... One of the output heads is randomly selected for approximate reconstruction of the target signal, and the output is the first one. Estimation of target signals ; The interference reconstruction branch takes the orthogonal redundancy features output by the orthogonal decoupling module as input and outputs residual components. The interference reconstruction branch uses an output channel with a number of... A one-dimensional convolutional layer with a kernel size of 5 and a single output head is implemented.

[0012] In one embodiment, the encoder, orthogonal decoupling module, and parallel decoder connected sequentially within the model are jointly optimized under the information bottleneck constraint, wherein the optimization objective of the information bottleneck constraint is: ; in, The latent feature representation of the encoder output, The input model contains aliased signals. To balance the parameters, For mutual information; The first term is a compression term, used to control the model to discard redundant information; the second term is used to maximize the mutual information with the target signal in order to retain the key features related to the target signal. To achieve the optimization goal constrained by information bottlenecks The variational information bottleneck principle is used to address the aliasing signals in the input model. With latent feature representation Constraints are imposed on mutual information between the two sides to construct variational information bottleneck loss. , represented as: ; in, To reconstruct the loss, For aliasing signals With target signal The mathematical expectation of the joint distribution; A standard Gaussian prior is used to guide compression. Indicates a Gaussian distribution. Represents the identity matrix; The latent distribution of the encoder output; combined with the structural design of two independent decoding branches in the parallel decoder, the reconstruction loss... It consists of two parts, one of which is the mean square error loss of the target signal decoding branch used for target signal reconstruction. , The first branch of the target signal decoding Output estimation of each output head For the corresponding number The source signal of a transmitter, The first is the total number of output heads; the second is the residual interference and noise constraint loss of the interference reconstruction branch. ,in, To interfere with the residual components of the reconstructed branch output; The KL divergence term is used to regularize the latent distribution, making it approximate the standard Gaussian prior, and is expressed as: ; in, and The potential distribution of the encoder output is respectively In the Mean and variance in dimension This represents the total number of dimensions.

[0013] In one embodiment, the total loss function, formed by the combination of scale-invariant signal-to-interference-plus-noise ratio loss, variational information bottleneck loss, and orthogonality loss, is expressed as: ; in, and These are hyperparameters used to balance the variational information bottleneck loss. and orthogonality loss Contribution; orthogonality loss , The target signal related features decomposed by the orthogonal decoupling module are indicated by the superscript. T Indicates transpose. The orthogonal redundant features are derived from the orthogonal decoupling module. The Frobenius norm is used to minimize the orthogonality loss, ensuring that the relevant features of the target signal and the orthogonal redundant features are as orthogonal as possible in the latent space; scale-invariant signal-to-interference-plus-noise ratio loss. Used to evaluate the quality of target signal estimation, denoted as: ; in, It is the first Output estimation of each output head In the corresponding source signal Projected components in the direction; This represents the noise component.

[0014] A wireless signal separation device, the device comprising: The signal acquisition unit is used to acquire aliased signals in co-frequency interference scenarios. The model building and training unit is used to build an orthogonal decoupling network model that incorporates information bottleneck constraints. During model training, a multi-task learning strategy is adopted to jointly optimize the encoder, orthogonal decoupling module, and parallel decoder connected sequentially within the model under the information bottleneck constraint. The total loss function of model training includes scale-invariant signal-to-interference-plus-noise ratio loss, variational information bottleneck loss, and orthogonality loss. The signal separation unit is used to input the aliased signal into the trained model. First, the encoder maps the aliased signal to a high-dimensional feature space to obtain a latent feature representation containing the target signal and redundant interference noise. Then, the orthogonal decoupling module performs an orthogonal mapping operation on the latent feature representation to explicitly decompose it into target signal related features and orthogonal redundant features. Finally, the parallel decoder reconstructs the time-domain waveform of the target signal based on the target signal related features to complete the wireless signal separation.

[0015] 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: Acquire aliased signals in co-channel interference scenarios; An orthogonal decoupling network model combining information bottleneck constraints is constructed. During model training, a multi-task learning strategy is adopted to jointly optimize the encoder, orthogonal decoupling module and parallel decoder connected sequentially within the model under the information bottleneck constraint. The total loss function of model training includes scale-invariant signal-to-interference-plus-noise ratio loss, variational information bottleneck loss and orthogonality loss. The aliased signal is input into the trained model. First, the encoder maps the aliased signal to a high-dimensional feature space to obtain a latent feature representation containing the target signal and redundant interference noise. Then, the orthogonal decoupling module performs an orthogonal mapping operation on the latent feature representation to explicitly decompose it into target signal related features and orthogonal redundant features. Finally, the parallel decoder reconstructs the time-domain waveform of the target signal based on the target signal related features to complete the wireless signal separation.

[0016] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Acquire aliased signals in co-channel interference scenarios; An orthogonal decoupling network model combining information bottleneck constraints is constructed. During model training, a multi-task learning strategy is adopted to jointly optimize the encoder, orthogonal decoupling module and parallel decoder connected sequentially within the model under the information bottleneck constraint. The total loss function of model training includes scale-invariant signal-to-interference-plus-noise ratio loss, variational information bottleneck loss and orthogonality loss. The aliased signal is input into the trained model. First, the encoder maps the aliased signal to a high-dimensional feature space to obtain a latent feature representation containing the target signal and redundant interference noise. Then, the orthogonal decoupling module performs an orthogonal mapping operation on the latent feature representation to explicitly decompose it into target signal related features and orthogonal redundant features. Finally, the parallel decoder reconstructs the time-domain waveform of the target signal based on the target signal related features to complete the wireless signal separation.

[0017] The aforementioned wireless signal separation method, apparatus, computer equipment, and storage medium construct an orthogonal decoupling network in an end-to-end manner, incorporating information bottleneck constraints. This effectively decouples and filters out information irrelevant to target signal reconstruction, retains the most critical features for target signal separation, and enhances the accuracy and robustness of target signal separation. Through collaborative optimization of multiple task loss functions, including scale-invariant signal-to-interference-plus-noise ratio loss, variational information bottleneck loss, and orthogonality loss, it simultaneously considers target signal estimation quality, key feature preservation, and feature orthogonality. Ultimately, it achieves adaptive extraction and accurate reconstruction of target signals from aliased signals without explicit interference modeling. This application effectively addresses interference factors of varying intensities, types, and multiple overlap modes, exhibiting excellent anti-interference capabilities and generalization performance, providing effective technical support for achieving highly reliable target signal separation in complex electromagnetic environments. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of a co-channel interference scenario in one embodiment; Figure 2 This is a schematic diagram of the overall architecture of an orthogonal decoupling network model in one embodiment; Figure 3 This is a schematic diagram of the deep residual shrinkage module in one embodiment; Figure 4 A schematic diagram illustrating how the scale-invariant signal-to-interference-plus-noise ratio loss changes with training epochs for different parameter settings; Figure 5 This is a schematic diagram illustrating the training and average performance under different model settings in one embodiment; wherein Figure 5 (a) shows the scale-invariant signal-to-interference-plus-noise ratio loss as a function of training epochs. Figure 5 (b) Average performance under different signal-to-interference ratios; Figure 6 This describes the model performance under different interference factors in one embodiment; where... Figure 6 (a) shows the scale-invariant signal-to-interference-plus-noise ratio (SI-SINR) as a function of the signal-to-interference-plus-noise ratio. Figure 6 (b) shows the change of the correlation coefficient with the information-interference ratio. Figure 6 (c) shows the variation of SI-SINR with frequency overlap rate. Figure 6 (d) shows the change of the correlation coefficient with frequency overlap rate. Figure 6 (e) shows the variation of SI-SINR with the number of interference sources. Figure 6 (f) shows how the correlation coefficient changes with the number of interference sources. Figure 6 (g) shows the change in SI-SINR overlap rate over time. Figure 6 (h) shows the change in the correlation coefficient over time; Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0019] 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.

[0020] In one embodiment, a wireless signal separation method is provided, comprising the following steps: Step 1: Obtain the aliasing signal under co-frequency interference scenario.

[0021] Step 2: Construct an orthogonal decoupling network model that incorporates information bottleneck constraints. During model training, a multi-task learning strategy is adopted to jointly optimize the encoder, orthogonal decoupling module, and parallel decoder connected sequentially within the model under the information bottleneck constraint. The total loss function of model training includes scale-invariant signal-to-interference-plus-noise ratio loss, variational information bottleneck loss, and orthogonality loss.

[0022] Step 3: Input the aliased signal into the trained model. First, the encoder maps the aliased signal to a high-dimensional feature space to obtain a latent feature representation containing the target signal and redundant interference noise. Then, the orthogonal decoupling module performs an orthogonal mapping operation on the latent feature representation to explicitly decompose it into target signal related features and orthogonal redundant features. Finally, the parallel decoder reconstructs the time-domain waveform of the target signal based on the target signal related features to complete the wireless signal separation.

[0023] The aforementioned wireless signal separation method employs an end-to-end architecture to construct an orthogonal decoupling network incorporating information bottleneck constraints. This efficiently decouples and eliminates redundant information that does not contribute to target signal reconstruction, accurately preserving the core features required for signal separation and significantly improving the accuracy and robustness against interference. Simultaneously, collaborative optimization is achieved through a multi-task joint loss function comprised of scale-invariant signal-to-interference-plus-noise ratio loss, variational information bottleneck loss, and orthogonality loss. This balances improving signal estimation quality, preserving key features, and ensuring feature orthogonality. This method does not require explicit modeling of interference; it adaptively extracts and accurately reconstructs the target component in aliased signals. It effectively adapts to interference scenarios of varying intensities, types, and superposition forms, exhibiting excellent anti-interference performance and generalization ability. This provides a reliable technical solution for highly robust and high-fidelity target signal separation in complex electromagnetic environments.

[0024] In this embodiment, a scenario of co-channel interference involving multiple transmitters is considered, such as... Figure 1As shown. The source signals from each transmitter superimpose during propagation and are received by a single receiver. Assuming both the transmitter and receiver are equipped with a single antenna, the... The source signal of each transmitter is represented as ,in , The total number of transmitters. This refers to the signal duration. The aliased signal received by the receiver. The result of the superposition of all transmitted signals can be expressed as: ; Considering that the source signals from each transmitter constitute co-channel interference CCI Therefore, aliasing signals in co-channel interference scenarios Further decomposed into the form of target signal and interference plus noise, it can be expressed as: ; in, and The first The transmitter and the first The transmitter-to-receiver channel attenuation coefficient; For the first The source signal from each transmitter is set as the target signal to be separated. SOI ; For the first The source signal of each transmitter, for the target signal SOI Constitutes co-channel interference CCI ; , The total number of transmitters; It is Gaussian white noise. Indicates a Gaussian distribution. Let be the variance of the Gaussian distribution. At this point, the TSS task aims to extract the specified target signal from the aliased observations, ultimately obtaining its estimate. : ; in This represents the function mapping for the TSS task. Given the high-dimensional function fitting capability and end-to-end learning mechanism of deep learning models, this embodiment designs an orthogonal decoupled network model, OD-IB-Net, which incorporates information bottleneck constraints, to implement the function mapping for the TSS task. The model structure is as follows... Figure 2 As shown, it mainly consists of three parts: an encoder, an orthogonal decoupling module, and a parallel decoder, which will be described in detail below.

[0025] To effectively extract the latent representation of the target signal from aliased signals, a network combining multi-scale parallel feature extraction branches is employed as the encoder structure. Since one-dimensional convolution has lower computational complexity and better performance than two-dimensional convolution when processing sequential data, the encoder uses one-dimensional convolutional layers as its basic component. The encoder's processing logic is as follows: The encoder input is of size The aliased signal is first processed by a convolution kernel with a size of 11 and an output channel number of... A one-dimensional convolutional layer yields a size of... Feature maps; then set the convolution kernel size. There are three parallel branches, numbered 3, 5, and 7 respectively, each branch consisting of... It consists of several Deep Residual Shrinkage Modules (DRSNs), with the first DRSN in each branch reducing the number of channels from... Expand to The following The number of output channels for each depth residual shrinkage module remains at [value]. Size is The feature maps are processed by three parallel branches to obtain three values ​​of size 1. The parallel outputs are stacked along the output channel dimension to obtain a size of [size missing]. The fused feature map is passed sequentially through two one-dimensional convolutional layers, each with a kernel size of 3, and the number of output channels is set to... and The final latent feature representation is obtained. As input for subsequent modules; among which, For the set of real numbers, The batch size for encoder input. The DRSN module structure is as follows: Figure 3 As shown, this module introduces a soft thresholding function based on the classic residual block to eliminate the influence of noise. The input of this module passes through two one-dimensional convolutional layers, and then is processed by the soft thresholding function before being added to the original input to obtain the module output. The soft thresholding function sets features with absolute values ​​below a set threshold to zero and shrinks other features towards zero. Its mathematical expression is: ; in, and These are the input features and the output features, respectively. This is an automatically set threshold. Specifically, when calculating the threshold, the value is... The feature map is processed by absolute value calculation and global average pooling to obtain a one-dimensional vector, which is then passed through two fully connected layers to generate channel weight vectors. And the weight values ​​are activated by the Sigmoid activation function. Constrain to the range of 0-1: ; Combine it with input features By multiplying the average values ​​of specific channels, the threshold for each channel can be obtained. and The first The weight values ​​before and after the constraints of each channel.

[0026] By explicitly applying orthogonal constraints, the orthogonal decoupling module can decouple target signal-related information from irrelevant components in the latent space. The processing logic of the orthogonal decoupling module is as follows: introduce a projection matrix. Represent the latent features output by the encoder The decomposition is divided into two complementary orthogonal subspaces, one of which preserves the correlation features of the target signal. The other subspace corresponds to redundant interference and noise components, represented by orthogonal redundancy features. ;in, It is an identity matrix, and .

[0027] The decoder comprises two independent decoding branches: a target signal decoding branch for reconstructing the target signal and an interference reconstruction branch for reconstructing residual interference and noise components. The target signal decoding branch takes the relevant characteristics of the target signal output from the orthogonal decoupling module as input, initially using a single output channel with... A one-dimensional convolutional layer with a kernel size of 5 performs initial decoding of the input features, followed by a convolutional layer consisting of... Multiple output heads composed of parallel fully connected layers reconstruct the signal during the model training phase. Each output head undergoes synchronous end-to-end training, and the loss function for the target signal decoding branch training is the average of the losses from each output head; during the inference phase, a random selection strategy is adopted, starting from... One of the output heads is randomly selected for approximate reconstruction of the target signal, and the output is the first one. Estimation of target signals The interference reconstruction branch takes the orthogonal redundancy features output by the orthogonal decoupling module as input and outputs residual components. The interference reconstruction branch uses an output channel with a number of... A one-dimensional convolutional layer with a kernel size of 5 and a single output head is implemented.

[0028] In this embodiment, the encoding, decoupling, and decoding stages are jointly trained end-to-end within the same network framework. Under the constraint of an information bottleneck, the encoder, orthogonal decoupling module, and parallel decoder connected sequentially within the model are jointly optimized. This enables the network to accurately estimate the target signal while suppressing interference, ultimately achieving robust separation of the target signal in complex electromagnetic environments. The information bottleneck constraint addresses the problem of the neural network potentially retaining redundant features unrelated to the target signal during forward propagation. The optimization objective of the information bottleneck constraint is: ; in, The latent feature representation of the encoder output, The input model contains aliased signals. To balance the parameters, For mutual information; The first term is a compression term, used to control the model to discard redundant information; the second term is used to maximize the mutual information with the target signal in order to retain the key features related to the target signal. To achieve the optimization goal constrained by information bottlenecks The variational information bottleneck (VIB) principle is used to address the aliasing signals in the input model. With latent feature representation Constraints are imposed on the mutual information between them, and the potential distribution of the encoder output is assumed to be... A Gaussian distribution with diagonal covariance is used for modeling. A variational information bottleneck loss is constructed. , represented as: ; in, To reconstruct the loss, For aliasing signals With target signal The mathematical expectation of the joint distribution; A standard Gaussian prior is used to guide compression. Indicates a Gaussian distribution. Represents the identity matrix; combining the structural design of two independent decoding branches in the parallel decoder, the reconstruction loss... It consists of two parts, one of which is the mean squared error (MSE) loss of the target signal decoding branch used for target signal reconstruction. , The first branch of the target signal decoding Output estimation of each output head For the corresponding number The source signal of a transmitter, The first is the total number of output heads; the second is the residual interference and noise constraint loss of the interference reconstruction branch. Since the interference signal is unknown and does not need to be recovered, only the residual components need to be preserved. and Orthogonality is used to ensure the effectiveness of feature decoupling. Specifically, the calculation... exist The projected energy on the surface satisfies the orthogonality condition when this energy is zero. Therefore, Defined as: ; The KL divergence term is used to regularize the latent distribution, making it approximate the standard Gaussian prior, and is expressed as: ; in, and The potential distribution of the encoder output is respectively In the Mean and variance in dimension The total number of dimensions. Based on this variational information bottleneck loss, the network can retain features highly correlated with the target signal while compressing redundant information, thereby improving the robustness and generalization ability of the TSS.

[0029] In this embodiment, to ensure the separation quality of the target signal under the combined effects of information bottleneck constraints and orthogonality constraints, this application employs a multi-task learning strategy to jointly train the network. The total loss function, formed by the combination of scale-invariant signal-to-interference-plus-noise ratio (SI-SINR) loss, variational information bottleneck loss, and orthogonality loss, is expressed as follows: ; in, and These are hyperparameters used to balance the variational information bottleneck loss. and orthogonality loss Contribution; orthogonality loss , The target signal related features decomposed by the orthogonal decoupling module are indicated by the superscript. T Indicates transpose. The orthogonal redundant features are derived from the orthogonal decoupling module. The Frobenius norm is used to minimize the orthogonality loss, ensuring that the relevant features of the target signal and the orthogonal redundant features are as orthogonal as possible in the latent space; scale-invariant signal-to-interference-plus-noise ratio loss. Used to evaluate the quality of target signal estimation, denoted as: ; in, It is the first Output estimation of each output head In the corresponding source signal Projected components in the direction; The noise component is represented by the aforementioned multi-task loss function. Through this function, the model can optimize the target signal separation accuracy while simultaneously considering the orthogonality constraints of feature decoupling and the compression constraints of information bottlenecks, achieving end-to-end joint optimization training of the network.

[0030] Furthermore, in order to thoroughly evaluate the performance of this method, a series of evaluation indicators were selected to comprehensively measure the similarity and quality difference between the separated signal and the original target signal from multiple perspectives, such as signal reconstruction accuracy, waveform fidelity, and robustness to different interference factors. This study explored and verified the actual effect and adaptability of this method in complex electromagnetic environments such as simultaneous and co-frequency interference.

[0031] First, let's introduce the evaluation indicators, which are as follows: (1) Scale-invariant signal-to-interference-plus-noise ratio (SI-SINR): SI-SINR measures the similarity between the separated signal and the original target signal. Its core characteristic is its insensitivity to signal amplitude scaling, objectively reflecting the fidelity of waveform reconstruction. It is a key indicator for evaluating the quality of target signal separation. The calculation formula is: ; in, Indicates the inner product. express Norm. First, the estimated signal... Projected onto real signal In terms of direction, the target component is obtained. Next, the noise components are calculated. Finally, calculate the logarithm of the ratio of the energy of these two components. The larger the SI-SINR value, the closer the waveform of the separated target signal is to the original signal, and the higher the separation quality.

[0032] (2) Correlation coefficient: The correlation coefficient is a measure of and The statistic of linear correlation between two signals reflects the degree of similarity in their waveforms. The calculation formula is: ; in, yes and covariance, and They are and The variance of the correlation coefficient. The range of values ​​for the correlation coefficient is within... Between, its physical meaning is: when When the two signals are perfectly positively correlated, their waveforms are completely identical; when... When the two signals are completely negatively correlated, their waveforms are completely opposite; when... This indicates that the two signals are uncorrelated and have no linear relationship. The closer the correlation coefficient is to 1, the more similar the waveform of the separated signal is to the target signal, and the better the separation effect.

[0033] To simulate the randomness and diversity of interference signals under complex electromagnetic environments, this application constructs a dataset covering various interference signal parameter configurations. The interference signals include BPSK (Binary Phase Shift Keying), QPSK (Quadrilateral Phase Shift Keying), 8PSK (Octal Phase Shift Keying), 16QAM (16-QAM), 64QAM (64-QAM), 256QAM (256-QAM), and 4PAM (Quadrilateral Pulse Amplitude Modulation). Specific parameter settings are shown in Table 1.

[0034] Table 1 Parameter settings for the dataset

[0035] in, Indicates the modulation type; Indicates signal duration, including target signal length. With the length of the interference signal ; Indicates frequency offset; The symbol rate is the ratio of the interference signal to the symbol rate of the target signal, where the target signal symbol rate is 512 Kbps; SIR and SNR represent the signal-to-interference ratio and signal-to-noise ratio, respectively. Sampling rate; This indicates the number of interfering signals contained in a single aliased sample; The starting time of the interference signal in the aliased samples is determined by... and Determined jointly.

[0036] The experiment focuses on combinations of controllable parameters A total of 1320 different parameter combinations were generated, with 100 samples generated for each combination. These samples were randomly divided into training, validation, and test sets in a ratio of 8:1:1.

[0037] The simulation experiment was implemented based on the PyTorch 1.12 deep learning framework. The number of training epochs was set to 200, the batch size to 64, and the learning rate to [value missing]. The Adam optimizer is used for optimization. An early stopping mechanism is introduced to prevent overfitting; training stops when the validation loss does not decrease for 10 consecutive epochs. The number of encoder output channels... Set to 64, the number of depth residual shrinkage modules. Set to 3, the number of multi-head outputs of the target signal decoder. Set it to 8. Loss trade-off coefficient It increases linearly with the number of training rounds.

[0038] To verify key hyperparameters and Sensitivity analysis was performed to assess the impact on model performance. For example... Figure 4 As shown, when and At that time, the model training process converges stably with minimal fluctuations. Increase to At that time, the reconstruction performance decreased slightly; while When the value is too small (0.1) or too large (1), significant oscillations occur in the early stages of training. Considering both convergence stability and reconstruction accuracy, the following value is selected: , This will be used as the default parameter configuration for subsequent experiments.

[0039] To further verify the effectiveness of information bottleneck constraints and orthogonal decoupling, the experiment set up three comparative model configurations: "w / o OD+IB" indicates that both orthogonal projection and information bottleneck constraints are removed simultaneously, and only the following configuration is used. As the total loss; "w / o OD" indicates that only the orthographic projection and its corresponding loss are removed. Loss, and in calculation Time As a reconstruction loss; and the complete "OD-IB-Net". For example... Figure 5 As shown in (a), OD-IB-Net has a faster convergence speed and lower final cost compared to the other two configurations. The value indicates that its optimization efficiency is higher and its generalization ability is stronger. From Figure 5 (b) shows that OD-IB-Net achieves significant improvements under different signal-to-interference ratios. These results demonstrate that the combination of orthogonal projection and information bottleneck constraints can effectively enhance the training dynamics and robustness of the model.

[0040] To verify the target signal separation performance of the proposed OD-IB-Net model, it is compared with the classic algorithm Conv-TasNet (Convolutional Temporal Audio Separation Network) and the recently proposed CTDCRN (Convolutional Temporal Decoupling Recurrent Network). Evaluation is based on scale-invariant signal-to-interference-plus-noise ratio (SI-SINR) and correlation coefficient. Two metrics were used to examine the model's performance in terms of signal-to-interference ratio (SIR) and frequency overlap rate. Number of interference sources Time domain overlap rate Performance under four interfering factors. Among them, Due to frequency offset Ratio to symbol rate This is jointly determined, indicating the overlap ratio between the interference spectrum and the target signal spectrum; Interference duration Number of interference sources The term "interference duration" is determined jointly and represents the proportion of the SOI duration.

[0041] Figure 6 The eight subgraphs in the diagram correspond to the performance under the four interference factors mentioned above. For example... Figure 6 As shown in (a), the SI-SINR of all methods monotonically increases with increasing SIR. Among the compared methods, OD-IB-Net consistently achieves the highest SI-SINR across the entire SIR range, especially under the extreme condition of -20dB SIR, where its SI-SINR remains close to 2dB, while CTDCRN and Conv-TasNet are both below -2dB. Figure 6 As shown in (b), the correlation coefficient The trend of change is consistent with that of SI-SINR, and OD-IB-Net still maintains the highest waveform fidelity under low SIR conditions. Figure 6 (c) and Figure 6 (d) shows the model performance as The changes are influenced by data distribution bias. In 0 The limited number of samples in the 0.1 interval resulted in lower initial performance, but OD-IB-Net performed well in all... It consistently outperforms the comparison method across all possible values. At 0.9 In the 1-interval range, OD-IB-Net's SI-SINR is approximately 2 dB, while the comparison methods are all below 0 dB. The indicators also show clear advantages. Figure 6 (e) to Figure 6 (h) gives the number of different interference sources. Overlap rates with different time domains Performance comparison of the models under two scenarios. SI-SINR of OD-IB-Net and... All indicators remained at their optimal levels.

[0042] In summary, the proposed OD-IB-Net consistently outperforms the comparative methods in all test scenarios, demonstrating strong robustness and good practical application value.

[0043] Table 2 presents a comparison of the computational complexity of different models. Compared to Conv-TasNet, the number of parameters in the proposed OD-IB-Net is reduced by an order of magnitude, although the inference time is slightly increased. Compared to the lightweight model CTDCRN, OD-IB-Net has a slight improvement in computational complexity, but this improvement is compensated by its significant gain in separation performance.

[0044] Table 2 Comparison of computational complexity of the models

[0045] It should be noted that the introduction of information bottleneck constraints and orthogonal decoupling primarily affects the training phase, guiding the network to learn more discriminative and decoupled feature representations through regularization terms and auxiliary losses. During the testing phase, the computational units corresponding to these constraints no longer participate in forward propagation, thus avoiding the introduction of additional inference overhead. The model only performs forward computations on the encoder and orthogonal projection target signal decoder during testing, maintaining overall efficiency. In summary, OD-IB-Net achieves a good balance between model complexity and separation performance, while also possessing real-time processing capabilities, making it suitable for practical deployment scenarios.

[0046] In summary, this method constructs an orthogonal decoupling network that integrates information bottleneck constraints in an end-to-end manner, achieving adaptive extraction and accurate reconstruction of target sources in aliased signals without the need for explicit interference modeling. This method effectively addresses interference factors of varying intensities, types, and overlap modes, exhibiting good anti-interference capabilities and generalization performance, thus providing effective technical support for achieving highly reliable TSS in complex electromagnetic environments.

[0047] In one embodiment, a wireless signal separation device is provided, comprising: The signal acquisition unit is used to acquire aliased signals in co-frequency interference scenarios. The model building and training unit is used to build an orthogonal decoupling network model that incorporates information bottleneck constraints. During model training, a multi-task learning strategy is adopted to jointly optimize the encoder, orthogonal decoupling module, and parallel decoder connected sequentially within the model under the information bottleneck constraint. The total loss function of model training includes scale-invariant signal-to-interference-plus-noise ratio loss, variational information bottleneck loss, and orthogonality loss. The signal separation unit is used to input the aliased signal into the trained model. First, the encoder maps the aliased signal to a high-dimensional feature space to obtain a latent feature representation containing the target signal and redundant interference noise. Then, the orthogonal decoupling module performs an orthogonal mapping operation on the latent feature representation to explicitly decompose it into target signal related features and orthogonal redundant features. Finally, the parallel decoder reconstructs the time-domain waveform of the target signal based on the target signal related features to complete the wireless signal separation.

[0048] Specific limitations regarding the wireless signal separation device can be found in the limitations of the wireless signal separation method described above, and will not be repeated here. Each module in the aforementioned wireless signal separation 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 operations corresponding to each module.

[0049] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 7 As 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 for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a wireless signal separation method. The display screen can be an LCD screen or an e-ink display screen. 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.

[0050] Those skilled in the art will understand that Figure 7 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.

[0051] 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: Acquire aliased signals in co-channel interference scenarios; An orthogonal decoupling network model combining information bottleneck constraints is constructed. During model training, a multi-task learning strategy is adopted to jointly optimize the encoder, orthogonal decoupling module and parallel decoder connected sequentially within the model under the information bottleneck constraint. The total loss function of model training includes scale-invariant signal-to-interference-plus-noise ratio loss, variational information bottleneck loss and orthogonality loss. The aliased signal is input into the trained model. First, the encoder maps the aliased signal to a high-dimensional feature space to obtain a latent feature representation containing the target signal and redundant interference noise. Then, the orthogonal decoupling module performs an orthogonal mapping operation on the latent feature representation to explicitly decompose it into target signal related features and orthogonal redundant features. Finally, the parallel decoder reconstructs the time-domain waveform of the target signal based on the target signal related features to complete the wireless signal separation.

[0052] 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: Acquire aliased signals in co-channel interference scenarios; An orthogonal decoupling network model combining information bottleneck constraints is constructed. During model training, a multi-task learning strategy is adopted to jointly optimize the encoder, orthogonal decoupling module and parallel decoder connected sequentially within the model under the information bottleneck constraint. The total loss function of model training includes scale-invariant signal-to-interference-plus-noise ratio loss, variational information bottleneck loss and orthogonality loss. The aliased signal is input into the trained model. First, the encoder maps the aliased signal to a high-dimensional feature space to obtain a latent feature representation containing the target signal and redundant interference noise. Then, the orthogonal decoupling module performs an orthogonal mapping operation on the latent feature representation to explicitly decompose it into target signal related features and orthogonal redundant features. Finally, the parallel decoder reconstructs the time-domain waveform of the target signal based on the target signal related features to complete the wireless signal separation.

[0053] 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.

[0054] 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.

[0055] 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 method for separating wireless signals, characterized in that, The method includes: Acquire aliased signals in co-channel interference scenarios; An orthogonal decoupling network model combining information bottleneck constraints is constructed. During model training, a multi-task learning strategy is adopted to jointly optimize the encoder, orthogonal decoupling module and parallel decoder connected sequentially within the model under the information bottleneck constraint. The total loss function of model training includes scale-invariant signal-to-interference-plus-noise ratio loss, variational information bottleneck loss and orthogonality loss. The aliased signal is input into the trained model. First, the encoder maps the aliased signal to a high-dimensional feature space to obtain a latent feature representation containing the target signal and redundant interference noise. Then, the orthogonal decoupling module performs an orthogonal mapping operation on the latent feature representation to explicitly decompose it into target signal related features and orthogonal redundant features. Finally, the parallel decoder reconstructs the time-domain waveform of the target signal based on the target signal related features to complete the wireless signal separation.

2. The wireless signal separation method according to claim 1, characterized in that, Aliasing signals in co-channel interference scenarios Represented as: ； in, and The first The transmitter and the first The transmitter-to-receiver channel attenuation coefficient; For the first The source signal from each transmitter is set as the target signal to be separated. SOI ; For the first The source signal of each transmitter, for the target signal SOI Constitutes co-channel interference CCI ; , The total number of transmitters; It is Gaussian white noise. Indicates a Gaussian distribution. Let be the variance of the Gaussian distribution.

3. The wireless signal separation method according to claim 1, characterized in that, The encoder employs a network structure that combines multi-scale parallel feature extraction branches, and the processing logic is as follows: The encoder input is of size The aliased signal is first processed by a convolutional kernel with a kernel size of 11 and an output channel number of... A one-dimensional convolutional layer yields a size of... The feature map; then set three parallel branches with kernel sizes of 3, 5 and 7 respectively, each branch consisting of... It consists of several deep residual shrinkage modules. The first deep residual shrinkage module in each branch reduces the number of channels from... Expand to The following The number of output channels for each depth residual shrinkage module remains at [value]. Size is The feature maps are processed by three parallel branches to obtain three values ​​of size 1. The parallel outputs are stacked along the output channel dimension to obtain a size of [size missing]. The fused feature map is passed sequentially through two one-dimensional convolutional layers, each with a kernel size of 3, and the number of output channels is set to... and The final latent feature representation is obtained. As input for subsequent modules; among which, For the set of real numbers, The batch size for encoder input.

4. The wireless signal separation method according to claim 1, characterized in that, The processing logic of the orthogonal decoupling module is as follows: Introducing the projection matrix Represent the latent features output by the encoder The decomposition is divided into two complementary orthogonal subspaces, one of which preserves the correlation features of the target signal. The other subspace corresponds to redundant interference and noise components, represented by orthogonal redundancy features. ;in, It is an identity matrix, and .

5. The wireless signal separation method according to claim 1, characterized in that, The decoder contains two independent decoding branches: a target signal decoding branch for reconstructing the target signal and an interference reconstruction branch for reconstructing residual interference and noise components. The target signal decoding branch takes the relevant features of the target signal output from the orthogonal decoupling module as input, and first uses an output channel with a number of... A one-dimensional convolutional layer with a kernel size of 5 performs initial decoding of the input features, followed by a convolutional layer consisting of... Multiple output heads composed of parallel fully connected layers reconstruct the signal during the model training phase. Each output head undergoes synchronous end-to-end training, and the loss function for the target signal decoding branch training is the average of the losses from each output head; during the inference phase, a random selection strategy is adopted, starting from... One of the output heads is randomly selected for approximate reconstruction of the target signal, and the output is the first one. Estimation of target signals ; The interference reconstruction branch takes the orthogonal redundancy features output by the orthogonal decoupling module as input and outputs residual components. The interference reconstruction branch uses an output channel with a number of... A one-dimensional convolutional layer with a kernel size of 5 and a single output head is implemented.

6. The wireless signal separation method according to claim 1, characterized in that, Under the constraint of information bottleneck, the encoder, orthogonal decoupling module, and parallel decoder connected sequentially within the model are jointly optimized. The optimization objective of the information bottleneck constraint is: in, The latent feature representation of the encoder output, The input model contains aliased signals. To balance the parameters, For mutual information; The first term is a compression term, used to control the model to discard redundant information; the second term is used to maximize the mutual information with the target signal in order to retain the key features related to the target signal. To achieve the optimization goal constrained by information bottlenecks The variational information bottleneck principle is used to address the aliasing signals in the input model. With latent feature representation Constraints are imposed on mutual information between the two sides to construct variational information bottleneck loss. , is represented as: ； in, To reconstruct the loss, For aliasing signals With target signal The mathematical expectation of the joint distribution; A standard Gaussian prior is used to guide compression. Indicates a Gaussian distribution. Represents the identity matrix; The latent distribution of the encoder output; combined with the structural design of two independent decoding branches in the parallel decoder, the reconstruction loss... It consists of two parts, one of which is the mean square error loss of the target signal decoding branch used for target signal reconstruction. , The first branch of the target signal decoding Output estimation of each output head For the corresponding number The source signal of a transmitter, The first is the total number of output heads; the second is the residual interference and noise constraint loss of the interference reconstruction branch. ,in, To interfere with the residual components of the reconstructed branch output; The KL divergence term is used to regularize the latent distribution, making it approximate the standard Gaussian prior, and is expressed as: ； in, and The potential distribution of the encoder output is respectively In the Mean and variance in dimension This represents the total number of dimensions.

7. The wireless signal separation method according to claim 6, characterized in that, The total loss function, formed by combining scale-invariant signal-to-interference-plus-noise ratio loss, variational information bottleneck loss, and orthogonality loss, is expressed as: ； in, and These are hyperparameters used to balance the variational information bottleneck loss. and orthogonality loss Contribution; orthogonality loss , The target signal related features decomposed by the orthogonal decoupling module are indicated by the superscript. T Indicates transpose. The orthogonal redundant features are derived from the orthogonal decoupling module. The Frobenius norm is used to minimize the orthogonality loss, ensuring that the relevant features of the target signal and the orthogonal redundant features are as orthogonal as possible in the latent space; scale-invariant signal-to-interference-plus-noise ratio loss. Used to evaluate the quality of target signal estimation, denoted as: ； in, It is the first Output estimation of each output head In the corresponding source signal Projected components in the direction; This represents the noise component.

8. A wireless signal separation device, characterized in that, The device includes: The signal acquisition unit is used to acquire aliased signals in co-frequency interference scenarios. The model building and training unit is used to build an orthogonal decoupling network model that incorporates information bottleneck constraints. During model training, a multi-task learning strategy is adopted to jointly optimize the encoder, orthogonal decoupling module, and parallel decoder connected sequentially within the model under the information bottleneck constraint. The total loss function of model training includes scale-invariant signal-to-interference-plus-noise ratio loss, variational information bottleneck loss, and orthogonality loss. The signal separation unit is used to input the aliased signal into the trained model. First, the encoder maps the aliased signal to a high-dimensional feature space to obtain a latent feature representation containing the target signal and redundant interference noise. Then, the orthogonal decoupling module performs an orthogonal mapping operation on the latent feature representation to explicitly decompose it into target signal related features and orthogonal redundant features. Finally, the parallel decoder reconstructs the time-domain waveform of the target signal based on the target signal related features to complete the wireless signal separation.

9. 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 7.

10. 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 7.

Images