A channel estimation method based on channel knowledge graph

Abstract

The present invention belongs to the field of information and communication technology, and relates to a channel estimation method based on a channel knowledge graph. We establish a channel graph (CKM) construction framework based on the historical received signals of the base station. The physical area is divided into spatial grids, and a set of common path power, delay and angle parameters are extracted from each grid according to the spatial consistency principle. We transform the CKM construction problem into a Bayesian inference problem, and develop a hybrid message passing algorithm to construct a CKM based on interference cancellation. Furthermore, we use the output of the CKM to design a channel estimator based on the minimum mean square error and interference suppression (MMSE-IRC) criterion. Numerical results show that the proposed CKM provides accurate channel parameters at relatively low signal-to-noise ratio (SINR), and the CKM-assisted channel estimator outperforms the existing state-of-the-art channel estimator.

Description

Technical Field

[0001] The present invention belongs to the field of information and communication technology, and in particular relates to a channel estimation method based on a channel knowledge graph. Background Art

[0002] The sixth generation of wireless communication networks (6G), expected to be available around 2030, is envisioned to provide ubiquitous connectivity (up to nodes per square kilometer), ultra-high throughput (peak rate of terabits per second), ultra-reliable low-latency communication (HRLLC), and high positioning accuracy (centimeter-level). To achieve these key performance indicators, academia and industry are conducting research on various promising technologies, including but not limited to ultra-massive MIMO (XL-MIMO), millimeter-wave communications, terahertz communications, and integrated interawareness communication (ISAC) technology.

[0003] With the application of XL-MIMO, millimeter wave, and terahertz technologies, access node density, array size, and communication bandwidth have increased significantly. This has led to a sharp increase in wireless channel response in both spatial and frequency domain dimensions. This higher-dimensional channel response poses greater challenges to physical layer communication modules, including the high computational burden of acquiring channel state information (CSI) and the high training overhead of beamforming. On the other hand, denser nodes (such as IoT sensors) provide a large amount of data. Combined with positioning algorithms, this data is location-based and reflects the characteristics of the local wireless propagation environment. This opportunity has shifted the communication paradigm from non-context-aware to context-aware communication.

[0004] Recently, the channel knowledge map (CKM) has been proposed as a digital twin of wireless channels. It can be understood as a priori database of channel model parameters at a specific user location, reflecting the inherent channel characteristics in the local propagation environment. The channel parameters provided by CKM can be designed according to specific requirements, including statistical information (such as LOS / NLOS existence) and large-scale parameters (such as path delay and path angle). In addition, since CKM acts as a mapping from user location to channel parameters, its implementation latency is significantly lower than existing ray tracing technology, thus enabling real-time physical layer signal processing.

[0005] Existing CKM construction methods include "Toward environment-aware 6G communications via channel knowledge map," "Channel knowledge map for environment-aware communications: EM athcalgorithm for map construction," and "Nerf2: Neural radio-frequency radiance fields." These methods utilize channel data at certain locations to interpolate or predict channel parameters at other locations, neglecting the process of extracting channel parameters from the collected data. Based on offline channel measurements, this assumption holds true by using high-precision channel detectors for channel measurement. However, channel measurements require relatively intensive training data and must be repeated as the propagation environment changes. To avoid the high measurement costs, an alternative strategy is to store and reuse historical received signals at the base station (such as pilot and synchronization signals) to extract channel parameters. However, the received signal is a nonlinear function of the channel parameters (e.g., the array response of the path angle as a complex exponential vector function) and often includes interference from neighboring cells. This means that the parameter extraction process significantly affects the accuracy of the CKM and cannot be ignored. Therefore, how to simultaneously extract model parameters and construct the CKM based on historical base station data has become an urgent problem and challenge. Summary of the Invention

[0006] To address the above issues, the present invention is aimed at uplink multiple-input multiple-output orthogonal frequency division multiplexing (MIMO-OFDM) systems in the presence of inter-cell interference, and focuses on how to effectively extract channel features and construct CKM using historical base station received signals. The present invention builds a CKM framework to unify the extraction and characterization of channel parameters. To ensure the accuracy of CKM, a channel knowledge graph structure of the interference cancellation mechanism is established in the Bayesian reasoning framework, and a high-performance, low-complexity channel estimator is designed using the output of CKM.

[0007] The technical solution of the present invention is:

[0008] A channel estimation method based on channel knowledge graph for MIMO-OFDM system, where the base station has M=M x ×M y antennas, with an antenna spacing of half a wavelength, and a base station communicating with a single target user via N subcarriers in the tth time slot; the method comprising:

[0009] S1. Establish the user receiving signal model as follows:

[0010]

[0011] in, and They represent the diagonal signal matrices of the user and the k-th jammer, respectively, and the vectors composed of their diagonal elements are x t and N t is an additive white Gaussian noise matrix whose elements independently follow Gaussian distribution User channel H t Interference Channel Respectively expressed as:

[0012]

[0013] Where L represents the number of multipaths of the target user, represents the number of multipath interference k, α t，l represents the lth complex random coefficient with unit energy, ρ t，l represents the path loss of the lth path, a i (ω),ω∈{τ t，l ,θ t，l ,φ t，l} represents the steering vector, and the expression is where i 2 =-1; and They represent the equivalent path delay, the directional component on the x-axis of the three-dimensional coordinate system, and the directional component on the y-axis, respectively. and represents the path delay, azimuth arrival angle, and elevation angle of the lth path in the tth time slot, and They represent the complex coefficient, path loss, equivalent delay, and directional components on the x-axis and y-axis of the interference k-th path respectively;

[0014] S2, receive signal matrix Y t Vectorize y t =vec(Y t ), construct an equivalent receiving signal model:

[0015]

[0016] in s t (τ t，l )=x t ⊙a N (τ t，l ),

[0017] S3. Define the mapping from time slot t to position q for:

[0018]

[0019] in, and denote the set of time slots and positions respectively;

[0020] S4. Set the channel parameter ρ at any position q. t，l , τ t，l ,θ t，l and φ t，l The changes within 2T consecutive time slots are negligible, where T is the update period of the channel knowledge graph; Obtained by the base station in each time slot; interference pilot It is known by the base station at the beginning of each channel knowledge graph update cycle; the channel knowledge graph is represented as a mapping:

[0021]

[0022] in, and They represent the number of equivalent channel paths, equivalent time delay, horizontal component, vertical component and path loss corresponding to the qth position of the spectrum output respectively;

[0023] S5. Channel knowledge graph mapping using S4 Re-characterize the channel matrix of S2 as:

[0024]

[0025] in is the equivalent complex coefficient of the lth article;

[0026] S6. Construct channel-based knowledge graph mapping The received signal model:

[0027]

[0028] in, Represents the equivalent noise formed by combining the channel characterization error and additive white Gaussian noise;

[0029] S7. Convert the channel knowledge graph construction problem into a Bayesian inference problem: define the vector ξ q Contains elements and Contains elements Given an observation ξ q and The posterior distribution of is decomposed into:.

[0030]

[0031] Among them, the user channel parameter prior probability model is The superscript "pri" is the abbreviation of "prior". represents the von Mises (VM) distribution; the prior probability model of the interference channel is and in and Probability density function representation of the joint Bernoulli-Gaussian model The probability of being a zero vector is 1-λ, and the model is used to describe the possible severe path loss and / or low transmit power of the interferer k;

[0032] S8. Construct a message passing factor graph: Use a factor graph to represent the posterior probability decomposition in S7. The factor nodes in the factor graph represent probability distributions, the variable nodes represent random variables, and the abbreviation of the probability density function is introduced, that is, and Abbreviated as and Use υ to uniformly represent variable nodes, and use f to uniformly represent factor nodes. represents the set of variables associated with factor node f; Representing a collection Element v is not included; represents the set of factor nodes associated with the variable node υ; Representing a collection Factor node g is not included; (or ) is represented as a message from a variable node v to a factor node f (or from a factor node f to a variable node v); b f (u f ) represents the factor node f with respect to the vector u f belief;

[0033] S9, message passing initialization: select a non-informative prior p(v), where Set and Make VM pdfp(v) tend to uniform pdf in [0, 2π), for μ in v , Random values are in the range [0, 2π), κ v Set to 105 ;use Least squares initialization of and The prior variances of and The sparsity λ is initialized to 1. After all parameters are updated once, λ is set to 0.5. The λ value in subsequent iterations is given by EM learning in S18. In addition, the initialization as well as

[0034] S10. Update slave factor nodes To variable node News

[0035]

[0036] where R t =Reshape(y t , [M, N]), and for have I x (·) represents the Bessel function of the xth order and the first kind;

[0037] S11. Approximate news In multiple time slots The product term on

[0038]

[0039] in

[0040]

[0041] in Representative function The inverse function of represent The local maximum point of For a single variable function, a heuristic search can be used to obtain the local maximum point

[0042] S12, Calculate Factor Node Beliefs in as follows

[0043]

[0044] in and and arg(·) represent the magnitude and angle operations respectively.

[0045] S13. Calculate from factor nodes To variable node News

[0046]

[0047] Then follow step S12 to get and The received signal model is symmetrical, so it is similar to For the message calculation of interference channel parameters, when Time τ t，k，l ,θ t，k，l and φ t，k，l Respectively and Symmetric, the corresponding message is updated by simplifying The message is obtained, that is, the multiplication of messages in different time slots in S11 is deleted.

[0048] S14, Calculate Factor Node About Belief as follows

[0049] in based on For variables Obtain, and

[0050]

[0051] S15. Calculate Factor Node To variable node β t，l and News

[0052]

[0053] in and

[0054] S16, Calculate Factor Node Beliefs in as follows

[0055]

[0056] in

[0057]

[0058] belief The corresponding mean and variance for

[0059]

[0060] S17. Calculate from factor nodes To variable node News in

[0061]

[0062] S18, EM learning updates prior parameters as follows

[0063]

[0064] S19, message passing reaches the maximum number of iterations and terminates, output As the channel knowledge graph information, Obtained by S13, and Obtained by S14, Obtained from S18.

[0065] S20, using the channel knowledge graph output to enable channel estimation, and obtaining the equivalent received signal model based on S1

[0066]

[0067] where r t =vec(X -1 Y t ),h t =vec(H t ), With w t =vec(X -1 N t ). vec(·) represents a matrix vectorized operation.

[0068] S21. Calculate the user channel frequency-spatial covariance matrix

[0069]

[0070] in and parameter Given by the constructed CKM, namely S19.

[0071] S22. Calculate the power delay spectrum of the user channel

[0072]

[0073] where ||·||2 represents the vector 2-norm.

[0074] S23. Estimation of interference spatial covariance matrix: Define P thres is the power threshold, and defines Represents The corresponding delay tap index set. Then estimate the sample covariance Right now

[0075]

[0076] S24, using MMSE-IRC channel estimation

[0077]

[0078] Based on the Khatri-Rao product property, The equivalent expression is in and and ⊙ represent Kronecker product and Hadamard product respectively.

[0079] S25, output As the channel h t Estimates.

[0080] The beneficial effects of the present invention are as follows: a channel knowledge graph (CKM) construction framework based on the historical received signals of the base station is established, and these signals are used as cost-free measurement data. By dividing the physical area into spatial grids, a set of common path powers, delays and angles are extracted from each grid according to the principle of spatial consistency. The difficulty lies in that the received signal is a nonlinear function of the channel parameters, and the bottleneck of construction accuracy is the potential interference source. To solve this problem, the CKM construction problem is transformed into a Bayesian inference problem. By characterizing the path loss / power difference of the interference as a block sparse prior, a hybrid message passing algorithm is developed to construct a CKM based on interference cancellation. In addition, a channel estimator based on the minimum mean square error and interference suppression (MMSE-IRC) criterion is designed using the output of the CKM. In particular, by utilizing the channel covariance structure provided by the CKM, the computational complexity of the MMSE-IRC is significantly reduced. Numerical results show that the proposed CKM provides accurate channel parameters at relatively low signal-to-noise ratio (SINR), and the CKM-assisted channel estimator outperforms the existing state-of-the-art channel estimator. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 This is a schematic diagram of a MIMO-OFDM system scenario under uplink inter-cell interference;

[0082] Figure 2 This is a schematic diagram of CKM construction;

[0083] Figure 3 is the message passing factor graph;

[0084] Figure 4 The graph accuracy of different algorithms changes with the number of reconstructed paths

[0085] Figure 5 This is the curve of the normalized mean square error (NMSE) of channel estimation using different algorithms changing with the signal-to-interference-and-noise ratio (SINR). DETAILED DESCRIPTION

[0086] The present invention is described in detail below with reference to the accompanying drawings.

[0087] The following parameter configuration is used: the base station is equipped with a 4×8 uniform rectangular array (URA) and serves one user on N=192 subcarriers. The subcarrier spacing is Δ f =30kHz, and the carrier frequency is 6.5GHz. The channel response is generated according to the UM NLOS scenario in the standard TR 38.901, where the number of channel clusters is set to 10 and each cluster has 20 subpaths. In addition, the channel parameters (including path delay and angle) drift at different positions. The heights of the user and the interference source are 1.5m respectively, and their initial positions are randomly generated, and the distance is randomly selected in the range of [100, 500]m. The user and the interference source move in random directions. The number of interference sources is 4 and the sparsity is 0.5. Pilots with random values of normal mode and phase in the range of [0, 2π] are used. The received signals collected along each straight line of the user movement are randomly divided into two sets with a ratio of 7:3, where the former is used for CKM construction and the latter is used for channel estimation evaluation. In CKM construction, the maximum number of iterations of the proposed algorithm is set to 40 and the number of interference path extraction is 15. In channel estimation, the proposed method estimates the interference covariance matrix, P thres The value is 10dB lower than the interference plus noise power.

[0088] According to the above parameter settings, the specific steps of the simulation are as follows:

[0089] S1. Communication system model and configuration: Consider the following Figure 1The uplink MIMO-OFDM system shown in Figure 1 is a MIMO-OFDM system. The base station (BS) is equipped with a uniform rectangular array (URA) consisting of M = 4 × 8 antennas. For simplicity, the antenna spacing is assumed to be half a wavelength. In existing cellular networks, base stations can serve hundreds of users through time division multiplexing (TDD) and frequency division multiplexing (FDD). Since the signal model of each user is the same, we focus on a single target user that communicates with the base station through N = 192 subcarriers in the tth time slot. Assume that the target user is equipped with an omnidirectional antenna and the subcarrier spacing is Δf = 30KHz. For the user, consider a multipath channel model, represented by L t Indicates the number of multipaths in the tth transmission time slot, and takes L t =200. and Denote the path delay, azimuth angle of arrival (AOA), and elevation angle AOA of the lth path in the tth time slot. and Denote the equivalent path delay, the directional component on the x-axis of the three-dimensional coordinates, and the directional component on the y-axis. In the t-th time slot, k is used to represent the interference index. Use and represents the diagonal signal matrix of the user and the kth jammer, where X t and The nth diagonal elements of t，n and Indicates. x t，n and The average power is P and P I ,Right now and Assume that the length of the cyclic prefix (CP) is greater than the maximum delay spread of the channel. After removing the CP and applying the discrete Fourier transform (DFT), the received signal in the frequency-space domain is expressed as

[0090]

[0091] N t is an additive white Gaussian noise (AWGN) matrix whose elements independently follow Gaussian distributions

[0092] User channel H t Interference Channel Respectively expressed as

[0093]

[0094] ρ t，l represents the path loss of the lth path, α t，l represents the lth complex random coefficient with unit energy, and ai (ω),ω∈{τ t，l ,θ t，l ,φ t，l} represents the steering vector, and the expression is where i 2 =-1; and They represent the complex coefficient, path loss, equivalent delay, and directional components on the x-axis and y-axis of the interference k-th path,

[0095] S2, receive signal matrix Y t Vectorize y t =vec(Y t ), construct the equivalent receiving signal model as follows

[0096] To illustrate the equivalence with the received signal model in S1, the following equation is given: t =vec(N t ),s t (τ t，l )=x t ⊙a N (τ t，l ),

[0097] S3. Construct the mapping from signal time slot t to user position q: The CKM update period is T, and the CKM update process in each period follows the same steps. To achieve CKM construction within a period, define the mapping from time slot t to position q as follows:

[0098]

[0099] in and Represents the set of time slots and positions respectively. Position q corresponds to a 2×2m 2 Note that It is a many-to-one mapping, i.e. multiple time slots are associated with the same position q. The following assumptions are made when constructing the CKM:

[0100] (a) At any position q, the channel parameter ρ t，l , τ t，l ,θ t，l and φ t，l The variation within 2T consecutive time slots can be ignored.

[0101] (b) Obtained by the base station in each time slot.

[0102] (c) Interference pilot Known by the base station at the beginning of each CKM update cycle.

[0103] The construction and update process of CKM is as follows Figure 2 Specifically, Denote the set of time slots associated with position q by using Will receive signal y t Associated with position q. In T time slots, the received signal from position q are collected as historical data, which can be reference signals, synchronization signals, and pilot signals of all users accessing the base station. These historical data are then used to update the CKM. In the next T time slots, the CKM is used to assist the communication of the target user at the qth position. Note that a base station can serve hundreds of users at different locations through TDD. Therefore, the received signal data of all users at the same position q can be used to construct the CKKM.

[0104] S4, based on S3 And three assumptions, the channel knowledge graph is represented as a mapping:

[0105]

[0106] in and They represent the number of equivalent channel paths, l equivalent delays, horizontal components, vertical components and path loss corresponding to the qth position of the spectrum output respectively.

[0107] S5. Channel knowledge graph mapping using S4 Re-characterize the channel matrix of S2 as

[0108]

[0109] S6. Construct channel-based knowledge graph mapping The received signal model

[0110]

[0111] in Represents the equivalent noise formed by the combination of channel characterization error and AWGN.

[0112] S7. Convert the channel knowledge graph construction problem into a Bayesian inference problem: define the vector Contains elements and Contains elements Given an observation and The posterior distribution of is decomposed into:.

[0113]

[0114] The user channel parameter prior probability model is: The superscript "pri" is the abbreviation of "prior". represents the von Mises (VM) distribution. The prior probability model of the interference channel is and in and The probability density function (pdf) of the joint Bernoulli-Gaussian model is expressed as The probability of being a zero vector is 1-λ, where λ is 0.5. This model describes the possible severe path loss and / or low transmit power of interferer k. Bayesian optimal estimation (e.g., minimum mean square error (MMSE) estimator) requires high-dimensional integration, which is computationally infeasible. In the following section, a message passing algorithm is developed to obtain an approximate solution.

[0115] S8. Construct message passing factor graph: as shown in the attached Figure 3 As shown in Figure 1, the factor graph is used to represent the posterior probability decomposition in S7, where rectangles (called factor nodes) represent probability distributions and circles (called variable nodes) represent random variables. For simplicity, the abbreviation of probability density function (pdf) is introduced, such as p(y t |·) is abbreviated as For any variable node v, introduce an alternative pdfb v (v) to approximate the posterior PDF of v. For example, the variable node at yes Approximation. At any factor node f and its associated variables (with vector u f Indicates) at the place where b is introduced f (u f ) to approximate u f The posterior pdf of . is the set of variables associated with factor node f; Representing a collection Element v is not included; is the set of factor nodes associated with the variable node v; Representing a collection Factor node f is not included; represents the projection of pdfb(·) onto a pdf that satisfies Gaussian or VM. (or ) is called a message from variable node v to factor node f (or from factor node f to variable node v).

[0116] S9, message passing initialization: select a non-informative prior p(v), where Set and So that VM pdfp(v) tends to a uniform pdf in [0, 2π). μ in v , Random values are in the region [0, 2π). Note that small concentration leads to Will κ v Set to a relatively large value, such as 10 5 Then, use Least squares initialization of β t，l and The prior variances of and In particular, the sparsity λ is initialized to 1, which means is considered as a non-sparse vector. This setting is due to the random initialization In the first iteration, the value deviates significantly from the true value, resulting in a N (·)and Due to this mismatch, The initial estimate of To avoid this problem, we set λ = 1 in the first iteration of message passing. Then, after all parameters have been updated once, we set λ = 0.5 to take advantage of the sparse prior, and the λ value in subsequent iterations is given by EM learning. as well as

[0117] S10. Update slave factor nodes To variable node News

[0118]

[0119] middle and Note that for have I x (·) represents the Bessel function of order x and first kind.

[0120] S11. Approximate news In multiple time slots The product term on

[0121]

[0122] in

[0123]

[0124] in Representative function The inverse function of represent The local maximum point of For a single variable function, a heuristic search can be used to obtain the local maximum point

[0125] S12, Calculate Factor Node Beliefs in as follows

[0126]

[0127] in and and arg(·) represent the magnitude and angle operations respectively.

[0128] S13. Calculate from factor nodes To variable node News

[0129]

[0130] Then follow step S12 to get Notice and It is symmetrical in the received signal model, so it can be similarly obtained For the calculation of the interference channel parameters, note that when Time τ t，k，l ,θ t，k，l and φ t，k，l Respectively and Symmetrically, these corresponding messages are updated by simplifying , that is, there is no need to multiply messages of different time slots in S11.

[0131] S14, Calculate Factor Node Beliefs in as follows

[0132]

[0133] in based on For variables Obtain, and

[0134]

[0135] S15. Calculate Factor Node To variable node β t，l News

[0136]

[0137] S16, Calculate Factor Node Beliefs in as follows

[0138]

[0139] in

[0140]

[0141] belief The corresponding mean and variance for

[0142]

[0143] S17. Calculate from factor nodes To variable node News in

[0144]

[0145] S18, EM learning updates prior parameters as follows

[0146]

[0147] S19: Message transmission is terminated when the maximum number of iterations is reached, and the channel knowledge graph is output. in Obtained by S13, and Obtained by S14, Obtained from S18.

[0148] S20, using the channel knowledge graph output to enable channel estimation, and obtaining the equivalent received signal model based on S1

[0149]

[0150] where rt =vec(X -1 Y t ),h t =vec(H t ), With w t =vec(X -1 N t ). vec(·) represents a matrix vectorized operation.

[0151] S21. Calculate the user channel frequency-spatial covariance matrix

[0152]

[0153] in and parameter Given by the constructed CKM, namely S19.

[0154] S22. Calculate the power delay spectrum of the user channel

[0155]

[0156] where ||·||2 represents the vector 2-norm.

[0157] S23. Estimation of interference spatial covariance matrix: Define P thres is the power threshold, and defines Represents The corresponding delay tap index set, P thres Take 10dB lower than the interference strength. Then, estimate the Right now

[0158]

[0159] S24, using MMSE-IRC channel estimation

[0160]

[0161] Based on the Khatri-Rao product property, The equivalent expression is in and and ⊙ represent Kronecker product and Hadamard product respectively.

[0162] Attachment Figure 4 It shows the CKM accuracy and The name “OMP-based CKM” corresponds to the paper “Orthogonal matching pursuit for sparse signal recovery with noise”, and the name “ICI-non-cognitive CKM” corresponds to the proposed message passing scheme ignoring the interference feature extraction and interference signal elimination steps. The CKM accuracy of the proposed scheme increases with and SINR increase. In particular, when When SINR = -5dB, the CKM accuracy of the proposed scheme reaches -20.5dB. When SINR = -5dB, the accuracy of the OMP-based ICI-non-cognitive CKM method is -8dB and -12.8dB, respectively, which is 12.5dB and 7.7dB higher than the proposed algorithm. These results verify the effectiveness of the proposed graph construction algorithm in interference cancellation and show that the proposed algorithm can achieve high accuracy at relatively low SINR.

[0163] Attachment Figure 5The channel estimation performance of the proposed MMSE-IRC is compared with baseline methods, with the CKM constructed at SINR = 0 dB. The names "OMP" correspond to the paper "Orthogonal matching pursuit for sparse signal recovery with noise," "TMP" correspond to the paper "Structured turbo compressed sensing for downlink massive MIMO-OFDM channel estimation," and "VBI-MMSE" correspond to the paper "Robust deep learning for uplink channel estimation in cellular network under inter-cell interference." OMP outperforms TMP and VBI-MMSE when SINR < -5 dB, but is surpassed by TMP and VBI-MMSE when SINR > -5 dB. This is because TMP and VBI-MMSE are designed based on a Bayesian framework, and the prior parameter learning is inaccurate at relatively low SINRs. The proposed scheme significantly outperforms the baseline methods. Specifically, to achieve MSE = -15dB, TMP and VBI-MMSE require SINR ≥ 5dB and SINR ≥ 2.5dB, respectively, while the proposed scheme only requires SINR ≥ -15dB, a reduction of at least 17.5dB. This result demonstrates that CKM can significantly save transmit power to achieve a specified MSE or significantly improve performance at a fixed SINR. Furthermore, MMSE-IRC based on the proposed CKM significantly outperforms MMSE-IRC based on OMP or ICI-non-cognitive CKM. This observation demonstrates the superiority of the proposed CKM construction approach in enhancing channel estimation.

Claims

1. A channel estimation method based on channel knowledge graph for MIMO-OFDM system, where the base station has M=M x ×M y The base station communicates with a single target user via N subcarriers in the tth time slot. The method comprises: S1. Establish the user receiving signal model as follows: in, and Denote the diagonal signal matrices of the user and the kth jammer, respectively, and the vectors composed of their diagonal elements are x t and N t is an additive white Gaussian noise matrix whose elements independently follow Gaussian distribution User channel H t Interference Channel Respectively expressed as: Where L represents the number of multipaths of the target user, represents the number of multipath interference k, α t，l represents the lth complex random coefficient with unit energy, ρ t，l represents the path loss of the lth path, a i (w), w∈{τ t，l ,θ t，l ,φ t，l } represents the steering vector, and the expression is where i 2 =-1; and They represent the equivalent path delay, the directional component on the x-axis of the three-dimensional coordinate system, and the directional component on the y-axis, respectively. and represents the path delay, azimuth arrival angle, and elevation angle of the lth path in the tth time slot, and They represent the complex coefficient, path loss, equivalent delay, and directional components on the x-axis and y-axis of the interference k-th path respectively; S2, receive signal matrix Y t Vectorize y t =vec(Y t ), construct an equivalent receiving signal model: among them s t (t t，l )=x t ⊙a N (t t，l ), S3. Define the mapping from time slot t to position q for: in, and denote the set of time slots and positions respectively; S4. Set the channel parameter ρ at any position q. t，l , τ t，l ,θ t，l and φ t，l The changes within 2T consecutive time slots are negligible, where T is the update period of the channel knowledge graph; Obtained by the base station in each time slot; interference pilot It is known by the base station at the beginning of each channel knowledge graph update cycle; the channel knowledge graph is represented as a mapping: in, and They represent the number of equivalent channel paths, equivalent time delay, horizontal component, vertical component and path loss corresponding to the qth position of the spectrum output respectively; S5. Channel knowledge graph mapping using S4 Re-characterize the channel matrix of S2 as: in is the equivalent complex coefficient of the lth article; S6. Construct channel-based knowledge graph mapping The received signal model: in, Represents the equivalent noise formed by combining the channel characterization error and additive white Gaussian noise; S7. Convert the channel knowledge graph construction problem into a Bayesian inference problem: define the vector ξ q Contains elements and Contains elements Given an observation ξ q and The posterior distribution of is decomposed into: Among them, the user channel parameter prior probability model is The superscript "pri" is the abbreviation of "prior". represents the von Mises (VM) distribution; the prior probability model of the interference channel is and in and Probability density function representation of the joint Bernoulli-Gaussian model The probability of being a zero vector is 1-λ, and the model is used to describe the possible severe path loss and / or low transmit power of the interferer k; S8. Construct a message passing factor graph: Use a factor graph to represent the posterior probability decomposition in S7. The factor nodes in the factor graph represent probability distributions, the variable nodes represent random variables, and the abbreviation of the probability density function is introduced, that is, p(β t，l ), and Abbreviated as and v is used to uniformly represent variable nodes, and f is used to uniformly represent factor nodes. represents the set of variables associated with factor node f; Representing a collection Element v is not included; Represents the set of factor nodes associated with the variable node v; Representing a collection Factor node f is not included; (or ) is represented as a message from a variable node v to a factor node f; b f (u f ) represents the factor node f with respect to the vector u f belief; S9, message passing initialization: select a non-informative prior p(v), where Set and Make VM pdf p(v) tend to uniform pdf in [0, 2π), for μ in v , Random values are in the range [0, 2π), κ v Set to 10 5 ;use Least squares initialization of β t，l and The prior variances of and The sparsity λ is initialized to 1. After all parameters are updated once, set λ = 0.5 and initialize as well as S10. Update slave factor nodes To variable node News where R t =Reshape(y t , [M, N]), and for have I x (·) represents the Bessel function of the xth order and the first kind; S11. Approximate news In multiple time slots The product term on : in in Representative function The inverse function of represent The local maximum point of For a single variable function, a heuristic search can be used to obtain the local maximum point S12, Calculate Factor Node Beliefs in in and and arg(·) represent the magnitude and angle calculations respectively; S13. Calculate from factor nodes To variable node News Then follow S12 to get and The received signal model is symmetrical, so it is similar to For the message calculation of interference channel parameters, when Time τ t，k，l ,θ t，k，l and φ t，k，l Respectively and Symmetric, the corresponding message is updated by simplifying The message is obtained, that is, the multiplication of messages in different time slots in S11 is deleted; S14, Calculate Factor Node About Belief as follows in based on For variables Obtain, and S15. Calculate Factor Node To variable node β t，l and News in and S16, Calculate Factor Node Beliefs in as follows in belief The corresponding mean and variance for S17. Calculate from factor nodes To variable node News in S18, EM learning updates prior parameters as follows S19, message passing reaches the maximum number of iterations and terminates, output As the channel knowledge graph information, Obtained by S13, and Obtained by S14, Obtained by S18; S20, using the channel knowledge graph output to enable channel estimation, and obtaining the equivalent received signal model based on S1 where r t =vec(X -1 Y t ),h t =vec(H t ), With w t =vec(X -1 N t ), vec(·) represents matrix vectorization operation; S21. Calculate the user channel frequency-space covariance matrix in and parameter Given by the constructed channel knowledge graph; S22. Calculate the power delay spectrum of the user channel Where ||·||2 represents the vector 2-norm; S23. Estimation of interference spatial covariance matrix: Define P thres is the power threshold, and defines Represents The corresponding index set of delay taps is then estimated by the sample covariance Right now S24, using MMSE-IRC channel estimation Based on the Khatri-Rao product property, The equivalent expression is in and and ⊙ represent Kronecker product and Hadamard product respectively; S25, output As the channel h t Estimates.

Images