A MIMO Channel Estimation Method Based on Pilot Subspace Refinement Flow
By refining the flow in the pilot subspace and using the complementary projection of the pilot subspace and null space with a generative flow matching model, the problem of inaccurate channel recovery under finite pilot conditions is solved, achieving efficient channel estimation. This method is applicable to baseband processors and general-purpose processors in communication systems.
Patent Information
- Application Number
- CN202611378690.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-07
- Publication Date
- 2026-11-17
AI Technical Summary
Under finite pilot conditions, existing MIMO channel estimation techniques struggle to effectively combine real-time pilot information with channel priors, leading to inaccurate channel recovery. This is especially problematic in high-dimensional channels where observational information is insufficient and noise has a significant impact. Existing methods also have limitations in representing complex channel structures and adapting to training configurations.
A MIMO channel estimation method based on pilot subspace refined flow is adopted. By constructing complementary projection operators of pilot subspace and pilot null space, and combining generative flow matching model and distributed sensing prior, inverse reasoning and local optimization are performed to achieve local coupling between physical observation and learned prior, thereby reducing the error in the invisible direction of pilot.
It improves channel recovery capability under finite pilot conditions, enhances the utilization of complex channel structures, reduces online computation overhead, and maintains measurement consistency in the generative inference process. It is suitable for implementation in communication baseband processors and general-purpose processors.
Smart Images

Figure CN123123079A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of wireless communication technology, specifically relating to a MIMO channel estimation method based on pilot subspace refinement flow. Background Technology
[0002] MIMO technology, by configuring multiple antennas at both the transmitting and receiving ends to create spatial degrees of freedom, is one of the key technologies for improving the spectral efficiency, coverage, and reliability of wireless systems. In actual baseband processing, functions such as coherent detection, beamforming, precoding, power allocation, user scheduling, and link adaptation all rely on accurate Channel State Information (CSI). Therefore, rapidly obtaining high-quality CSI under limited time-frequency and computational resources is a fundamental problem in large-scale MIMO and future high-dimensional wireless systems.
[0003] As the number of transmit and receive antennas increases, the channel matrix dimension expands accordingly, while the coherence time and bandwidth of the wireless channel do not increase proportionally. If strictly orthogonal pilot sequences are desired across all transmit spatial dimensions, the training overhead typically needs to increase with the transmit dimension, thus consuming more coherence block resources and compressing the effective data transmission time. In scenarios with short coherence times, such as high-speed mobile, short-packet communication, or high-frequency communication, the conflict between pilot overhead and data transmission efficiency becomes even more pronounced. Therefore, practical systems often need to complete channel estimation with a pilot number smaller than the transmit dimension.
[0004] Meanwhile, the development of millimeter waves, terahertz waves, and very large-scale arrays has further increased the difficulty of channel training. On the one hand, high-frequency propagation may experience significant path loss, blocking loss, and beam mismatch, reducing the effective signal-to-noise ratio at the receiver for channel training. On the other hand, large-scale arrays introduce stronger spatial correlation, angle resolution, array structure dependence, and mixed near-field and far-field propagation phenomena. Even though the embodiments of this invention employ a narrowband point-to-point MIMO model, the above trends still illustrate that as the dimensionality and propagation complexity of wireless systems increase, simply increasing the number of pilots to improve estimation accuracy often leads to unacceptable resource overhead.
[0005] The statistical structure of real-world wireless channels cannot always be accurately described by a simple independent and identically distributed Gaussian model. Multipath clustering, angular leakage, spatial correlation, line-of-sight / non-line-of-sight (LOS / NLOS) switching, array aperture, scatterer distribution, and scene variations can all alter the joint distribution of the channel. In such cases, relying solely on limited prior information such as the mean, covariance, or a fixed sparse dictionary may not be sufficient to fully represent the complex correlations present in high-dimensional channels. Especially when pilots are few, the information provided by observations decreases, making the estimation results more sensitive to the quality of the prior model.
[0006] Consider having One transmitting antenna and A narrowband point-to-point MIMO system with one receiving antenna transmits within a coherent block. One pilot symbol. When In practice, pilot observation operators are generally not one-to-one mappings; received pilots can only directly constrain a portion of the transmission spatial directions, leaving the remaining directions unidentifiable under pilot observation. Therefore, low-pilot channel estimation is not merely about improving fitting accuracy under noisy conditions, but rather requires simultaneously utilizing physical pilot observations and channel priors to address the underdetermined recovery problem.
[0007] From the perspective of the linear inverse problem, the above-mentioned finite pilot case involves two interrelated difficulties: first, noise introduces errors into the finite observations themselves; second, when the pilot observation operator is not full-dimensional, channel degrees of freedom that cannot be directly identified from the current pilot may appear. The latter means that even if the received noise approaches zero, multiple candidate channels may produce the same pilot response. Therefore, finite pilot channel estimation cannot be simply understood as "improving the fitting accuracy under low signal-to-noise ratio," but requires simultaneously solving the problems of observation consistency and unobservable degrees of freedom recovery.
[0008] Traditional channel estimation methods primarily include Least Squares (LS) and Maximum Likelihood (ML) methods. These methods mainly rely on real-time pilot observations, do not require explicit channel priors, and have clear and easy-to-implement models. When there are sufficient pilots and the observation matrix is well-conditional, LS / ML can obtain stable estimates; however, when the number of pilots is significantly smaller than the transmit dimension, the observation equation tends to be underdetermined, and LS usually only obtains solutions that satisfy a certain minimum norm condition, rather than determining the complete channel from the pilots themselves. At low signal-to-noise ratios, pseudo-inverse operations may also amplify the effects of noise.
[0009] Another classic approach improves estimation by introducing channel statistical priors, including Linear Minimum Mean Square Error (LMMSE), Minimum Mean Square Error (MMSE), and Maximum A Posteriori (MAP) estimation. LMMSE primarily uses first- and second-order statistics such as the channel mean and covariance, and its online computation structure is relatively well-defined. However, when the actual channel distribution has a significantly non-Gaussian or multimodal structure, the second-order statistical information is insufficient to fully represent the prior. MMSE and MAP can theoretically use a more complete prior distribution, but in complex high-dimensional channels, accurately obtaining the prior probability density, calculating the posterior integral, or solving high-dimensional non-convex optimization problems is often quite difficult.
[0010] Furthermore, methods such as compressed sensing, sparse representation, LASSO, atomic norm, and AMP / D-AMP often utilize the structural sparse characteristics of millimeter-wave or massive MIMO channels in the angular, time-delay, or beam domains to recover the channel with fewer pilots. These methods exhibit good model interpretability when finite paths and significant sparse structure are present, but their performance depends on the chosen dictionary, grid, and sparsity assumptions. When the actual channel exhibits angular leakage, grid mismatch, rich scattering, or array structure variations, the pre-defined sparse model may deviate from the actual channel.
[0011] In recent years, deep learning has been used for MIMO channel estimation. Traditional data-driven methods typically use structures such as convolutional neural networks, fully connected networks, attention networks, or Transformers to directly learn the nonlinear mapping from received pilots, coarse estimates, or other features to the channel matrix. Model-driven methods, on the other hand, expand iterative thresholding, AMP, denoising, or optimization steps into trainable networks. Their advantage lies in their ability to learn nonlinear structures that are difficult to express using traditional analytical models using historical data, and to achieve faster online inference during training configuration matching. However, pure end-to-end mapping often incorporates the pilot matrix, noise level, and channel distribution into the network parameters, which may lead to distribution shifts when faced with changes in pilot configuration, antenna dimension, signal-to-noise ratio, or propagation environment; some methods also lack explicit guarantees for the consistency of physical pilots.
[0012] Building upon this foundation, generative AI methods such as diffusion / score-based generative models and flow matching have begun to be used to learn the high-dimensional probabilistic structure of wireless channels. Unlike discriminative methods, which primarily learn deterministic input-output mappings, generative methods emphasize learning the channel prior distribution, score function, or continuous transmission field from a large number of noise-free or high-quality CSI samples, and then using actual pilot observations to conditionally constrain the generation process in the online phase. Their potential advantage lies in the fact that propagation statistics can be learned offline, while measurement information such as pilot matrices and noise levels can be explicitly injected during online inference, thus forming a posterior recovery framework of "learning priors + communication physical observations."
[0013] For diffusion / scoring models, a common approach is to guide samples towards directions that satisfy the observations during the stepwise denoising process using pilot likelihood, gradient, or projection operations. For streaming matching methods, the continuous transmission direction from noisy states to clean channel states can be learned, and a local clean channel representation can be formed during the reverse generation process. Compared to complex random diffusion trajectories, streaming methods can employ deterministic or fewer steps of continuous transmission and can be combined with closed-loop linear observation updates, thus possessing good engineering implementation potential.
[0014] However, the ability of generative priors to represent complex channel structures does not automatically solve the finite pilot posterior problem. Online estimation still needs to answer two key questions: first, whether the local channel representation provided by the learned prior in the current intermediate state is accurate enough, especially for channel directions where pilots cannot be directly identified; and second, whether the generated trajectory can still maintain consistency with actual pilot observations after pilot consistency updates, time-level transformations, or re-embedding. Furthermore, using residual correction that uniformly processes all channel directions may cause unnecessary mutual interference between physical observations and learned priors in pilot-invisible directions.
[0015] Therefore, existing finite-pilot MIMO channel estimation techniques generally face the following common contradictions: pilot observations only contain finite-dimensional real-time information, while complete high-dimensional channel recovery requires full utilization of complex propagation priors; traditional analytical priors have limited expressive power, and the physical consistency and cross-configuration adaptability of pure data-driven mappings are insufficient. Although generative methods provide stronger prior modeling capabilities, how to stably and with low complexity couple the prior information and pilot information according to the channel direction that can be identified by actual observations is still a technical problem that needs to be further solved.
[0016] The disadvantages of existing technology are specifically described as follows:
[0017] 1. The main drawback of the LS / ML method based on no explicit channel prior is that the information available to it comes entirely from real-time pilot observations. When the number of pilots is less than the effective degrees of freedom of the channel to be estimated, the observation equation may become an underdetermined problem. Multiple candidate channels can produce the same or similar pilot responses, so relying solely on pilot residuals cannot uniquely determine the complete channel. Under low signal-to-noise ratio conditions, pseudo-inverses or residual fitting may further amplify the noise effect.
[0018] 2. The main drawback of LMMSE / MMSE / MAP methods based on statistical priors lies in the expression and solution of complex channel priors. LMMSE mainly relies on first- and second-order statistics such as mean and covariance. When the real channel has a non-Gaussian, multimodal, spatially strongly correlated, or significantly scene-dependent structure, the second-order statistical information is difficult to fully describe its distribution. Although MMSE and MAP can utilize more complete priors, the probability distribution of high-dimensional channels is often difficult to obtain accurately and explicitly. The posterior integral required for MMSE and general MAP optimization may also have high computational complexity.
[0019] 3. The main drawback of structure-prior methods based on sparse or compressed sensing is their sensitivity to pre-defined propagation structures. These methods typically require the assumption that the channel has sufficient sparsity in a specific domain and rely on the design of dictionaries, lattices, or atomic sets. When the actual channel has angular leakage, lattice mismatch, rich scattering, spatial correlation enhancement, or array structure changes, the actual channel may deviate from the predetermined structural model, thereby reducing the recovery quality under finite pilot conditions.
[0020] 4. A major drawback of end-to-end methods based on discriminative deep learning is that network mapping is easily coupled with the training configuration. If the mapping from received pilots to the complete channel is learned directly, the network may simultaneously absorb the pilot matrix, pilot density, noise range, antenna size, and channel data distribution from the training phase. When the actual system configuration changes, the joint input-output distribution changes accordingly, which may require retraining, fine-tuning, or the addition of additional conditional coding. Furthermore, pure end-to-end prediction does not inherently guarantee that the output strictly satisfies the actual pilot observations.
[0021] 5. While model-driven deep learning or deep unrolling methods retain some physical models, their main drawback lies in the fact that learnable modules and iterative structures are usually tied to specific forward models, noise assumptions, and training distributions. When the pilot matrix, array size, or propagation statistics change significantly, the original learnable parameters may not remain applicable. Furthermore, multiple layers or iterations increase network calls and online iterations, leading to additional computational and storage overhead in high-dimensional MIMO scenarios.
[0022] 6. The main shortcoming of generative AI prior-based methods has shifted from "lack of complex priors" to "how to stably couple learned priors with finite pilot observations posteriorly." For existing streaming schemes, if only a single local clean channel center represents the prior near the current state, the error of this center in the pilot-unidentifiable direction cannot be directly corrected by the pilots. Closed or near-end pilot consistency updates may generate pilot domain residuals again after re-embedding into the next time layer. If a unified residual correction is further applied to the entire channel space, the anisotropy of pilot observations acting only in certain directions may be ignored. Adding an independent measurement correction network would increase the overhead of training data, model parameters, and online network calls. Therefore, generative methods still need to address the issues of prior center quality, pilot consistency in trajectory evolution, and information division between observable and unobservable directions.
[0023] 7. Existing MIMO channel estimation methods based on sparse geometric channel models recover the channel by estimating parameters such as the angle and gain of the propagation path. These methods have a clear physical model foundation, but they are typically dependent on the channel sparsity, array structure, and the accuracy of path parameter estimation. Their recovery capability may be affected by complex scattering, non-ideal sparsity, or deviations between the propagation model and the actual environment. Furthermore, these methods emphasize using a pre-defined structural model to describe the channel, making it difficult to fully utilize the complex high-dimensional statistical features contained in historical CSI data, and they lack specialized processing mechanisms for the differences in the identifiability of different channel directions under finite pilot conditions.
[0024] In summary, although the aforementioned technical approaches employ different prior forms and solution mechanisms, they all need to address the same fundamental contradiction in high-dimensional MIMO scenarios with finite pilots: real-time pilots can only provide finite-dimensional observation information that is affected by noise, while complete channel recovery relies on prior inferences about insufficient observation degrees of freedom. Existing methods are either limited by prior representation capabilities, training configuration adaptability, or have not yet fully resolved the stable coordination problem between learned priors and physical pilots in different identifiable directions. This common problem constitutes the technical background that subsequent inventions aim to further address. Summary of the Invention
[0025] To address the aforementioned problems in the prior art, the present invention aims to provide a MIMO channel estimation method applicable to finite pilot conditions, which improves the recovery capability of high-dimensional channels without relying on increasing the number of orthogonal pilots, and enables real-time pilot information and propagation priors learned from historical channel data to complement each other.
[0026] The main objective of this invention is to solve the following technical problems:
[0027] 1. To address the problem of insufficient observation information and the difficulty in uniquely determining the complete channel from the received pilot under finite pilot conditions, and to improve the ability to recover channel degrees of freedom that cannot be directly identified by pilots.
[0028] 2. Improve the ability to utilize the statistical structures of complex, non-Gaussian, and difficult-to-parameterize MIMO channels, and reduce the potential lack of prior representation when only second-order statistics or fixed sparse models are used.
[0029] 3. Improve the synergy between the learned channel prior and the actual pilot observation, so that the channel estimation formed in the generative inference process is continuously constrained by the real communication observation, and reduce the measurement domain bias caused by intermediate state evolution.
[0030] 4. Reduce mutual interference between physical pilot information and learned priors in unidentifiable channel directions, so that different information sources can play their role according to the observation range that the actual pilots can provide.
[0031] 5. While achieving the above estimation objectives, control the overhead of online computation and model invocation, so that the relevant processing can be completed by communication baseband processors, general-purpose processors or dedicated acceleration hardware with feasible matrix operations and finite number of neural network forward computations.
[0032] The technical solution of this invention is as follows:
[0033] A MIMO channel estimation method based on pilot subspace refinement flow, for applications with... One transmitting antenna and A narrowband block fading point-to-point MIMO system with multiple receiving antennas, wherein the refined flow is a flow matching generation model that incorporates a distributed-aware prior refinement (DPR) mechanism; the method includes:
[0034] Acquire receive pilot block Given pilot matrix And the pre-trained refined flow model, the physical observation model is , The MIMO channel matrix to be estimated is... This is the noise matrix;
[0035] According to the pilot matrix Construct complementary projection operators to decompose the channel matrix into a first channel subspace that can be directly identified by the current pilot and a second channel subspace that cannot be directly identified by the current pilot.
[0036] In the reverse reasoning process of the refined flow model, for each reverse time step:
[0037] The current interpolation state is input into the pre-trained refined flow model, and the DPR noiseless channel prior center is obtained by extrapolating the mean of the output transmission direction. ;
[0038] Constructing a pilot data consistency term and a priori center The local quadratic optimization problem of the prior neighbor term is solved to obtain a closed-form update solution, realizing the local coupling of physical observation and learning prior.
[0039] The closed-form update solution is re-embedded to obtain the near-end interpolation state of the next time layer, and a pilot domain bridging target matching the time layer is constructed.
[0040] A posteriori bridge-guided PBG correction is performed on the near-end interpolation state according to the bridging target, wherein the correction increment is located in the observable direction of the first channel subspace;
[0041] Extract the first channel subspace component from the corrected state, extract the second channel subspace component from the prior interpolation state, and combine them in a directional manner to form the interpolation state of the next time layer.
[0042] Repeat the above reverse time step until scheduling ends, and output the final MIMO channel estimation matrix. .
[0043] Furthermore, the method for obtaining the pre-trained refined flow model is as follows:
[0044] Acquiring noiseless MIMO channel samples and the above Same-dimensional complex Gaussian noise samples Construct a linear interpolation path from the clean channel to the noise anchor point:
[0045] ,
[0046] in For continuous interpolation time, For time The interpolation state, For displacement target;
[0047] Train the flow matching model so that it simultaneously outputs the mean of the transmission direction. and positive scale And from these two, construct the conditional distribution of the displacement target:
[0048] ;
[0049] Selecting on the same interpolation path that satisfies High time layer and lower time layer ,make , will the high time layer state Input model to obtain and Construct a conditional Gaussian distribution for the low-time state:
[0050] ;
[0051] With real low-time state The negative log-likelihood under this conditional distribution constructs a distribution that senses the prior refinement loss:
[0052] ,
[0053] By minimizing Update the model parameters to obtain the pre-trained refined flow model;
[0054] Among them, positive scale To construct the distribution-aware loss, the mean value of the transmission direction is used instead of this scale for random sampling during the online inference stage. A priori center is formed.
[0055] Furthermore, according to the pilot matrix The specific steps for constructing complementary projection operators are:
[0056] Computation of pilot subspace projection operator and pilot null projection operator ,in for Moore-Penrose pseudoreverse, for An identity matrix of order 1;
[0057] The first channel subspace is composed of The projection yields the second channel subspace, which is composed of... Obtained by projection.
[0058] Furthermore, the DPR noiseless channel prior center is obtained by extrapolating the mean of the output transmission direction. The specific method is:
[0059] Based on the current interpolation state of the reverse time layer and current time variable As input, the above Input the refined flow model to obtain the mean value of the transmission direction output by the model. The DPR noiseless channel prior center is obtained by extrapolation using the following formula:
[0060] ,
[0061] in, This represents the noiseless channel estimate obtained by extrapolating the current interpolation state along the transmission direction.
[0062] Furthermore, the local quadratic optimization problem is as follows:
[0063] ,
[0064] Its closed-form update solution is:
[0065] ,
[0066] in, These are the pilot / prior balance weights for the current time layer; in practice, they are calculated using Cholesky decomposition or by solving a system of linear equations, without the need to explicitly form the matrix inverse.
[0067] Furthermore, the trajectory re-embedding includes:
[0068] Using the same fixed noise anchor point Construct the next time layer respectively Proximal interpolation state and prior interpolation state ;
[0069] The pilot domain bridging target is:
[0070] ,
[0071] in, For a bounded scaling factor, .
[0072] Furthermore, the post-bridge guided PBG correction includes performing... Next iteration:
[0073] ,
[0074] ,
[0075] in, , For time-related step size, And the correction amount Right multiplication The value is then zero, thus not modifying the second channel subspace component.
[0076] Furthermore, the directional combination includes:
[0077] ,
[0078] ,
[0079] in, The combination coefficient is used to control the degree to which the second channel subspace component regresses to the prior trajectory.
[0080] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method of the present invention.
[0081] The beneficial effects of this invention are as follows:
[0082] 1. Improve the complete channel recovery capability under finite pilot conditions. Compared with estimation methods that rely solely on pilot residual fitting, this scheme can supplement the information that pilots cannot directly provide by utilizing learned propagation priors when real-time observations are insufficient. Therefore, it is more suitable for underdetermined channel estimation scenarios where the number of pilots is less than the transmission dimension.
[0083] 2. Enhance the ability to express and utilize complex propagation priors. Compared with schemes that mainly rely on fixed covariance, simple probability distributions, or pre-set sparse dictionaries, this scheme learns high-dimensional channel structures from CSI samples through generative models, and can utilize complex statistical features such as multipath, spatial correlation, array correlation, and LOS / NLOS without explicitly providing the complete analytical probability density.
[0084] 3. Improve the rationality of the fusion of physical pilot information and learned priors. The scheme organizes two types of information sources according to the range that can be identified by actual pilot observations, so that physical observations mainly act on the channel directions that they can directly constrain, and the pilot invisible directions are mainly recovered by learned priors, thereby reducing the mutual coverage or miscorrection of information sources caused by unified full-space correction.
[0085] 4. Improve measurement consistency during generative inference. The scheme continuously considers the actual pilot response during trajectory evolution, which can suppress pilot domain deviations caused by finite time steps, state reconstruction, or trajectory re-embedding, ensuring that the intermediate state remains consistent with real communication observations as it evolves towards a clean channel.
[0086] 5. While maintaining strong prior modeling capabilities, online computational overhead is controlled. Observation-side processing mainly consists of closed-loop linear algebra, projection, and lightweight residual updates, without requiring additional training of independent measurement correction networks. For fixed or repeatedly used pilot matrices, relevant pseudo-inverses, projections, Gram matrices, and scale quantities can be pre-computed or cached, which helps reduce repetitive inference overhead.
[0087] 6. It has good implementation and expansion flexibility. The generative network backbone, the number of time steps, the linear equation solver, and the deployment hardware can all be adjusted without changing the core information division of labor. Therefore, different implementation methods can be selected according to the computing power and latency constraints of base stations, terminals, centralized baseband units, or edge nodes. Attached Figure Description
[0088] Figure 1 This is a schematic diagram of the MIMO pilot channel estimation system model to which this invention applies;
[0089] Figure 2 This is the main flowchart of the MIMO channel estimation method of the present invention;
[0090] Figure 3 This is a schematic diagram of the overall structure of the pilot subspace refinement flow within a single reverse time step;
[0091] Figure 4 This is a schematic diagram of the decomposition of the pilot subspace and pilot null space defined by the pilot matrix;
[0092] Figure 5This is a flowchart of the offline training process of the flow matching model used in this invention;
[0093] Figure 6 This is a schematic diagram of the process of combining post-bridge guidance with pilot subspace / pilot null space orientation. Detailed Implementation
[0094] The present invention will be further described in detail below with reference to the accompanying drawings.
[0095] The core technical concept of this invention is: based on a known pilot matrix Construct two complementary transmit space projections. For those that can influence the predicted pilot block... The components are received using pilot signals. Perform physical consistency updates; for those... Components that result in zero after multiplication and cannot be directly identified by the pilots are primarily retained based on the channel priors learned by the pre-trained flow matching model. Around this subspace division of labor, in each inverse time step, the following are executed sequentially: Distribution-Aware Prior Refinement (DPR), Closed Pilot Consistency Secondary Update, Trajectory Re-embedding, Posterior Bridge Guidance (PBG), and Pilot Subspace / Pilot Nullspace Orientation Combination.
[0096] like Figure 1 As shown, the present invention considers having One transmitting antenna and A narrowband block fading point-to-point MIMO system with one receiving antenna, specifically having The transmitting end of each transmitting antenna transmits a full-rank pilot matrix. The signal passes through a narrowband block fading MIMO channel. Noise is superimposed at the receiving end during transmission. ,have The receiving ends of each receiving antenna form a receiving pilot block. The channel matrix remains unchanged within a coherent block, while the pilot phase occupies [a certain amount of space / time]. One channel is used. The pilot matrix is known. It is a full column rank matrix, which satisfies the following in the low pilot embodiment: The receiver obtains the received pilot block Y, and the physical observation model is as follows:
[0097] (1)
[0098] In equation (1), To receive pilot blocks; Here is the MIMO channel matrix to be estimated; Given the pilot matrix; It is an additive noise matrix; Represents the pilot observation operator. This formula corresponds to Figure 1 "Transmitter-Full Rank Pilot Matrix" —MIMO channel —The signal chain of the receiving end, and satisfies the markings in the figure. .
[0099] Physically, Equation (1) defines MIMO channel estimation as an observation problem of "known excitation - unknown linear channel - random noise". The pilot matrix determines which directions in the transmission space are actually excited, and HP represents the noiseless reception response that the candidate channel should produce under the excitation of the pilot. The receiver can only infer H from the relationship between Y and HP. When the number of pilot columns decreases, the real-time information contained in Y will not cover all degrees of freedom of H, so additional priors need to be introduced.
[0100] because Observation operator Generally, it's not a one-to-one mapping. To clearly distinguish between transmit spatial directions directly constrained by the receive pilot and those that cannot be directly constrained, based on... Construct complementary projection operators:
[0101] (2)
[0102] In equation (2), For pilot subspace projection operators; For pilot null space projection operator; for Moore-Penrose pseudo-reverse; for An identity matrix of order 1. Both are... The launch space projection.
[0103] From a geometric perspective, equation (2) is not simply for calculating the projection matrix, but rather for characterizing "which channel directions are visible to the current pilot". The pilot subspace projection preserves the direction that can change the predicted pilot response, while the pilot null space projection preserves the direction that is orthogonal to the current pilot and does not change the direction of the predicted pilot response. The complementary relationship between the two determines the information recovery range that subsequent physical observations and learned priors should each undertake.
[0104] For any candidate channel matrix, right multiplication is performed respectively. and Extract the observable components and null components of the pilot signals. For a real channel... ,have:
[0105] (3)
[0106] In equation (3), Represents the pilot subspace components. This represents the zero-space component of the pilot frequency. Because... Null space components do not change the received pilot prediction; while This indicates that pilot observations depend only directly on observable components. Figure 4 This complementary relationship is represented by two parallel regions: the left side is the pilot subspace, using... When projected, its components can affect Therefore, it can be received by pilot signals. Direct constraints; the right side is the pilot null space, using... When projected, its components satisfy multiplication by The remainder is zero, therefore it is supplemented by the prior knowledge of the learned channel. Subsequent PBG and subspace combinations are all based on this geometric decomposition.
[0107] The key physical meaning of equation (3) is that the channel can be decomposed into observable components and null space components, but the received pilot only directly depends on the former. It can be seen that even if the observable components are accurately recovered under finite pilots, the complete channel may still have a large error due to the unknown null space components; on the contrary, any improvement to the null space components must come from propagation priors or other additional information, and cannot be claimed to come from the current pilot itself.
[0108] This invention acquires noiseless MIMO channel samples in an offline phase. and generate with Same-dimensional complex Gaussian noise samples , specifically Figure 5 Input noiseless MIMO channel samples Same-dimensional complex Gaussian noise samples Construct a linear interpolation path based on formula (4) and select from the same path The mean of the flow matching model output and positive scale Using formula (6), a conditional Gaussian distribution of the low-time-layer state is constructed, and then the model parameters are updated according to the DPR conditional negative log-likelihood of formula (7) to obtain a pre-trained flow matching model for use in the online stage. In order to use a real-valued neural network, the real and imaginary parts of the complex channel matrix can be stacked along the channel dimension to form a real-valued representation. Flow matching training adopts a linear interpolation path from the clean channel to the noise anchor point:
[0109] (4)
[0110] In equation (4), For continuous interpolation time; For time The interpolation state; The target displacement is along the interpolation path. Corresponding to the clean channel endpoint, Corresponding noise anchor points. Figure 5 "Noise-free MIMO channel sample" on the left "and same-dimensional complex Gaussian noise samples" "Together, we enter the "Constructing a Linear Interpolation Path" module, which corresponds to formula (4).
[0111] Equation (4) defines a continuous coordinate system from the real channel to the noise anchor point in the generative model. The time variable t can be understood as the degree to which the current state is close to the clean channel endpoint: the closer t is to 0, the closer the state is to the actual channel; the closer t is to 1, the closer the state is to random noise. Online reverse reasoning is precisely about gradually reducing the random component and restoring the structure that conforms to the channel distribution in the opposite direction.
[0112] Unlike predicting only deterministic velocities, DPR enables the flow matching model to output the mean of the transmission direction simultaneously during the training phase. and positive scale To construct the conditional distribution of the displacement target:
[0113] (5)
[0114] In equation (5), For parameters The corresponding conditional displacement distribution; Output the average value in the transmission direction; The output should be positive-scaled. Positive scales can be generated using softplus, exponential mapping, or other monotonic mappings with positive lower bounds. The model backbone only needs to receive the current state matrix and temporal embedding and output the mean and scale that match the displacement target dimension; there is no requirement to limit the specific number or width of U-Net layers.
[0115] Equation (5) enables the network to provide not only a deterministic transmission direction, but also a local scale description of the vicinity of that direction. The mean characterizes the main trend of the current state moving towards the clean channel, while the positive scale reflects the dispersion of the training samples during finite-time migration. This scale serves distributed-aware training and is not equivalent to the true posterior covariance after pilot conditionalization.
[0116] To enable the model to learn local uncertainties over finite time steps, a higher time layer is selected on the same interpolation path. and lower time layer ,satisfy and order .Will and Input the flow matching model to be trained to obtain and Then the conditional distribution of the lower time-level states can be written as:
[0117] (6)
[0118] Equation (6) represents the finite-step transmission from the higher time layer to the lower time layer as a conditional Gaussian distribution, where For conditional mean, For conditional covariance. Based on the actual low-time state. The negative log-likelihood under this conditional distribution is used to define the DPR training loss:
[0119] The physical function of Equation (6) is to transform "how the current high-time-layer state should evolve to the lower-time-layer state" into a learnable conditional transfer distribution, rather than simply requiring the network to predict the velocity at a single instant. This allows the network to utilize the difference in the actual state between two time layers on the same path to constrain the network to learn a more stable local channel evolution pattern with a finite step size.
[0120] (7)
[0121] In equation (7), To refine the prior knowledge loss for distribution perception; The dimension of the real-valued representation of the complex channel after stacking the real and imaginary parts; This represents the Frobenius norm. During training, the model parameters are updated by minimizing this loss. Figure 5 "flow matching model output" and positive scale "and "DPR conditional negative log-likelihood update model parameters" correspond to formulas (6) and (7) respectively.
[0122] Equation (7) is used to perform maximum likelihood training on the aforementioned conditional transfer distribution. The error term makes the conditional mean closer to the true low-time-layer state, while the scale-related term prevents the network from escaping mean error by infinitely increasing the prediction variance. Its engineering purpose is not to increase online random sampling, but to improve the clean channel prior center used online through more sufficient offline distribution learning.
[0123] Special note: Positive scale The distributed sensing loss is used to construct the loss during the training phase, and random sampling based on this scale is not required during the online channel estimation phase. Online deterministic estimation primarily uses the mean of the transmission direction. A clean channel prior center is formed, so the online overhead of DPR can be shared with one forward computation of the original flow model without the need for additional network calls.
[0124] The main process of the online phase and Figure 2 To maintain consistency, execute steps S1 through S7: S1 retrieves... , S1 Model and reverse scheduling and initialization; S2 Predict transmission direction and form DPR prior refined estimate; S3 Solve pilot-prior quadratic problem to obtain S4 Construct the near-end / prior state and pilot domain bridging target; S5 Perform K a posteriori bridging corrections; S6 Directional combine the pilot subspace and pilot null space and update the state; If the reverse time step scheduling has not ended, return to the next reverse time step, and after the end, S7 outputs the MIMO channel estimation matrix. Figure 3 Furthermore, the state transition relationship within a single reverse time step is given: current interpolation state Entering the flow matching model yields , DPR according to Forming a noise-free prior center ; With the receiving pilot block Pilot matrix Together they entered the closed-loop secondary update to obtain Subsequently, the corrected state is obtained through trajectory re-embedding and PBG. Then, through subspace directional combination, we obtain And proceed to the next reverse time step. Figure 6 The process of combining PBG with subspace orientation is specifically given: the upper branch interpolates the state from the proximal end. Starting point, using pilot frequency domain bridging target To obtain the result, perform PBG K times. And extract its pilot subspace components; the lower branch starts from the prior interpolation state. Starting from this point, its pilot null space component is extracted. The two are then combined in a subspace module to form... And updated according to the combination coefficient ρ The details are as follows.
[0125] S1: Obtain the receive pilot block Pilot matrix Pre-trained flow matching model and inverse time step scheduling Calculate the correlation projection between the pilot subspace and the pilot null space according to formula (2); in the pilot matrix... If the result remains unchanged, it can be pre-calculated before the reverse reasoning begins. Two subspace projections as well as At least one of the following: Generate complex Gaussian noise anchors of the same dimension as the channel to be estimated. In a complete reverse reasoning process, the noise anchor point is fixed, and with Initialize the current interpolation state .
[0126] S2: For the current reverse time and the next time The current interpolation state and time Input the pre-trained flow matching model to obtain the mean value of the transmission direction. and positive scale Online estimation uses The DPR noiseless channel prior center is formed according to formula (8).
[0127] (8)
[0128] In equation (8), This is the DPR prior refined noiseless channel estimation matrix. This matrix is the center of the subsequent closed pilot-prior secondary update. Since the pilot null space is not directly constrained by observations, improving this prior center can directly improve the prior components inherited from the null space in the secondary update.
[0129] Equation (8) extrapolates the current noisy intermediate state along the learned average transmission direction to the clean time end, which can be regarded as the "prior suggestion value" given by the flow model near the current state. Since the pilot null space lacks direct observation, the accuracy of this suggestion value in the null space will directly determine the unobservable channel information that subsequent closed-loop updates can inherit.
[0130] S3: To receive pilot blocks With candidate channels The pilot fitting error between them is used as a physical observation term, with candidate channels With DPR Priority Center Using the proximity error between them as a priori term, we construct a quadratic optimization problem:
[0131] (9)
[0132] In equation (9), These are the pilot / prior balance weights for the current time layer. This problem is a convex quadratic problem, and its closed-form update is:
[0133] From a posterior inference perspective, equation (9) retains two sources simultaneously: the first requires the candidate channel to be able to explain the actual received pilot, reflecting data consistency; the second requires the candidate channel not to deviate excessively from the generative prior center, reflecting propagation prior constraints. Weights It determines whether the current time layer is more inclined to real-time measurement or learning priors, and its function is similar to the intensity of local regularization.
[0134] (10)
[0135] In equation (10), for The conjugate transpose of . Due to , Since it is a positive definite matrix, the closed-form solution can be obtained directly. In practical engineering implementations, to improve numerical stability, the same closed-form update can be achieved by Cholesky decomposition or solving the linear equation system, without the need to explicitly form the matrix inverse. Figure 3 "Closed-loop secondary update" "That is, the corresponding formulas (9) and (10).
[0136] Equation (10) provides the analytical solution to the aforementioned local posterior proxy problem, enabling the fusion of physical observations and learned priors without requiring the training of a separate network. Its matrix structure includes... The received pilot information is fed back to the transmit channel dimension. This indicates the contribution of the prior center, and These factors collectively determine the weighted sum and numerical stability in different launch directions.
[0137] From the perspective of the pilot subspace, the pilot term in equation (9) depends only on the pilot observable component of X, while its null space component is determined by the neighboring term. Therefore, the closed update does not use non-existent pilot information to determine the null space out of thin air, but rather makes this part inherit the learning prior center provided by DPR.
[0138] S4: Closed Update Located at the clean time end, while the next flow model calculation needs to be located at the next time layer. The interpolation state. To maintain consistency with the training interpolation path, the same fixed noise sample is used. Construct the proximal interpolation state and the prior interpolation state respectively:
[0139] (11)
[0140] In equation (11), Used for subsequent pilot residual correction. Used to store the next time-level state controlled by DPR priors. Since both use the same noise anchor point. This can avoid unnecessary state differences introduced by different random anchor points.
[0141] The physical meaning of Equation (11) is to put the clean end estimates from two different sources back into the next time layer coordinates consistent with the training of the flow model. The near-end state carries the result after pilot coupling, while the prior state retains the generated trajectory corresponding to the DPR center. Both use the same noise anchor point, which allows subsequent comparisons and subspace combinations to reflect only the difference between "pilot correction and prior preservation" without additional random anchor point differences.
[0142] Even after trajectory re-embedding, It exhibits good pilot consistency at the clean end. The pilot response still includes the noise anchor point. Therefore, a pilot-domain bridging target is constructed based on the next time layer:
[0143] (12)
[0144] In equation (12), For pilot domain bridging target; It is a bounded scaling factor; The bridging target will receive pilot signals along the next time layer. Interpolate with the pilot response at the anchor point to avoid... The noiseless pilot signal is directly forced to be satisfied in the noisy intermediate state.
[0145] Equation (12) does not require that the intermediate noisy state be directly equal to the noiseless received pilot, but rather constructs a pilot domain reference trajectory that matches the current time layer. As t′ moves from the noisy end to the clean end, the bridging target gradually transitions from the pilot response on the anchor side to the actual received pilot, so that the measurement constraints are consistent with the time layer in which the generated state is located.
[0146] S5: Future bridge inspection guidance Perform a light pilot residual correction. Let ,right Execute in sequence:
[0147] (13)
[0148] In equation (13), For the first Pilot residuals of the sub-PBG; The Lipschitz scale is determined by the pilot matrix; K represents the time-dependent step size; K is a positive integer. This update is first calculated in the receive pilot domain. and The residual, multiplied on the right. It maps back to the channel space, thus eliminating the need for additional training of the measurement correction network. Figure 6 Upper branch "Proximal interpolation state → (K)th PBG correction → extraction" The pilot subspace component corresponds to this step.
[0149] Equation (13) can be understood as performing one or more normalized gradient corrections on the pilot prediction error of the current intermediate state. The residual is first formed in the received pilot domain, and then multiplied by... The pilot is redirected back into the channel space, so the correction direction is determined by the actual pilot operator, rather than being freely generated by an additional network. Used to control based on gradient scale, This determines the correction strength for the current time layer.
[0150] Depend on achievable Therefore, the correction amount of equation (13) Right multiplication The remainder is zero. In other words, the increment of PBG is naturally located in the observable direction supported by the pilot and does not directly modify the component obtained by the pilot null space projection. This property makes PBG similar to... Figure 4 The physical meaning of the subspace shown is consistent.
[0151] S6: After completing K PBG cycles, from Extract pilot observable components from... Extract the pilot null space components and perform directional combination and state update:
[0152] (14)
[0153] In equation (14), This represents the state of subspace composition. These are the combination coefficients; This is the updated interpolation state for the next time layer. Because... use As an observable part, used As a null space component, the combination process retains the pilot support component obtained by PBG and introduces the null space information of the DPR prior trajectory into the next state. Figure 6 The lower branch's "prior interpolation state → extraction of pilot null space components of prior state" and the upper branch's "formation" and according to The "Update" module is combined, which corresponds to formula (14).
[0154] Equation (14) achieves directional recombination of information sources, rather than simply averaging two complete channels. The observable portion is taken from the state after PBG correction to retain information consistent with the actual pilot; the pilot null space portion is taken from the prior trajectory to avoid forcibly using non-existent pilot information for unidentifiable directions. The combination coefficients mainly adjust the degree of null space prior injection, thereby controlling the prior regression strength while maintaining the pilot correction results.
[0155] Furthermore, Pilot observable projection and The pilot observable projections are the same, therefore the final result is based on... Mixing does not change the observable components already obtained by PBG; The main control is over the degree to which the null component regresses to the DPR prior trajectory. This achieves a division of labor in directional information, where the pilot subspace is corrected using physical observations and the pilot null space is supplemented using learned priors.
[0156] S7: Will As the current state of the next time layer, repeat steps S2 to S6 according to the reverse time step scheduling T until the scheduling ends, and output the final MIMO channel estimation matrix. If T contains only one reverse time step, it forms a one-step inference approach; if T contains multiple reverse time steps, it forms a multi-step inference approach. Each reverse time step only requires calling the stream matching network once.
[0157] The core protection of this invention does not depend on a fixed set of hyperparameters. To illustrate feasibility, preferred settings verified in simulations can be used: the DPR finite migration time layer can be selected. ,in PBG is acceptable. , , Time-dependent step size is acceptable. ,in For pilot density, Let V be the pilot noise variance. The weights for closed-loop updates can be taken as... ,in .
[0158] In terms of online complexity, DPR shares a single forward computation with the flow matching network, without adding extra neural network calls. Each PBG mainly includes one... Multiplication and first degree The complexity of the main matrix operations for K-fold multiplication and correction varies with... Growth. If a coherent block or a set of estimation tasks... If it remains unchanged, then and It can be pre-calculated and reused.
[0159] This method can be validated using a dataset with known truth channels. The pilot matrix, pilot density, signal-to-noise ratio, test samples, and normalization method should be fixed across different methods, and the normalized mean square error should be calculated. It can also simultaneously statistically analyze the normalized pilot measurement residuals. The model parameters, training time, inference time, peak storage, inverse time steps, number of network calls, and K value are all considered.
[0160] Simulation verification uses 3GPP CDL-A / B / C / D channels. Furthermore, in the master setting with SNR ranging from −10 dB to 30 dB, the SRF achieved NMSE improvement relative to the compared RC-Flow at 8 out of 9 SNR points, with an average gain of approximately 0.29 dB; at 0 dB, it was approximately 0.07 dB lower. The additional gain of the PBG was most significant in the noise-limited regions of −10 dB and −5 dB; at medium to high SNR, the DPR-only version was close to the full SRF. These results are intended to demonstrate the feasibility of the above structure under the corresponding simulation settings and should not be extrapolated to deterministic performance in all propagation environments, real-world air interfaces, or hardware deployments.
[0161] In a reproducible experimental embodiment, the main settings adopt pilot density The SNR values are -10 dB, -5 dB, 0 dB, 5 dB, 10 dB, 15 dB, 20 dB, 25 dB, and 30 dB. Pilot matrix elements use independent QPSK symbols, and each CDL test scenario contains a fixed number of independent channel implementations. To ensure fair comparison between different algorithms, all methods should use the same test channels, pilot matrices, noise implementation methods, and normalization rules.
[0162] Comparable baselines include: LMMSE, representing second-order statistical prior methods; LASSO and fsAD, representing sparse / structured prior methods; LDAMP, representing model-driven deep unrolling / learning-based denoising methods; DMCE, representing diffusion-generative channel estimation; and RC-Flow, representing existing streaming-generative channel estimation. An ablation version can also be configured to enable only DPR and disable PBG to differentiate the impact of prior refinement and posterior bridge correction on different SNR ranges.
[0163] In addition to the main dimension, it can also be changed , And scalability verification with pilot density, such as in larger antenna dimensions or The variation of NMSE with SNR was observed at a lower pilot ratio. The purpose of this verification is not to limit the invention to a specific array dimension, but to confirm that the proposed information division of "observable directions are corrected by physical pilots and unobservable directions are supplemented by learned priors" is still computable when the pilot null space dimension is expanded.
[0164] During project acceptance, in addition to NMSE, it is recommended to simultaneously measure the pilot domain normalized residual, the average inference time of a single channel estimation, peak GPU memory or on-chip memory usage, the number of network forward calls, the number of PBG iterations, and the computation time saved by the amount of pre-computed matrices under fixed pilots. For actual deployment, latency and storage overhead should also be remeasured on the target processing platform; GPU simulation runtime alone cannot replace baseband real-time performance evaluation.
[0165] This method can be implemented by a processor executing a computer program, or it can be mapped to a functional module in a baseband processing device. An exemplary device includes: a data acquisition module for acquiring... , Pre-trained flow matching model and inverse time step scheduling; projection construction module, used to construct and The module initializes fixed noise anchor points; a priori refinement module is used to generate DPR noise-free prior centers; a closed update module is used to solve formula (9) and calculate formula (10); a bridging correction module is used to construct formulas (11) and (12) and execute formula (13); a subspace combination module is used to execute formula (14); and a channel output module is used to output channel estimates after the reverse time scheduling ends.
[0166] The aforementioned modules can be implemented using a general-purpose CPU, digital signal processor (DSP), graphics processing unit (GPU), field-programmable gate array (FPGA), application-specific integrated circuit (ASIC), or a combination thereof. The memory can be used to store pre-trained model parameters, pre-computed quantities related to the pilot matrix, inverse time step scheduling, and intermediate state matrices. The device can be deployed in base stations, user terminals, centralized baseband processing units, or other communication nodes with pilot reception and baseband computing capabilities.
[0167] The specific embodiments of the present invention have been described above. Those skilled in the art will understand that the accompanying drawings are merely illustrative of preferred embodiments, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention. Obviously, those skilled in the art can make various modifications and variations to the present invention without departing from its spirit and scope. If these modifications and variations fall within the scope of the claims of the present invention and their equivalents, the present invention also intends to include these modifications and variations, for example:
[0168] 1. The computation methods for the pilot observable subspace and pilot null space can be replaced, except for explicit computation. and In addition, it can also be achieved through singular value decomposition (SVD), QR decomposition, orthogonal basis construction, null basis, eigenvalue decomposition, or other methods that can obtain the pilot operator's domain / null equivalent basis. Specifically, the implementation can either form a complete projection matrix or be calculated directly in the subspace basis coordinates.
[0169] 2. The numerical solution method of pilot-prior local coupling can be replaced. The closed-form result of the quadratic problem can be obtained by explicit matrix inversion or by Cholesky decomposition, QR decomposition, conjugate gradient, preconditioning iteration or other linear equation solvers; as long as it still contains the pilot data consistency term and the proximity constraint represented by the generative prior, its basic function is not changed.
[0170] 3. Local prior surrogates are not limited to isotropic Gaussian forms. DPR can employ diagonal covariance Gaussians, low-rank covariance Gaussians, Gaussian mixtures, heavy-tailed distributions, mixed density networks, or other parameterized distributions capable of describing the uncertainty of finite-time transitions. In the online phase, deterministic prior centers can be formed using only their mean / mode, or finite sampling or multi-candidate aggregation can be performed when the computational budget allows.
[0171] 4. The parameterization method for the positive scale or other distribution parameters of DPR can be replaced. The positive scale can be generated by softplus, exponential mapping, square mapping with a positive lower bound, or other functions that guarantee effective scale; the complex channel can be processed by stacking real / imaginary channels, amplitude / phase representation, real block matrix representation, or complex-valued neural network processing.
[0172] 5. The generative network backbone can be replaced. U-Net, ResNet, convolutional networks, Transformer, DiT, state-space networks, graph neural networks, complex-valued networks, or other models that can receive the current channel state and time conditions and output the required transmission parameters can be used; the number of network layers, width, attention form, and normalization method are not core limitations.
[0173] 6. The specific continuous model of generative priors can be extended. While maintaining the premise of "learning channel priors + explicit pilot posterior constraints + subspace orientation processing", the flow matching backbone can be replaced with an equivalent continuous normalized flow, probabilistic flow ODE, diffusion / score generation model, or other generative models that can provide local clean channel representation or local prior direction, and the time variables and state update rules can be adjusted accordingly.
[0174] 7. Trajectory re-embedding and pilot domain bridging targets can employ other time interpolation methods. In addition to linear clean-to-noise interpolation, nonlinear time scheduling, variance-preserving / variance-explosive noise paths, or reparameterized continuous trajectories can be used depending on the training path; bridging targets can also employ time-dependent, SNR-dependent, pilot density-dependent, or noise variance-dependent weights, as long as the target matches the current generation time layer.
[0175] 8. PBG is not limited to a single fixed-step gradient update. Multiple Landweber iterations, projected gradients, proximal gradients, normalized residual updates, conjugate direction updates, or other lightweight iterations based on pilot residuals can be used; the correction amount can be explicitly multiplied. Back-projection can also be calculated directly on the pilot observable subspace basis, as long as its main range of action remains in the pilot support direction.
[0176] 9. Subspace-oriented combination methods can be replaced. In addition to directly splicing "the observable component of the corrected state + the null component of the prior state" and performing convex combination, time-related soft weights, SNR-related weights, layer-by-layer gating, subspace-by-subspace energy normalization, or other fusion methods that do not disrupt the division of observable / unobservable information can also be used.
[0177] 10. The number of reverse time steps and scheduling can be interchanged. Single-step, few-step, or multi-step inference can be used; time points can be uniformly or non-uniformly discrete, adaptively selected based on error, or determined by computational budget. (Under fixed pilot conditions) Subspace basis , The equivalent of the Lipschitz scale can be pre-calculated and reused before inference.
[0178] 11. The observation model can be extended to broadband OFDM, multi-subcarrier, multi-slot, frequency-selective, or other linear MIMO pilot models. When extending, the following should be considered: Replace it with the corresponding linear pilot observation operator, and redefine the observable and unobservable directions according to the range and null space of the operator; the core is still that physical observation mainly corrects the identifiable part, and learning prior mainly supplements the observation null space part.
[0179] 12. The device deployment and application organization methods can be replaced. This method can be deployed on base stations, user terminals, centralized / distributed baseband processing units, edge computing nodes, or cloud-side training-edge-side inference architectures. Computation can be performed by CPUs, DSPs, GPUs, FPGAs, ASICs, or heterogeneous hardware. The DPR offline training method and the online SRF channel estimation method can be disclosed as a complete technology chain in the same application, or they can be organized separately as the subject matter of claims such as training methods, estimation methods, devices, electronic devices, and computer-readable storage media based on unity of invention and search results.
Claims
1. A MIMO channel estimation method based on pilot subspace refinement flow, used for channels with... One transmitting antenna and A narrowband block fading point-to-point MIMO system with one receiving antenna, characterized in that, The refined flow is a flow matching generation model that introduces a distribution-aware prior refinement (DPR) mechanism; the method includes: Acquire receive pilot block Given pilot matrix And the pre-trained refined flow model, the physical observation model is , The MIMO channel matrix to be estimated is... This is the noise matrix; According to the pilot matrix Construct complementary projection operators to decompose the channel matrix into a first channel subspace that can be directly identified by the current pilot and a second channel subspace that cannot be directly identified by the current pilot. In the reverse reasoning process of the refined flow model, for each reverse time step: The current interpolation state is input into the pre-trained refined flow model, and the DPR noiseless channel prior center is obtained by extrapolating the mean of the output transmission direction. ; Constructing a pilot data consistency term and a priori center The local quadratic optimization problem of the prior neighbor term is solved to obtain a closed-form update solution, realizing the local coupling of physical observation and learning prior. The closed-form update solution is re-embedded to obtain the near-end interpolation state of the next time layer, and a pilot domain bridging target matching the time layer is constructed. A posteriori bridge-guided PBG correction is performed on the near-end interpolation state according to the bridging target, wherein the correction increment is located in the observable direction of the first channel subspace; Extract the first channel subspace component from the corrected state, extract the second channel subspace component from the prior interpolation state, and combine them in a directional manner to form the interpolation state of the next time layer. Repeat the above reverse time step until scheduling ends, and output the final MIMO channel estimation matrix. .
2. The MIMO channel estimation method based on pilot subspace refinement flow according to claim 1, characterized in that, The method for obtaining the pre-trained refined flow model is as follows: Acquiring noiseless MIMO channel samples and the above Same-dimensional complex Gaussian noise samples Construct a linear interpolation path from the clean channel to the noise anchor point: , in For continuous interpolation time, For time The interpolation state, For displacement target; Train the flow matching model so that it simultaneously outputs the mean of the transmission direction. and positive scale And from these two, construct the conditional distribution of the displacement target: ; Selecting on the same interpolation path that satisfies High time layer and lower time layer ,make , will the high time layer state Input model to obtain and Construct a conditional Gaussian distribution for the low-time state: ; With real low-time state The negative log-likelihood under this conditional distribution constructs a distribution that senses the prior refinement loss: , By minimizing Update the model parameters to obtain the pre-trained refined flow model; Among them, positive scale To construct the distribution-aware loss, the mean value of the transmission direction is used instead of this scale for random sampling during the online inference stage. A priori center is formed.
3. The MIMO channel estimation method based on pilot subspace refinement flow according to claim 2, characterized in that, According to the pilot matrix The specific steps for constructing complementary projection operators are: Computation of pilot subspace projection operator and pilot null projection operator ,in for Moore-Penrose pseudoreverse, for An identity matrix of order 1; The first channel subspace is composed of The projection yields the second channel subspace, which is composed of... Obtained by projection.
4. The MIMO channel estimation method based on pilot subspace refinement flow according to claim 3, characterized in that, The DPR noiseless channel prior center is obtained by extrapolating the mean of the output transmission direction. The specific method is: Based on the current interpolation state of the reverse time layer and current time variable As input, the above Input the refined flow model to obtain the mean value of the transmission direction output by the model. The DPR noiseless channel prior center is obtained by extrapolation using the following formula: , in, This represents the noiseless channel estimate obtained by extrapolating the current interpolation state along the transmission direction.
5. The MIMO channel estimation method based on pilot subspace refinement flow according to claim 4, characterized in that, The local quadratic optimization problem is: , Its closed-form update solution is: , in, These are the pilot / prior balance weights for the current time layer; in practice, they are calculated using Cholesky decomposition or by solving a system of linear equations, without the need to explicitly form the matrix inverse.
6. The MIMO channel estimation method based on pilot subspace refinement flow according to claim 5, characterized in that, The trajectory re-embedding includes: Using the same fixed noise anchor point Construct the next time layer respectively Proximal interpolation state and prior interpolation state ; The pilot domain bridging target is: , in, For a bounded scaling factor, .
7. The MIMO channel estimation method based on pilot subspace refinement flow according to claim 6, characterized in that, The post-bridge guided PBG correction includes performing Next iteration: , , in, , For time-related step size, And the correction amount Right multiplication The value is then zero, thus not modifying the second channel subspace component.
8. The MIMO channel estimation method based on pilot subspace refinement flow according to claim 7, characterized in that, The directional combination includes: , , in, The combination coefficient is used to control the degree to which the second channel subspace component regresses to the prior trajectory.
9. A computer-readable storage medium, characterized in that, It stores a computer program thereon, which, when executed by a processor, implements the method of any one of claims 1 to 8.
