A Single-Channel Blind Source Separation Method and System Based on Multi-Scale Feature Fusion Complex Neural Network

By using an end-to-end processing architecture based on multi-scale feature fusion complex neural networks, the problem of poor separation performance of single-channel blind source separation methods in low signal-to-noise ratio dynamic wireless environments is solved. This achieves high-precision and robust separation of complex communication signals, preserves the phase information of the signals, and reduces computational complexity.

CN123123967APending Publication Date: 2026-11-17NAT UNIV OF DEFENSE TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611358004.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-03
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing single-channel blind source separation methods perform poorly in low signal-to-noise ratio dynamic wireless environments, rely heavily on statistical assumptions, and current deep learning networks cannot natively process complex communication signals, resulting in severe loss of phase information and making it difficult to achieve high-precision and robust signal separation in complex electromagnetic environments.

Method used

An end-to-end processing architecture based on multi-scale feature fusion complex neural networks is adopted. Through full complex domain operations, combined with multi-scale feature fusion masking networks and complex decoders, ModReLU activation function is designed to achieve efficient separation of complex communication signals while preserving phase information. Furthermore, time-frequency dynamic characteristics at different granularities are modeled synchronously through a dual-path iterative structure and multi-scale complex deep convolution modules.

Benefits of technology

High-precision and robust signal separation was achieved in complex electromagnetic environments, preserving the phase integrity of physical layer signals, reducing modeling and computational complexity and the number of parameters, and improving separation performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123967A_ABST
    Figure CN123123967A_ABST
Patent Text Reader

Abstract

This invention provides a single-channel blind source separation method and system based on a multi-scale feature fusion complex neural network, relating to the fields of wireless communication physical layer security and digital signal processing technology. The method includes: using a test set to perform separation evaluation testing on a trained multi-scale feature fusion complex neural network; constructing a hardware test platform to generate a mixed signal formed by the target signal and interference signal through a vector signal generator, and collecting measured signal data; segmenting the collected measured signal data to obtain a fixed-length complex-form aliased signal; inputting the complex-form aliased signal into the trained multi-scale feature fusion complex neural network to separate the target signal from the complex-form aliased signal, thus completing single-channel blind source separation. This invention can achieve high-precision and highly robust blind separation of wireless communication interference signals in complex electromagnetic environments.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of wireless communication physical layer security and digital signal processing technology, and in particular to a single-channel blind source separation method and system based on a complex neural network with multi-scale feature fusion. Background Technology

[0002] With the rapid development of wireless communication technology towards high density and high dynamics, the electromagnetic spectrum is becoming increasingly crowded, and communication systems are facing unprecedented interference threats. Due to the open broadcast characteristics of wireless channels, transmitted signals are highly susceptible to environmental noise and interference from nearby devices operating on the same frequency. Especially in complex electromagnetic environments, high-power unintentional radiation sources or malicious jammers often generate strong interference that highly overlaps with the target signal in both the time and frequency domains, severely damaging the integrity of data reception and posing a substantial challenge to the reliability of physical layer communication.

[0003] To ensure effective wireless link connectivity, the industry has deployed anti-interference strategies such as frequency hopping spread spectrum, beamforming, and active interference detection. However, these traditional anti-interference methods often fail when the target signal and the interfering signal deeply overlap in both the time and frequency domains. Blind source separation technology aims to recover the source signals from the mixed observation signals, becoming a feasible technical solution in the above-mentioned scenarios. Among them, multi-channel blind source separation algorithms (such as blind separation algorithms based on antenna arrays) can improve separation performance by utilizing spatial diversity, but their performance is highly dependent on the number of sensors and the array manifold structure, making them difficult to deploy in single-antenna receiving nodes with limited size, power consumption, and cost. Therefore, single-channel blind source separation technology, which recovers the source signal using only a single-channel observation signal, has become the core defense means to ensure the anti-interference security of such receiving nodes.

[0004] Existing single-channel blind source separation methods are mostly based on statistical frameworks such as independent component analysis and negative matrix factorization, which have significant limitations in low signal-to-noise ratio dynamic wireless environments and strongly rely on the basis statistical assumptions of the source signal. In recent years, driven by the progress of deep learning speech separation technology, researchers have introduced models such as fully convolutional temporal separation networks, dual-path iterative structure networks, and Transformer architectures for signal separation. However, the above models are all designed for real-valued speech signals, and directly transferring them to the field of wireless communication has inherent defects: on the one hand, most models cannot natively process complex signals. If the in-phase and quadrature components of the communication signal are simply separated into independent real channels for processing, the complex algebraic relationship between the two will be severed, seriously damaging the phase integrity of the physical layer signal; on the other hand, some complex recursive networks exhibit a sharp degradation in separation performance when faced with non-ideal channel impairments such as carrier frequency offset and sudden interruptions; in addition, communication signals have multi-scale time-frequency dynamic characteristics, such as microscopic instantaneous phase fluctuations and macroscopic symbol envelopes, while existing networks use single-scale convolutional layers, making it difficult to simultaneously model the evolution laws of signals at these different granularities. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a single-channel blind source separation method and system based on multi-scale feature fusion complex neural network, which can efficiently compress features while perfectly preserving phase information, thereby achieving high-precision and high-robust blind separation of wireless communication interference signals in complex electromagnetic environments.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0007] Firstly, a single-channel blind source separation method based on a multi-scale feature fusion complex neural network includes a training phase and an application phase, wherein the training phase includes the following steps:

[0008] Create a mixed signal dataset. The mixed signal in the dataset is a single-channel superimposed signal of target signal, interference signal and environmental noise. Carrier frequency offset and channel impairment due to sudden interruption are introduced into the mixed signal. The mixed signal dataset is divided into training set, validation set and test set.

[0009] A fully complex algebraic computation environment is constructed, and a multi-scale feature fusion complex neural network is built. The multi-scale feature fusion complex neural network includes a complex encoder, a multi-scale feature fusion mask network, and a complex decoder.

[0010] Create a network training loss function, which consists of a weighted sum of scale-invariant signal-to-noise ratio and mean square error;

[0011] The multi-scale feature fusion complex neural network was trained and validated using the training set and validation set to obtain the trained multi-scale feature fusion complex neural network.

[0012] The application phase includes the following steps:

[0013] The trained multi-scale feature fusion complex neural network was subjected to a separation evaluation test using the test set.

[0014] A hardware testing platform was constructed, which generates a mixed signal of target signal and interference signal through a vector signal generator, and collects measured signal data.

[0015] The collected measured signal data is segmented to obtain a complex aliased signal of fixed length;

[0016] The target signal is separated from the complex aliased signal by inputting it into a trained multi-scale feature fusion complex neural network, thus completing single-channel blind source separation.

[0017] Secondly, a single-channel blind source separation system based on a multi-scale feature fusion complex neural network is provided, the system comprising a training module and an application module, wherein:

[0018] The training module includes a dataset creation unit, a network building unit, a loss function creation unit, and a network training unit, among which:

[0019] The dataset creation unit is used to create a dataset of single-channel aliased signals. The dataset includes complex baseband signals resulting from the mixing of the target signal and the interference signal using different modulation methods.

[0020] The network construction unit is used to build a multi-scale feature fusion complex neural network, which includes a complex encoder, a multi-scale feature fusion mask network, and a complex decoder.

[0021] The loss function creation unit is used to create the network training loss function;

[0022] The network training unit is used to train and validate the multi-scale feature fusion complex neural network obtained by the network building unit based on the network training loss function and the dataset obtained by the dataset creation unit. The network parameters are iteratively updated through backpropagation to obtain the trained multi-scale feature fusion complex neural network.

[0023] The application module includes a signal acquisition unit, an aliasing signal processing unit, and a blind separation unit, wherein:

[0024] The signal acquisition unit is used to acquire the time-frequency aliasing signal to be separated through simulation software and a vector signal generator;

[0025] The aliasing signal processing unit is used to segment the time-frequency aliased signal to obtain a baseband aliased signal of fixed length in complex form.

[0026] The above-described solution of the present invention has at least the following beneficial effects:

[0027] This application constructs an end-to-end processing architecture in the full complex domain, which can directly model the in-phase and quadrature components of wireless communication signals. During the training process, it makes full use of the inherent phase correlation in the complex domain, avoiding the phase information loss caused by splitting the in-phase and quadrature components separately in traditional real-valued methods. It completely protects the phase integrity of the physical layer signal at the architecture level.

[0028] The multi-scale feature fusion masking network designed in this application adopts a dual-path iterative structure. By segmenting and recombining the encoded features and alternately executing intra-segment modeling and inter-segment modeling, it fully explores the local temporal features and global dependencies of the mixed signal. At the same time, intra-segment modeling and inter-segment modeling share the parameters of the same multi-scale complex Conformer module, which significantly reduces the computational complexity of modeling long temporal sequences and the number of network parameters while maintaining modeling capabilities.

[0029] The multi-scale complex depth convolution module designed in this application employs three parallel complex channel-wise depth convolution branches with different kernel sizes to capture the low-frequency, mid-frequency, and high-frequency features of communication signals, respectively. This enables synchronous modeling of the time-frequency dynamic characteristics of communication signals at different granularities, such as microscopic instantaneous phase fluctuations and macroscopic symbol envelopes, overcoming the shortcomings of existing single-scale convolutional layers that cannot take into account the features of different time scales.

[0030] This application employs the ModReLU activation function in the network. This activation function only performs nonlinear scaling on the magnitude of complex features while fully preserving the phase angle, fundamentally eliminating the phase distortion introduced by the nonlinear activation process, and further ensuring the phase accuracy of complex features during deep network transmission. Attached Figure Description

[0031] Figure 1 This is a flowchart of the multi-scale feature fusion network based on the present invention.

[0032] Figure 2 This is a schematic diagram of the core architecture of the multi-scale feature fusion masking network of the present invention.

[0033] Figure 3 This is a schematic diagram of the multi-scale complex depth convolution architecture of the present invention.

[0034] Figure 4 This is a graph showing the symbol error rate of the signal separation in this invention. Detailed Implementation

[0035] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0036] To address the inherent limitations of traditional single-channel blind source separation methods, such as poor separation performance in low signal-to-noise ratio dynamic wireless environments and high dependence on statistical assumptions, as well as the technical shortcomings of existing deep learning-based speech separation networks that are designed for real-valued signals and cannot natively adapt to complex communication signals, and whose phase information is severely damaged when splitting in-phase and quadrature components, this embodiment establishes a fully complex domain end-to-end trainable multi-scale feature fusion complex neural network specifically for single-channel blind source separation of complex communication signals.

[0037] It should be noted that single-channel blind source separation refers to the process where the receiver obtains only one observation signal through a single antenna. This observation signal is superimposed with the target signal, interference signals, and environmental noise. Under the condition that the prior information of the source signal is unknown, the target signal is recovered solely from this single observation signal. Communication signals are typically represented in complex baseband form, consisting of in-phase and quadrature components. The in-phase component is the projection of the signal onto the real axis, and the quadrature component is the projection of the signal onto the imaginary axis. Together, they are called I / Q components, and the amplitude and phase correlation between them together carry the complete modulation information of the communication signal.

[0038] like Figure 1 As shown, the single-channel blind source separation method based on multi-scale feature fusion complex neural network in this embodiment includes a training phase and an application phase. The training phase includes steps S1 to S4, and the application phase includes steps S5 to S8. The specific implementation process of each step is as follows.

[0039] The specific implementation process of the training phase is as follows:

[0040] Step S1, create a mixed signal dataset:

[0041] The purpose of step S1 is to construct single-channel aliasing signal samples that can cover both ideal and non-ideal channel conditions, providing a data foundation for network training and performance verification. In this embodiment, the target signal is a QPSK modulated communication signal, the interference signals include BPSK modulated, 8PSK modulated, and 16QAM modulated communication signals, and the environmental noise is Gaussian white noise. Gaussian white noise refers to random noise with a power spectral density uniformly distributed across the entire frequency domain and an amplitude following a Gaussian distribution.

[0042] A time-frequency aliasing communication signal model under ideal channel conditions is established. The generation rule is as follows: the target signal is multiplied point-by-point by the first complex gain channel mixing coefficient to obtain the first mixed component; the interference signal is multiplied point-by-point by the second complex gain channel mixing coefficient to obtain the second mixed component; the first and second mixed components are added point-by-point, and then Gaussian white noise is superimposed point-by-point to obtain the single-channel time-frequency aliasing communication signal under ideal channel conditions. The complex gain channel mixing coefficient is used to characterize the amplitude attenuation and phase rotation effects of the wireless channel on the corresponding source signal; its value is an unknown complex number to simulate the condition of missing prior channel information in blind separation scenarios.

[0043] To fully verify the robustness of the separation model under non-ideal channel conditions, the above ideal mixed signal model is further extended into a generalized mixed signal model, whose generation rules are as follows:

[0044] The first step is to introduce carrier frequency offsets for both the target signal and the interfering signal. Carrier frequency offset refers to the deviation of the received signal's carrier frequency relative to the nominal frequency caused by the inconsistency of the local oscillator frequencies at the transmitting and receiving ends and the Doppler effect. The method is as follows: the product of the discrete sampling number and the normalized carrier frequency offset is used as the phase step. This phase step is added to the initial carrier phase of the corresponding source signal to obtain the total phase of the current sampling point. A power operation is performed with the natural constant e as the base and the product of the total phase and the imaginary unit as the exponent to obtain the rotation factor for the current sampling point. Each sampling point of the source signal is multiplied by the corresponding rotation factor to complete the introduction of the carrier frequency offset for the source signal.

[0045] The second step involves introducing sudden interruption channel impairments to both the target and interfering signals. Sudden interruptions refer to the intermittent damage to the signal during transmission caused by momentary obstruction from obstacles or intermittent transmission from interfering sources. This is achieved by constructing a discrete characteristic function, where the value is either 0 or 1. A value of 1 indicates normal signal transmission during the current sampling period, while a value of 0 indicates signal interruption. The source signal, after carrier frequency offset processing, is then multiplied point-by-point by this discrete characteristic function, thus simulating the sudden interruption channel impairment.

[0046] The third step involves multiplying the target signal after carrier frequency offset and burst interruption processing with the first complex gain channel mixing coefficient point by point, multiplying the interference signal after carrier frequency offset and burst interruption processing with the second complex gain channel mixing coefficient point by point, adding the two product results point by point, and then superimposing Gaussian white noise point by point to obtain the generalized mixed signal.

[0047] In terms of simulated signal evaluation, computer simulation software generates communication signals with various modulation schemes as target and interference signals. These signals are randomly mixed to form an aliased signal, and noise is added to generate a time-frequency aliased communication signal. Simulation parameters are set as follows: signal-to-noise ratio (SNR) range of -5 dB to 20 dB, interference-to-signal ratio (ISR) range of -5 dB to 10 dB, carrier frequency offset range of -0.01 to 0.01, and burst interruption length of one-quarter to two-thirds of the total signal length. Here, SNR refers to the ratio of signal power to noise power, and ISR refers to the ratio of interference signal power to target signal power. The discrete sampling length for all signals is set to 1024 sampling points, generating 1500 samples under each SNR condition, for a total of 9000 samples. The generated dataset is randomly divided into training, validation, and test sets in an 8:1:1 ratio. The training set is used for learning and updating network parameters, the validation set is used for periodic evaluation of network performance during training, and the test set is used for the final evaluation of network separation performance after training. All complex baseband signal samples are stored in a stacked form of in-phase and quadrature components. Each sample constitutes a two-dimensional matrix. The first dimension of the matrix is ​​the component dimension, which contains two elements: the in-phase component and the quadrature component. The second dimension of the matrix is ​​the time dimension, which contains 1024 discrete sampling points. After adding a batch dimension before the component dimension, a batch of samples forms a three-dimensional tensor data consisting of the batch dimension, the component dimension, and the time dimension, which serves as the input to the network.

[0048] In terms of evaluating the measured signal, a hardware test platform was built, in which a vector signal generator produced a mixed signal formed by the target signal and the interference signal, with the interference-to-signal ratio set to 0 dB to 20 dB; a software-defined radio platform collected data of the mixed signal through a single antenna, and after segmentation processing, formed a measured test set.

[0049] Step S2, construct a multi-scale feature fusion complex neural network:

[0050] The purpose of step S2 is to construct an end-to-end signal separation network in which all operations are performed in the complex domain. This embodiment constructs a fully complex algebraic computation environment within the PyTorch deep learning framework. The constructed multi-scale feature fusion complex neural network mainly consists of three parts: a complex encoder, a multi-scale feature fusion masking network, and a complex decoder. Figure 2As shown, the overall signal processing flow of the network is as follows: the complex encoder extracts features from the input single-channel complex-valued time-domain aliased signal and maps it to a high-dimensional complex feature space; the multi-scale feature fusion mask network performs deep modeling of the high-dimensional complex features and generates a complex mask, which is applied to the encoded features through point-by-point complex multiplication to obtain the enhanced features corresponding to the target signal; the complex decoder reconstructs the enhanced features layer by layer to recover the time-domain complex waveform of the target signal. The specific construction process of each part is as follows.

[0051] A complex encoder is constructed, comprising a complex one-dimensional convolutional layer. In this embodiment, the kernel size of the complex one-dimensional convolutional layer is set to 16, the stride to 8, and the number of output feature channels to 64. The input single-channel complex time-domain signal consists of in-phase and quadrature components. The complex one-dimensional convolutional layer performs joint parallel convolution operations on the in-phase and quadrature components according to the rules of complex algebra. The specific operation rules are as follows:

[0052] The real part of the complex one-dimensional convolution kernel is convolved with the in-phase component of the input signal to obtain the first convolution result; the imaginary part of the complex one-dimensional convolution kernel is convolved with the quadrature component of the input signal to obtain the second convolution result; the first convolution result and the second convolution result are subtracted point by point, and the difference is taken as the in-phase component of the convolution output.

[0053] The real part of the complex one-dimensional convolution kernel is convolved with the orthogonal component of the input signal to obtain the third convolution result; the imaginary part of the complex one-dimensional convolution kernel is convolved with the in-phase component of the input signal to obtain the fourth convolution result; the third convolution result and the fourth convolution result are added point by point, and the sum is used as the orthogonal component of the convolution output.

[0054] The rules for one-dimensional convolution are as follows: the convolution kernel slides along the time dimension of the signal with a preset stride. At each sliding position, each element of the convolution kernel is multiplied point by point with each element in the corresponding window of the signal, and the sum is taken as the convolution output value at that position. The in-phase and quadrature components of the convolution output together constitute the coding features in the complex field.

[0055] After the above joint convolution operation, the input signal is mapped to a high-dimensional complex feature space. The output encoded feature is a four-dimensional tensor with four dimensions: batch dimension, feature channel dimension, feature length dimension, and complex component dimension. The feature channel dimension contains 64 feature channels, the complex component dimension contains two elements: in-phase component and quadrature component, and the length of the feature length dimension is determined by the sampling length of the input signal, the convolution kernel size, and the stride.

[0056] A multi-scale feature fusion masking network is constructed, which consists of a cascaded dual-path iterative structure with a preset number of layers. In this embodiment, the preferred number of layers in the dual-path iterative structure is two. The processing flow of the input features in each layer of the dual-path iterative structure is as follows: segmentation and recombination, intra-segment modeling, and inter-segment modeling.

[0057] The first step is segmentation and recombination. Zero-padding is performed on the end of the time dimension of the input features, that is, zero-valued elements are added to the end of the feature sequence so that the feature length can be divided by a preset segment length. In this embodiment, the preset segment length is preferably 32. The features are recombined from a four-dimensional tensor consisting of batch dimension, feature channel dimension, feature length dimension and complex component dimension into a five-dimensional tensor consisting of batch dimension, segment number dimension, segment length dimension, feature channel dimension and complex component dimension. The size of the segment number dimension is the quotient of the padded feature length and the preset segment length, and the size of the segment length dimension is equal to the preset segment length. After recombination, the feature sequence is divided into multiple local sub-blocks that are connected end to end.

[0058] The second step is intra-segment modeling. The batch dimension and segment number dimension of the segmented five-dimensional tensor are merged and rearranged into a four-dimensional tensor consisting of the merged batch dimension, intra-segment length dimension, feature channel dimension, and complex component dimension. This tensor is then input into the multi-scale complex Conformer module, which models and enhances the temporal correlation within each local sub-block and outputs intra-segment enhancement features, thereby extracting the local temporal features of the mixed signal.

[0059] The third step is inter-segment modeling. The features output from intra-segment modeling are restored to a five-dimensional tensor consisting of batch dimension, segment number dimension, intra-segment length dimension, feature channel dimension, and complex component dimension. Then, the intra-segment length dimension and segment number dimension are interchanged through dimension permutation. Subsequently, the batch dimension and intra-segment length dimension are merged to obtain a four-dimensional tensor consisting of the merged batch dimension, segment number dimension, feature channel dimension, and complex component dimension. This four-dimensional tensor is then input into the same multi-scale complex Conformer module that shares parameters with the intra-segment modeling. This module models the long-range dependencies between different local sub-blocks and outputs inter-segment enhancement features, thereby extracting the global dependencies of the mixed signal.

[0060] Intra-segment modeling and inter-segment modeling are executed alternately once in each layer of the dual-path iterative structure, using the same multi-scale complex Conformer module with shared parameters, thus completing a full local and global joint modeling process. The significance of adopting the dual-path iterative structure is that direct global modeling of long time-series signals involves enormous computational costs. However, segmenting the signal first and then modeling on both intra-segment and inter-segment paths can capture both local details and global contextual information of the signal with a computationally much lower complexity than direct global modeling.

[0061] A multi-scale complex Conformer module is constructed, which adopts a cascaded sandwich residual structure. The input features pass through four sub-layers in sequence: the first half-step feedforward network, the multi-head self-attention mechanism, the multi-scale complex depthwise convolution, and the second half-step feedforward network. The specific processing of each sub-layer is as follows.

[0062] The first sublayer, the first half-step feedforward network, first processes the input features through a complex layer normalization process. The normalization rule is as follows: for each complex element of the feature, the mean and variance are calculated according to the feature channel dimension. The mean is subtracted from each complex element, and then the result is divided by the standard deviation to obtain the normalized features, stabilizing the feature distribution. The normalized features are then fed into the complex feedforward network, which consists of two levels of complex linear transformation and inter-level nonlinear activation, used for nonlinear transformation and channel dimension mapping of the features. The output of the complex feedforward network is multiplied by a scaling factor of 0.5 and then subjected to Dropout regularization. The result is then added point-by-point to the input features of the multi-scale complex Conformer module to form residual connections, achieving initial feature enhancement and stabilizing training. The Dropout regularization rule is as follows: during training, some elements in the features are randomly set to zero according to a preset dropout probability, and the remaining elements are scaled by the reciprocal of the retention probability to prevent overfitting; during testing, no features are dropped.

[0063] The second sub-layer employs a multi-head self-attention mechanism. After feature normalization, positional encoding is introduced to preserve temporal order information. Positional encoding is achieved by generating a unique encoding vector for each position in the feature sequence, adding the encoding vector point-by-point to the feature at that position, enabling the network to perceive the sequential order of feature elements. Subsequently, features carrying positional information are fed into the multi-head self-attention mechanism. Its operational rules are as follows: features are mapped to query features, key features, and value features through three independent linear transformations; these features are equally divided into multiple heads along the feature channel dimension; within each head, the similarity between the query feature and the key features at each position is calculated to obtain attention weights; these attention weights are normalized and then weighted and summed with the value features to obtain the attention output for that head; the attention outputs of each head are concatenated along the feature channel dimension and then mapped again through a linear transformation to obtain the multi-head self-attention output; after Dropout regularization, the multi-head self-attention output is added point-by-point to the current feature to form residual connections, achieving global information fusion.

[0064] The third sublayer uses multi-scale complex depthwise convolutions, such as... Figure 3As shown, the features after complex layer normalization are input in parallel into three complex one-dimensional channel-wise depthwise convolution branches. The kernel sizes of the three branches are set to 3, 7, and 15, respectively, to extract local features under different receptive fields. The three kernel sizes correspond to low-frequency feature extraction, mid-frequency feature extraction, and high-frequency feature extraction of the communication signal, respectively. Channel-wise depthwise convolution means that each feature channel uses an independent one-dimensional convolution kernel for convolution operation, without information mixing between channels, thus extracting temporal local features within each channel with extremely low parameter count. The convolution operations of the three branches all follow the aforementioned rules for complex joint parallel convolution operations. The three convolution outputs are concatenated along the feature channel dimension to form a high-density feature map. Subsequently, the high-density feature map is fed into a complex one-dimensional convolution layer for cross-scale feature fusion, allowing full interaction of feature information at the three scales. The fused features are then processed by the ModReLU activation function and Dropout regularization, and then added point-by-point with the current feature to form residual connections, achieving multi-scale information fusion.

[0065] The ModReLU activation function operates as follows: The modulus of the complex feature is calculated, which is the square root of the sum of the squares of the in-phase and quadrature components of the complex feature. This modulus is then added to the learnable bias parameter to obtain the bias modulus. If the bias modulus is greater than zero, it is used as the correction modulus; otherwise, zero is used. The ratio of the correction modulus to the original modulus is used as the scaling factor. This scaling factor is then multiplied point-by-point by the original complex feature to obtain the activation output. This activation function only performs nonlinear scaling on the modulus of the complex feature, completely preserving the phase angle and fundamentally eliminating phase distortion.

[0066] The fourth sub-layer is the second half-step feedforward network. After the features are normalized by the complex layer, they are fed into the complex feedforward network again for nonlinear transformation and channel dimension mapping. The output is multiplied by a scaling factor of 0.5 and then processed by Dropout regularization. It is then added point by point with the current feature to form a residual connection, thus completing the full forward propagation of this multi-scale complex Conformer module.

[0067] The mask generation layer and complex mask are applied after the dual-path iterative structure with all preset layers completes feature modeling. The output features are then fed into a mask generation layer consisting of a single complex one-dimensional convolutional layer to generate a complex mask. A complex mask is a complex weighted matrix corresponding to the dimension of the encoded features; the value of each element represents the degree to which the complex features at the corresponding position and channel belong to the target signal. Subsequently, the complex mask is applied to the encoded features output by the complex encoder using pointwise complex multiplication to obtain the enhanced features corresponding to the target signal. The specific rules for pointwise complex multiplication are as follows:

[0068] The in-phase component of the complex mask is multiplied point by point with the in-phase component of the coded feature to obtain the first mask product; the quadrature component of the complex mask is multiplied point by point with the quadrature component of the coded feature to obtain the second mask product; the first mask product and the second mask product are subtracted point by point, and the difference is taken as the in-phase component of the enhanced feature.

[0069] The in-phase component of the complex mask is multiplied point by point with the orthogonal component of the coding feature to obtain the third mask product; the orthogonal component of the complex mask is multiplied point by point with the in-phase component of the coding feature to obtain the fourth mask product; the third mask product and the fourth mask product are added point by point, and the sum is used as the orthogonal component of the enhanced feature.

[0070] A complex decoder is constructed, employing a structure corresponding to the complex encoder, including a one-dimensional complex transposed convolutional layer. In this embodiment, the kernel size and stride of this one-dimensional complex transposed convolutional layer are set to 16 and 8, respectively. The operation rule of the one-dimensional complex transposed convolution is as follows: each element of the input feature is multiplied by the convolution kernel to obtain several weighted local sequences. These local sequences are then stacked onto the corresponding positions of the output sequence according to the stride, thereby expanding and restoring the short high-dimensional feature sequence into a long time-domain sequence. Based on the rules of complex algebra operations, the complex decoder jointly processes the in-phase and quadrature components of the enhanced features, gradually mapping the high-dimensional complex features back to the time domain, reconstructing the single-channel complex-valued time-domain waveform of the target signal. The output estimated signal is a two-dimensional matrix, with the first dimension being the time dimension and the second dimension being the complex component dimension, containing two elements: the in-phase component and the quadrature component.

[0071] Step S3: Create the network training loss function. The network training loss function created in this embodiment takes the difference between the average mean square error of the in-phase and quadrature components of the complex signal and the average scale-invariant signal-to-noise ratio of the in-phase and quadrature components as the optimization objective. The specific construction process is as follows.

[0072] Calculate the scale-invariant signal-to-noise ratio (SNR); the scale-invariant SNR is the ratio of the energy of the target component to the energy of the distortion component in the estimated signal after eliminating the effects of overall signal amplitude scaling. It is used to measure the fidelity of the separated signal, and its calculation rules are as follows:

[0073] The first dot product is obtained by multiplying the estimated signal and the source signal point by point and summing the results. The second dot product is obtained by multiplying the source signal and itself point by point and summing the results. The projection coefficients are obtained by dividing the first dot product and the second dot product. The target component is obtained by multiplying the projection coefficients and the source signal point by point. The target component is the effective component in the estimated signal whose waveform is consistent with that of the source signal. The noise component is obtained by subtracting the estimated signal and the target component point by point. The noise component is the residual interference and distortion components in the estimated signal. The energy ratio is obtained by dividing the square root of the sum of squares of the elements of the target component by the square root of the sum of squares of the elements of the noise component. The scale-invariant signal-to-noise ratio is obtained by taking the logarithm to base 10 of the energy ratio and multiplying the resulting logarithm by 10. The unit is decibel.

[0074] Calculate the mean square error; the mean square error is used to measure the point-by-point deviation between the estimated signal and the source signal. The calculation rule is as follows: subtract the estimated signal from the source signal point by point to obtain the error sequence; square each element of the error sequence and take the average value to obtain the mean square error.

[0075] The scale-invariant signal-to-noise ratio (SNR) is calculated for both the in-phase and quadrature components of the complex signal according to the above rules, and the two calculation results are averaged to obtain the two-channel average scale-invariant SNR. The mean square error (MSE) is also calculated for both the in-phase and quadrature components according to the above rules, and the two calculation results are averaged to obtain the two-channel average MSE. The difference between the two-channel average MSE and the two-channel average scale-invariant SNR is used as the network training loss function. Since a larger scale-invariant SNR indicates higher separation quality and a smaller MSE indicates smaller separation deviation, minimizing this difference as the loss function is equivalent to simultaneously maximizing the scale-invariant SNR of the separated signal and minimizing the point-by-point estimation deviation, thereby ensuring both overall waveform fidelity and sampling-point-level reconstruction accuracy.

[0076] Step S4: Network training and validation; parameter optimization of the multi-scale feature fusion complex neural network built in step S2. The specific implementation rules are as follows:

[0077] The complex baseband signal samples in the training and validation sets are processed in batches according to a preset batch size. The complex tensor data of each batch are used as the input of the multi-scale feature fusion complex neural network. The network performs forward propagation on the input batches, outputs the estimated signal of each sample, and calculates the loss value between the estimated signal and the source signal according to the network training loss function created in step S3. The gradient of the loss value with respect to each trainable parameter of the network is calculated through backpropagation, and the network parameters are iteratively updated using an adaptive moment estimation optimizer. The adaptive moment estimation optimizer is an optimization algorithm that adaptively adjusts the update step size of each parameter according to the first-order moment estimation and second-order moment estimation of the gradient. In this embodiment, an adaptive moment estimation optimizer combined with weight decay is used, that is, an additional decay term proportional to the parameter value is applied when updating the parameters to suppress overfitting.

[0078] During training, a learning rate scheduler is used to dynamically adjust the learning rate. Specifically, when the evaluation metric on the validation set plateaus, the current learning rate is multiplied by a preset decay factor to obtain a decayed learning rate, allowing the network to converge more finely with smaller step sizes in the later stages of training. An early stopping strategy is employed to prevent overfitting. Specifically, if the network fails to achieve a positive improvement in the evaluation metric on the validation set for a preset number of consecutive epochs, training is immediately terminated, and the model weights at their optimal performance on the validation set are saved. Furthermore, during each backpropagation, the gradients of all trainable parameters are pruned using the maximum norm. Specifically, the total norm of the vector formed by the gradients of all parameters is calculated. When the total norm exceeds a preset pruning threshold, all gradients are reduced proportionally to the ratio of the pruning threshold to the total norm to ensure numerical stability during training and suppress gradient explosion.

[0079] The specific implementation in the application phase includes the following:

[0080] Step S5: Load the optimal model weight parameters saved in Step S4, and feed all test samples from the simulation test set reserved in Step S1 into the trained network for forward propagation to obtain the separation estimation signals for each sample. Based on the estimated signals and the real source signals, calculate quantitative evaluation indicators such as Pearson correlation coefficient, symbol error rate, signal distortion ratio, and error vector magnitude. Specifically, the Pearson correlation coefficient measures the linear correlation between the waveforms of the separated signal and the source signal; the symbol error rate measures the proportion of symbol errors in the separated signal after demodulation and decision; the signal distortion ratio measures the ratio of effective signal energy to total distortion energy in the separated signal; and the error vector magnitude measures the deviation of the constellation points of the separated signal from the ideal constellation points. All indicators are averaged across all samples in the test set to evaluate the model's generalization ability and robustness under different signal-to-noise ratios, different carrier frequency offsets, and different sudden interruption conditions. Figure 4The symbol error rate curve for signal separation in this embodiment is given, showing that the network of the present invention still maintains a low symbol error rate under conditions of low signal-to-noise ratio and non-ideal channel impairment.

[0081] Step S6: Construct a hardware test platform. A vector signal generator is used to generate the target QPSK signal and the co-channel interference signal. The carrier frequency of both signals is set to 600 MHz, the symbol rate to 1 MHz, and both are baseband pulse-shaped using a root-raised cosine filter with a roll-off factor of 0.35. The root-raised cosine filter is a pulse-shaping filter whose frequency response is the square root of the raised cosine characteristic, used to limit signal bandwidth and suppress inter-symbol interference. The generated two signals are then power-adjusted and combined into a single-channel time-frequency aliasing signal. In this embodiment, the interference-to-signal ratio is set to 0 dB, representing the most stringent aliasing condition where the target signal and interference signal have equal power. A software-defined radio platform receives this mixed signal through a single antenna. After down-conversion and analog-to-digital conversion, a digital baseband I / Q discrete sequence is acquired and saved as a complex baseband format file for subsequent separation testing.

[0082] Step S7: The measured time-frequency aliasing signal acquired in step S6 is segmented into segments of a fixed length. In this embodiment, the fixed length is set to 1024 sampling points, consistent with the signal length in the training phase. For segments shorter than 1024 sampling points, zero-padding is used to pad to 1024 sampling points, i.e., zero-valued sampling points are added to the end of the segment. Each signal segment is stored in a stacked form of in-phase and quadrature components, forming a three-dimensional tensor composed of batch dimension, component dimension, and time dimension, ensuring complete consistency with the data format in the training phase. All signal segments after segmentation constitute the measured test set.

[0083] Step S8: Using the multi-scale feature fusion complex neural network trained in step S4, the measured complex-form aliased signal obtained in step S7 is separated and tested. Each measured signal segment is fed into the network for forward propagation: the complex encoder extracts high-dimensional complex encoded features; the multi-scale feature fusion mask network performs dual-path iterative modeling on the encoded features and generates a complex mask; feature weighting is performed through complex point-by-point multiplication to obtain the enhanced features corresponding to the target signal; the complex decoder reconstructs the enhanced features into a time-domain complex-valued waveform and outputs the target signal estimation result corresponding to that signal segment. The separation results of all signal segments are collected and reassembled in the original time sequence to restore the complete recovered signal. By plotting constellation diagrams before and after separation and calculating indicators such as error vector magnitude, the separation accuracy under actual hardware conditions is intuitively evaluated.

[0084] The end-to-end multi-scale feature fusion complex neural network in the full complex domain established in this embodiment achieves stable and consistent separation performance in non-ideal channel impairment environments such as strong carrier frequency offset and sudden interruption, requiring only 0.93 megabytes of lightweight parameters. This fully verifies the advancement, adaptability and superiority of the proposed full complex domain multi-scale feature fusion architecture compared with existing schemes in wireless communication anti-interference blind separation tasks, and realizes high-precision and effective separation of mixed communication signals under single-antenna reception conditions.

[0085] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A single-channel blind source separation method based on a multi-scale feature fusion complex neural network, characterized in that, It includes a training phase and an application phase, wherein the training phase includes the following steps: Create a mixed signal dataset. The mixed signal in the dataset is a single-channel superimposed signal of target signal, interference signal and environmental noise. Carrier frequency offset and channel impairment due to sudden interruption are introduced into the mixed signal. The mixed signal dataset is divided into training set, validation set and test set. A fully complex algebraic computation environment is constructed, and a multi-scale feature fusion complex neural network is built. The multi-scale feature fusion complex neural network includes a complex encoder, a multi-scale feature fusion mask network, and a complex decoder. Create a network training loss function, which consists of a weighted sum of scale-invariant signal-to-noise ratio and mean square error; The multi-scale feature fusion complex neural network was trained and validated using the training set and validation set to obtain the trained multi-scale feature fusion complex neural network. The application phase includes the following steps: The trained multi-scale feature fusion complex neural network was subjected to a separation evaluation test using the test set. A hardware testing platform was constructed, which generates a mixed signal of target signal and interference signal through a vector signal generator, and collects measured signal data. The collected measured signal data is segmented to obtain a complex aliased signal of fixed length; The target signal is separated from the complex aliased signal by inputting the trained multi-scale feature fusion complex neural network, thus completing single-channel blind source separation.

2. The single-channel blind source separation method based on multi-scale feature fusion complex neural network according to claim 1, characterized in that, The creation of mixed-signal datasets includes the creation of simulated signal datasets and measured signal datasets, wherein: The simulation signal dataset is created as follows: target signals and interference signals with different modulation methods are generated by computer simulation; the target signals and interference signals are randomly mixed to form aliased signals; noise is added to the aliased signals to generate time-frequency aliased communication signals; the dataset composed of the time-frequency aliased communication signals is divided into training set, validation set and test set. The measured signal dataset was created as follows: A test platform was built in a microwave anechoic chamber, with four antennas deployed. Two transmitting antennas were connected to two vector signal generators, respectively, to radiate the target signal and the interference signal. Two receiving antennas were connected to a software-defined radio platform to receive the spatially synthesized mixed signal. The two signal generators or external attenuators were adjusted, and the transmitted and received mixed signals were calibrated and observed using a spectrum analyzer and an oscilloscope to assist in verifying the channel impairment injection effects such as sudden interruptions. Two mixed signals were acquired by the software-defined radio platform, either of which could be used as a single-channel aliasing signal. The acquired complex baseband aliasing signal was segmented and processed to obtain a fixed-length complex aliasing signal, which served as the test set.

3. The single-channel blind source separation method based on a multi-scale feature fusion complex neural network according to claim 1, characterized in that, The complex encoder includes a complex one-dimensional convolutional layer, and the input single-channel complex time-domain signal consists of in-phase components and quadrature components; The complex one-dimensional convolutional layer performs joint parallel convolution operations on the in-phase and quadrature components of the single-channel complex-valued time-domain signal according to the rules of complex algebra, mapping the single-channel complex-valued time-domain signal to a high-dimensional complex feature space to obtain encoded features; the joint parallel convolution operation specifically includes: The real part of the complex one-dimensional convolution kernel is convolved with the in-phase component of the single-channel complex-valued time-domain signal to obtain a first convolution result. The imaginary part of the complex one-dimensional convolution kernel is convolved with the orthogonal component of the single-channel complex-valued time-domain signal to obtain a second convolution result. The difference between the first convolution result and the second convolution result is used as the in-phase component of the encoded feature. The real part of the complex one-dimensional convolution kernel is convolved with the orthogonal component of the single-channel complex-valued time-domain signal to obtain the third convolution result. The imaginary part of the complex one-dimensional convolution kernel is convolved with the in-phase component of the single-channel complex-valued time-domain signal to obtain the fourth convolution result. The sum of the third convolution result and the fourth convolution result is used as the orthogonal component of the encoded feature.

4. The single-channel blind source separation method based on multi-scale feature fusion complex neural network according to claim 1, characterized in that, The multi-scale feature fusion masking network consists of a cascaded dual-path iterative structure with a preset number of layers. The processing procedure for the input features in each dual-path iterative structure is as follows: The input features are segmented and reorganized into multiple local sub-blocks; The recombined features are input into the multi-scale complex Conformer module to perform intra-segment modeling, model the temporal correlation within each local sub-block, extract local temporal features, and obtain intra-segment enhanced features; After dimensional permutation, the intra-segment enhancement features are input again into the same multi-scale complex Conformer module that shares parameters with the intra-segment modeling to perform inter-segment modeling. This modeling of long-range dependencies between different local sub-blocks is used to extract global dependencies and obtain inter-segment enhancement features. In this process, the intra-segment modeling and the inter-segment modeling are executed alternately once in each layer of the dual-path iterative structure, completing a joint local and global modeling process.

5. The single-channel blind source separation method based on multi-scale feature fusion complex neural network according to claim 4, characterized in that, The multi-scale complex Conformer module adopts a cascaded sandwich residual structure. The input features sequentially pass through a first feedforward network, a multi-head self-attention module, a multi-scale complex depthwise convolution module, and a second feedforward network, wherein: The first feedforward network performs nonlinear transformation and channel dimension mapping on the input features. Its output is multiplied by a scaling factor and then regularized before being added to the residual of the module input features. The multi-head self-attention module introduces complex positional encoding to preserve temporal order information and capture the global long-range dependencies of the feature sequence. Its output is regularized and then added to the residual of the current feature. The multi-scale complex depth convolution module is used to extract local features at different time scales, and its output is added to the residual of the current feature; The second feedforward network performs nonlinear transformation and channel dimension mapping on the features again. Its output is multiplied by a scaling factor and then regularized. The residual is added to the current feature to complete the forward propagation of the multi-scale complex Conformer module.

6. The single-channel blind source separation method based on a multi-scale feature fusion complex neural network according to claim 5, characterized in that, The multi-scale complex depth convolution module adopts a parallel structure, dividing the input features into three paths, which are respectively fed into three complex one-dimensional channel-wise depth convolution branches with different convolution kernel sizes for feature extraction. The three different convolution kernel sizes correspond to low-frequency feature extraction, mid-frequency feature extraction and high-frequency feature extraction of communication signals, respectively. The output features of the three complex one-dimensional channel-wise depthwise convolutional branches are concatenated along the channel dimension to obtain concatenated features; cross-scale feature fusion is performed on the concatenated features through a complex one-dimensional convolutional layer; the fused features are then processed by the ModReLU activation function and Dropout regularization in sequence to obtain the fused multi-scale complex feature representation.

7. The single-channel blind source separation method based on multi-scale feature fusion complex neural network according to claim 4, characterized in that, After the dual-path iterative structure with all preset layers completes feature modeling, the output features are input into a mask generation layer composed of complex one-dimensional convolutional layers to generate a complex mask. The complex mask is then applied to the encoded features output by the complex encoder using point-by-point complex multiplication to obtain the enhanced features corresponding to the target signal. Specifically, the point-by-point complex multiplication is as follows: The in-phase component of the enhanced feature is obtained by subtracting the product of the in-phase component of the complex mask and the in-phase component of the coded feature from the product of the in-phase component of the complex mask and the product of the quadrature component of the coded feature. The orthogonal component of the enhanced feature is obtained by multiplying the in-phase component of the complex mask with the quadrature component of the coded feature, and then adding the product of the quadrature component of the complex mask with the in-phase component of the coded feature.

8. The single-channel blind source separation method based on a multi-scale feature fusion complex neural network according to claim 1, characterized in that, The complex decoder corresponds to the complex encoder structure and includes a complex one-dimensional transposed convolutional layer. The complex decoder, based on the rules of complex algebra operations, performs joint processing on the in-phase and quadrature components of the enhanced features, gradually mapping the high-dimensional complex features back to the time domain space, and reconstructing the single-channel complex-valued time-domain waveform of the target signal.

9. The single-channel blind source separation method based on a multi-scale feature fusion complex neural network according to claim 1, characterized in that, The network training loss function is the difference between the average mean square error of the in-phase and quadrature components of the complex signal and the average scale-invariant signal-to-noise ratio of the in-phase and quadrature components. The training and validation process is as follows: The complex baseband signal samples in the training set and validation set are processed in batches according to a preset batch size, and the complex tensor data of each batch is used as the input of the multi-scale feature fusion complex neural network. An adaptive moment estimation optimizer is used to iteratively update the network parameters, and a learning rate scheduler is used to dynamically decay the learning rate. An early stopping strategy is adopted, and training is terminated when the network's evaluation metric on the validation set fails to improve for a preset number of consecutive rounds. During backpropagation, the gradients of the network parameters are pruned using the maximum norm to suppress gradient explosion.

10. A single-channel blind source separation system based on a multi-scale feature fusion complex neural network, characterized in that, The system is used to implement the single-channel blind source separation method based on a multi-scale feature fusion complex neural network as described in any one of claims 1 to 9. The system includes a training module and an application module, wherein: The training module includes a dataset creation unit, a network building unit, a loss function creation unit, and a network training unit, among which: The dataset creation unit is used to create a dataset of single-channel aliased signals. The dataset includes complex baseband signals resulting from the mixing of the target signal and the interference signal using different modulation methods. The network construction unit is used to build a multi-scale feature fusion complex neural network, which includes a complex encoder, a multi-scale feature fusion mask network, and a complex decoder. The loss function creation unit is used to create the network training loss function; The network training unit is used to train and validate the multi-scale feature fusion complex neural network obtained by the network building unit based on the network training loss function and the dataset obtained by the dataset creation unit. The network parameters are iteratively updated through backpropagation to obtain the trained multi-scale feature fusion complex neural network. The application module includes a signal acquisition unit, an aliasing signal processing unit, and a blind separation unit, wherein: The signal acquisition unit is used to acquire the time-frequency aliasing signal to be separated through simulation software and a vector signal generator; The aliasing signal processing unit is used to segment the time-frequency aliased signal to obtain a baseband aliased signal of fixed length in complex form.