Attention-based method for separating signals of the same frequency, apparatus, equipment and medium

Abstract

This application relates to a method, apparatus, device, and medium for separating co-channel signals based on an attention mechanism, belonging to the field of wireless communication technology. The method includes: acquiring a mixed signal under co-channel interference scenarios and simulating the generation of multiple types of modulated signals as prior pure signals; inputting all prior pure signals into a prior signal feature extraction network trained based on the Arcface loss function to obtain the embedding centers of various modulated signals as prior information; constructing and training a target signal separation network, in which mixed signal features are extracted according to a mixed signal encoder, modulation features are encoded and extracted from the embedding centers according to a modulation feature encoding module, and bidirectional cross-attention calculation is performed on the mixed signal features and modulation features according to a cross-attention module to obtain fused features focused on the target signal, which are then decoded to reconstruct and output a target estimation signal. This method can achieve accurate separation of target signals under co-channel interference scenarios.

Description

Technical Field

[0001] This application relates to the field of wireless communication technology, and in particular to a method, apparatus, device and medium for separating co-frequency signals based on an attention mechanism. Background Technology

[0002] The rapid development of IoT, 5G, and Beyond 5G (B6G) technologies has led to an exponential increase in the number of wireless access devices and a sharp rise in demand for spectrum resources, making the electromagnetic spectrum increasingly congested. Wireless communication networks are facing unprecedented spectrum resource pressure. The scarcity of spectrum resources has prompted researchers to explore more efficient spectrum reuse techniques, which inevitably introduces serious co-channel interference problems. Specifically, the overlap of co-channel signals in the time and frequency domains leads to distortion of the target signal at the receiver and a decrease in the signal-to-noise ratio, thus affecting communication quality and data transmission reliability, becoming a key bottleneck restricting the performance improvement of wireless communication systems. Therefore, accurately separating target signals and ensuring communication link stability in complex co-channel interference environments has become one of the core technical challenges that urgently need to be overcome in the field of wireless communication.

[0003] Against this backdrop, accurately and efficiently separating information from a specific source from complex mixed received signals—that is, target signal separation—is of great significance for improving spectrum utilization and ensuring communication reliability and security. The target signal separation problem is essentially highly similar to the classic "cocktail party problem" in the field of speech. In the "cocktail party problem," one must focus on the speech information of a specific speaker in an environment with noisy voices and complex background interference. In the field of wireless communication, with the increasing number of frequency-using devices, target signals and interference signals overlap significantly in both the time and frequency domains, resulting in simultaneous co-channel interference. However, traditional communication methods struggle to quickly sense and adaptively adjust to such complex electromagnetic environments, which significantly limits their performance in such scenarios.

[0004] Traditional anti-jamming techniques such as frequency hopping, spread spectrum, and filtering, as well as methods like blind source separation and adaptive beamforming, are essentially external countermeasures. Their effectiveness largely depends on the mastery of prior knowledge of the interfering signal and the accurate construction of the interference model. However, when facing non-cooperative and dynamically changing simultaneous co-channel interference scenarios, these conditions based on prior assumptions often fail to hold, causing a sharp decline in the processing power of traditional communication methods. The fundamental reason is the inherent lack of anti-jamming performance in these methods, which greatly limits their robustness and applicability in dynamic and adversarial real-world wireless communication environments. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, apparatus, device, and medium for separating co-frequency signals based on an attention mechanism to address the aforementioned technical problems.

[0006] A method for separating co-frequency signals based on an attention mechanism, the method comprising:

[0007] Acquire mixed signals under simultaneous and same-frequency interference scenarios. The mixed signals include the target signal and the interference signal that is simultaneous and same-frequency but has a different modulation type. Simulate and generate multiple types of modulation signals as prior pure signals.

[0008] All prior pure signals are input into a prior signal feature extraction network trained based on the Arcface loss function for feature extraction, and embedding vectors of various modulation signals are obtained. The embedding centers of various modulation signals are obtained by summing and averaging the embedding vectors, which serve as prior information for target signal separation.

[0009] Construct and train a target signal separation network that includes a mixed-signal encoder, a modulation feature encoding module, a cross-attention module, and a mixed-signal decoder;

[0010] The mixed signal is input into the trained target signal separation network. The mixed signal encoder extracts the mixed signal features, the modulation feature encoding module encodes and extracts the modulation features from the embedding center, and the cross attention module performs bidirectional cross attention calculation on the mixed signal features and modulation features to obtain the fused features focused on the target signal. The fused features are then decoded by the mixed signal decoder to reconstruct the output target estimation signal.

[0011] In one embodiment, the prior signal feature extraction network includes three one-dimensional convolutional layers, two residual blocks, and one one-dimensional convolutional layer connected in sequence. The first three convolutional layers are followed by batch normalization layers and Tanh activation functions, and the last convolutional layer is followed by batch normalization layers and Tanh activation functions.

[0012] The prior signal feature extraction network performs feature extraction, including: first, it sequentially performs wide-range temporal correlation capture, local feature deepening learning and fine-grained feature extraction on the input signal through three layers of one-dimensional convolutional layers; then, it performs feature enhancement through two layers of residual blocks; and finally, it performs size constraint through one layer of one-dimensional convolutional layer, and outputs the embedding vectors of various modulation signals of the input.

[0013] The main convolutional path of the residual block combines two one-dimensional convolutional layers, each followed by a batch normalization layer. The first one-dimensional convolutional layer is also followed by a ReLU activation function. The main convolutional path is used to extract and transform features from the input channels to the output channels. The skip connection path of the residual block dynamically selects its processing method based on whether the input and output channels are consistent. When the number of input and output channels of the main convolutional path is different, the skip connection path uses a one-dimensional convolutional layer to adjust the number of input channels to match the output dimension of the main convolutional path. When the number of input and output channels of the main convolutional path is the same, the skip connection path uses an identity mapping to directly pass the original input. The input processed by the skip connection path is added to the output of the main convolutional path to achieve residual fusion, and finally, the result is output after passing through the ReLU activation function.

[0014] In one embodiment, all prior pure signals are input into a prior signal feature extraction network trained based on the Arcface loss function for feature extraction to obtain embedding vectors for various modulation signals. The embedding centers for various modulation signals are then obtained by summing and averaging these embedding vectors, including:

[0015] The embedding vector output from the prior signal feature extraction network is flattened and then fed into the Arcface layer to obtain the output label. The network is then optimized and trained by calculating the Arcface loss function to obtain the trained prior signal feature extraction network. ;

[0016] Prior pure signal dataset enter Feature extraction is performed to obtain the embedding vector set. , represented as:

[0017] ;

[0018] in, This represents the total number of samples in the prior pure signal dataset. express Corresponding modulation tags, express go through The embedded vector obtained after extraction; Indicates the first Signal data of each sample; through the embedded vector set The embedding center set is obtained by summing and averaging the embedding vectors of each type of modulated signal. ,in It is the number of modulation types, the first The embedding center of a modulated signal is represented as:

[0019] ;

[0020] in, Indicates the first The number of samples in the modulated signal Indicates the first Modulation tag The embedding vector.

[0021] In one embodiment, the mixed signal encoder includes a one-dimensional convolutional layer, a residual shrinking and self-attention module, and two one-dimensional convolutional layers connected in sequence. The residual shrinking and self-attention module includes a residual shrinking bottleneck unit, a self-attention module, two residual shrinking bottleneck units, a self-attention module, and two residual shrinking bottleneck units connected in sequence. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function.

[0022] Extracting mixed signal features from a mixed signal encoder includes:

[0023] The mixed signal is input into the mixed signal encoder. First, it passes through a one-dimensional convolutional layer to perform preliminary feature extraction and dimension adjustment. Then, it enters the residual shrinkage and self-attention module. The residual shrinkage bottleneck unit eliminates the influence of noise and learns the deep features of the signal. The self-attention module captures the long-distance dependencies of the signal and mines the correlation between features. Finally, it passes through two one-dimensional convolutional layers to deepen the signal features, and finally completes the encoding of the mixed signal and outputs the mixed signal features.

[0024] In the residual shrinkage bottleneck unit, the input data first passes through two one-dimensional convolutional layers, then through a soft thresholding function to obtain an output, and finally adds the output to the original input data to obtain the final output. The input-output relationship of the soft thresholding function satisfies:

[0025] ;

[0026] in, and These represent the input features and the output features, respectively. The threshold is represented by the following method: First, for a value of... The feature map is subjected to absolute value extraction and global average pooling to obtain a one-dimensional vector. This one-dimensional vector is then passed through two fully connected layers to obtain the channel weight vectors. Next, the weight vector Scale it to its maximum value using the Sigmoid activation function. Range, represented as , Representing the Channel weights, , The total number of channels. This represents the obtained scaling value; ultimately, it represents the scaling value for all channels. Multiply The average value is used to obtain the threshold. ,in, and These represent the widths of the feature maps. and the ordinal index of height;

[0027] In the self-attention module, the correlation between each position in the input data sequence and all other positions is calculated, and then a weighted sum is performed based on these correlations. The calculation formula is as follows:

[0028] ;

[0029] Where, it is assumed If the input data sequence is a query matrix, then... Key matrix Value matrix , , and It is a learnable weight matrix, which is essentially the model parameters of a one-dimensional convolution. Key matrix Dimensions.

[0030] In one embodiment, the modulation feature coding module adopts a three-layer one-dimensional convolutional neural network structure, with each convolutional layer followed by a batch normalization layer and a ReLU activation function;

[0031] Modulation features are extracted from the embedding center based on the modulation feature coding module, including:

[0032] All embedding centers are input into the modulation feature encoding module. First, they are encoded through the first one-dimensional convolutional layer to capture extensive local context information. Then, they are input into the second one-dimensional convolutional layer to further extract features by reducing the kernel size and increasing the number of feature channels. Finally, a third one-dimensional convolutional layer is used to extract more refined feature representations. Ultimately, modulation features of various modulation signals are encoded and extracted from the embedding centers.

[0033] In one embodiment, the cross-attention module includes a first cross-attention unit and a second cross-attention unit; the mixed signal decoder adopts a four-layer one-dimensional convolutional neural network structure, wherein the first three convolutional layers are followed by a batch normalization layer and a ReLU activation function.

[0034] Based on the bidirectional cross-attention calculation of the mixed signal features and modulation features by the cross-attention module, the fused features focused on the target signal are obtained. Then, the fused features are decoded by the mixed signal decoder to reconstruct the output target estimation signal, including:

[0035] The mixed-signal features extracted by the mixed-signal encoder and the modulation features extracted by the modulation feature encoding module are fed into the cross-attention module. In the first cross-attention unit, a query matrix is ​​constructed based on the mixed-signal features, and a key matrix and a value matrix are constructed based on the modulation features. Cross-attention calculation is used to focus on the parts of the prior information that are related to the target signal. In the second cross-attention unit, a query matrix is ​​constructed based on the modulation features, and a key matrix and a value matrix are constructed based on the mixed-signal features. Cross-attention calculation is used to focus on the features of the mixed-signal features that match the prior information. Finally, the features output by the two cross-attention calculations are added together to obtain the fused features focused on the target signal, which are then input into the mixed-signal decoder. A four-layer one-dimensional convolutional neural network structure is used to perform stepwise dimensional transformation and feature extraction operations on the fused features to output the target estimation signal.

[0036] In one embodiment, the target signal separation network is trained using a strategy of combining multiple target loss functions, with the total loss function... Represented as:

[0037] ;

[0038] in, This is the scale-invariant signal-to-interference-plus-noise ratio loss. For scale-invariant signal-to-interference-plus-noise ratio. For mean square error loss, For the sample size, For the first The reference signal for each sample, For the first The estimated signal for each sample, This is the loss weighting coefficient.

[0039] A device for separating co-frequency signals based on an attention mechanism, the device comprising:

[0040] The signal acquisition module is used to acquire mixed signals under simultaneous and same-frequency interference scenarios. The mixed signals include the target signal and the interference signals that are at the same frequency but have different modulation types. It also simulates and generates multiple types of modulation signals as prior pure signals.

[0041] The prior information extraction module is used to input all prior pure signals into the prior signal feature extraction network trained based on the Arcface loss function for feature extraction, to obtain the embedding vectors of various modulation signals, and to obtain the embedding center of various modulation signals as the prior information for target signal separation by summing and averaging the various embedding vectors.

[0042] The network construction and training module is used to build and train a target signal separation network that includes a mixed-signal encoder, a modulation feature encoding module, a cross-attention module, and a mixed-signal decoder.

[0043] The target signal separation module is used to input the mixed signal into the trained target signal separation network, extract the mixed signal features according to the mixed signal encoder, extract the modulation features from the embedding center according to the modulation feature encoding module, and perform bidirectional cross-attention calculation on the mixed signal features and modulation features according to the cross-attention module to obtain the fused features focused on the target signal. Then, the fused features are decoded by the mixed signal decoder to reconstruct the output target estimation signal.

[0044] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the above-described attention mechanism-based method for separating co-frequency signals.

[0045] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described attention mechanism-based method for separating co-frequency signals.

[0046] The aforementioned attention-based method, apparatus, device, and medium for separating co-frequency signals explore the problem of target signal separation in simultaneous co-frequency interference scenarios, where the interference signal and the target signal differ only in modulation type. In the signal separation process, all simulated prior clean signals are first input into a prior signal feature extraction network trained using the Arcface loss function. This yields embedding vectors for various modulation signals, which are then summed and averaged to obtain embedding centers for each type of modulation signal. Each embedding center possesses strong class discriminative power, intra-class compactness, and inter-class discreteness, serving as prior information for distinguishing different signals. Next, the mixed signal is input into the trained target signal separation network. The mixed signal encoder and modulation feature encoder modules extract the mixed signal features and modulation features from the embedding centers, respectively. Then, a cross-attention module performs bidirectional cross-attention calculations on the mixed signal features and modulation features, focusing on features useful for target signal separation. Finally, the fusion features output by the cross-attention module, focused on the target signal, are input into the mixed signal decoder for decoding, reconstructing and outputting the estimated target signal. Compared with existing technologies, this application breaks through the limitation of traditional technologies that require explicit perception and modeling of interference signals. By introducing the prior modulation features of the target signal and making full use of the powerful nonlinear fitting and feature learning capabilities of deep learning models, it can form an adaptive perception and extraction capability for interference noise conditions of different intensities and types. It can accurately capture the unique characteristics of the target signal in complex electromagnetic environments, thereby efficiently achieving the separation of the target source signal. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the overall architecture of a method for separating co-frequency signals based on an attention mechanism in one embodiment;

[0048] Figure 2 This is a schematic diagram of the residual block structure in one embodiment;

[0049] Figure 3 This is a schematic diagram of the target signal separation network structure in one embodiment;

[0050] Figure 4 This is a schematic diagram of the residual shrinkage bottleneck unit structure in one embodiment;

[0051] Figure 5 This is a schematic diagram illustrating the t-SNE visualization of embedding vectors and embedding centers obtained using different activation functions in one embodiment; wherein, Figure 5 (a) A visual representation of the embedding vectors and embedding centers obtained using the Softmax activation function in t-SNE. Figure 5 (b) A t-SNE visualization of the embedding vectors and embedding centers obtained using the Arcface activation function;

[0052] Figure 6 This is a schematic diagram illustrating the training and fitting process in one embodiment;

[0053] Figure 7 This is a schematic diagram illustrating ablation results under different SINR conditions in one embodiment; wherein, Figure 7 (a) is a schematic diagram of ablation results under different SINR conditions in dataset 1. Figure 7 (b) is a schematic diagram of ablation results under different SINR conditions in dataset 2;

[0054] Figure 8 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0055] 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.

[0056] In one embodiment, such as Figure 1 As shown, a method for separating co-frequency signals based on an attention mechanism is provided, including:

[0057] 1. Signal Acquisition: Acquire mixed signals under simultaneous and same-frequency interference scenarios. The mixed signals include the target signal and the interference signals that are simultaneously and at the same frequency but with different modulation types. Simulate and generate multiple types of modulation signals as prior pure signals.

[0058] 2. Prior information extraction: Input all prior clean signals into the prior signal feature extraction network trained based on the Arcface loss function to extract features, obtain the embedding vectors of various modulation signals, and obtain the embedding center of various modulation signals as the prior information for target signal separation by summing and averaging the embedding vectors of various types.

[0059] This section comprises two parts: prior signal feature extraction and embedding space learning. The prior signal feature extraction network can process prior clean signals (…). Feature extraction is performed, and redundant information from signal samples is removed to obtain embedding vectors containing modulation features. The center of the embedding vectors for signals of the same modulation type serves as prior information for target signal separation; therefore, the embedding center should possess significant representativeness and inter-class discriminability. To this end, the embedding vectors need further optimization of the network structure using a loss function with an Arcface layer. This loss function, by introducing a marginal penalty mechanism in the angle space, effectively enhances the aggregation of embedding vectors for similar modulation signals and the discreteness of embedding vectors for dissimilar modulation signals, ultimately constructing a pure signal feature embedding space with stronger discriminative power. This embedding space includes various embedding centers, providing a reliable prior feature foundation for subsequent target signal separation tasks.

[0060] Specifically, for sequential data, one-dimensional convolution has lower computational complexity and is more efficient than two-dimensional convolution. Therefore, this embodiment designs a multi-scale feature learning network structure using one-dimensional convolution as a prior signal feature extraction network to capture complex features in signal samples for IQ time-series input signals of size 2×1024. The overall structure is shown in Table 1.

[0061] The prior signal feature extraction network performs feature extraction, including: firstly, feature dimensionality enhancement and abstraction are achieved through three one-dimensional convolutional layers. The first layer uses a 7×1 convolution to capture a wide range of temporal correlations, expanding the number of channels from 2 to 64; the second layer uses a 5×1 convolution to deepen local feature learning, increasing the number of channels to 128; the third layer uses a 3×1 convolution to further extract fine-grained features, reaching 256 channels. Each convolutional layer is combined with a batch normalization (BN) layer and a rectified linear unit (ReLU) activation function to enhance nonlinear expressive power and stabilize the training process. Then, feature enhancement is performed through two residual blocks. The first residual block maintains the 256-channel dimension and alleviates the gradient vanishing problem in deep networks through skip connections, enhancing feature transfer efficiency; the second residual block compresses the number of channels to 128, retaining key discriminative features while controlling model complexity. Finally, a 1×1 convolution maps the number of feature channels back from 128 to 2. Batch normalization and the Tanh (hyperbolic tangent) activation function constrain the output to the [-1, 1] interval, resulting in an output with the same size as the input. This convolution strategy ensures sufficient capture of multi-scale features of the temporal signal and improves the training stability of deep networks through a residual mechanism, ultimately achieving an end-to-end mapping from the input signal to the embedding vector.

[0062] Table 1. Network Structure for Prior Signal Feature Extraction

[0063]

[0064] Among them, the residual block structure is as follows Figure 2 As shown, the main convolutional path of this residual block combines two one-dimensional convolutional layers (Conv1d), each followed by a batch normalization layer. The first one-dimensional convolutional layer is followed by a ReLU activation function. The kernel size is 3, padded to 1 to maintain the temporal length, achieving feature extraction and transformation from the input channel to the output channel. The main convolutional path is used for feature extraction and transformation from the input channel to the output channel. The skip connection path of the residual block dynamically selects its processing method based on whether the input and output channels are consistent. When the number of input and output channels of the main convolutional path is different, the skip connection path uses a 1×1 convolution of the one-dimensional convolutional layer to adjust the number of input channels to match the output dimension of the main convolutional path. When the number of input and output channels of the main convolutional path is the same, the skip connection path uses an identity mapping to directly pass the original input. The input processed by the skip connection path is added to the output of the main convolutional path to achieve residual fusion, and finally, the result is output after passing through the ReLU activation function.

[0065] The 2×1024 embedding vector obtained from the aforementioned prior signal feature extraction network structure is flattened and further passed through the Arcface layer to obtain the output label. Then, the Arcface loss function (additive angular residual loss) is calculated to optimize the extraction of the embedding center. Compared with the traditional fully connected layer plus Softmax layer for classification, the Arcface layer acts as both a classifier that maps high-dimensional features to modulation categories and a feature optimizer that enhances inter-class separation in angular space.

[0066] The traditional Softmax activation function can be expressed as:

[0067] ;

[0068] in, Represents the input vector. Indicates the first The original score of the target category, The total number of categories, This indicates that the current iteration has reached the [number]th [item]. The original score value of the target category. In the recognition and classification task, the output of this activation function is further processed by the cross-entropy loss function to obtain the task loss. The complete loss function formula is:

[0069] ;

[0070] in, It is the sample size. It is the number of categories. Representing the The feature vector of each sample Representing the The true class labels of the samples, and the length of the feature vector is . , yes The List, and bias terms Together, they form a fully connected layer that obtains feature vectors. yes The weight matrix, The length is First, let Then perform the following conversion:

[0071] ;

[0072] in It is an eigenvector Compared with the true class weight vector The angle between them, this inner product can be seen as the decision boundary of the classifier. Let where and The complete loss function is then further transformed into:

[0073] .

[0074] However, in classification tasks, the Softmax-based loss function primarily achieves classification by maximizing the probability difference between the correct and incorrect classes, focusing more on the separability between classes and neglecting the discriminative learning of features. To address this issue, Arcface modifies the decision boundary in the angle space as follows:

[0075] ;

[0076] in, It is an additional decision boundary. Its core innovation lies in making the decision boundary more stringent, changing the condition for the classifier to correctly classify the data as follows:

[0077] ;

[0078] Equivalent to ;

[0079] Furthermore, the complete loss function based on Arcface is obtained as follows:

[0080] ;

[0081] During the optimization process, the objective of this loss function is to minimize the loss for the target class, i.e., the true class. Orient the feature vectors toward the weight vectors of the true class; for non-target classes, maximize... Make the feature vectors deviate from the directions of the weight vectors of other classes. Wherein, Its function is to create a safety margin in angle space, requiring that the angle of the true class must not only be smaller than that of other classes, but also at least smaller than that of other classes. The radian property enhances inter-class separability and the robustness of decision boundaries.

[0082] The prior signal feature extraction network structure and Arcface loss function designed above can effectively extract the modulation features of signal samples, so that the input signal dataset exhibits clear geometric distribution characteristics in the embedding space. The embedding vectors of the same type of modulation signal will be closely clustered in the hyperspherical local region with its category center as the origin, while the embedding vector clusters of different modulation types maintain sufficient angular spacing.

[0083] Based on the aforementioned constructed and optimized prior signal feature extraction network, The prior pure signal dataset enter Feature extraction is performed to obtain the embedding vector set. , represented as:

[0084] ;

[0085] in, This represents the total number of samples in the prior pure signal dataset. express Corresponding modulation tags, express go through The embedded vector obtained after extraction; Indicates the first Signal data of each sample; through the embedded vector set The embedding center set is obtained by summing and averaging the embedding vectors of each type of modulated signal. ,in It is the number of modulation types, the first The embedding center of a modulated signal is represented as:

[0086] ;

[0087] in, Indicates the first The number of samples in the modulated signal Indicates the first Modulation tag The embedding vector. The embedding center set This serves as prior information for the subsequent separation of the target signal.

[0088] 3. Target Signal Separation: Construct and train a target signal separation network including a mixed signal encoder, a modulation feature encoding module, a cross-attention module, and a mixed signal decoder. Input the mixed signal into the trained target signal separation network, extract the mixed signal features according to the mixed signal encoder, extract the modulation features from the embedding center according to the modulation feature encoding module, and perform bidirectional cross-attention calculation on the mixed signal features and modulation features according to the cross-attention module to obtain the fused features focused on the target signal. Then, decode the fused features through the mixed signal decoder to reconstruct the output target estimation signal.

[0089] The structure of the target signal separation network is as follows: Figure 3As shown, the mixed-signal encoder includes a one-dimensional convolutional layer, a residual shrinking and self-attention module, and two one-dimensional convolutional layers connected in sequence. The residual shrinking and self-attention module includes a one-dimensional residual shrinking bottleneck unit (RSBU_CW), a one-dimensional self-attention module, two residual shrinking bottleneck units, a one-dimensional self-attention module, and two residual shrinking bottleneck units connected in sequence. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function.

[0090] Extracting mixed signal features from a mixed signal encoder includes: converting the mixed signal... The input signal is fed into the mixed signal encoder. First, it passes through a one-dimensional convolutional layer with a kernel size K of 5 and padding P of 2 to perform preliminary feature extraction and dimensionality adjustment. Then, it enters the residual shrinkage and self-attention module. The residual shrinkage bottleneck unit combines residual and adaptive thresholding mechanisms to eliminate noise and learn deep signal features. The self-attention module captures long-distance dependencies in the signal and mines the correlations between features. Finally, the signal features are deepened through two one-dimensional convolutional layers, thus completing the encoding of the mixed signal and outputting the mixed signal features. This provides a fully extracted and processed feature representation for subsequent decoding operations.

[0091] The residual shrinkage bottleneck unit structure in a mixed-signal encoder is as follows: Figure 4 As shown, this method introduces a soft thresholding function into the classic residual module to optimize model efficiency and performance. By dynamically calculating and applying an adaptive threshold, it accurately suppresses and removes small-amplitude feature components introduced by interference or noise in the feature map. In the residual shrinkage bottleneck unit, the input data first passes through two one-dimensional convolutional layers, then through a soft thresholding function to obtain an output. The output is then added to the original input data to obtain the final output. The soft thresholding function is a key step in the residual shrinkage bottleneck unit to eliminate the influence of noise. It sets features with absolute values ​​below a certain threshold to zero and adjusts other features towards zero. The input-output relationship of the soft thresholding function satisfies:

[0092] ;

[0093] in, and These represent the input features and the output features, respectively. Representing a threshold, unlike traditional signal denoising algorithms that require manually setting the threshold, this residual shrinkage bottleneck unit can generate the threshold corresponding to the feature map of each channel. The threshold is determined as follows: First, for a size of... The feature map is subjected to absolute value extraction and global average pooling (GAP) to obtain a one-dimensional vector. This one-dimensional vector is then passed through two fully connected layers (FC) to obtain the channel weight vectors. Next, the weight vector Scale it to its maximum value using the Sigmoid activation function. Range, represented as , Representing the Channel weights, , The total number of channels. This represents the obtained scaling value; ultimately, it represents the scaling value for all channels. Multiply The average value is used to obtain the threshold. ,in, and These represent the widths of the feature maps. And the height index.

[0094] The self-attention mechanism in this mixed-signal encoder can directly establish dependencies between elements at any position in the sequence, capturing global information of the signal. In the self-attention module, the correlation between each position in the input data sequence and all other positions is calculated, and then a weighted sum is performed based on these correlations. The calculation formula is as follows:

[0095] ;

[0096] Where, it is assumed If the input data sequence is a query matrix, then... Key matrix Value matrix , , and It is a learnable weight matrix, which is essentially the model parameters of a one-dimensional convolution. Key matrix The dimension. Through dot product. The attention score is calculated, representing the relevance between the query and the key. Higher relevance receives greater weight and is further utilized. To scale the dot product result and prevent dimensionality issues. When the values ​​are large, the gradient of the Softmax function vanishes. Softmax ensures that the weights are positive and sum to 1, achieving a soft selection of information at different locations. The normalized weights are then summed with... Multiplying these results yields the final attention output. For example, the input vector of the first self-attention module in a mixed-signal encoder. The matrix is ​​64×1024 in size and is obtained after passing through three different one-dimensional convolutional layers. , and Following the self-attention calculation process described above, we can obtain output features that incorporate global dependencies of the sequence, providing a more comprehensive and representative feature representation for the encoding of mixed signals.

[0097] As the embedding center for prior information, it needs to pass through a modulation feature coding module to generate a feature representation that can more effectively assist in the separation of the target signal, such as... Figure 3 As shown, the modulation feature encoding module adopts a three-layer one-dimensional convolutional neural network structure. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function to ensure the stability and effectiveness of the feature encoding process. The modulation feature encoding module extracts modulation features from the embedding centers, including: inputting all embedding centers into the modulation feature encoding module, first encoding them through the first one-dimensional convolutional layer using a large convolutional kernel. To capture extensive local contextual information, the output channel count is 64; then, the second one-dimensional convolutional layer is input, and the kernel size is reduced to... The number of feature channels is increased to 128 to further extract features. Finally, a more refined feature representation is generated in the highest dimension space with 256 output channels through a third one-dimensional convolutional layer. Finally, the modulation features of various modulation signals are encoded and extracted from the embedding center.

[0098] The cross-attention module aims to guide the attention of the target signal's modulation features to the features of the mixed signal, thereby focusing on the target signal components. It can be used to fuse information from two different feature sequences. The core principle is establishing a correlation between the two sequences, allowing one sequence (the querying party) to selectively focus on important information in the other sequence (the queryed party), achieving cross-feature information interaction and fusion. The calculation formula for the cross-attention mechanism is the same as that for the self-attention mechanism; the difference lies in the query matrix. ,in This represents the first feature sequence, while the key matrix... Value matrix ,in This represents the first characteristic sequence. For example... Figure 3 As shown, the cross-attention module includes a first cross-attention unit and a second cross-attention unit.

[0099] The cross-attention module performs bidirectional cross-attention calculations on the mixed signal features and modulation features to obtain fused features focused on the target signal. This includes: feeding the mixed signal features extracted by the mixed signal encoder and the modulation features (256×1024) extracted by the modulation feature encoding module into the cross-attention module, where the modulation features serve as guiding information; in the first cross-attention unit, a query matrix is ​​constructed based on the mixed signal features, and a key matrix and value matrix are constructed based on the modulation features. Cross-attention calculations focus on the parts of the prior information related to the target signal, specifically enhancing the expression of the target signal features; in the second cross-attention unit, a query matrix is ​​constructed based on the modulation features... The query matrix constructs a key matrix and a value matrix based on the features of the mixed signal. Through cross-attention calculation, it focuses on the features in the mixed signal that match the prior information, further clarifying the feature boundaries of the target signal and more accurately separating the target signal from the mixed signal, suppressing the influence of interference and noise signals. This bidirectional cross-attention mechanism allows both sides to supplement their own shortcomings by leveraging each other's strengths. Finally, by adding the features output from the two cross-attention calculations, a fused feature focused on the target signal is obtained. This enables the model to dynamically adjust the way it utilizes prior information based on new mixed signal features, while also allowing prior information to better adapt to new mixed signal situations, ensuring the accuracy of target signal separation and improving the model's adaptability to complex scenarios.

[0100] like Figure 3 As shown, the mixed-signal decoder employs a four-layer one-dimensional convolutional neural network structure, where the first three convolutional layers are followed by batch normalization (BN) layers and ReLU activation functions. This decoder aims to decode the output of the cross-attention module to reconstruct the target signal. By decoding the fused features through the mixed-signal decoder, it reconstructs the output target estimation signal. This includes: the first three convolutional layers have 128, 128, and 64 output channels respectively, with a kernel size of 3×1 for each layer, followed immediately by a BN layer and a ReLU activation function to achieve non-linear transformation and distribution normalization of the features; the fourth convolutional layer has 2 output channels and a kernel size of 1×1, primarily used for the final dimensionality mapping. By performing the aforementioned progressive dimensionality transformation and feature extraction operations on the input fused features, the decoder can output an estimated target signal. .

[0101] Furthermore, the aforementioned target signal separation network is trained using a multi-objective loss function combination strategy, simultaneously optimizing the separation quality and temporal waveform accuracy of the target signal. The loss function consists of two parts: scale-invariant source-to-interference-noise ratio (SI-SINR) loss and mean squared error (MSE) loss. SI-SINR is a metric used to measure signal quality, particularly in signal separation tasks. By comparing the energy ratio of the separated signal and the reference signal, ignoring amplitude differences, it assesses the degree of distortion of the target signal. The formula for calculating SI-SINR is:

[0102] ;

[0103] in, It is an estimated signal. It is a reference signal. This indicates an element-wise multiplication and summation operation. yes The SI-SINR norm. A higher SI-SINR value indicates higher quality of the separated target signal and better separation of the signal from interference noise. The SI-SINR loss is calculated by taking the negative value of SI-SINR. The purpose is to transform the problem of maximizing SI-SINR into the problem of minimizing the loss function, thus adapting to the standard optimization paradigm of deep learning frameworks. It is defined as: .

[0104] For wireless communication systems, the signal amplitude directly affects the signal-to-interference-plus-noise ratio (SIR), which has significant practical implications in engineering. MSE loss provides fine-grained time-domain constraints, ensuring the accuracy of waveform reconstruction, and is directly sensitive to the signal waveform amplitude. The MSE loss function is defined as: ,in For the sample size, For the first The reference signal for each sample, For the first The estimated signal for each sample.

[0105] The final total loss function is a weighted combination of the two:

[0106] ;

[0107] in, The loss weighting coefficient is used to adjust the relative weight of the MSE loss term in the total loss, balancing the influence of the two loss functions on the optimization direction of the model. This dual constraint mechanism enables the model to generate target signals with better separation while meeting the stringent signal quality requirements of communication systems, ensuring robustness and practicality in real-world deployment environments.

[0108] To thoroughly evaluate the performance of the attention-based signal separation method provided in this application, a series of evaluation metrics were selected to reflect the similarity and quality difference between the separated signal and the original target signal from different perspectives, exploring and verifying the actual effectiveness of the method under simultaneous co-channel interference conditions. The evaluation metrics include:

[0109] (1) Scale-invariant signal-to-interference-plus-noise growth (SI-SINRi): SI-SINRi measures the improvement in SI-SINR of the signal before and after processing by the target signal separation algorithm, and is defined as:

[0110] ;

[0111] in, The SI-SINR value represents the original mixed signal, reflecting the initial proportional relationship between the target signal and interference / noise in the original mixed signal; This represents the SI-SINR value of the estimated signal after processing by the target signal separation method. This value reflects the algorithm's ability to extract the target signal and its suppression effect on interference and noise.

[0112] (2) Signal-to-Interference-plus-Noise Ratio (SINR): SINR is an important indicator for measuring the performance of a communication system. It refers to the ratio of the strength of the target signal in the received signal to the strength of the interfering signal and noise. The calculation formula is:

[0113] ;

[0114] in, Indicates the power of the target signal. This represents the residual interference power after separation. This indicates the residual noise power.

[0115] (3) Correlation coefficient: The correlation coefficient is a statistic that measures the linear correlation between two signals, reflecting the degree of similarity in waveform between the two signals. The calculation formula is:

[0116] ;

[0117] in, yes and covariance, and They are and The variance. For discrete complex baseband IQ signals, the calculation formula is:

[0118] ;

[0119] in, and They are and The mean, This represents the complex conjugate operation. The correlation coefficient ranges from... 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.

[0120] Experimental Scenario Construction: The dataset for the proposed method was generated through Python simulation, covering five commonly used wireless communication digital modulation signals: BPSK (Binary Phase Shift Keying), QPSK (Quadrature Phase Shift Keying), 16QAM (16th-order Quadrature Amplitude Modulation), 64QAM (64th-order Quadrature Amplitude Modulation), and 4PAM (4-level Pulse Amplitude Modulation). Referring to typical signal configurations in 5G communication scenarios, the symbol rate and sampling rate were set to 80MHz and 122.86MHz, respectively. All modulation signals underwent pulse shaping using a root-raised cosine filter (PRC) with a roll-off factor of 0.35. When generating mixed signals, one modulation signal from each of the five modulation types was used as the target signal, and the remaining four modulation signals were used as co-channel interference signals, resulting in a total of 20 mixing scenarios. The SINR range for each mixed signal was set to... At 2dB intervals, 1200 mixed signal samples were generated for each SINR condition, and divided into training, validation, and test sets in an 8:1:1 ratio. The final training set had 192,000 signal samples, while the validation and test sets each had 24,000 signal samples. The power ratio of interference signal to noise for each signal sample was set to... Random values ​​between these ranges are used to simulate the dynamic changes in interference and noise intensity in a real communication environment. The data format is as follows: The simulation generated 1600 prior pure signals for each of the five modulation types for prior information extraction. These prior pure signals contain no interference or noise and are pure complex baseband IQ signals.

[0121] Dataset 1: BPSK is used as the target signal, and the other four modulation types are used as interference signals, resulting in four mixed scenarios. The training set contains 38,400 signal samples, and the validation and test sets each contain 4,800 signal samples.

[0122] Dataset 2: All five modulation types can be used as target signals, resulting in 20 possible combinations. The training set contains 192,000 signal samples, while the validation and test sets each contain 24,000 signal samples.

[0123] The specific signal parameter settings are shown in Table 2:

[0124] Table 2 Signal Parameter Settings

[0125]

[0126] The simulation experiment used PyTorch 1.12 as the deep learning framework, employing the Adam optimizer to update model parameters, with a learning rate set to 0.001. A dynamic learning rate adjustment strategy was also configured: when the monitored metric showed no improvement for five consecutive epochs, the learning rate was decayed by a factor of 0.5, thus achieving adaptive adjustment of the learning rate and optimizing the model training process. (Loss weighting coefficients are also mentioned.) The value is 1.

[0127] First, we analyze the effectiveness of prior information extraction: To verify the feature representation performance of the embedding vectors using the Arcface activation function and the traditional Softmax activation function, we obtain the distribution of the embedding vectors and embedding centers in the high-dimensional space under the two activation functions, and visualize their representations using t-SNE dimensionality reduction. Figure 5 As shown. By Figure 5 It can be seen that, compared with the obvious overlap of 16QAM, 64QAM and 4PAM in the feature space of the Softmax network, the scatter groups of different modulation types in the embedding space obtained by the ArcFace network are separated from each other, and the same type of modulation signals are closely clustered while the different types of signals are clearly separated. This shows that the ArcFace network used to generate embedding vectors can effectively learn an embedding space that can effectively distinguish different modulation signals, and has good intra-class clustering and inter-class distinguishability.

[0128] Secondly, an analysis of the training and fitting performance of the target signal separation network was conducted: Figure 6This paper demonstrates how the loss value of the target signal separation network changes with the number of training epochs under two dataset conditions. It can be seen that for the task of extracting BPSK target signals, the loss values ​​of both the training and validation sets decrease rapidly with increasing training epochs, and the final loss value is low and tends to stabilize, indicating that the model fits well on this single-target signal separation task. However, for the task of extracting five target signals (Allinone), although the loss values ​​of both the training and validation sets also decrease with increasing training epochs, the overall loss value is higher than that of the BPSK target signal extraction task, and the rate of decrease is relatively slower. This may be because the multi-target signal separation task is more complex, and the model needs to learn more features of different types of target signals, making it more difficult to reduce the loss. However, judging from the overall trend of loss reduction, the model is gradually adapting to the needs of multi-target signal separation.

[0129] Next, the performance of the target signal separation network was evaluated: To verify the robustness of the proposed method under different interference intensities, Table 3 shows the numerical test results of different performance indicators for the two datasets under different SINR conditions for the mixed signal, and also provides the average values ​​of each indicator. It can be observed that for the BPSK single-target signal separation task, SI-SINRi initially increases and then slightly decreases, while the estimated signal's SINR gradually increases and tends to stabilize at around 48. The correlation coefficient... When the SINR reaches 8 dB or higher, it reaches a value of 1, with an average SI-SINRi of 30.4222 dB and an average SINR of 37.1083 dB. The mean SINR is 0.9784. Compared to the true average SINR, the estimated signal's SINR is improved by 28 dB, indicating a significant and highly effective improvement in separating a single BPSK target signal. However, for dataset 2 containing five target signals, both SI-SINRi and SINR increase continuously with increasing SINR. But compared to dataset 1, at the same SINR, the values ​​of SI-SINRi and SINR are lower, and the correlation coefficient is also lower. Although it also gradually approaches 1 as SINR increases, and the SINR of its estimated signal is improved by nearly 12dB, the overall value and the approach speed are slower than those of dataset 1. This reflects that in the complex task of extracting five target signals, although the signal separation performance improves with the increase of signal-to-interference-plus-noise ratio, the SI-SINRi, SINR and correlation performance of dataset 1, which only extracts a single BPSK target signal, are weaker than those of dataset 1, which involves multiple types of target signals. This shows that the complexity of multi-target signal separation task has a certain impact on the signal separation effect.

[0130] Table 3 Performance indicators under different SINR conditions

[0131]

[0132] Furthermore, ablation experiments were conducted. To verify the overall improvement effect of the attention mechanism on target signal separation, Table 4 shows the average performance index achieved by the dataset under different combinations of attention mechanisms. It can be observed that for dataset 1, when neither cross-attention nor self-attention is used, the average SINR index of the estimated signal is ( The optimal performance is 38 dB. The lowest average performance is achieved when only cross-attention is used. When both cross-attention and self-attention are used, the average SI-SINRi ( ) and average correlation coefficient ( The optimal values ​​are 30.4222 dB and 0.9784 dB, respectively. For dataset 2, the lowest average performance index is achieved when neither cross-attention nor self-attention mechanisms are used. At this point, the average SINR of the estimated signal is improved by 9.5 dB compared to the mixed signal. As the attention mechanism evolves from none to only cross-attention, and then to a combination of cross-attention and self-attention, all performance indicators show a gradual optimization trend. The optimal average SINR can improve by nearly 12 dB compared to the original mixed signal. Overall, whether it is a single-target or multi-target signal scenario, the combined use of cross-attention and self-attention improves all three indicators to varying degrees compared to using attention mechanisms alone or without them. For the multi-target dataset 2, this combination of attention mechanisms has a more significant effect on performance improvement and better reflects its advantages in complex multi-target signal processing.

[0133] Table 4 Ablation Experiment

[0134]

[0135] Figure 7 This demonstrates the SINR and correlation coefficient of the estimated signal under different SINR conditions and different combinations of attention mechanisms. The influence of different attention mechanisms is illustrated in the figure, where the blue curves represent SINR changes and the orange curves represent correlation coefficients. The changes in SINR and SINR of the two datasets under various attention combination mechanisms can be observed. All show an upward trend. For dataset 1, at high SINR, the estimated signal SINR corresponding to no attention is better, while at low SINR, the correlation coefficients corresponding to the joint cross-attention and self-attention mechanisms are higher. Superior performance; for dataset 2, and under the same SINR, the combined cross-attention and self-attention mechanism exhibits the best performance, with the highest SINR and correlation coefficient. It is also closer to 1, and the performance under the no-attention mechanism is the worst. This shows that the combined use of cross-attention and self-attention can more effectively improve the quality and correlation of complex multi-target signal separation.

[0136] In practical applications, different combinations of attention mechanisms can be selected to meet the needs of different scenarios. In particular, for single-target extraction tasks, a combination without attention mechanisms can achieve optimal performance across various metrics. In more complex scenarios such as multi-target modulation signal separation, the combined use of cross-attention and self-attention can more significantly optimize the average SI-SINRi, average SINR, and average correlation coefficient. These key indicators better meet the performance requirements of signal separation.

[0137] In summary, the method proposed in this application breaks through the limitation of traditional techniques that require explicit perception and modeling of interference signals. By introducing the prior modulation features of the target signal and making full use of the powerful nonlinear fitting and feature learning capabilities of deep learning models, it can form an adaptive perception and extraction capability for interference noise conditions of different intensities and types. It can accurately capture the unique characteristics of the target signal in complex electromagnetic environments, thereby efficiently achieving the separation of the target source signal.

[0138] In one embodiment, an attention-based signal separation device is provided, comprising:

[0139] The signal acquisition module is used to acquire mixed signals under simultaneous and same-frequency interference scenarios. The mixed signals include the target signal and the interference signals that are at the same frequency but have different modulation types. It also simulates and generates multiple types of modulation signals as prior pure signals.

[0140] The prior information extraction module is used to input all prior pure signals into the prior signal feature extraction network trained based on the Arcface loss function for feature extraction, to obtain the embedding vectors of various modulation signals, and to obtain the embedding center of various modulation signals as the prior information for target signal separation by summing and averaging the various embedding vectors.

[0141] The network construction and training module is used to build and train a target signal separation network that includes a mixed-signal encoder, a modulation feature encoding module, a cross-attention module, and a mixed-signal decoder.

[0142] The target signal separation module is used to input the mixed signal into the trained target signal separation network, extract the mixed signal features according to the mixed signal encoder, extract the modulation features from the embedding center according to the modulation feature encoding module, and perform bidirectional cross-attention calculation on the mixed signal features and modulation features according to the cross-attention module to obtain the fused features focused on the target signal. Then, the fused features are decoded by the mixed signal decoder to reconstruct the output target estimation signal.

[0143] Specific limitations regarding the attention-based signal separation device can be found in the limitations of the attention-based signal separation method described above, and will not be repeated here. Each module in the aforementioned attention-based signal separation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0144] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 8 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 to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for separating signals of the same frequency 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.

[0145] Those skilled in the art will understand that Figure 8 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.

[0146] 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:

[0147] Acquire mixed signals under simultaneous and same-frequency interference scenarios. The mixed signals include the target signal and the interference signal that is simultaneous and same-frequency but has a different modulation type. Simulate and generate multiple types of modulation signals as prior pure signals.

[0148] All prior pure signals are input into a prior signal feature extraction network trained based on the Arcface loss function for feature extraction, and embedding vectors of various modulation signals are obtained. The embedding centers of various modulation signals are obtained by summing and averaging the embedding vectors, which serve as prior information for target signal separation.

[0149] Construct and train a target signal separation network that includes a mixed-signal encoder, a modulation feature encoding module, a cross-attention module, and a mixed-signal decoder;

[0150] The mixed signal is input into the trained target signal separation network. The mixed signal encoder extracts the mixed signal features, the modulation feature encoding module encodes and extracts the modulation features from the embedding center, and the cross attention module performs bidirectional cross attention calculation on the mixed signal features and modulation features to obtain the fused features focused on the target signal. The fused features are then decoded by the mixed signal decoder to reconstruct the output target estimation signal.

[0151] 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:

[0152] Acquire mixed signals under simultaneous and same-frequency interference scenarios. The mixed signals include the target signal and the interference signal that is simultaneous and same-frequency but has a different modulation type. Simulate and generate multiple types of modulation signals as prior pure signals.

[0153] All prior pure signals are input into a prior signal feature extraction network trained based on the Arcface loss function for feature extraction, and embedding vectors of various modulation signals are obtained. The embedding centers of various modulation signals are obtained by summing and averaging the embedding vectors, which serve as prior information for target signal separation.

[0154] Construct and train a target signal separation network that includes a mixed-signal encoder, a modulation feature encoding module, a cross-attention module, and a mixed-signal decoder;

[0155] The mixed signal is input into the trained target signal separation network. The mixed signal encoder extracts the mixed signal features, the modulation feature encoding module encodes and extracts the modulation features from the embedding center, and the cross attention module performs bidirectional cross attention calculation on the mixed signal features and modulation features to obtain the fused features focused on the target signal. The fused features are then decoded by the mixed signal decoder to reconstruct the output target estimation signal.

[0156] 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, and when executed, it 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 can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various 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.

[0157] 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.

[0158] 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. An attention mechanism-based co-frequency signal separation method, characterized in that, The method comprises: acquiring a mixed signal in a co-time and co-frequency interference scenario, the mixed signal comprising a target signal and an interference signal with different modulation types from the target signal, and generating a plurality of modulation signals as prior pure signals by simulation; inputting all prior pure signals into a prior signal feature extraction network trained based on an Arcface loss function to perform feature extraction, obtaining embedding vectors of the modulation signals, and obtaining embedding centers of the modulation signals as prior information for target signal separation by summing and averaging the embedding vectors; constructing and training a target signal separation network comprising a mixed signal encoder, a modulation feature encoding module, a cross-attention module, and a mixed signal decoder; inputting the mixed signal into the trained target signal separation network, extracting mixed signal features according to the mixed signal encoder, extracting modulation features from the embedding centers according to the modulation feature encoding module, performing bidirectional cross-attention calculation on the mixed signal features and the modulation features according to the cross-attention module, obtaining fusion features focused on the target signal, and then decoding the fusion features by the mixed signal decoder to reconstruct and output a target estimated signal; wherein all prior pure signals are input into a prior signal feature extraction network trained based on an Arcface loss function to perform feature extraction, obtain embedding vectors of the modulation signals, and obtain embedding centers of the modulation signals by summing and averaging the embedding vectors, comprising: The embedding vector output by the prior signal feature extraction network is flattened and input into an Arcface layer to obtain an output label, and network optimization training is performed by calculating an Arcface loss function, to obtain a trained prior signal feature extraction network ; a prior pure signal dataset input perform feature extraction to obtain an embedding vector set is represented as: ； in, This represents the total number of samples in the prior pure signal dataset. express Corresponding modulation tags, express go through The embedded vector obtained after extraction; Indicates the first Signal data of one sample; By summing and averaging the embedding vectors of each type of modulated signal in the embedding vector set , an embedding center set is obtained, where is the number of modulation types, and the embedding center of the th type of modulated signal is represented as: ； wherein, represents the number of samples of a class of modulated signals, represents the embedding vector of a modulated label .

2. The method of claim 1, wherein, the prior signal feature extraction network comprises three one-dimensional convolution layers, two residual blocks, and one one-dimensional convolution layer connected in sequence, and the first three convolution layers are followed by batch normalization layers and Tanh activation functions, and the last convolution layer is followed by a batch normalization layer and a Tanh activation function; the prior signal feature extraction network performs feature extraction by sequentially capturing temporal correlations, deepening local features, and extracting fine-grained features through the three one-dimensional convolution layers, strengthening features through the two residual blocks, and performing size constraint through the one one-dimensional convolution layer to output embedding vectors of the input modulation signals; wherein the main convolution path of the residual block combines two one-dimensional convolution layers, each followed by a batch normalization layer, and the first one-dimensional convolution layer is further followed by a ReLU activation function, and the main convolution path is used to realize feature extraction and conversion from input channels to output channels; the skip connection path of the residual block dynamically selects a processing mode according to whether the input channels and the output channels are consistent, when the input and output channel numbers of the main convolution path are different, the skip connection path adjusts the input channel number by using a one-dimensional convolution layer to match the output dimension of the main convolution path; when the input and output channel numbers of the main convolution path are the same, the skip connection path directly transmits the original input by identity mapping; the input processed by the skip connection path is added to the output of the main convolution path to realize residual fusion, and finally the ReLU activation function outputs the result. 3.The method of claim 1, wherein, The mixed signal encoder comprises a one-dimensional convolution layer, a residual shrinkage and self-attention module and two one-dimensional convolution layers connected in sequence, the residual shrinkage and self-attention module comprises a residual shrinkage bottleneck unit, a self-attention module, two residual shrinkage bottleneck units, a self-attention module and two residual shrinkage bottleneck units connected in sequence; wherein each layer of convolution is followed by a batch normalization layer and a ReLU activation function; According to the mixed signal encoder, the mixed signal features are extracted, comprising: The mixed signal is input into the mixed signal encoder, first passes through a one-dimensional convolution layer for preliminary feature extraction and dimension adjustment, then enters the residual shrinkage and self-attention module, eliminates the influence of noise according to the residual shrinkage bottleneck unit and learns the deep features of the signal, captures the long-distance dependence of the signal according to the self-attention module, excavates the correlation between the features, and finally passes through two one-dimensional convolution layers to deepen the signal features, finally completes the encoding of the mixed signal and outputs the mixed signal features; In the residual shrinkage bottleneck unit, the input data first passes through two one-dimensional convolution layers, then passes through a soft threshold function to obtain an output, and adds the output and the original input data to obtain the final output; wherein the input-output relationship of the soft threshold function satisfies: ； wherein, and respectively represent input features and output features, represents a threshold value, and the threshold value is determined in the following manner: first, an absolute value and a global average pooling are performed on a feature map with a size of to obtain a one-dimensional vector, then the one-dimensional vector passes through two fully connected layers to obtain a channel weight vector , then the weight vector is scaled to a range of by a Sigmoid activation function, and is represented as , represents the weight of the th channel, , is the total number of channels, represents the obtained scaling value; finally, the scaling values of all channels are multiplied by the average value of to obtain the threshold value , wherein, and respectively represent the width and the height of the feature map. In the self-attention module, the correlation degree of each position in the input data sequence with all positions is calculated, and then weighted summation is performed based on the correlation degrees, and the calculation formula is: ； wherein it is assumed that is an input data sequence, then the query matrix is a key matrix is a value matrix , , and are learnable weight matrices, which are essentially model parameters of one-dimensional convolution, denotes the dimension of the key matrix .

4. The method of claim 1, wherein, The modulation feature encoding module adopts a three-layer one-dimensional convolutional neural network structure, and each layer of convolution is followed by a batch normalization layer and a ReLU activation function; According to the modulation feature encoding module, the modulation features are encoded and extracted from the embedding center, comprising: All embedding centers are input into the modulation feature encoding module, first encoded by the first one-dimensional convolution layer to capture extensive local context information, then input into the second one-dimensional convolution layer to extract features by reducing the size of the convolution kernel and increasing the number of feature channels, and finally extract finer feature representations through the third one-dimensional convolution layer, and finally encode and extract modulation features of various modulation signals from the embedding center.

5. The method of claim 1, wherein, The cross-attention module comprises a first cross-attention unit and a second cross-attention unit; the mixed signal decoder adopts a four-layer one-dimensional convolutional neural network structure, wherein the first three layers of convolution are followed by a batch normalization layer and a ReLU activation function; After the cross-attention module performs bidirectional cross-attention calculation on the mixed signal features and the modulation features to obtain the fusion features focused on the target signal, the mixed signal decoder decodes the fusion features to reconstruct and output the target estimated signal, comprising: The mixed signal features extracted by the mixed signal encoder and the modulated feature encoded by the modulated feature encoding module are fed into the cross attention module. In the first cross attention unit, a query matrix is constructed according to the mixed signal features, a key matrix and a value matrix are constructed according to the modulated features, and the cross attention is used to calculate the part of the prior information related to the target signal. In the second cross attention unit, a query matrix is constructed according to the modulated features, a key matrix and a value matrix are constructed according to the mixed signal features, and the cross attention is used to calculate the features in the mixed signal features that match the prior information. Finally, the features output by the two cross attention calculations are added to obtain the fusion features focused on the target signal and input into the mixed signal decoder. The fusion features are subjected to step-by-step dimension transformation and feature extraction operations by the four-layer one-dimensional convolutional neural network structure, and the target estimated signal is output.

6. The method of claim 1, wherein, The target signal separation network adopts a multi-target loss function combination strategy for training, and a total loss function is represented as: ； wherein, is a scale-invariant signal-to-jamming-noise ratio loss, is a scale-invariant signal-to-jamming-noise ratio, is a mean square error loss, is a number of samples, is a reference signal for the th sample, is an estimated signal for the th sample, is a loss weighting coefficient.

7. An attention mechanism based co-frequency signal separation device, characterized in that, The device comprises: A signal acquisition module is configured to acquire a mixed signal in a co-channel interference scenario, wherein the mixed signal comprises a target signal and an interference signal with different modulation types, and a plurality of modulation signals are simulated and generated as prior pure signals; A prior information extraction module is configured to input all prior pure signals into a prior signal feature extraction network trained based on an Arcface loss function to extract features, obtain embedding vectors of the modulation signals, and obtain embedding centers of the modulation signals as prior information for target signal separation by summing and averaging the embedding vectors; A network construction and training module is configured to construct and train a target signal separation network comprising a mixed signal encoder, a modulated feature encoding module, a cross attention module, and a mixed signal decoder; A target signal separation module is configured to input the mixed signal into the trained target signal separation network, extract mixed signal features according to the mixed signal encoder, extract modulated features from the embedding centers according to the modulated feature encoding module, and perform bidirectional cross attention calculation on the mixed signal features and the modulated features according to the cross attention module to obtain fusion features focused on the target signal, and then decode the fusion features by the mixed signal decoder to reconstruct and output a target estimated signal. The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 6. The embedding vector output by the prior signal feature extraction network is flattened and input into an Arcface layer to obtain an output label, and network optimization training is performed by calculating an Arcface loss function, to obtain a trained prior signal feature extraction network ; a prior pure signal dataset input perform feature extraction to obtain an embedding vector set is represented as: ； wherein, is the total number of samples in the prior pure signal dataset, represents the corresponding modulation label, represents the signal data of the i-th sample after the embedding vector extracted therefrom; represents the signal data of the i-th sample; th sample. The embedding center set is obtained by sum average of embedding vectors of each type of modulation signal in the embedding vector set wherein is the number of modulation types, and the embedding center of the mth type of modulation signal is represented as: ​​ ； wherein, represents the number of samples of a class of modulated signals, represents the embedding vector of a modulated label . 8.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-7. The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 6.

Images