Blind separation method for multi-time-frequency overlapping mixed signals based on diffusion model

Abstract

The application provides a multi-time-frequency overlapping mixed signal blind separation method based on a diffusion model, and relates to the technical field of mixed signal separation. First, the multi-time-frequency overlapping mixed signal is preprocessed to obtain initial estimation information of each source signal; a cascaded mixed signal separation model is constructed, the model comprises multiple parallel diffusion generation modules and consistency constraint modules, and the estimation information of the source signal is iterated and interacted between the modules; in the iteration process, the score estimation is calculated by using the diffusion generation module to provide an optimization direction for recovering the signal conforming to the prior distribution; the consistency constraint module is used to jointly optimize the reference signal based on the score estimation to obtain the maximum posterior, and the remaining source signal estimation value is cooperatively corrected; and the final separation result of the source signal is obtained based on the estimation result from different reference signals. Through the parallel cascading and iterative cooperative correction mechanism, the multi-time-frequency overlapping signal separation with strong generalization ability and high reliability is realized.

Description

Technical Field

[0001] This invention relates to the technical field of mixed signal separation, and more specifically, to a blind separation method for multi-time-frequency overlapping mixed signals based on a diffusion model. Background Technology

[0002] In modern digital communication systems, Single-Channel Blind Source Separation (SCBSS) is a key technology in signal processing. Unlike multi-channel separation, which utilizes multi-dimensional observation information in the spatial and angular domains to separate multiple unknown independent source signals aliased in the time and frequency domains, SCBSS uses only the mixed signal captured by a single receiving channel. Based on prior knowledge and physical characteristics of each source signal, it achieves signal separation and reconstruction in different representations or transform domains. This technology has significant application value in scenarios such as spectrum monitoring, cognitive communication, and non-cooperative communication.

[0003] Traditional SCBSS methods primarily rely on mathematical modeling of the signal's intrinsic structure and specific prior knowledge. This involves mapping the mixed signals to a suitable feature space, establishing reasonable separation criteria based on prior knowledge, constructing a corresponding separation optimization model, and finally achieving signal separation and extraction through filtering, matrix factorization, and estimation theory. However, this approach heavily depends on manually designed feature extraction mechanisms and strong prior assumptions about signal characteristics. In real-world communication scenarios, due to the high time-frequency overlap of multiple signals, the time-varying and nonlinear nature of various link channels, and the prevalence of complex noise and interference, these assumptions are difficult to meet, leading to a significant deterioration in separation performance.

[0004] In recent years, significant progress has been made in deep learning-based SCBSS (Separate Channel Broadcasting System) methods for communication signals. These methods leverage the powerful nonlinear mapping and feature extraction capabilities of deep neural networks to automatically learn the intrinsic structure of signals and channel mixing patterns from massive training samples, thereby achieving signal separation and fitting in an end-to-end manner and significantly improving the quality of signal separation and reconstruction. However, the separation performance of these methods is highly dependent on supervised model training on large-scale, high-quality training datasets. In practical applications, simultaneously acquiring sufficient mixed signals and corresponding single-source signals is difficult and infeasible, leading to incomplete or skewed training data. Furthermore, when faced with separation tasks involving multiple different signal combinations, the models often struggle to generalize effectively, resulting in decreased separation performance. Summary of the Invention

[0005] To address the shortcomings of existing technologies in separating multi-time-frequency overlapping signals, this invention proposes a blind separation method for multi-time-frequency overlapping mixed signals based on a diffusion model, achieving separation of multi-time-frequency overlapping signals with strong generalization ability and high reliability.

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

[0007] In a first aspect, this application proposes a blind separation method for multi-time-frequency overlapping mixed signals based on a diffusion model, comprising the following steps:

[0008] S1. Preprocess the multi-time-frequency overlapping mixed signal to obtain the current estimation information of each source signal in the multi-time-frequency overlapping mixed signal; the number of signal sources in the multi-time-frequency overlapping mixed signal is known;

[0009] S2. Construct a cascaded mixed-signal separation model and pre-train it. The cascaded mixed-signal separation model includes... A parallel diffusion generation module and A parallel consistency constraint module, A diffusion generation module and The consistency constraint modules interact and iterate to obtain the estimation information of each source signal. , It is a positive integer;

[0010] S3. Order Set the maximum number of iterations to Execute S4;

[0011] S4. Proceed to the next step. The second iteration: [The second iteration is missing from the original text] The current estimation information of the source signal is input to the first... The diffusion generation module obtains the first... The score estimates of the current estimation information of each source signal are obtained in total. The score estimate is the current estimation information of the source signal; the score estimate is the score estimate from the first... Recovering a signal from a source signal that conforms to the prior distribution of that source signal provides an optimization direction;

[0012] S5. In the In the consistency constraint module, the first one The source signal is used as the reference signal, based on The score estimate of the current estimation information of each source signal is used to jointly optimize the reference signal to obtain the score estimate of the current source signal. The maximum a posteriori of the source signals, and using the maximum a posteriori to analyze the remaining signals. Coordinated correction of the estimated values ​​of individual source signals;

[0013] S6. Based on the correction results, obtain The estimation results from different reference signals are based on the... The estimation results from different reference signals yielded the first group. Estimates of the source signal;

[0014] S7. Determine if the maximum number of iterations has been reached. If so, then the first The estimated value of the source signal is used as the first... The final separation result of the source signals; otherwise, let The value is increased by 1, and the estimated value obtained in S6 is used as the first... The current estimation information of each source signal is returned to S4.

[0015] In this technical solution, the multi-time-frequency overlapping mixed signals are first preprocessed to obtain initial estimation information for each source signal. A cascaded mixed signal separation model is then constructed, comprising a diffusion generation module and a consistency constraint module. These modules iteratively exchange the estimation information of the source signals. During the iteration process, the diffusion generation module calculates score estimates, providing optimization directions for recovering signals conforming to the prior distribution. The consistency constraint module performs joint optimization on the reference signal based on the score estimates and collaboratively corrects the remaining source signal estimates. The final separation result of the source signals is obtained based on the joint optimization and correction results. This method significantly improves the separation accuracy and robustness of multi-time-frequency overlapping signals through parallel cascading and iterative collaborative correction mechanisms.

[0016] Preferably, the preprocessing process for the multi-time-frequency overlapping mixed signal is as follows:

[0017] S11. Order Determine the received multi-time-frequency overlapping mixed signal Source signal in , Indicates the order of the source signals;

[0018] S12. For the source signal Blind synchronization processing is performed to obtain a multi-time-frequency overlapping mixed signal after blind synchronization. ;

[0019] S13. For the multi-time-frequency overlapping mixed signal after blind synchronization Perform deconstellation mapping and remodulation to obtain the source signal. Current estimation information ;

[0020] S14. Transfer the source signal Current estimation information Multi-time-frequency overlapping mixed signal after blind synchronization Perform cross-correlation alignment and subtraction to obtain the residual signal. ;

[0021] S15. Order The value increases by 1;

[0022] S16. Judgment Is it less than or equal to? If so, the residual signal As input, return S12; otherwise, output... Individual source signal Current estimation information .

[0023] Preferably, the blind synchronization process is as follows: determining the mixed signal using blind symbol rate estimation of a single-carrier signal. Symbol rate, based on symbol rate for multi-time-frequency overlapping mixed signals Sampling is performed, and blind carrier synchronization and blind symbol timing synchronization are executed on the sampling results.

[0024] Preferably, the modulation type of each source signal is known in the multi-time-frequency overlapping mixed signal, and each of the diffusion generation modules is loaded with a pre-trained diffusion model corresponding to the modulation type of the source signal.

[0025] Preferably, the diffusion generation module is used to obtain the first... The score estimation process for the current estimated information of each source signal is as follows:

[0026] S41. Order Set the maximum number of noise-added samples to 1. Execute S42;

[0027] S42. Based on the random sampling noise vector and the random sampling time step, perform noisy sampling on each source signal to obtain noisy source signal samples. The expression is:

[0028]

[0029] in, Indicates the random sampling time step. Indicates the number of samplings with added noise. This represents the signal attenuation coefficient corresponding to the time step. This represents a randomly sampled noise vector that follows a standard complex normal distribution.

[0030] S43. Using a diffusion model, estimate the noise in the noisy sampling based on the noisy source signal samples and their time steps. ;

[0031] S44. Determine whether the number of noise samplings has reached the preset maximum number of noise samplings. If not, then let Increment the value by 1 and return S42; if so, calculate the 1st noise level smoothing based on multi-noise level smoothing. Current estimation information of individual source signals Score estimation The calculation expression is:

[0032]

[0033] in, This represents the standard deviation of the noise vector added to the source signal. Represents all random sampling time steps The expectation.

[0034] Preferably, the joint optimization of the reference signal in step S5 includes:

[0035] Using the score estimate as a probability gradient, based on a preset iterative learning rate... Update the current estimation information of the reference signal, and the update expression is:

[0036]

[0037] in, This represents the updated reference signal estimate. The mixing coefficients of the reference signal are represented. This represents the score estimate of the reference signal. Indicates the remainder The mixing coefficients of the source signals, Indicates the remainder Score estimation of individual source signals.

[0038] Preferably, the use of the first in step S5 The maximum posterior pair of the source signals remains The estimated values ​​of each source signal are used for collaborative correction, and the calculation expression is as follows:

[0039]

[0040] in, Indicates the remainder A single source signal among multiple source signals Indicates the remainder Interference signals in the source signal Indicates the remainder The mixing coefficient of interference signals in a source signal.

[0041] Preferably, S6 is based on the The estimation results from different reference signals yielded the first group. The estimated values ​​of the source signals include:

[0042] Based on preset quality indicators, from Selected from group estimation results The group estimation results, the ;

[0043] Based on the above In the group estimation results, the calculation of the first group is performed. The average of the estimation results of the n source signals is used as the nth The estimated value of the source signal.

[0044] Secondly, this application also proposes a computer device, which includes a memory, a processor, and a computer program stored in the memory that can be run by the processor. The processor executes the computer program to implement the blind separation method for multi-time-frequency overlapping mixed signals based on the diffusion model.

[0045] Thirdly, this application also proposes a computer storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, implement the blind separation method for multi-time-frequency overlapping mixed signals based on a diffusion model.

[0046] Compared with the prior art, the beneficial effects of the present invention are:

[0047] This invention proposes a blind separation method for multi-time-frequency overlapping mixed signals based on a diffusion model. First, the multi-time-frequency overlapping mixed signals are preprocessed to obtain initial estimates for each source signal. A cascaded mixed signal separation model is constructed, comprising multiple parallel diffusion generation modules and a consistency constraint module. These modules iteratively exchange the source signal estimates. During the iteration process, the diffusion generation module calculates score estimates, providing optimization directions for recovering signals conforming to the prior distribution. The consistency constraint module performs joint optimization on the reference signals based on the score estimates to obtain the maximum a posteriori (MAP), and collaboratively corrects the remaining source signal estimates. Finally, the final separation result of the source signals is obtained by fusing the estimation results from different reference signals. This method significantly improves the separation accuracy and robustness of multi-time-frequency overlapping signals through parallel cascading and iterative collaborative correction mechanisms. Attached Figure Description

[0048] Figure 1 This is a flowchart illustrating the blind separation method for multi-time-frequency overlapping mixed signals based on a diffusion model proposed in Embodiment 1 of the present invention.

[0049] Figure 2 This is a schematic diagram illustrating the preprocessing process for multi-time-frequency overlapping mixed signals proposed in Embodiment 2 of the present invention;

[0050] Figure 3 This indicates that the diffusion generation module proposed in Embodiment 2 of the present invention is used to obtain the first... A flowchart illustrating the score estimation process for the current estimated information of each source signal;

[0051] Figure 4 This diagram illustrates the process of multi-signal blind separation based on a diffusion model proposed in Embodiment 2 of the present invention.

[0052] Figure 5 A schematic diagram showing the convergence curve of the mean square error (MSE) as a function of the number of iterations when separating time-frequency overlapping mixed signals as proposed in Embodiment 3 of the present invention;

[0053] Figure 6 A schematic diagram showing the convergence curve of the bit error rate (BER) as a function of the number of iterations when separating time-frequency overlapping mixed signals as proposed in Embodiment 3 of the present invention;

[0054] Figure 7 A schematic diagram showing the convergence curve of the mean square error (MSE) as a function of the signal frequency (SNR) when separating time-frequency overlapping mixed signals as proposed in Embodiment 3 of the present invention;

[0055] Figure 8 A schematic diagram showing the convergence curve of the bit error rate (BER) as a function of SNR when separating time-frequency overlapping mixed signals as proposed in Embodiment 3 of the present invention;

[0056] Figure 9 This is a schematic diagram of the structure of the computer device proposed in Embodiment 4 of the present invention. Detailed Implementation

[0057] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.

[0058] To better illustrate this embodiment, some parts of the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions;

[0059] It is understandable to those skilled in the art that some well-known details may be omitted from the accompanying drawings.

[0060] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0061] The positional relationships depicted in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent.

[0062] Example 1

[0063] This embodiment proposes a blind separation method for multi-time-frequency overlapping mixed signals based on a diffusion model. A flowchart of this method can be found here. Figure 1 This includes the following steps:

[0064] S1. Preprocess the multi-time-frequency overlapping mixed signal to obtain the current estimation information of each source signal in the multi-time-frequency overlapping mixed signal; the number of signal sources in the multi-time-frequency overlapping mixed signal is known;

[0065] S2. Construct a cascaded mixed-signal separation model and pre-train it. The cascaded mixed-signal separation model includes... A parallel diffusion generation module and A parallel consistency constraint module, A diffusion generation module and The consistency constraint modules interact and iterate to obtain the estimation information of each source signal. , It is a positive integer;

[0066] S3. Order Set the maximum number of iterations to Execute S4;

[0067] S4. Proceed to the next step. The second iteration: [The second iteration is missing from the original text] The current estimation information of the source signal is input to the first... The diffusion generation module obtains the first... The score estimates of the current estimation information of each source signal are obtained in total. The score estimate is the current estimation information of the source signal; the score estimate is the score estimate from the first... Recovering a signal from a source signal that conforms to the prior distribution of that source signal provides an optimization direction;

[0068] S5. In the In the consistency constraint module, the first one The source signal is used as the reference signal, based on The score estimate of the current estimation information of each source signal is used to jointly optimize the reference signal to obtain the score estimate of the current source signal. The maximum a posteriori of the source signals, and using the maximum a posteriori to analyze the remaining signals. Coordinated correction of the estimated values ​​of individual source signals;

[0069] S6. Based on the correction results, obtain The estimation results from different reference signals are based on the... The estimation results from different reference signals yielded the first group. Estimates of the source signal;

[0070] S7. Determine if the maximum number of iterations has been reached. If so, then the first The estimated value of the source signal is used as the first... The final separation result of the source signals; otherwise, let The value is increased by 1, and the estimated value obtained in S6 is used as the first... The current estimation information of each source signal is returned to S4.

[0071] In this embodiment, the multi-time-frequency overlapping mixed signal is first preprocessed to obtain initial estimation information for each source signal. A cascaded mixed signal separation model is then constructed, comprising a diffusion generation module and a consistency constraint module. These modules iteratively exchange the estimation information of the source signals. During the iteration process, the diffusion generation module calculates score estimates, providing optimization directions for recovering signals conforming to the prior distribution. The consistency constraint module performs joint optimization on the reference signal based on the score estimates and collaboratively corrects the remaining source signal estimates. The final separation result of the source signals is obtained based on the joint optimization and correction results. This method significantly improves the separation accuracy and robustness of multi-time-frequency overlapping signals through parallel cascading and iterative collaborative correction mechanisms.

[0072] Specifically, the multi-time-frequency overlapping mixed signal Depend on A number of independent and energy-normalized unknown source signals The linear superposition yields the following mathematical model, which can be expressed as an underdetermined linear mixture model:

[0073]

[0074] in, For the first Individual source signal The corresponding mixing coefficients. To avoid confusion in the order of the estimation results from the various source signals, the following settings are used: ,at this time Main energy signal, This is a secondary energy signal. The goal of solving this hybrid model is to address the issue of the number of signal sources. Mixing coefficient Under the conditions of modulation type parameters of each source signal, for multi-time-frequency overlapping mixed signals The source signals are estimated by performing separation processing.

[0075] When multiple time-frequency overlapping mixed signals When the source signal contains additive noise, There exists an energy-normalized noise signal. Its elements follow a normal distribution.

[0076] Example 2

[0077] In this embodiment, the flowchart for preprocessing multi-time-frequency overlapping mixed signals is shown below. Figure 2 As shown, the process is as follows:

[0078] S11. Order Determine the received multi-time-frequency overlapping mixed signal Source signal in , Indicates the order of the source signals;

[0079] S12. For the source signal Blind synchronization processing is performed to obtain a multi-time-frequency overlapping mixed signal after blind synchronization. ;

[0080] S13. For the multi-time-frequency overlapping mixed signal after blind synchronization Perform deconstellation mapping and remodulation to obtain the source signal. Current estimation information ;

[0081] S14. Transfer the source signal Current estimation information Multi-time-frequency overlapping mixed signal after blind synchronization Perform cross-correlation alignment and subtraction to obtain the residual signal. ;

[0082] S15. Order The value increases by 1;

[0083] S16. Judgment Is it less than or equal to? If so, the residual signal As input, return S12; otherwise, output... Individual source signal Current estimation information .

[0084] Specifically, the purpose of blind synchronization processing is to eliminate multi-time-frequency overlapping mixed signals. Middle main energy signal The frequency offset and sampling timing error are used to obtain the synchronized mixed signal. At this point, the synchronized mixed signal The symbol sampling rate is 2 or 4 times that of the main energy signal; based on the main energy signal The modulation type, for the synchronized mixed signal After performing constellation mapping to obtain the information bits, the main energy signal is then re-applied. modulation type and The symbol sampling rate is used for secondary modulation to obtain the remodulated main energy signal. This serves as the initial estimate of the main energy signal for subsequent diffusion models; the remodulated main energy signal is cross-correlated and aligned with the synchronized mixed signal, and the synchronized mixed signal is used as the initial estimate. Subtract the remodulated main energy signal The residual signal was obtained. ; Residual signal For secondary energy signals Blind synchronization and remodulation are performed again to obtain an initial estimate of the remodulated sub-energy signal. Repeat this process to obtain each remodulation signal in sequence. .

[0085] Specifically, when multiple time-frequency overlapping mixed signals There are noise signals that need to be separated. At that time, skip the noise signal. The blind synchronization and remodulation process is performed, and the residual signal is directly assigned to the initial estimate of the noise signal and the residual signal of the next round. The calculation expression is as follows:

[0086]

[0087] In this embodiment, the blind synchronization process is as follows: the mixed signal is determined using blind symbol rate estimation of a single-carrier signal. Symbol rate, based on symbol rate for multi-time-frequency overlapping mixed signals Sampling is performed, and blind carrier synchronization and blind symbol timing synchronization are executed on the sampling results.

[0088] Specifically, the single-carrier signal blind symbol rate estimation includes blind symbol rate estimation based on cyclic spectrum estimation; the blind synchronization method also includes: upsampling method, downsampling method, blind carrier synchronization method (such as blind carrier synchronization based on cyclic spectrum estimation), and blind symbol timing synchronization method (such as Gardner symbol timing synchronization).

[0089] In this embodiment, the modulation type of each source signal is known in the multi-time-frequency overlapping mixed signal, and each diffusion generation module is loaded with a pre-trained diffusion model corresponding to the modulation type of the source signal.

[0090] Specifically, the diffusion model A unified approach is adopted, utilizing the mature diffusion model framework Diffwave from the field of speech synthesis, and extended to the complex domain. This expands the original single-channel dimensional structure for processing time-domain sound waves to a dual-channel dimensional structure supporting the processing of both the real and imaginary parts of complex IQ signals in the time domain. This allows for the effective extraction and learning of the inherent discrete structure and statistical distribution characteristics of complex digital communication signal waveforms. By independently pre-training diffusion models for different modulation types, the large-scale pairwise training data required by traditional deep learning methods is eliminated, enhancing the generalization ability and adaptability for separating mixed signals with arbitrary modulation type combinations.

[0091] During the model training phase, the corresponding source signals are used. Complex IQ signal datasets of modulation type The diffusion model is trained following a standard forward noise addition and backward generation process. The training objective is to minimize... Predicted noise vs. actual added noise The mean squared error between them, the expression for the loss function is:

[0092]

[0093] in, This represents the signal attenuation coefficient determined by the noise scheduling strategy. Expressing expectations, This represents the time step of random sampling during the forward noise addition process.

[0094] In this embodiment, the diffusion generation module is used to obtain the first... A flowchart illustrating the score estimation process for the current estimated information of a single source signal is shown below. Figure 3 As shown, the process is as follows:

[0095] S41. Order Set the maximum number of noise-added samples to 1. Execute S42;

[0096] S42. Based on the random sampling noise vector and the random sampling time step, perform noisy sampling on each source signal to obtain noisy source signal samples. The expression is:

[0097]

[0098] in, Indicates the random sampling time step. Indicates the number of samplings with added noise. This represents the signal attenuation coefficient corresponding to the time step. This represents a randomly sampled noise vector that follows a standard complex normal distribution.

[0099] S43. Using a diffusion model, estimate the noise in the noisy sampling based on the noisy source signal samples and their time steps. ;

[0100] S44. Determine whether the number of noise samplings has reached the preset maximum number of noise samplings. If not, then let Increment the value by 1 and return S42; if so, calculate the 1st noise level smoothing based on multi-noise level smoothing. Current estimation information of individual source signals Score estimation The calculation expression is:

[0101]

[0102] in, This represents the standard deviation of the noise vector added to the source signal. Represents all random sampling time steps The expectation.

[0103] Specifically, the expression for the noise vector added to the source signal is:

[0104]

[0105] The The expression is:

[0106] .

[0107] In this embodiment, the joint optimization of the reference signal described in S5 includes:

[0108] Using the score estimate as a probability gradient, based on a preset iterative learning rate... Update the current estimation information of the reference signal, and the update expression is:

[0109]

[0110] in, This represents the updated reference signal estimate. The mixing coefficients of the reference signal are represented. This represents the score estimate of the reference signal. Indicates the remainder The mixing coefficients of the source signals, Indicates the remainder Score estimation of individual source signals.

[0111] Specifically, the reference signal Joint optimization is performed, including in the consistency constraint module. In, a module With the corresponding As a reference signal, the score estimate is based on the received scores from each diffusion generation module. ,right Perform maximum a posteriori estimation; for the reference signal The expression for maximum a posteriori estimation is:

[0112]

[0113]

[0114]

[0115] Based on diffusion model theory, let the variable to be estimated... As The new estimate, and using the score estimate as the probability gradient, is expressed as:

[0116]

[0117] .

[0118] In this embodiment, S5 utilizes the first The maximum posterior pair of the source signals remains The estimated values ​​of each source signal are used for collaborative correction, and the calculation expression is as follows:

[0119]

[0120] in, Indicates the remainder A single source signal among multiple source signals Indicates the remainder Interference signals in the source signal Indicates the remainder The mixing coefficient of interference signals in a source signal.

[0121] Specifically, based on the calculation expression, for the remainder The process of co-correcting the estimated values ​​of individual source signals is as follows: based on the physical constraint of the linear mixing model of the single-channel signal, a model is constructed for the remaining signals. Conditional observation of a single source signal, i.e., from the received multi-time-frequency overlapping mixed signal In, subtract the updated reference signal components And except for the current signal to be corrected Estimated components of other interference signals The calculated residual is used as the signal to be corrected. The corrected new estimate is used to physically isolate the coupling interference between signals, thereby correcting and optimizing the estimates of the remaining source signals.

[0122] Specifically, the schematic diagram of the multi-signal blind separation process based on the diffusion model is as follows: Figure 4 As shown, the Figure 4 The cascaded hybrid signal separation model based on the diffusion model integrates multiple parallel diffusion generation modules and multiple parallel consistency constraint modules. The two types of modules iteratively interact with the estimation information of each source signal to achieve incremental optimization. Figure 4 In A parallel diffusion generation module In, a module Based on the corresponding source signal Pre-trained diffusion model The current signal is estimated through an iterative multi-noise level smoothing estimation process. Calculate score estimation The score is estimated. To estimate from a noisy signal The source signal was gradually recovered. The prior distribution signal provides direction for optimization. The score estimation... Consistency constraint module passed to the other side of the graph Each consistency constraint module uses one of the source signals as a reference, performs joint gradient optimization (i.e., maximum a posteriori estimation) on the reference signal based on the received score estimate, and uses the updated reference signal to collaboratively correct the estimates of the remaining signals. Through this process, each consistency constraint module... A new set of signal estimation results were obtained. The diffusion generation module and the consistency constraint module iteratively interact with each score estimate and the updated signal estimate through the path indicated by the bidirectional arrow, forming a closed-loop optimization process. The signal processing and information transmission between modules are similar to the Turbo-style "external information interaction" process, which realizes effective decoupling between modules, effectively alleviates the estimation bias caused by multi-signal coupling, and improves the fidelity and robustness of multi-signal separation and reconstruction.

[0123] In this embodiment, S6 is based on the The estimation results from different reference signals yielded the first group. The estimated values ​​of the source signals include:

[0124] Based on preset quality indicators, from Selected from group estimation results The group estimation results, the ;

[0125] Based on the above In the group estimation results, calculate the value of the first group. The average of the estimation results of the n source signals is used as the nth The estimated value of the source signal.

[0126] Specifically, the calculation regarding the first The expression for the average value of the estimation results of the individual source signals is:

[0127]

[0128] in, This is the number of the consistency constraint module. For a high-quality estimation index set, the preset quality indicators include relevant evaluation indicators such as the quality distribution or aggregation degree of the estimation results. .

[0129] Example 3

[0130] In this embodiment, the blind separation method for multi-time-frequency overlapping mixed signals based on the diffusion model is used to perform blind separation of the time-frequency overlapping mixed signals. The time-frequency overlapping mixed signals include three source signals with modulation types of BPSK, QPSK, and QPSK, symbol rates of 400kHz, 600kHz, and 800kHz, and mixing coefficients of 5.6, 3.2, and 1, respectively. The corresponding signal energy ratio is 15:10:1, the center frequency range is [199.99MHz, 200.01MHz], the sampling rate is 2MHz, and the length of the signal acquired and separated in a single session is... .

[0131] Based on the average results of 100 independent experiments, the convergence curve of the mean square error (MSE) of separating the three source signals as a function of the number of iterations is shown below. Figure 5 As shown, the convergence curve of the bit error rate (BER) of separating the three source signals as a function of the number of iterations is as follows: Figure 6 As shown in the curves, the blind separation method for multi-time-frequency overlapping mixed signals based on the diffusion model described in this application exhibits extremely fast convergence speed. All three curves show a sharp downward trend in the initial 50 iterations, and then stabilize and approach zero after approximately 100 to 150 iterations. This indicates that the blind separation method for multi-time-frequency overlapping mixed signals based on the diffusion model can quickly lock signal features and remove interference. It is noteworthy that although signal 3 has the weakest energy and is subject to the strongest interference, resulting in a higher initial error, its error curve still successfully converges to zero as iterations proceed. This demonstrates that the method, utilizing the synergistic correction mechanism of "diffusion generation" and "consistency constraint," can not only efficiently process strong signals but also accurately separate weak signals under strong interference environments, exhibiting extremely high separation accuracy and robustness.

[0132] Specifically, the blind separation method for multi-time-frequency overlapping mixed signals based on the diffusion model is used to perform blind separation of the mixed signal. The mixed signal is a linear mixture of three source signals, including two modulation signals and a Gaussian white noise signal. The modulation types of the two modulation signals are BPSK and QPSK, respectively, with symbol rates of 600kHz and 800kHz. The mixing coefficients of modulation signal 1 and modulation signal 2 are 1.8 and 1, respectively, corresponding to a signal-to-energy ratio of 5:1. Other simulation parameters are consistent with the previous experiment. SNR is defined as the signal-to-noise ratio of modulation signal 1 to the white noise signal.

[0133] Based on the average results of 100 independent experiments, the convergence curve of the mean square error (MSE) of separating the three signals as a function of SNR is shown below. Figure 7 As shown, the convergence curves of the bit error rate (BER) of the three separated signals as a function of SNR are as follows: Figure 8 As shown, by Figure 7 and Figure 8 It can be seen that as the SNR increases, both the MSE and BER of the two signals exhibit a monotonically decreasing trend, indicating that the separation accuracy significantly improves with the improvement of channel quality. More importantly, even in low signal-to-noise ratio (SNR) environments (e.g., 5dB), this method can still maintain a low error level, while at high SNR (25dB), the MSE drops even further. The fact that the signal is orders of magnitude lower than the target strongly demonstrates the robustness of this technical solution in complex noise and interference environments, enabling high-precision separation and reconstruction of multiple time-frequency overlapping signals.

[0134] Example 4

[0135] In this embodiment, a computer device 100 is proposed, which includes a memory 101, a processor 102, and a computer program stored in the memory 101 that can be executed by the processor. The processor 102 executes the computer program to implement the blind separation method for multi-time-frequency overlapping mixed signals based on the diffusion model. A schematic diagram of the device is shown below. Figure 9 As shown.

[0136] In this embodiment, a computer storage medium is also proposed, on which a computer program is stored. The computer program includes program instructions, which, when executed by a processor, implement the blind separation method for multi-time-frequency overlapping mixed signals based on a diffusion model.

[0137] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A blind separation method for multi-time-frequency overlapping mixed signals based on a diffusion model, characterized in that, Includes the following steps: S1. Preprocess the multi-time-frequency overlapping mixed signal to obtain the current estimation information of each source signal in the multi-time-frequency overlapping mixed signal; the number of signal sources in the multi-time-frequency overlapping mixed signal is known; S2. Construct a cascaded mixed-signal separation model and pre-train it. The cascaded mixed-signal separation model includes... A parallel diffusion generation module and A parallel consistency constraint module, A diffusion generation module and The consistency constraint modules interact and iterate to obtain the estimation information of each source signal. , It is a positive integer; S3. Order Set the maximum number of iterations to Execute S4; S4. Proceed to the next step. The next iteration: [The number of iterations is missing from the original text.] The current estimation information of the source signal is input to the first... The diffusion generation module obtains the first... The score estimates of the current estimation information of each source signal are obtained in total. The score estimate is the current estimation information of the source signal; the score estimate is the score estimate from the first... Recovering a signal from a source signal that conforms to the prior distribution of that source signal provides an optimization direction; The first step is obtained by utilizing the diffusion generation module. The score estimation process for the current estimated information of each source signal is as follows: S41. Order Set the maximum number of noise-added samples to 1. Execute S42; S42. Based on the random sampling noise vector and the random sampling time step, perform noisy sampling on each source signal to obtain noisy source signal samples. The expression is: in, Indicates the random sampling time step. Indicates the number of samplings with added noise. This represents the signal attenuation coefficient corresponding to the time step. This represents a randomly sampled noise vector that follows a standard complex normal distribution. S43. Based on the noisy source signal samples and random sampling time steps, estimate the noise in the noisy sampling using a diffusion model. ; S44. Determine whether the number of noise samplings has reached the preset maximum number of noise samplings. If not, then let Increment the value by 1 and return S42; if so, calculate the 1st noise level smoothing based on multi-noise level smoothing. Current estimation information of individual source signals Score estimation The calculation expression is: in, This represents the standard deviation of the noise vector added to the source signal. Represents all random sampling time steps Expectations; S5. In the In the consistency constraint module, the first one The source signal is used as the reference signal, based on The score estimate of the current estimation information of each source signal is used to jointly optimize the reference signal to obtain the score estimate of the current source signal. The maximum a posteriori of the source signals, and using the maximum a posteriori to analyze the remaining signals. Coordinated correction of the estimated values ​​of individual source signals; The joint optimization of the reference signal includes: Using the score estimate as a probability gradient, based on a preset iterative learning rate... Update the current estimation information of the reference signal, and the update expression is: in, This represents the updated reference signal estimate. The mixing coefficients of the reference signal are represented. This represents the score estimate of the reference signal. Indicates the remainder The mixing coefficients of the source signals, Indicates the remainder Score estimation of individual source signals; The use of the first The maximum posterior pair of the source signals remains The estimated values ​​of each source signal are used for collaborative correction, and the calculation expression is as follows: in, Indicates the remainder A single source signal among multiple source signals Indicates the remainder Interference signals in the source signal Indicates the remainder The mixing coefficient of interference signals in a source signal; S6. Based on the correction results, obtain The estimation results from different reference signals are based on the... The estimation results from different reference signals yielded the first group. Estimates of the source signal; S7. Determine if the maximum number of iterations has been reached. N If so, then the first The estimated value of the source signal is used as the first... The final separation result of the source signals; otherwise, let The value is increased by 1, and the estimated value obtained in S6 is used as the first... The current estimation information of each source signal is returned to S4.

2. The blind separation method for multi-time-frequency overlapping mixed signals based on a diffusion model according to claim 1, characterized in that, The preprocessing process for the multi-time-frequency overlapping mixed signal is as follows: S11. Order Determine the received multi-time-frequency overlapping mixed signal Source signal in , Indicates the order of the source signals; S12. For the source signal Blind synchronization processing is performed to obtain a multi-time-frequency overlapping mixed signal after blind synchronization. ; S13. For the multi-time-frequency overlapping mixed signal after blind synchronization Perform deconstellation mapping and remodulation to obtain the source signal. Current estimation information ; S14. Transfer the source signal Current estimation information Multi-time-frequency overlapping mixed signal after blind synchronization Perform cross-correlation alignment and subtraction to obtain the residual signal. ; S15. Order The value increases by 1; S16. Judgment Is it less than or equal to? If so, the residual signal As input, return S12; Otherwise, output Individual source signal Current estimation information .

3. The blind separation method for multi-time-frequency overlapping mixed signals based on a diffusion model according to claim 2, characterized in that, The blind synchronization process is as follows: The mixed signal is determined using blind symbol rate estimation of a single-carrier signal. Symbol rate, based on symbol rate for multi-time-frequency overlapping mixed signals Sampling is performed, and blind carrier synchronization and blind symbol timing synchronization are executed on the sampling results.

4. The blind separation method for multi-time-frequency overlapping mixed signals based on the diffusion model according to claim 3, characterized in that, The multi-time-frequency overlapping mixed signal also has known modulation types of each source signal, and each of the diffusion generation modules is loaded with a pre-trained diffusion model corresponding to the modulation type of the source signal.

5. The blind separation method for multi-time-frequency overlapping mixed signals based on the diffusion model according to claim 4, characterized in that, S6 is based on the The estimation results from different reference signals yielded the first group. The estimated values ​​of the source signals include: Based on preset quality indicators, from Selected from group estimation results The group estimation results, the ; Based on the above In the group estimation results, the calculation of the first group is performed. The average of the estimation results of the n source signals is used as the nth The estimated value of the source signal.

6. A computer device, characterized in that, The computer device includes a memory, a processor, and a computer program stored in the memory that can be executed by the processor, wherein the processor executes the computer program to implement the method according to any one of claims 1 to 5.

7. A computer storage medium, characterized in that, It stores a computer program, which includes program instructions that, when executed by a processor, implement the method described in any one of claims 1 to 5.

Images