Signal restoration method and device, equipment, storage medium and product

Abstract

The invention relates to the technical field of wireless communication, and discloses a signal restoration method and device, computer equipment and a storage medium, target diffusion time and a target received signal obtained by a receiving end are determined, and then the target received signal and a pure noise signal are utilized to carry out data prediction processing operation based on expected propagation, so that a target diffusion starting point is obtained. According to the method, a noise-containing receiving signal is pre-estimated through an expected propagation algorithm to obtain a better initial state, so that a starting point with more information content is provided for reverse sampling, and the reverse sampling is guided to converge to a reasonable data manifold more quickly. And finally, performing reverse diffusion sampling operation based on the target diffusion starting point, iteratively optimizing signal estimation in the diffusion sampling process, and finally obtaining a more accurate target original signal.

Description

Technical Field

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

[0002] Large-scale multiple-input multiple-output (LS-MIMO) technology is key to achieving high spectral efficiency and high energy efficiency in fifth-generation (5G) and future sixth-generation (6G) wireless communication systems. With its core advantages of significantly improving system capacity and connection reliability, it is widely used in 5G / 6G communication base stations and user equipment.

[0003] However, due to the large-scale nature of MIMO systems, the methods for accurately separating and recovering the original transmitted signal from the mixed and noisy received signal in related technologies still have certain limitations. A signal restoration method that can balance complexity and detection accuracy is needed. Summary of the Invention

[0004] This application aims to at least partially solve one of the technical problems in related technologies. To this end, this application proposes a signal restoration method, apparatus, computer equipment, storage medium, and product. The main technical solutions adopted in this application include:

[0005] In a first aspect, embodiments of this application provide a signal restoration method applied to the receiver of a large-scale multiple-input multiple-output communication system. The method includes: determining the target diffusion time and the target received signal obtained by the receiver; performing data prediction processing based on expected propagation using the target received signal, the target diffusion time, and a pure noise signal to obtain the target diffusion start point; wherein the target diffusion time is determined based on a preset noise scheduling parameter; and performing a reverse diffusion sampling operation based on the target diffusion start point to obtain the original target signal.

[0006] Secondly, embodiments of this application provide a signal restoration device applied to the receiver of a large-scale multiple-input multiple-output communication system. The device includes: a signal determination module for determining the target diffusion time and the target received signal obtained by the receiver; a starting point determination module for performing data prediction processing based on desired propagation using the target received signal, the target diffusion time, and a pure noise signal to obtain the target diffusion starting point; wherein the target diffusion time is determined based on a preset noise scheduling parameter; and a reverse diffusion module for performing reverse diffusion sampling based on the target diffusion starting point to obtain the original target signal.

[0007] Thirdly, this application also provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above methods.

[0008] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the methods described above.

[0009] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of any of the above methods.

[0010] In the above embodiments, by determining the target diffusion time and the target received signal at the receiver, and using the target received signal, the target diffusion time, and the pure noise signal to perform data prediction processing based on expectation propagation, the target diffusion starting point is obtained. The noisy received signal is pre-estimated using the expectation propagation algorithm to obtain a better initial state, thus providing a more informative starting point for backsampling and guiding it to converge more quickly to a reasonable data manifold. Finally, a backsampling sampling operation is performed based on this target diffusion starting point, and the signal estimation is iteratively optimized during the diffusion sampling process to ultimately obtain a more accurate original target signal. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0012] Figure 1 This is a flowchart of a signal restoration method according to an embodiment of this application;

[0013] Figure 2a This is a flowchart of a method for performing a reverse diffusion sampling operation according to an embodiment of this application;

[0014] Figure 2b A diagram illustrating an iterative algorithm for a data prediction model provided according to an embodiment of this application;

[0015] Figure 2c A diagram of a reverse diffusion sampling iterative algorithm provided according to an embodiment of this application;

[0016] Figure 3a This is a block diagram of a desired propagation-guided diffusion detector according to one embodiment of this application;

[0017] Figure 3b This is a graph showing the performance comparison results provided according to one embodiment of this application;

[0018] Figure 3c This is a graph showing the performance comparison results according to another embodiment of this application;

[0019] Figure 4 This is a structural block diagram of a signal restoration apparatus according to an embodiment of this application;

[0020] Figure 5 This is an internal structural diagram of a computer device provided according to an embodiment of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] Large-Scale Multiple-Input Multiple-Output (LS-MIMO) technology is widely used in 5G / 6G communication base stations and user equipment. However, to fully realize the potential of LS-MIMO, it is crucial to develop high-precision and computationally feasible data detection algorithms.

[0023] However, the LS-MIMO detectors in related technologies still have limitations: Maximum Likelihood (ML) has the best performance but its complexity is exponential, making it unsuitable for large-scale scenarios; traditional linear detectors such as Zero-Forcing (ZF) detectors and Minimum Mean-Square Error (MMSE) detectors have low complexity but suffer significant performance loss under high loads or high-order modulation. Therefore, nonlinear detection schemes have been developed instead.

[0024] For high-order modulation scenarios, the Expectation Propagation (EP) algorithm exhibits excellent detection accuracy. Furthermore, Generative Diffusion Models (GDMs) demonstrate powerful capabilities for modeling complex high-dimensional data through forward noise addition and inverse data recovery. Higher-order diffusion solvers such as DPM-Solver++ can reduce sampling steps and improve efficiency, making diffusion model-based MIMO detection potentially achieve both high performance and low complexity, enhancing its practical application feasibility. However, ALD diffusion detectors based on Annealed Langevin Dynamics (ALD) within the diffusion model framework require a large number of sampling and iteration steps, resulting in high latency. The EP detector based on expectation propagation experiences significantly increased computational complexity in LS-MIMO with an increased number of antennas.

[0025] To address the above technical problems, this application provides a signal restoration method embodiment. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0026] This embodiment provides a signal restoration method applied to the receiver of a large-scale multiple-input multiple-output communication system. Figure 1 This is a flowchart of a signal restoration method according to an embodiment of this application, such as... Figure 1 As shown, the process includes the following steps:

[0027] S110. Determine the target diffusion time and the target received signal obtained by the receiver.

[0028] It should be noted that the receiver in a large-scale multiple-input multiple-output (MIMO) communication system can be either the base station side or the user equipment side, corresponding to the transmitter side. The target received signal refers to the signal data acquired by the receiver, which contains information about the transmitted signal to be reconstructed. It can serve as a conditional input for the entire signal reconstruction process.

[0029] Specifically, the target received signal can be determined in the following ways:

[0030] First, determine the initial received signal obtained by the receiver. The initial received signal can refer to the original signal directly received by the receiving antenna array, represented by a complex domain channel model. For example, consider a MIMO system configured with N transmit antennas and M receive antennas, where the transmit symbol vector... Contains N items from constellation sets Quadrature Amplitude Modulation (QAM) symbols. Specifically, constellation sets. Each symbol in The average power is So the initial received signal It can then be expressed as follows:

[0031]

[0032] in, It is the initial received signal in the complex domain; It is a vector of additive white Gaussian noise (AWGN) in the complex field, which can be represented as Let I be the complex domain variance of the noise. M The matrix is ​​the identity matrix, and M is the number of receiving antennas; This is the initial channel matrix, where M is the number of receive antennas and N is the number of transmit antennas; It is the initial transmitted signal of the complex domain to be restored.

[0033] It should be noted that the initial channel matrix refers to the data describing the signal transmission link characteristics between each transmitting antenna at the transmitter and each receiving antenna at the receiver, which is obtained in advance through channel estimation.

[0034] Furthermore, to facilitate analysis and processing, the complex domain channel model shown above can be subjected to an equivalent real-valued transformation. Specifically, to perform an equivalent real-valued transformation on the above channel model, the complex domain channel model can first be decomposed into its real part. and the virtual part Two parts. For example, first define the transmitted signal to be restored as... The initial received signal is Additive white Gaussian noise is And the channel matrix is The equivalent model in the real domain can then be derived from the following equation:

[0035] y = Hx + n

[0036] Where y represents the equivalent target received signal, which is essentially a real-domain signal; H represents the equivalent real-domain channel matrix; and n represents the equivalent real-domain additive white Gaussian noise, which also has... I M Let M be the identity matrix, and M be the number of receiving antennas. Let be the real-domain variance of the noise, and how does it relate to the complex-domain variance of the noise? x represents the equivalent target transmitted signal.

[0037] It should be noted that, since the real and imaginary parts of a QAM symbol are orthogonal, after splitting the QAM symbol into its real and imaginary parts, the symbols in vector x now all belong to the Pulse Amplitude Modulation (PAM) constellation set Ω, and the constellation size... Where |Ω| represents the cardinality of the PAM constellation set, Denotes the cardinality of the QAM constellation set, x∈Ω 2N Each symbol x in the constellation base Ω b The average energy is In other words, the PAM constellation set can be viewed as a decomposition of the QAM constellation set in the real domain, and the two parts after decomposition can correspond to the symbols in the PAM constellation set respectively. For example, after 64QAM is decomposed into real and imaginary parts, each part corresponds to an 8PAM constellation set. Furthermore, since the discrete points of the PAM constellation set are distributed only on a single real axis, different symbols are distinguished only by amplitude differences, making it more suitable for the operational logic of the real domain.

[0038] In this case, if we assume that the receiver has perfect Channel State Information (CSI), then the MIMO detection task can be regarded as recovering the original transmitted signal x based on the known target received signal y and the equivalent real-valued channel matrix H.

[0039] Thus, by converting the complex domain channel model to the real domain, the high-dimensional complex matrix operations originally performed in the complex domain are broken down into low-dimensional real matrix operations involving the real and imaginary parts. This allows all subsequent detection, inversion, and iterative calculations to be performed entirely in the real space. This not only significantly reduces computational complexity but also maintains the integrity of the channel information, providing a faster and more hardware-efficient numerical foundation for subsequent signal reconstruction steps.

[0040] S120. Using the target received signal, target diffusion time, and pure noise signal, perform data prediction processing based on desired propagation to obtain the target diffusion starting point.

[0041] It should be noted that if you want to gradually recover the original signal from the noisy signal, you can use a diffusion probability model (DPMs).

[0042] Specifically, the diffusion probability model is a model that restores a signal by first adding noise and then denoising it. Its core lies in defining a forward diffusion process within a preset time range (time step) and a learnable reverse process. In the forward diffusion process, the noisy signal at each time step follows a Gaussian distribution, and the noise intensity of this signal is controlled by predetermined noise scheduling parameters. The reverse diffusion process, on the other hand, learns denoising rules by solving differential equations to reverse the forward process.

[0043] For example, the forward diffusion process {x t} t∈[0，T] It can be viewed as a random variable x0∈Ω 2N Starting from a noisy signal x at any time step t t The conditional distribution p(x) t |x0) follows a mean of α t x0, variance is Real-valued Gaussian distribution The process. Where [0, T] represents the preset time range, and T represents the target iteration time step; x t Let x0 represent the current noisy signal at the current time step t. 2N In this context, x0 refers to the initial state of the diffusion process, i.e., the diffusion start point; Ω represents the PAM constellation set; and N is the number of transmitting antennas.

[0044] The mean is α t x0, variance is A real-valued Gaussian random variable, α t σ represents the scaling factor. t Let I represent the noise standard deviation at the current time step t, and let I be the identity matrix.

[0045] It is understandable that the intensity variation of forward diffusion noise in the diffusion probability model, i.e., noise scheduling, is determined by the scaling factor α. t and noise standard deviation σ t These two parameters, jointly determined by the time step, directly influence the noise addition ratio at each time step, thus determining the noisy signal x at each step. t Can a balance be struck between preserving the original signal characteristics and introducing reasonable noise? For example, the scaling factor α... t and noise standard deviation σ t Relationship is acceptable

[0046] Furthermore, due to the scaling factor α t This can be understood as the coefficient that preserves the characteristics of the original signal during forward diffusion, α tThe closer it is to 1, the better the current noisy signal x at time step t is. t The closer it is to the original signal x, the better. And the noise standard deviation σ... t This can be understood as an index that introduces noise intensity, σ t The larger the value, the more significant the interference of noise on the original signal. Therefore, the α value corresponding to time step t can also be determined by the following formula for calculating the parameters of noise scheduling. t value:

[0047]

[0048] In the formula, α t β0 represents the scaling factor; t represents the current time step; β0 and β1 are both noise scheduling parameters, and can be β0 = 0.1 and β1 = 20.

[0049] Furthermore, to avoid boundary issues in subsequent calculations, such as the parameter definition at time step t=0, a scaling factor α can be specifically defined. t The boundary value is α -1 =0, to handle boundary conditions in subsequent calculations.

[0050] The sampling process (i.e., reverse diffusion) of the diffusion probability model is achieved by solving the corresponding ODE. Specifically, time t can be reversed from the end point T of the forward diffusion to the starting point 0 of the original signal, which means discretizing the following ODE:

[0051]

[0052] Where t represents the current time step; x t σ represents the current noisy signal at the current time step t; t This represents the noise standard deviation at the current time step t. The noise variance at the current time step t is represented by c; c is the condition variable; the coefficient function f(t) is defined as follows: Where, σ t This represents the scaling factor for the current time step t. This represents the logarithmic derivative of the scaling factor with respect to the current time step; g 2 The coefficient function of (t) is defined as This represents a conditional data prediction model, which aims to predict the current noisy signal x based on the previous time step t. t We use condition c to estimate the initial state of diffusion, i.e., the diffusion start point x0.

[0053] It should be noted that, since the reverse diffusion of the traditional diffusion model usually starts with the pure noise signal at the target iteration time step T, that is, the starting point is set at t0=1. I is the identity matrix. However, according to the parameter formula for noise scheduling mentioned earlier, at t=1, α t ≠0 and σ t The noise distribution of the noisy signal at the current time step t is not strictly a standard normal distribution. Theoretically, as the diffusion time t→∞, a purely noisy signal x can be obtained. inf α inf =0 and σ inf =1, and x can be obtained from the parameter formula of noise scheduling mentioned above. inf Indeed, it has converged to Since the time step t cannot reach infinity in practice, the initial state needs to be optimized and improved. Furthermore, excessive noise can cause the loss of effective information from the original signal, making it difficult for back diffusion to converge to the true signal. Therefore, to guide back diffusion to more efficiently approximate the true signal distribution, a better initial state needs to be determined as the target diffusion starting point.

[0054] Specifically, by using the target received signal, the target diffusion time, and the pure noise signal to perform data prediction processing based on the desired propagation, the starting point of the target diffusion can be obtained. First, signal prediction can be performed using the target received signal and the pure noise signal to obtain the initial estimate of the original transmitted signal.

[0055] The target diffusion time can refer to a smaller time value chosen during the forward diffusion process to control the degree of noise introduction, thereby providing a more favorable initial state for reverse sampling. Since the purpose of the target diffusion time is to avoid excessive degradation of the information contained in the predicted value during the initial forward diffusion process, the target diffusion time can be determined empirically or based on preset noise scheduling parameters. The pure noise signal can refer to a random noise vector that does not contain any effective transmitted signal information and follows a Gaussian distribution. It should be noted that the pure noise signal here is used to help distinguish between effective information and interference noise in the target received signal; it is essentially an interference reference quantity unrelated to the real signal, and is not the same as the additive white Gaussian noise mentioned earlier due to channel characteristics. The initial estimate can refer to a linear estimate that is close to the original transmitted signal and is possibly optimal. Specifically, the target received signal y, which contains original transmitted signal information but is mixed with interference, can be compared with the pure noise signal x, which serves only as an interference reference and has no effective information. infFeature comparison is performed. Based on the interference characteristics of pure noise signals, the portion of the target received signal that differs significantly from the pure noise characteristics is selected. This portion is highly likely to be valid transmitted signal information, thus eliminating interference components consistent with pure noise characteristics. Then, combined with known channel characteristics, such as signal attenuation or phase change patterns from the transmitting antenna to the receiving antenna, the selected valid information portion is adjusted in reverse to counteract signal distortion caused by channel transmission. Finally, based on the PAM constellation set to which the original transmitted signal belongs, the adjusted signal values ​​are corrected to the valid discrete points of the constellation set, thus obtaining an initial estimate of the original transmitted signal that effectively approximates the original transmitted signal. Due to this initial estimate This represents a predicted value of the original data, and therefore can be used as an effective starting point for the subsequent short-term forward diffusion process.

[0056] Subsequently, the initial estimate is subjected to a short-term forward diffusion process using the target diffusion time to determine the target diffusion starting point.

[0057] Short-time forward diffusion processing refers to the process of substituting the obtained initial estimate into a small target diffusion time according to the forward diffusion formula to generate a new signal with finite noise. For example, the initial sampling state of the reverse sampling process can first be derived as follows:

[0058]

[0059] In the formula, This represents the initial noise state (sampling state), i.e., the starting point of target diffusion; t0 represents the target diffusion time; x represents the noise standard deviation at the current time step after the target diffusion time; inf This represents a pure noise signal; This represents the noise standard deviation of the initial estimate; h0 represents the scaling factor corresponding to the current time step after the target diffusion time; h0 represents the change in the logarithmic signal-to-noise ratio (log-SNR) at the current time step after the target diffusion time, which can be regarded as an intermediate variable for simplified calculation. This represents the initial estimate obtained after signal prediction.

[0060] Understandably, due to the initial estimate The noise standard deviation has Furthermore, since the change in its logarithmic signal-to-noise ratio h0→∞, the above formula can also be simplified to:

[0061]

[0062] in, Indicates the starting point of target diffusion; The current time step represents the noise standard deviation after the target diffusion time; t0 represents the target diffusion time; x inf This represents a pure noise signal; This represents the scaling factor corresponding to the current time step; This represents the initial estimate.

[0063] Based on the forward diffusion principle, this equation physically represents the initial estimate. The process of diffusion to the selected target diffusion time t0, to the target diffusion starting point.

[0064] Optionally, to ensure that only a limited amount of noise is introduced, the target diffusion time can be set to 0.3. It should be noted that this target diffusion time is a continuous mathematical parameter, ranging from 0 to 1. It does not refer to clock time, but rather a progress scale used to continuously control the degree of noise addition. Progress t = 0 represents the starting point (clean data), and progress t = 1 represents the ending point (pure noise). A target diffusion time of 0.3 balances the amount of noise added with the degree of preservation of the original signal, ensuring the stability and accuracy of the subsequent reverse diffusion process.

[0065] Therefore, by first using the target received signal and the pure noise signal, the effective component with the greatest difference from the pure noise characteristics is filtered out. Then, the distortion is inverted according to the channel characteristics and mapped back to the legitimate constellation point to obtain a highly reliable initial estimate. Subsequently, the estimate is transformed into the target diffusion starting point with finite noise in only a very small diffusion time. This ensures that the inverse sampling starts in a high-probability region close to the true solution from the beginning. This avoids a large number of blind iterations starting from zero and minimizes the introduced noise, significantly compressing the number of convergence steps and computational delay. At the same time, it maintains the equivalent structure of the traditional algorithm and directly reduces the symbol error rate under the same complexity budget.

[0066] S130. Perform reverse diffusion sampling based on the target diffusion starting point to obtain the original target signal.

[0067] The reverse diffusion sampling operation refers to a processing method that uses the target diffusion starting point as the initial input and gradually reverses the forward diffusion process by discretizing and solving the diffusion ordinary differential equation (ODE) or by gradient-guided iterative adjustment to finally obtain the original target signal. For example, this can be achieved by breaking down the continuous time step into discrete small steps, with each step directly updating the signal's basic Euler method based on the rate of change of the current signal (ODE derivative). Alternatively, it can be achieved by combining the signal's posterior gradient information with gradually decreasing noise intensity, iteratively adjusting the noisy signal to gradually converge to the ALD of the original signal distribution. It can also be implemented using the Diffusion Probabilistic Models Solver++ (DPM-Solver++) framework.

[0068] Specifically, a reverse diffusion sampling operation is performed based on the target diffusion starting point to obtain the original target signal, including:

[0069] First, the starting point of the target diffusion is taken as T. L The sampling state corresponding to the time step.

[0070] Among them, T L The time step can refer to the current time step in the reverse diffusion sampling process. Specifically, if T... L The time step serves as the starting time step for reverse diffusion sampling. It corresponds to the time node after a small amount of noise has been added during forward diffusion, such as the time step number corresponding to the target diffusion time in forward diffusion. Its value is greater than 0 and less than the maximum time step T of the forward diffusion. The purpose is to start reverse denoising from a signal containing a small amount of noise, rather than starting from pure noise. And T... L The sampling state corresponding to the time step refers to T L The signal containing a small amount of noise at the time step is essentially the target diffusion starting point, which retains the key information of the original transmitted signal and conforms to the signal distribution requirements of the diffusion model. Specifically, the target diffusion starting point obtained through short-time forward diffusion processing in the aforementioned steps can be directly assigned to the inverse sampling algorithm at time step T. L The state variables required for the time step.

[0071] Subsequently, at the current time step T L In the case of time step, based on T L The sampling state corresponding to the time step is subjected to a reverse diffusion sampling operation to obtain T. L The denoised signal corresponding to the time step.

[0072] Among them, T LThe denoised signal corresponding to the time step can refer to the output result after one back-diffusion update calculation. It represents the denoised signal value that is closer to the real signal after denoising the current sampling state once. L represents the number of iterations in the back-diffusion sampling operation. Specifically, this can be achieved by calling a higher-order ODE solver, such as DPM-Solver++, to denoise the current time step T. L and the current time step T L The current sampling state, the target received signal, and predefined noise scheduling parameters are taken as inputs. The solver then calculates and outputs an updated, less noisy signal estimate based on the inverse ordinary differential equation of the diffusion model, i.e., performing a reverse diffusion sampling operation to obtain T. L The denoised signal corresponding to the time step.

[0073] Next, update the current time step to T. L-1 Time step.

[0074] Among them, T L-1 The time step is T L The next time step after the previous time step. It's important to note that T... L-1 The time step is T L The time step is the previous time step on the time axis. Because back diffusion is a process starting from the moment of maximum noise T... L The process involves backtracking towards time T0 when the noise is zero, so the time step index is decreasing.

[0075] Then, T L The denoised signal corresponding to the time step is used as T L-1 The sampling state corresponding to the time step is used to perform the above reverse diffusion sampling operation again to obtain T. L-1 The denoised signal corresponding to the time step.

[0076] Similarly, it is understandable that T L-1 The sampling state corresponding to the time step is the previous time step (T) L The denoising result of T is shown. Specifically, the implementation process is exactly the same as the previous step, that is, T... L The denoised signal of the time step is used as T L-1 The input sampling state of the time step is used to call the higher-order ODE solver again, inputting the new current time step (T). L-1 ) and other parameters, perform a reverse update calculation, and the output result is T. L-1 The denoised signal corresponding to the time step.

[0077] Finally, the iterative process of the reverse diffusion sampling operation is repeated. When the current time step reaches the target time step, the denoised signal corresponding to the target time step is used as the target denoised signal to determine the target original signal.

[0078] The target time step can refer to the termination time step of the reverse diffusion process, that is, the time step in the forward diffusion that is finally set to have no added noise. Its value can be 0, corresponding to the initial state of the forward diffusion. The target denoised signal can refer to the denoised signal at the target time step, where the noise intensity has been reduced to the minimum, or even can be regarded as almost noiseless, completely preserving the information of the original transmitted signal.

[0079] It should be noted that although the target original signal refers to the original symbol signal sent by the transmitter without being transmitted through the channel and subjected to noise interference, after the iterative process of the reverse diffusion sampling operation, it can be approximated that the final denoising result calculated at the target time step is infinitely close to the original signal of the transmitted symbol vector to be restored. In other words, the target denoised signal corresponding to the target time step can be regarded as the target original signal.

[0080] In other words, the iterative process can be executed according to the following logic: Each time, first determine whether the current time step is the target time step. If not, continue executing the loop process of using the denoised signal of the current time step as the sampling state of the next time step, denoising based on the sampling state, obtaining the denoised signal of the next time step, and updating the time step to the current time step - 1. If the current time step has reached the target time step, stop the iteration, and use the denoised signal at this time (the denoised signal corresponding to the target time step) as the target original signal that has restored the original signal from the transmitting end.

[0081] Thus far, by using T with a small amount of noise L Starting with a time step, the process of gradually decreasing the time step to repeatedly denoise avoids the problems of traditional reverse diffusion starting from pure noise and losing a lot of information. By combining the reference information of the target received signal at each step, the denoising direction is ensured to be accurate. In the end, the original target signal can be restored efficiently and accurately, which is suitable for the requirements of large-scale MIMO systems for high signal restoration accuracy and fast convergence speed.

[0082] In the above implementation, the target diffusion time and the target received signal are determined, and the target received signal, the target diffusion time, and the pure noise signal are used to perform data prediction processing based on expectation propagation to obtain the target diffusion starting point. The noisy received signal is pre-estimated using the expectation propagation algorithm to obtain a better initial state, thus providing a more informative starting point for backsampling and guiding it to converge more quickly to a reasonable data manifold. Finally, a backsampling sampling operation is performed based on this target diffusion starting point, and the signal estimation is iteratively optimized during the diffusion sampling process to obtain a more accurate original target signal.

[0083] In some implementation methods, please refer to the appendix. Figure 2a The reverse diffusion sampling operation is performed in the following manner:

[0084] S210. For the current time step, perform data prediction processing based on the target received signal and the sampling state of the current time step to obtain the current signal estimate corresponding to the current time step.

[0085] The current time step can refer to a specific time node being processed during the reverse diffusion sampling process. The sampling state at the current time step refers to the intermediate result of the signal after partial denoising at that time step. It includes signal information inherited from the previous time step and gradually purified. The current signal estimate corresponding to the current time step is an optimal linear estimate that is closer to the original transmitted signal, obtained by further denoising the current sampling state through data prediction processing. It is usually represented by the mean of a Gaussian distribution.

[0086] It should be noted that regardless of whether the current time step is the starting T L Or the T in the middle L-1 As long as the target time step has not been reached, the operation of performing expected propagation processing based on the current sampling state and the target received signal to obtain the current signal estimate will be repeated, ensuring that noise is gradually reduced and the signal is approximated at each step.

[0087] For example, data prediction processing based on expectation propagation can be performed in the following ways:

[0088] First, based on the sampling state at the current time step and the target received signal, construct the current posterior probability distribution corresponding to the current prediction round.

[0089] The posterior probability distribution includes a variational prior distribution. It can be understood that the current posterior probability distribution describes the probability distribution of the original transmitted signal's possible values ​​under the current sampling state, requiring consideration of the target received signal's reference information and the current noise intensity. The variational prior distribution included is a Gaussian distribution assumed for computational simplification; it can be described using only the mean and variance, and can be further decomposed into a small Gaussian distribution for each transmitted symbol, reducing complexity.

[0090] Subsequently, given that the current prediction round is L1, a moment matching update operation is performed on the variational prior distribution of the L1 prediction round to obtain the optimized posterior probability distribution corresponding to the L1 prediction round.

[0091] The specific moment matching update operation includes: First, temporarily removing the influence of a single transmitted symbol node in the variational prior distribution to obtain the cavity edge distribution, thus eliminating interference from the initial guess of that node. Second, using an indicator function, only valid signal values ​​of the PAM constellation set to which the original transmitted signal belongs are accepted; invalid values ​​are invalid. This, along with the transmission pattern of the target received signal, replaces the interference components in the cavity edge distribution, resulting in a distribution closer to the real signal and also approximately a Gaussian edge posterior distribution. Third, adjusting the mean and variance of the edge posterior distribution to be completely consistent with the mean and variance of the previously removed variational prior distribution nodes, i.e., performing the moment matching update operation, ultimately integrating to obtain the optimized posterior probability distribution for the L1 prediction rounds.

[0092] Next, update the current prediction round to the L2 prediction round, and use the optimized posterior probability distribution corresponding to the L1 prediction round as the variational prior distribution of the L2 prediction round. Then, perform the above moment matching update operation again to obtain the optimized posterior probability distribution corresponding to the L2 prediction round.

[0093] In this step, the L2 prediction round is the next prediction round after the L1 prediction round. This step uses the distribution optimized in the previous round as a new initial guess, and then processes it again by removing nodes, replacing legal signals, and performing moment matching operations to match the mean and variance, so that the optimized posterior probability distribution is closer to the true original signal distribution.

[0094] Furthermore, repeat the iterative process of the moment matching update operation described above. If the current prediction round reaches the target prediction round, use the optimized posterior probability distribution corresponding to the current prediction round as the target posterior probability distribution.

[0095] The target prediction round can refer to the prediction round determined based on experience. It should be noted that the target prediction round can be 2 to 3 times. Experiments have shown that the distribution at this number of rounds is close enough to the true posterior. Increasing the number of rounds will have limited improvement on accuracy and will also increase the amount of computation.

[0096] Ultimately, since the target posterior probability distribution is a Gaussian distribution, its mean is the value that is most likely to be close to the original transmitted signal. It can retain the effective signal to the greatest extent and eliminate noise. Therefore, the target mean data of the target posterior probability distribution can be used as the current signal estimate corresponding to the current time step.

[0097] For example, the process can be referred to Figure 2b As shown, data prediction processing based on expectation propagation can be implemented through a data prediction model, and the sampling state at the current time step... The current time step t is determined from the target received signal y. i Corresponding current signal estimate It can be represented as:

[0098]

[0099] In the formula, Represents the sampling state at the current time step; t i y represents the current time step; y represents the target received signal. This represents the data prediction model at the current time step, and represents the estimated signal value obtained by performing data prediction processing based on expectation propagation.

[0100] Specifically, the first step is to construct the current posterior probability distribution corresponding to the current prediction round based on the Maximum A Posterior (MAP) criterion. It can be represented as:

[0101]

[0102] Among them, t i Represents the current time step; This represents the current signal estimate; ∝ represents the sampling state corresponding to the current time step; ∝ represents proportional to, ignoring the normalization constant; y represents the target received signal.

[0103] It should be noted that, Represents an indicator function, and when (that is to mean) When the value is the b-th element, it takes the value 1; otherwise, it takes the value 0. b is the symbol number in the constellation set Ω to which the original signal belongs. The constellation set Ω has a total of 2N elements, where N is the number of transmitting antennas. That is, b represents the current signal estimate. The element number in the constellation set Ω.

[0104] In the given Under the given conditions, the target received signal y and the sampling state corresponding to the current time step It is conditionally independent, and Understandable Represents prior probability. Represents the joint likelihood probability. This represents the likelihood probability of the current signal estimate. This represents conditional probability.

[0105] The mean μ and variance Σ of this approximate Gaussian distribution can then be expressed as:

[0106]

[0107] In the formula, μ represents the mean; Σ represents the variance; H represents the real-domain channel matrix; H Ty represents the transpose of the real-valued channel matrix; y represents the target received signal. This represents the noise variance at the current time step; This represents the scaling factor corresponding to the current time step; This indicates the sampling state corresponding to the current time step; I represents the noise variance at the starting point of the target diffusion; I represents the identity matrix.

[0108] It is important to reiterate that directly calculating the posterior probability typically involves high computational complexity. To address this, a novel and improved EP algorithm can be employed to approximate the formula with acceptable computational complexity. Specifically, this model aims to approximate the true current posterior probability distribution by minimizing the KL (Kullback-Leibler) divergence (i.e., relative entropy) on each marginal distribution. It approximates a Gaussian distribution. To achieve this, the data prediction model can iteratively construct a tractable approximate posterior probability by matching the mean and variance of the true fore and posterior probability distributions.

[0109] Specifically, we can first assume the current signal estimate. Let b be the variational prior distribution (i.e., the approximate Gaussian distribution) of the b-th element in the constellation set Ω given the condition. Adopt the following form:

[0110]

[0111] In the formula, ∝ represents proportional to; b represents the current signal estimate; b represents the current signal estimate. The element number in the constellation set Ω; γ b The ratio of the mean to the variance of the b-th element in the variational prior distribution that characterizes this approximately Gaussian distribution; Λ b The inverse of the variance of the b-th element in the variational prior distribution that characterizes this approximate Gaussian distribution.

[0112] Subsequently, the approximate posterior probability It can be given by the following decomposition:

[0113]

[0114] Furthermore, its covariance χ and mean η are respectively:

[0115]

[0116] Among them, γ = [γ1, γ2,..., γ 2N ],Λ=diag(Λ1,Λ2,...,Λ 2Nb∈[1,2,...,2N]. For initialization, you can set... Independently, the marginal posterior distribution of each emitted symbol is characterized in Gaussian form. Where, χ bb It is the b-th diagonal element of the covariance χ.

[0117] For example, if the current prediction round is an L1 prediction round, i.e., l is an L1 prediction round, then the approximate posterior probability at the current prediction round l is: Then, each variational prior distribution is updated independently. Obtain the approximate posterior probability of the next prediction round with refinement This yields the optimized posterior probability distribution for the next prediction round (L2 prediction round).

[0118] Next, update the current prediction round to the L2 prediction round, and repeat the above iterative process of moment matching update operation until the current prediction round l reaches the target prediction round L.

[0119] Specifically, the iterative steps first require calculating the cavity edge distribution:

[0120]

[0121] in, This represents the current signal estimate at the l-th iteration. The approximate posterior probability of the b-th element represents the probability estimate of the possible values ​​of the b-th symbol of the original transmitted signal in the l-th iteration, before the influence of the initial assumptions of the b-th element in the variational prior is eliminated; This represents the current signal estimate after removing the influence of the b-th element in the l-th iteration. The cavity probability of the b-th element, where cavity means temporarily excluding the interference of the initial guess of the b-th element in the variational prior; Indicates the current signal estimate. The b-th element; It represents the reciprocal of the variance of the b-th element in the variational prior distribution at the l-th iteration; This represents the ratio of the mean to the variance of the b-th element in the variational prior distribution at the l-th iteration.

[0122] Understandable This indicates that the edge distribution of the cavity can be approximated by a Gaussian distribution. This represents the mean of the Gaussian distribution, that is, the mean of the cavity edge distribution at the l-th iteration.

[0123] Subsequently, after obtaining the expression for the cavity edge distribution, expectation-based prediction can be performed using the following three parallel inter-symbol computation formulas to determine the current time step t. i Corresponding current signal estimate

[0124] It should be noted that in the expression for the cavity edge distribution... Let represent the variance of the cavity edge distribution during the l-th iteration, and we have:

[0125]

[0126] in, Let b be the diagonal element of the covariance matrix of the approximate posterior probability at the l-th iteration. This represents the mean value of the cavity edge distribution during the l-th iteration; Let represent the mean of the approximate posterior probability at the l-th iteration; It represents the reciprocal of the variance of the b-th element in the variational prior distribution at the l-th iteration; This represents the ratio of the mean to the variance of the b-th element in the variational prior distribution at the l-th iteration.

[0127] Understandably, this step represents removing element b from the variational prior distribution to strip away the b dimension from the initial guess, leaving space for updating the true posterior information. After removal, the variational prior distribution can be viewed as a marginal posterior distribution that approximates a Gaussian distribution.

[0128] Then, when calculating the current signal estimate in the l-th iteration... The marginal posterior distribution of the b-th element and mean and variance

[0129]

[0130] In the formula, Indicates the current signal estimate. The b-th element; This represents the current signal estimate after removing the influence of the b-th element in the l-th iteration. The cavity probability of the b-th element.

[0131] This step is based on the cavity edge distribution after removing element b. It is first multiplied by the observation likelihood, diffusion likelihood and constellation indicator function in the true posterior, so as to realize element replacement, update the edge distribution in dimension b, and obtain an edge distribution that approximates the true posterior.

[0132] Subsequently, the parameters of the variational prior are updated. and Ensure the adjusted marginal probabilities, i.e., the updated variational prior distribution. Approximate marginal posterior distribution Matching the expected value and variance, and performing moment matching, can be achieved using the following formula:

[0133]

[0134] In the formula, It represents the reciprocal of the variance of the b-th element in the variational prior distribution at iteration l+1; This represents the ratio of the mean to the variance of the b-th element in the variational prior distribution at the (l+1)-th iteration. This represents the mean value of the cavity edge distribution during the l-th iteration; This represents the variance of the cavity edge distribution during the l-th iteration; Let represent the mean of the marginal posterior distribution at the l-th iteration; Let represent the variance of the marginal posterior distribution at the l-th iteration.

[0135] By using the moment matching operation in the above formula, we can force the marginal moments of the approximate distribution to be consistent with the marginal moments of the true posterior, ensuring that it can be as close as possible to the statistical characteristics of the true posterior probability distribution, while maintaining the tractability of the Gaussian distribution.

[0136] It is important to note that after L iterations and according to the MAP criterion, when the target prediction round is reached, the output of the current time step t is... i Corresponding current signal estimate The obtained Gaussian distribution We can approximate the calculation using the mean of the given value, that is:

[0137] Thus, the expected propagation data prediction operation, through continuous iteration of prediction rounds, approximates the high-dimensional discrete posterior as a symbol-level parallel Gaussian distribution. While maintaining constellation constraints, it makes the mean and variance consistent with the true posterior, reducing computational complexity to linear, significantly reducing the symbol error rate, and providing a low-noise and high-precision initial estimate for subsequent back diffusion, thereby shortening the overall detection latency.

[0138] S220. Based on the sampling state of the current time step, the current signal estimate, and the previous signal estimate, perform denoising processing to obtain the denoised signal corresponding to the current time step.

[0139] The previous signal estimate is obtained by performing data prediction processing based on expected propagation in the previous time step. Specifically, the above data prediction model can be used to determine the previous signal estimate corresponding to the previous time step based on the sampling state and the target received signal in the previous time step.

[0140] For example, taking the DPM-Solver++ framework, a high-order numerical solver for efficiently solving diffusion ODEs, as an example, the denoising process is as follows:

[0141] It should be noted that, since the reverse sampling process of the DPM-Solver++ framework is an iterative process, it starts from the initial noise state (i.e., the initial sampling state) corresponding to the current time step t0 = T. (Starting from the current time step t) i Gradually decrease to t I =0, and progressively calculate the sampling state corresponding to each subsequent time step. The update formula for the first time step i=1 can be shown below:

[0142]

[0143] in, This represents the inverse diffusion state at the first time step, i.e., the sampling state at the first time step; The initial noise state (sampling state) can be the initial estimate of the original transmitted signal; t0 represents the initial time step; c is a condition variable. This represents the noise standard deviation at the first time step; The noise standard deviation represents the initial time step; This represents the scaling factor corresponding to the first time step; This represents the conditional data prediction model for the initial time step. h1 represents the change in the logarithmic signal-to-noise ratio (log-SNR) at the first time step, which can be considered as an intermediate variable for simplified calculation.

[0144] Similarly, for subsequent iteration steps i≥2, we can use the second-order multi-step method in DPM-Solver++, utilizing the sampled states from the first two steps. and The corresponding conditional data prediction results are used to calculate, in order to achieve higher precision sampling, as shown in the following formula:

[0145]

[0146] Where c is the condition variable; h i This indicates that in the i-th time step of the reverse diffusion process, the current time step t... i The change in log-SNR is This represents the scaling factor corresponding to the i-th time step. r represents the noise standard deviation at the current time step i. iIt is the ratio of the change in log-SNR at the previous time step to the change in log-SNR at the current time step, r i =h i-1 / h i ;

[0147] Indicates the previous time step t i-1 Data prediction models; Indicates the first two time steps t i-2 Data prediction model.

[0148] Further, please refer to Figure 2c To perform a reverse diffusion sampling operation, the target received signal y, the real-valued channel matrix H, and the noise variance at the target diffusion initiation point must first be input. Scaling factor corresponding to the current time step Pure noise signal x that follows a Gaussian distribution inf ~(0,I) and the set of time steps for the inverse diffusion iteration in, It contains all time points from the initial time step (i=0) to the final target time step (i=1).

[0149] Then through The first data prediction is performed, which involves iteratively matching the mean and variance of the posterior probability distribution to estimate a result closer to the original signal from the noisy signal. It's important to note that, as this is the initialization phase, this step can be understood as using the target received signal y and the pure noise signal x... inf Signal prediction is performed to obtain an initial estimate of the original transmitted signal. The process.

[0150] Subsequently passed Initial estimate The process of diffusion to the selected target diffusion time t0 generates the initial state of inverse sampling, which is the starting point of the target diffusion. It is important to note that this also utilizes a scaling factor representing the current time step after the target diffusion time has elapsed. For the initial estimate Perform a linear combination to allow the generated target to diffuse from its starting point. It retains the statistical characteristics of noise distribution while incorporating effective information from preliminary noise reduction.

[0151] Next, the process enters a loop where the current time step i iterates from the initial time step 1 to the target time step I. This corresponds to the core logic of decreasing the current time step from large to small, with each step optimizing the sampled state based on historical information. Specifically, firstly, through r... i ←h i-1 / h iCalculate the ratio of the change in log-SNR, and this ratio r i It can be regarded as an intermediate variable, representing the ratio of the log-SNR change of the previous time step i-1 to the current time step i. It is used to fuse the denoising information of the previous two steps to improve accuracy, but it does not have a specific physical meaning.

[0152] Then pass Perform denoising prediction at the current time step. Where t i-1 Indicates the previous time step; This indicates the sampling state at the previous time step i-1.

[0153] Furthermore, by integrating the information from the first two steps, through Calculate intermediate quantity D i .in, This indicates the signal estimates for the first two time steps, output by the EP_Data_Predictor from an earlier round.

[0154] Then use Update the sampling state at the current time step. Where, represents The noise standard deviation at the current time step; This represents the noise standard deviation of the previous time step. The ratio of the two is used to scale the sampling state of the previous time step so that its noise intensity matches the current time step. This represents the scaling factor for the current time step, while This represents a coefficient derived from the change in log-SNR at the current time step, controlling the magnitude of the denoising adjustment. Through this step, based on the sampling state of the previous time step, and combined with the intermediate quantity D from the second-order multi-step fusion,... i Calculate the current time step t using parameters such as noise and scaling factor. i Sampling status This further reduces the noise in the denoised signal, making it closer to the original transmitted signal.

[0155] Repeat the above steps. When the iteration reaches the target time step, the result of gradually recovering the original signal from the initial noisy signal to a value close to the original signal of the transmitting end can be obtained. This is the denoised signal corresponding to the target time step, which is the target original signal that restores the original signal of the transmitting end.

[0156] In the above implementation, the inverse diffusion iteration of DPM-Solver++ and the second-order multi-step method are used as the framework, and supplemented by a data prediction model based on expectation propagation. By determining the initial starting point, iterating denoising prediction, and updating the sampling state in the second-order multi-step process, the advantages of DPM-Solver++ of fewer iterations and high precision are utilized, and the accuracy of denoising is improved by expectation propagation. Finally, the goal of efficient and accurate recovery from noisy received signal to original transmitted signal is achieved.

[0157] In some implementations, this signal restoration method can be viewed as proposing an EP-Guided Diffusion (EPGD) detector, such as... Figure 3a As shown, the EPGD detector's workflow consists of two core modules: initialization and reverse sampling. Through the collaboration of the predictor, sampler, and FIFO buffer, it restores the original transmitted signal from the noisy received signal, as detailed below:

[0158] First, the initialization module receives the target received signal y and the pure noise signal x. inf This is used to generate the initial sampling state for inverse diffusion. The predictor is used to perform EP-based data prediction processing, i.e., the EP_Data_Predictor processing mentioned earlier, to obtain the initial estimate. The sampler is used to calculate the initial estimate. Combined with the noise variance of the target diffusion initiation point Scaling factor corresponding to the current time step Parameters such as these are used to generate the target diffusion starting point. As the starting input signal for the reverse sampling module.

[0159] The reverse sampling module is used for iterative denoising to gradually restore the original signal. Each iteration relies on the denoising result of the predictor, the sampler's state update, and the historical information cached by the FIFO buffer. The FIFO stores the signal estimate for the current time step and the signal estimate for the previous time step in the previous iteration. The sampler receives parameters such as the sampling state, intermediate values, noise standard deviation, and scaling factor from the previous time step, performs a sampling state update to obtain the denoised signal for the current time step, and uses this as the sampling state for the next time step, repeating the iterative cycle.

[0160] Furthermore, the performance of this EPGD detector can be evaluated. Taking all experiments as an example, which were conducted on an Nvidia GeForce RTX 4070Ti Super, we first evaluate the performance of the EPGD detector in reducing the symbol error rate (SER) under signal-to-noise ratio (SNR) conditions, and compare it with detectors in related technologies.

[0161] For example, Figure 3b and Figure 3cThe results comparing the symbol error rate performance of different detectors under different signal-to-noise ratios are presented. Specifically, Figure 3b The results of using 64QAM in the MIMO system are shown. Figure 3c The results of the MIMO system using 256QAM are presented. As can be seen from the figures, the proposed EPGD detector has a significant advantage in SER compared to the MMSE detector. Compared to the K-best detector, the proposed EPGD detector... Figure 3c and Figure 3b The SER values ​​were improved by at least 2.8 dB and 6.2 dB, respectively. With similar iteration counts (computational complexity), this EPGD detector showed a significant advantage, particularly in the high SNR region, compared to the EP detector. Furthermore, the SER of this EPGD detector decreased more sharply with increasing SNR. Compared to the ALDMIMO diffuse detector, this EPGD detector not only reduced the number of sampling steps but also achieved at least a 6 dB and 5.6 dB gain in SER performance when the target was met, respectively, in 64QAM and 256QAM configurations.

[0162] Further analysis of the complexity of this EPGD detector is performed, measured by the number of real-valued multiplication (RVM) operations. The total computational complexity of this EPGD detector is... The runtime required to detect a single symbol vector is then given, summarizing all comparison metrics. The results are shown in Table 1 below:

[0163] Table 1. Comparison of the complexity of different detectors

[0164]

[0165]

[0166] Example 1 illustrates a MIMO system employing 64-QAM modulation where K = 32. L EP =9, (L=3, I=2), the average RVMs and runtime (seconds) required to detect one symbol vector. Example 2 shows the average RVMs and runtime (seconds) required to detect one symbol vector in a MIMO system using 256-QAM modulation when K=128, L EP =15, (L=5, I=2), the average number of real-valued multiplications and running time required to detect a symbol vector.

[0167] Based on the performance analysis of SER and complexity above, it can be found that under the same number of iterations, the RVM and runtime of the proposed EPGD detector are similar to those of the EP detector, but the performance gain of the proposed EPGD detector is superior. At 64-QAM, the RVM and runtime of the proposed EPGD detector are reduced compared to ALD. At 256-QAM, although the RVM of the ALD detector is comparable to that of EPGD, the latency is two orders of magnitude higher due to the overhead of double iterations. The K-best detector has low computational complexity, but its serial processing results in high latency; in both cases, the runtime is 1.96 times and 4.45 times that of the proposed EPGD detector, respectively. Furthermore, although the proposed EPGD detector has higher complexity and longer runtime than the MMSE detector, it achieves a significant performance improvement.

[0168] This specification also provides a signal restoration device 400, applied to the receiving end of a large-scale multiple-input multiple-output communication system, such as... Figure 4 As shown, it includes: a signal determination module 410, a starting point determination module 420, and a reverse diffusion module 430, wherein:

[0169] The signal determination module 410 is used to determine the target diffusion time and the target received signal obtained by the receiver.

[0170] The starting point determination module 420 is used to perform data prediction processing based on the desired propagation using the target received signal, the target diffusion time, and the pure noise signal to obtain the target diffusion starting point; wherein, the target diffusion time is determined based on preset noise scheduling parameters.

[0171] The reverse diffusion module 430 is used to perform reverse diffusion sampling based on the target diffusion starting point to obtain the target original signal.

[0172] In some implementations, the starting point determination module 420 is also used to perform signal prediction using the target received signal and the pure noise signal to obtain an initial estimate of the original transmitted signal; and to perform short-time forward diffusion processing on the initial estimate using the target diffusion time to determine the target diffusion starting point.

[0173] In some implementations, the reverse diffusion module 430 is further configured to use the target diffusion origin as T. L The sampling state corresponding to the time step; at the current time step T L In the case of time step, based on T L The sampling state corresponding to the time step is subjected to a reverse diffusion sampling operation to obtain T. L The denoised signal corresponding to the time step; update the current time step to T. L-1 Time step, where T L-1 The time step is T L The next time step of time step; will T LThe denoised signal corresponding to the time step is used as T L-1 The sampling state corresponding to the time step is used to perform the above reverse diffusion sampling operation again to obtain T. L-1 The denoised signal corresponding to the time step; repeat the iterative process of the above reverse diffusion sampling operation, and when the current time step reaches the target time step, take the denoised signal corresponding to the target time step as the target denoised signal to determine the target original signal.

[0174] In some implementations, the reverse diffusion module 430 is further configured to perform a data prediction processing operation based on the target received signal and the sampling state of the current time step to obtain the current signal estimate corresponding to the current time step; and perform denoising processing based on the sampling state of the current time step, the current signal estimate, and the previous signal estimate to obtain the denoised signal corresponding to the current time step; wherein the previous signal estimate is obtained by performing the data prediction processing operation based on the expected propagation in the previous time step.

[0175] In some embodiments, a signal restoration apparatus 400 further includes a data prediction module for constructing a current posterior probability distribution corresponding to the current prediction round based on the sampling state and the target received signal at the current time step; wherein the posterior probability distribution includes a variational prior distribution; when the current prediction round is L1 prediction round, a moment matching update operation is performed on the variational prior distribution of the L1 prediction round to obtain an optimized posterior probability distribution corresponding to the L1 prediction round; the current prediction round is updated to L2 prediction round, and the optimized posterior probability distribution corresponding to the L1 prediction round is used as the variational prior distribution of the L2 prediction round, and the moment matching update operation is performed again to obtain an optimized posterior probability distribution corresponding to the L2 prediction round; wherein the L2 prediction round is the next prediction round of the L1 prediction round; the iterative process of the moment matching update operation is repeated, and when the current prediction round reaches the target prediction round, the optimized posterior probability distribution corresponding to the current prediction round is used as the target posterior probability distribution; the target mean data of the target posterior probability distribution is used as the current signal estimate corresponding to the current time step.

[0176] In some implementations, the signal determination module 410 is further configured to determine the initial received signal obtained by the receiver; wherein the initial received signal is represented by a complex domain channel model; and the complex domain channel model is subjected to real-valued equivalent transformation processing to obtain the target received signal.

[0177] For specific limitations regarding a signal restoration device, please refer to the limitations of a signal restoration method described above, which will not be repeated here. Each module in the aforementioned signal restoration device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module. In this embodiment, the signal restoration device is presented in the form of a functional unit. Here, a unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the aforementioned functions.

[0178] Please see Figure 5 , Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application, such as... Figure 5 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 5Taking a processor 10 as an example, the processor 10 can be a central processing unit, a network processor, or a combination thereof. The processor 10 may further include a hardware chip. This hardware chip can be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device can be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GPRS), or any combination thereof. The memory 20 stores instructions executable by at least one processor 10 to cause at least one processor 10 to perform the methods shown in the above embodiments. The memory 20 may include a stored program area and a stored data area. The stored program area may store the operating system and an application program required for at least one function; the stored data area may store data created based on the use of the computer device, etc. Furthermore, the memory 20 may include high-speed random access memory (RAM) and may also include non-transient memory, such as at least one disk storage device, flash memory device, or other non-transient solid-state storage device. In some optional embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories can be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks (LANs), mobile communication networks, and combinations thereof. The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk, or solid-state drive; the memory 20 may also include combinations of the above types of memory. The computer device also includes an input device 30 and an output device 40. The processor 10, memory 20, input device 30, and output device 40 may be connected via a bus or other means. Figure 5 Taking a bus connection as an example, input device 30 can receive input numerical or character information, and generate key signal inputs related to user settings and function control of the computer device, such as a touch screen, keypad, mouse, trackpad, touchpad, joystick, one or more mouse buttons, trackball, joystick, etc. Output device 40 may include display devices, auxiliary lighting devices (e.g., LEDs), and haptic feedback devices (e.g., vibration motors). The aforementioned display devices include, but are not limited to, liquid crystal displays, light-emitting diodes, displays, and plasma displays. In some optional embodiments, the display device may be a touch screen.

[0179] This application also provides a computer-readable storage medium. The methods described in this application can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code downloaded over a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and subsequently stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the methods shown in the above embodiments are implemented.

[0180] This application provides a computer program product including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the method of any embodiment of this application.

[0181] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and all such modifications and variations fall within the scope defined by the appended claims. It is understood that in the specific embodiments of this application, data related to user information, location information, navigation data, etc., are involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices. For ease of description, the above devices are described separately by function as various units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0182] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. The various embodiments in this specification are described in a progressive manner, and similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for system embodiments, since they are substantially similar to method embodiments, the description is relatively simple, and relevant parts can be referred to the description of the method embodiments.

[0183] The above description is merely an embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of this application should be included within the scope of the claims of this application. Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A signal restoration method characterized by, The method is applied to a receiving end of a large-scale multiple-input multiple-output communication system, and comprises the following steps: determining a target diffusion time and a target received signal obtained by the receiving end; performing a data prediction processing operation based on expectation propagation by using the target received signal, the target diffusion time and a pure noise signal to obtain a target diffusion starting point; wherein the target diffusion time is determined based on a preset noise scheduling parameter; performing a reverse diffusion sampling operation based on the target diffusion starting point to obtain a target original signal.

2. The method of claim 1, wherein, The data prediction processing operation based on expectation propagation by using the target received signal, the target diffusion time and a pure noise signal to obtain a target diffusion starting point comprises the following steps: performing signal prediction by using the target received signal and the pure noise signal to obtain an initial estimated value of an original transmitted signal; performing a short-time forward diffusion processing operation on the initial estimated value by using the target diffusion time to determine a target diffusion starting point.

3. The method of claim 1, wherein, The reverse diffusion sampling operation based on the target diffusion starting point to obtain a target original signal comprises the following steps: The target diffusion starting point is taken as T L The sampling state corresponding to the time step; At the current time step, the T L At the T L At the T L At the T corresponding to the current time step, the T corresponding to the current time step, the T corresponding to the current time step, the T corresponding to the current time step, the T updating the current time step to T L-1 time step; wherein the T L-1 time step is the T L next time step of the T T L The denoised signal corresponding to the time step is used as T L-1 The sampling state corresponding to the time step is used to perform the above reverse diffusion sampling operation again to obtain the T. L-1 The denoised signal corresponding to the time step; repeating an iteration process of the reverse diffusion sampling operation, and in a case where a target time step is reached at the current time step, taking a de-noised signal corresponding to the target time step as a target de-noised signal to determine the target original signal.

4. The method of claim 3, wherein, The reverse diffusion sampling operation comprises the following steps: for a current time step, performing a data prediction processing operation based on expectation propagation based on the target received signal and a sampling state of the current time step to obtain a current signal estimated value corresponding to the current time step; performing de-noising processing based on the sampling state of the current time step, the current signal estimated value and a previous signal estimated value to obtain a de-noised signal corresponding to the current time step; wherein the previous signal estimated value is obtained by performing a data prediction processing operation based on expectation propagation in a previous time step.

5. The method of claim 4, wherein, The data prediction processing operation based on expectation propagation comprises the following steps: constructing a current posterior probability distribution corresponding to a current prediction round based on a sampling state of the current time step and the target received signal; wherein the posterior probability distribution contains a variational prior distribution; in a case where the current prediction round is an L1 prediction round, performing a moment matching update operation on the variational prior distribution of the L1 prediction round to obtain an optimized posterior probability distribution corresponding to the L1 prediction round; updating the current prediction round to an L2 prediction round, taking the optimized posterior probability distribution corresponding to the L1 prediction round as a variational prior distribution of the L2 prediction round, and performing the moment matching update operation again to obtain an optimized posterior probability distribution corresponding to the L2 prediction round; wherein the L2 prediction round is a next prediction round of the L1 prediction round; repeating an iteration process of the moment matching update operation, and in a case where a target prediction round is reached at the current prediction round, taking an optimized posterior probability distribution corresponding to the current prediction round as a target posterior probability distribution; taking a target mean value data of the target posterior probability distribution as the current signal estimated value corresponding to the current time step.

6. The method of claim 1, wherein, The target received signal is determined by the following steps: determining an initial received signal obtained by the receiving end; wherein the initial received signal is represented by a complex domain channel model; performing real value domain equivalent change processing on the complex domain channel model to obtain the target received signal.

7. A signal restoration apparatus characterized by comprising: The application discloses a receiving end applied to a large-scale multiple-input multiple-output communication system. A signal determining module is configured to determine a target diffusion time and a target received signal obtained by the receiving end. A starting point determining module is configured to perform data prediction processing operation based on expectation propagation by using the target received signal, the target diffusion time and a pure noise signal to obtain a target diffusion starting point. An inverse diffusion module is configured to perform inverse diffusion sampling operation based on the target diffusion starting point to obtain a target original signal.

8. A computer device, comprising: The application further discloses a computer readable storage medium and a computer program. The computer readable storage medium stores computer instructions, and the computer instructions are used to make a computer execute the method in any one of claims 1 to 6.

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

10. A computer program product comprising a computer program, characterized in that, ​

Images