Multipath channel parameter reconstruction and beam coverage prediction method

Abstract

The invention discloses a multipath channel parameter reconstruction and beam coverage prediction method, and relates to the technical field of wireless communication. Aiming at the problems of low data utilization efficiency, limited environment modeling precision, strong coupling of measured data and antenna beam configuration and the like in the prior art, the method comprises the steps of configuring antenna beams and orientation parameters, collecting reference signal receiving strength, performing deterministic channel modeling simulation, establishing a mapping model, performing sparse recovery, constructing a channel propagation model, evaluating a channel map and the like. In combination with a coherence perception weighting network (CARE-Net) algorithm and ray tracing (RT) physical prior guidance, accurate reconstruction from a low-dimensional reference signal receiving strength (RSRP) measurement value to a high-dimensional multipath parameter is realized. The method has the advantages that the reconstructed multi-path parameters are decoupled from the antenna configuration, the method has the cross-beam / cross-configuration generalization prediction capability, the channel coverage performance under different antenna configurations can be accurately predicted, a reliable basis is provided for wireless network optimization, and the method is low in cost, high in precision and strong in generalization.

Description

A method for multipath channel parameter reconstruction and beam coverage prediction Technical Field

[0001] This invention relates to the field of wireless communication technology, specifically to channel modeling and data fusion technology in complex propagation environments, and in particular to a multipath channel parameter reconstruction and beam coverage prediction method that utilizes physical simulation priors to assist sparse measurement data for high-precision channel parameter extraction and global radio map construction. Background Technology

[0002] A wireless channel is a physical description of the communication environment between a signal transmitter and receiver, and its multipath spatiotemporal characteristics determine the transmission quality of the communication system. As 5G-Advanced and 6G networks evolve towards higher frequency bands such as millimeter waves, beamforming technology has become crucial. This requires the network to accurately perceive the multipath angles, intensities, and spatial distributions in the environment, constructing high-precision radio maps to enable beam management, base station site selection, and integrated sensing applications. Currently, constructing radio maps and obtaining channel parameters mainly relies on the following three methods:

[0003] The first category is the interpolation method based on measured data, which is currently the most commonly used method for operator network optimization. Reference Signal Received Power (RSRP) and Channel State Information (CSI) are collected within a specific area through drive testing or user crowdsourcing. Spatial interpolation algorithms are then used to extend the discrete measurement points to the entire area, and combined with location information to construct a radio map. This is the most intuitive method for constructing radio maps, based on measurements from the real physical environment, and can reflect the current actual channel state. However, it is costly, time-consuming, relies on high-density sampling data, and the interpolation algorithm ignores the physical propagation mechanism of signals in the angular domain, resulting in larger prediction errors in non-line-of-sight (NLoS) areas.

[0004] The second category is ray tracing (RT) methods based on deterministic modeling. These methods perform electromagnetic calculations using electronic environment maps and can obtain the multipath structure of the channel across the entire domain. However, this method is highly dependent on the geometric and material accuracy of the building models. In real-world complex environments, due to dynamic environmental changes or missing map information, simulation results often exhibit systematic biases and are difficult to directly reflect the true channel state.

[0005] The third category is sparse recovery methods based on Compressed Sensing (CS). In large-scale MIMO systems, CSI is typically sparse due to spatial coherence and the limited scattering characteristics of the environment. To reduce measurement costs and correct simulation errors, existing techniques establish a mapping relationship between the received signal and the channel matrix, retrieving channel parameters from a limited number of measurements.

[0006] To obtain high-precision radio maps, existing technologies mainly include:

[0007] Chinese patent CN117914426A proposes a method for constructing a channel knowledge map based on a coarse physical map and radio frequency data. This method extracts coarse physical information such as building locations and land area outlines from a satellite map of the region of interest. Based on the regional environment type, an empirical building height distribution model (such as Rayleigh distribution) is selected to calculate the line-of-sight (LOS) link probability between the base station and the center point of a 3D grid, constructing an initial probabilistic line-of-sight channel knowledge map (CKM). Using measured received signal strength (RSS) from a UAV, the link status—Line of Sight (LoS) / NLoS—is determined through hypothesis testing. Based on the link status, the upper and lower limits of building height and the probability distribution are updated, iteratively optimizing the CKM until the UAV completes data collection.

[0008] Chinese patent CN113938935A discloses a method for constructing a channel knowledge map based on the expectation-maximization algorithm. It acquires multi-source channel knowledge data (such as path loss and angle of arrival) through offline RT simulation, offline field measurement, and online real-time measurement. It iteratively estimates model parameters using the expectation-maximization algorithm, that is, corrects the parameters in the statistical channel model with actual channel data, and constructs a channel map that maps "location to model parameters". It uses the inverse distance weighted interpolation algorithm to find the current channel information based on the user's location and the channel map.

[0009] Chinese patent CN116155412A proposes a localized statistical channel modeling method based on beamspace RSRP measurements. It establishes a linear mapping between low-dimensional RSRP measurement data and the statistical characteristics of high-dimensional wireless channel angular power spectra (APS). Leveraging the sparsity of multipath channels in the angular domain, it employs either the Weighted Non-negative Orthogonal Matching Pursuit (WNOMP) method or the Sparse Bayesian Learning (SBL) method to solve the sparse optimization problem. Specifically, WNOMP reduces the influence of large-amplitude columns in the coefficient matrix by designing dynamic weights, thereby estimating the multipath departure angle and path gain from single-grid RSRP measurements, enabling rapid assessment and modeling of wireless channel quality.

[0010] Currently, various methods for constructing high-precision radio maps in complex environments generally struggle to achieve an effective balance between the three core requirements of "lightweight," "high accuracy," and "generalizability." Specific problems manifest themselves in the following aspects:

[0011] 1. Low data utilization efficiency and high acquisition cost. Traditional drive tests mostly collect channel parameters such as RSS or RSRP as channel state characteristics, or store CSI in large quantities. It is difficult to use the data efficiently, and the data storage scale requirement is too high. Each round can only measure channel information under a single antenna configuration, which consumes a lot of human and material resources and cannot meet the needs of dynamic network optimization.

[0012] 2. Limited accuracy of environmental modeling. Pure physical simulation using RT technology can generate global radio maps with various antenna parameter configurations, but its prediction accuracy is highly dependent on the geometric accuracy of environmental modeling and the accuracy of material electromagnetic parameters. The simulation results often deviate from the real environment and are difficult to correct through simple parameter adjustments, resulting in maps that rely solely on simulation being unable to support high-precision network optimization.

[0013] 3. Measured data is strongly coupled with antenna beam configuration, lacking generalization capability across antenna configurations. Existing measurement results are essentially scalar results resulting from the combined effect of spatial multipath channel gain and specific antenna patterns. They are observations from a specific beam perspective, rather than environmental characteristics independent of the device. Due to the lack of effective decoupling mechanisms in current technology, it is difficult to reconstruct antenna-independent physical multipath parameters from sparse measurement data. This means that measurement data is bound to the current antenna weights, downtilt angle, or codebook depth. Once the antenna configuration changes, the original data becomes meaningless, requiring a complete re-acquisition of the entire domain. This prevents the "one measurement, multiple scenario simulations" approach, limiting the application potential of radio maps in network planning. Summary of the Invention

[0014] The purpose of this invention is to address the problems of low data utilization efficiency, limited environmental modeling accuracy, and strong coupling between measured data and antenna beam configuration in existing technologies. This invention provides a multipath channel parameter reconstruction and beam coverage prediction method that combines physical simulation priors with advanced sparse recovery algorithms to achieve low-cost, high-precision, and highly generalizable channel parameter reconstruction and beam coverage prediction.

[0015] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0016] A method for multipath channel parameter reconstruction and beam coverage prediction includes the following steps:

[0017] 1) Configure antenna beam and orientation parameters and transmit reference signals: Configure the antenna array parameters of the transmitting end and construct a discrete Fourier transform beamforming codebook to control the transmitting end to transmit reference signals in the corresponding beam directions in sequence;

[0018] 2) The receiving end collects the reference signal received strength under different beam conditions: The user terminal receives reference signals from different beam directions and measures their reference signal received strength (RSRP) to obtain the macroscopic energy statistics characteristics of the location under a specific beam configuration;

[0019] 3) Deterministic channel modeling and simulation to generate channel multipath parameters: Deterministic ray tracing technology is used to simulate the electromagnetic propagation environment of the target area, obtain coarse-grained multipath distribution information, construct a model that reflects the prior probability of channel multipath parameters, and indicates the probability of the existence of paths in each angular direction;

[0020] 4) Establish a mapping model between the reference signal received strength statistics and the channel multipath parameters: Based on the antenna pattern of the antenna array, the discrete Fourier transform codebook and other configuration parameters, calculate the theoretical power gain of each beam on each transmission angle grid, establish a linear mapping relationship from the macroscopic low-dimensional measurement domain to the high-dimensional angle domain, and construct a joint sensing dictionary matrix of RSRP measurement values ​​and APS.

[0021] 5) Deterministic channel modeling and simulation prior-guided sparse recovery of channel multipath parameters: Guided by the prior probability of channel multipath parameters, the sparse recovery algorithm is used to solve the underdetermined equations and invert the high-dimensional sparse APS from the low-dimensional RSRP measurement vector;

[0022] 6) Obtain the actual environment channel propagation model from the channel multipath parameters: Extract non-zero elements from the reconstructed sparse APS to obtain the actual environment multipath parameters, and construct an actual environment channel propagation model independent of the antenna configuration, reflecting the inherent physical propagation properties of the environment.

[0023] 7) Evaluate the environmental channel map under different antenna configurations: Based on the reconstructed actual environmental channel propagation model, predict and evaluate the channel coverage performance of base station antennas under different physical orientations or beam configurations, and verify the accuracy, physical authenticity and generalization ability of the channel model.

[0024] In step 1), the transmitter uses a uniform planar array (UPA) as the base station transmitter model, and the antenna array consists of... It consists of several array elements, among which The number of elements in the horizontal direction. The number of array elements in the vertical direction is set to the element spacing. The working center frequency is The corresponding free space wavenumber is The base station is simulated based on the Discrete Fourier Transform (DFT) codebook. Each beam, generated by adjusting the phase shifter, can cover a specific spatial area;

[0025] No. beams ( The shaped vector of ) Shaped vectors of the horizontal dimension and the shaping vector in the vertical dimension Decide:

[0026]

[0027] in, For the shape vector of the horizontal dimension, The frequency is the spatial frequency in the horizontal dimension, j is the imaginary unit, and T represents the transpose operation. It is the base of the natural index; The shaping vector refers to the vertical dimension. Spatial frequency in the vertical dimension; It is the first The beamforming vector, The total number of elements in the antenna array (i.e. ), It is the horizontal dimension shaping vector corresponding to the m-th beam. It is the vertical dimension shaping vector corresponding to the m-th beam. This represents the Kronecker product operation;

[0028] At the transmitting end, based on the discretization concept, the continuous AoD spatial angle domain is transformed into a finite mesh;

[0029]

[0030] Among them, tilt angle With azimuth angle They are uniformly discretized in free space, and respectively obtained An angle of inclination and Azimuth angle, The vertical angle sampling interval, The horizontal angle sampling interval, the total number of angle grids .

[0031] In step 4), a linear mapping relationship is established from the macroscopic low-dimensional measurement domain to the high-dimensional angular domain:

[0032]

[0033] in, Number of beam directions This represents the total number of angle grids, which is also the candidate direction of the path; Characterized the first The DFT beam in the first Power response in each spatial angular direction, Indicates the first The path power along each discrete angular grid direction corresponds to the complex gain of that path in the physical channel impulse response. The square of the modulus, that is ;

[0034]

[0035] Introducing measured antenna pattern data, ,in, It is the antenna element in Total power gain in the direction, It is the antenna element in the direction Vertical polarization complex gain, It is the antenna element in the direction Horizontal polarization complex gain; It is the tilt angle corresponding to the nth angular direction; It is the azimuth angle corresponding to the nth angular direction; Represents the modulo square operation for complex numbers;

[0036] The array beam gain is obtained by the combined effect of the codebook beam and the array response factor, where the array response factor is:

[0037]

[0038] in, The array response factor vector in the horizontal dimension, where j is the imaginary unit, π is pi, and λ is the signal wavelength. It is the spacing between antenna elements. It is the azimuth angle corresponding to the j-th beam. It is the tilt angle corresponding to the i-th beam. The number of elements in the horizontal direction. The number of elements in the vertical direction. It is the base of the natural exponent, and T represents the transpose of a vector; This is the array response factor vector along the vertical dimension. It is the total array response factor vector corresponding to the nth angular direction. It is the horizontal dimension array response factor vector corresponding to the nth angular direction. It is the vertical dimension array response factor vector corresponding to the nth angular direction. Represents the Kronecker product operation;

[0039] In step 4), the construction of the joint sensing dictionary matrix of RSRP measurements and sparse angular power spectrum is first transformed into a sparse linear inverse problem through a linear mapping model between beam RSRP and APS:

[0040]

[0041] For RSRP measurement vectors, For the APS to be reconstructed, this vector is sparse. ,Right now The vast majority of elements are 0, only Each position is a non-zero positive value. ; For the joint perception dictionary matrix.

[0042] In step 5), the deterministic channel modeling simulation prior guides the sparse recovery of channel multipath parameters, and models the channel sparse parameter reconstruction process as a series of iteratively solved convex optimization subproblems, specifically including three core stages: preprocessing normalization, coherence-aware iterative weighted optimization, and non-negative least squares debiasing, to achieve high-precision parameter reconstruction under ill-conditioned measurement matrices.

[0043] 5.1) In the preprocessing normalization stage, to address the issue of significant differences in gain between the main lobe and side lobes of the antenna beam, a normalized measurement space is constructed to eliminate search bias caused by uneven antenna beam gain; the measurement vector... and joint perception dictionary matrix Perform norm normalization to construct a standardized measurement model:

[0044]

[0045]

[0046] in, For a normalized dictionary matrix, satisfying This step ensures that all beam directions have equal energy weights in the initial optimization phase, eliminates scale bias in beam gain, and ensures that the algorithm searches all angular directions fairly.

[0047] 5.2) In the iterative optimization phase, a dynamic closed loop of "probe-feedback-coherence suppression" is constructed;

[0048] The iterative process aims to gradually adjust the regularization penalty weights; in the algorithm's... The next iteration ( In this process, the coefficients of all angle grids are evaluated simultaneously, and dynamic weights are constructed based on the estimation results of each round. This globally balances the data fit and the sparsity of the solution, avoiding the local optima of greedy algorithms. Each iteration solves the following weighted non-negative elastic network subproblem to obtain the distribution of intermediate solutions under the current weights:

[0049]

[0050] in, yes Regularization coefficients are used to control sparsity. When multiple features are highly correlated, the algorithm tends to randomly select one feature and suppress the others; therefore, regularization coefficients are introduced. Regularization coefficients ensure that highly correlated adjacent beams acquire similar non-zero coefficients in the early stages of iteration, avoiding simplistic approaches. Regularization mitigates the risk of randomly discarding highly coherent paths while ensuring the strict convexity and uniqueness of the solution under underdetermined conditions; whereby... yes Regularization coefficients are used to control numerical stability and prevent noise amplification caused by normalization operations in the preprocessing stage.

[0051] An iterative weighting mechanism based on a dynamic coherence repulsion field is used to introduce coherence-aware repulsion weights into the weight construction. By integrating sparsity promotion and coherence repulsion, it actively identifies strong paths and imposes penalties on their neighborhoods, effectively suppressing false sidelobes and achieving accurate single-path resolution in ultra-high resolution grids. Defined by the following formula:

[0052]

[0053] in, To prevent small constants with a denominator of zero (e.g.) ), This is the coherence penalty coefficient. Let be the normalization factor, ensuring the weight mean is 1 to maintain the stability of the regularization strength; let For the first The set of strong paths identified by round iteration, if grid If an element does not belong to the set of strong paths, then the maximum cross-correlation between it and atoms in the set of strong paths is calculated as the repulsion value. ;

[0054]

[0055] If grid If the beam direction is similar to the currently discovered strong paths, then the penalty weight is increased, making... Increase in size, thus optimizing Suppression is set to zero, thus inhibiting the generation of false splitting paths;

[0056] 5.3) In the non-negative least squares bias removal stage, because Regularization term This will cause a contraction effect on the signal amplitude, and the normalization in the preprocessing stage changes the physical dimensions of the signal, resulting in a smaller reconstructed RSRP power. Therefore, after iterative convergence, a debiasing operation needs to be performed to restore the true physical power.

[0057] Using the non-zero support set obtained in the final iteration Solving the nonnegative least squares problem based on the unnormalized original measurement model:

[0058]

[0059] Supporting sub-matrix Reference from the original perception dictionary In the middle, only the support set is extracted. The column-full rank submatrix reconstructed from the column vectors corresponding to each index; when ,in To determine the total number of identified paths, then: The dimension of this matrix is ,in To measure the number of beams, The number of paths, usually At the found angular position, the least squares method is used for unbiased estimation to solve the amplitude contraction problem caused by sparse regularization, accurately recover the power gain of the multipath signal, and ensure the physical accuracy of the reconstructed APS.

[0060] In step 5), guided by the prior probability of channel multipath parameters, this invention proposes physical prior information to guide grid search. When the number of measurements is limited or the signal-to-noise ratio is low, the reconstruction accuracy is limited. Coarse-grained multipath distribution information of the target scene is obtained through RT simulation, physical environment information is introduced, and physical prior is used to guide channel reconstruction, guiding the algorithm to prioritize the angle region that is physically likely to exist when searching for the path.

[0061] Specifically, discrete RT simulation data Transform it into a continuous probability weight vector covering all angle grids. ;No. Prior probability weights of each angle grid Defined as:

[0062]

[0063] in, It is the first APS parameters for each angle grid. It is the first APS parameters for each angle grid. It is a distance metric. It is the tunable diffusion coefficient in the Gaussian kernel;

[0064] exist The regularization coefficients take into account prior probability weights. If the first... If a grid at a given angle has prior support, then the angle of the grid will decrease. Regularization coefficients penalize weights, increasing the probability of selection; weights The construction formula that combines sparsity and prior knowledge is as follows:

[0065]

[0066] The physical consistency and generalization ability of the channel multipath parameter reconstruction method will be evaluated through antenna array rotation experiments, so as to achieve cross-configuration beam coverage prediction based on environment-hardware decoupling.

[0067] In step 7), the accuracy, physical authenticity, and generalization ability of the verification channel model are assessed, wherein the error formula for evaluating the generalization ability is:

[0068]

[0069] in, The total number of samples involved in the calculation; the index of the sample (from 1 to n); This represents the true value corresponding to the i-th sample; Let be the predicted value corresponding to the i-th sample.

[0070] The core innovations of this invention are as follows:

[0071] 1. Coherence-aware iterative weighting mechanism

[0072] To obtain high-precision angle estimates, the preset angle grid is usually extremely dense, resulting in a high correlation between adjacent column vectors in the perception dictionary. Traditional algorithms such as Lasso and NNOMP cannot distinguish between real paths and their adjacent spurious noise paths, and are prone to mistakenly identifying a real path as multiple adjacent weak paths. This leads to the loss of the reconstructed APS physical environment structure information, which fails to reflect the true physical propagation model.

[0073] This invention proposes CARE-Net, which establishes an iterative weighting mechanism based on a dynamic coherence repulsion field, unlike traditional weighting methods. The algorithm weights signals only based on their amplitude, and introduces coherence-aware repulsive weights into the weight construction. During the iteration process, the algorithm identifies the current set of strong paths in real time and calculates the coherent interference distribution of these strong paths on the entire grid. In the next round of weight updates, a high penalty is imposed on grids that are highly correlated with strong paths but are not the strong paths themselves. This method helps maintain the independence of solutions in dense grids, actively removes highly coherent spurious sidelobes, effectively suppresses spurious path and sidelobe errors caused by dictionary coherence, and significantly improves the physical resolution and sparsity of angle estimation.

[0074] 2. Raytracing Prior Guidance Based on Physical Environment Model

[0075] Traditional sparse recovery algorithms are usually data-driven blind estimations. When measurement data is limited or the signal-to-noise ratio is low, pure data-driven algorithms are prone to getting trapped in local optima and recovering noisy spurious paths.

[0076] APS itself reflects the actual environment channel propagation model and implicitly contains physical environment structure information. This invention utilizes known building geometry information and introduces deterministic ray tracing simulation data as an aid to construct a continuous physical prior probability field that reflects the environmental geometry. This factor is then embedded as a regulating factor into the regularization weights to guide the sparse recovery grid search. The specific construction is as follows: This approach lowers the penalty threshold in angular regions where a path is physically likely to exist, and raises the penalty barrier in impossible regions. By effectively integrating the static prior probability of real-time analysis (RT) with dynamic local corrections from measurement data, the solution space of the problem is significantly narrowed, ensuring that the algorithm can quickly locate the true physical path region even under low signal-to-noise ratio or extremely underdetermined conditions, thus significantly improving robustness.

[0077] 3. Cross-beam / cross-configuration generalization prediction capability based on "environment-hardware decoupling"

[0078] The data obtained from current channel measurements is a product of the coupling between the environmental channel and the antenna configuration. Once the base station adjusts the antenna's mechanical downtilt angle, azimuth angle, or changes the beam codebook, it is usually necessary to re-perform drive testing, which results in extremely high network optimization costs.

[0079] The CARE-Net algorithm proposed in this invention has a three-stage structure of "energy normalization - coherence-aware weighting - debiasing and energy reconstruction". It reverse-engineers the APS (Average Power Span) independent of the antenna hardware from the RSRP (Resonance Reflection Point) scalar data affected by a specific beam, characterizing the inherent electromagnetic propagation properties of the environment and achieving complete decoupling between environmental characteristics and measurement hardware. Based on this model, coverage performance of base station antennas under any new configuration, such as mechanical rotation, downtilt adjustment, or beamforming changes, can be accurately predicted without further drive testing.

[0080] Compared with the prior art, the present invention has the following beneficial effects:

[0081] 1. High-precision reconstruction: Through the coherence sensing iterative weighting mechanism, the problem of large energy difference between beam main lobe and side lobe and high coherence of dense grid is effectively overcome, which improves the physical resolution under dense angle grid. The reconstructed multipath parameters have extremely high physical fidelity and can reflect the signal propagation path and real environmental structure information.

[0082] 2. Strong robustness: The prior guidance constructed using RT simulation data provides global search navigation for sparse recovery, which solves the problem of algorithm robustness in scenarios with limited measurement data or low signal-to-noise ratio, and ensures that the algorithm can quickly lock the real physical path area.

[0083] 3. Strong generalization ability: The reconstructed multipath parameters are decoupled from the antenna configuration, and have generalization ability independent of the measurement hardware. It can accurately predict the coverage performance of the base station antenna under different new configurations, providing a reliable theoretical basis for beam management and coverage optimization of wireless networks and reducing network optimization costs.

[0084] 4. Low cost: Utilizing RSRP measurement data commonly used by existing UEs, there is no need for expensive dedicated equipment to collect underlying CSI data. The data acquisition cost is low and the convenience is high, making it easy to carry out high-density, long-term data collection across the entire network. Attached Figure Description

[0085] Figure 1 is a flowchart of the multipath channel parameter reconstruction and beam coverage prediction method provided in an embodiment of the present invention;

[0086] Figure 2 is a schematic diagram of the antenna array reference coordinate system and multipath channel provided in an embodiment of the present invention;

[0087] Figure 3 is a dictionary matrix diagram of the observed antenna orientation provided in an embodiment of the present invention;

[0088] Figure 4 is a distribution map of buildings in a target scene provided by an embodiment of the present invention; wherein, grayscale represents the height of the buildings;

[0089] Figure 5 is a flowchart of the verification process for the generalization capability of the sparse recovery method for channel multipath parameter reconstruction provided in an embodiment of the present invention.

[0090] Figure 6 is a dictionary matrix diagram of the observation antenna orientation 2 provided in an embodiment of the present invention;

[0091] Figure 7 shows the beam RSRP recovery curve and error at a single grid point under antenna orientation 1 provided in the embodiment of the present invention;

[0092] Figure 8 shows the beam RSRP prediction curve and error at a single grid point under antenna orientation 2 provided in the embodiment of the present invention;

[0093] Figure 9 shows the RSRP prediction error CDF curve for antenna orientation 2 provided in the embodiment of the present invention. Detailed Implementation

[0094] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0095] This invention proposes a method for multipath channel parameter reconstruction and beam coverage prediction. Channel multipath parameters can be obtained by acquiring complete Channel Impulse Response (CIR) or channel matrices. However, on the one hand, storing and transmitting complete CSI data requires extremely high bandwidth resources, making high-density, long-term acquisition across the entire network difficult; on the other hand, existing User Equipment (UEs) are limited by hardware costs and protocol stack constraints, typically only supporting the reporting of signal strength statistics, and obtaining the underlying CSI requires expensive dedicated equipment. Therefore, there is an urgent need for a low-cost technical solution that can utilize commonly used measurement data from existing UEs to invert channel multipath parameters.

[0096] RSRP is the best alternative to solving the aforementioned data acquisition challenges. As a mandatory measurement quantity defined in 5G and LTE standard protocols, RSRP can be directly extracted from the Synchronization Signal Block (SSB) or the pilot of the Channel-State-Information Reference Signal (CSI-RS). It can be collected using existing equipment. Furthermore, as a scalar statistical value, RSRP has a very small data volume, is easy to store, and possesses extremely high data acquisition convenience and low cost advantages. Although RSRP is a low-dimensional statistic, it is essentially the superposition of multipath signal energy in the beam domain, and has a definite physical mapping relationship with the angle and intensity of the multipath. By utilizing the statistical mapping relationship between RSRP and APS, combined with the angular sparsity of the channel, reverse reconstruction from low-dimensional measurements to high-dimensional multipath parameters can be achieved.

[0097] Current traditional 5G network optimization methods rely on engineering experience and multiple rounds of driving tests, requiring significant investment of manpower and resources. Offline optimization, on the other hand, employs the concept of digital twins, using channel models such as ray tracing simulations to achieve modeling and simulation without manual intervention; however, the accuracy of the environmental model is limited. To bridge the performance gap between offline optimization and real networks, as shown in Figure 1, this invention proposes to construct a channel propagation model for the actual environment based on RSRP measurements of beam scanning under a single antenna orientation, achieving accurate calculations under different antenna orientations and beams. The steps include:

[0098] ① Configure antenna beam and orientation parameters, and transmit reference signals: Configure the antenna array parameters of the transmitting end and construct a discrete Fourier transform beamforming codebook, and control the transmitting end to transmit reference signals in the corresponding beam directions in sequence.

[0099] ② The receiving end collects the received strength of the reference signal under different beam conditions: The user terminal receives the reference signal from different beam directions and measures its RSRP to obtain the macroscopic energy statistics characteristics of the location under a specific beam configuration.

[0100] ③ Deterministic channel modeling and simulation to generate channel multipath parameters: Deterministic ray tracing technology is used to simulate the electromagnetic propagation environment of the target area, obtain coarse-grained multipath distribution information, construct a prior probability reflecting the channel multipath parameters, and indicate the probability of the existence of paths in each angular direction.

[0101] ④ Establish a mapping model between the reference signal received strength statistics and the channel multipath parameters: Based on the antenna pattern of the antenna array, the discrete Fourier transform codebook and other configuration parameters, calculate the theoretical power gain of each beam on each transmit angle grid, establish a linear mapping relationship from the macroscopic low-dimensional measurement domain to the high-dimensional angle domain, and construct a joint sensing dictionary matrix of RSRP measurement values ​​and sparse angle power spectrum.

[0102] ⑤ Deterministic channel modeling and simulation prior-guided sparse recovery of channel multipath parameters: Guided by the prior probability of channel multipath parameters, the sparse recovery algorithm is used to solve the underdetermined equations and invert the high-dimensional sparse angular power spectrum from the low-dimensional RSRP measurement vector.

[0103] ⑥ Obtain the actual environment channel propagation model from the channel multipath parameters: Extract non-zero elements from the reconstructed sparse angular power spectrum to obtain the actual environment multipath parameters, and construct an actual environment channel propagation model independent of the antenna configuration, reflecting the inherent physical propagation properties of the environment.

[0104] ⑦ Evaluate the environmental channel map under different antenna configurations: Based on the reconstructed actual environmental channel propagation model, predict and evaluate the channel coverage performance of base station antennas under different physical orientations or beam configurations to verify the accuracy, physical authenticity and generalization ability of the channel model.

[0105] As shown in Figure 2, a uniform planar array (UPA) is used as the base station transmitter model. The antenna array consists of... It consists of several array elements, among which The number of elements in the horizontal direction. This represents the number of array elements in the vertical direction. The element spacing is set to... (usually half wavelength) The working center frequency is The corresponding free space wavenumber is The base station employs an analog design based on a Discrete Fourier Transform (DFT) codebook. Each beam, generated by adjusting the phase shifter, points to cover a specific spatial area.

[0106] No. beams ( The shaped vector of ) From horizontal and vertical The spatial frequency is determined by two dimensions.

[0107]

[0108] Among them, spatial frequency The UE receiver uses a single omnidirectional antenna, making it difficult to estimate the angle of arrival (AoA) and path delay. Therefore, the APS is described by the angle of departure (AoD) and gain of each path. At the transmitter, based on the discretization concept, the continuous AoD spatial angle domain is transformed into a finite grid.

[0109]

[0110] Among them, tilt angle With azimuth angle They are uniformly discretized in free space, and respectively obtained An angle of inclination and Azimuth angle, The vertical angle sampling interval, The horizontal angle sampling interval, the total number of angle grids .

[0111] In practical wireless communication environments, the RSRP under each beam is a superposition of signals from multiple propagation paths, which can be modeled as a linear projection of the specific antenna beam pattern onto the APS. Under the above antenna configuration, a linear mapping model between RSRP and APS is established:

[0112]

[0113] in, Number of beam directions This represents the total number of angle grids, which is also the candidate direction for the path. Characterized the first The DFT beam in the first Power response in each spatial angular direction, Indicates the first The path power along each discrete angular grid direction corresponds to the complex gain of that path in the physical channel impulse response. The square of the modulus, that is .

[0114]

[0115] Introducing measured antenna pattern data, ,in, It is the antenna element in Total power gain in the direction, It is the antenna element in the direction Vertical polarization complex gain, It is the antenna element in the direction Horizontal polarization complex gain; It is the tilt angle corresponding to the nth angular direction; It is the azimuth angle corresponding to the nth angular direction; Represents the modulo square operation for complex numbers;

[0116] The array beam gain is obtained by the combined effect of the codebook beam and the array response factor, where the array response factor is:

[0117]

[0118] in, The array response factor vector in the horizontal dimension, where j is the imaginary unit, π is pi, and λ is the signal wavelength. It is the spacing between antenna elements. It is the azimuth angle corresponding to the j-th beam. It is the tilt angle corresponding to the i-th beam. The number of elements in the horizontal direction. The number of elements in the vertical direction. It is the base of the natural exponent, and T represents the transpose of a vector; This is the array response factor vector along the vertical dimension. It is the total array response factor vector corresponding to the nth angular direction. It is the horizontal dimension array response factor vector corresponding to the nth angular direction. It is the vertical dimension array response factor vector corresponding to the nth angular direction. This represents the Kronecker product operation.

[0119] In real-world wireless communication environments, signals emitted from base stations are often obstructed by obstacles such as buildings, and the number of objects capable of reflecting the signal is limited. Especially in millimeter wave and mid-to-high frequency bands of 5G, the signal wavelength is short, and the diffraction capability is weak. The signal reaches the user primarily through LoS propagation and a few reflection paths, with very little energy in other directions. This means that among countless possible AOD (Aspect-Oriented Distance) propagation directions, the number of truly effective energy paths is extremely limited. The energy is extremely low, with zero energy in most directions, resulting in a sparse energy distribution in the angular domain. Therefore, by using a linear mapping model between beam RSRP and APS, channel parameter reconstruction can be transformed into a sparse linear inverse problem.

[0120]

[0121] For RSRP measurement vectors, For the APS to be reconstructed, this vector is sparse. ,Right now The vast majority of elements are 0, only Each position is a non-zero positive value. . Let be the joint sensing dictionary matrix. Therefore, this problem is reduced to a standard nonnegative sparse recovery problem, from which a high-dimensional sparse APS vector can be accurately recovered with extremely high probability from a small number of RSRP measurements. .

[0122]

[0123] in, This indicates the maximum number of non-zero elements.

[0124] Figure 3 is a schematic diagram of the dictionary matrix for the observed antenna orientation. The z-axis represents the beam gain, indicating the signal gain of the antenna under a preset orientation for different codebook beams on the angle grid. This means that the dictionary matrix is ​​closely related to the antenna pattern and codebook beam. The beam RSRP is affected by the main lobe and side lobes, resulting in extremely uneven column energy. The amplitude of the main lobe column is much larger than that of the side lobes, and the side lobe paths are easily ignored. At the same time, in order to obtain high-precision angle estimation, the spatial angle grid is divided into extremely dense sections. This results in adjacent column vectors in the dictionary matrix being approximately parallel in height. The most basic method for sparse recovery is to use Non-Negative Orthogonal Matching Pursuit (NNOMP), a classic non-negative greedy iterative algorithm.

[0125]

[0126] In each iteration, the correlation between the residual and each column of the sensing matrix is ​​calculated, and the column with the highest correlation, i.e. the beam direction with the strongest energy, is selected to be added to the support set. This causes NNOMP to prioritize the main lobe column when applying multipath channel reconstruction and cannot distinguish highly correlated columns.

[0127] To mitigate the susceptibility of NNOMP to beam gain unevenness (main lobe masking side lobes), existing techniques (Zhang S, Ning X, Zheng X, et al. A physics-based and data-driven approach for localized statistical channel modeling[J]. IEEE Transactions on Wireless Communications, 2023, 23(6): 5409-5424) propose WNOMP. WNOMP introduces column normalization and dynamic energy weights in the atom selection of the NNOMP iterative solution. By weighting different columns, selection bias is corrected, and the influence of the main lobe and side lobes of the beam is balanced to alleviate the problem of uneven energy distribution in the sensing matrix columns.

[0128] WNOMP Algorithm Flowchart

[0129] enter: sparsity

[0130] Construct a column-normalized perceptual dictionary.

[0131] 1) Sparse recovery initialization: Supports collection residual

[0132] 2) Sub-iteration atom selection and update:

[0133] 3) Calculation

[0134] 4) Select index The first term, normalized correlation, is used to overcome the problem of the main lobe masking the side lobes. The second term, dynamically adjusting the weights, integrates the original column energy information to select the path.

[0135] 5) Update the support set

[0136] 6) Solve the nonnegative least squares problem; the sparse solution is: And make the unselected paths , The selected column of atoms.

[0137] 7) Optimize the reset support set based on the results of nonnegative least squares.

[0138] 8) Update residuals

[0139] 9) Iteration counting

[0140] 10) Until

[0141] Output: Path APS

[0142] While WNOMP alleviates the problem of uneven column energy distribution to some extent, it does not change the inherent nature of the greedy algorithm. Faced with densely packed grids with extremely high coherence, it will still fall into local optima traps. This invention shifts to a solution approach based on global convex optimization, seeking the sparsest solution and using the fewest paths to interpret the measurement data, i.e., solving... Norm minimization problem However, due to its non-convexity and discontinuity, it belongs to the non-deterministic polynomial-time hard (NP-hard) problem. Mathematically, a convex relaxation strategy is usually adopted to transform it into a problem. The regularization problem, namely the classic Least Absolute Shrinkage and Selection Operator (LASO) model:

[0143]

[0144] pass The geometric properties of the norm transform the optimization problem into a strictly convex optimization problem. The regularization strength represents the regularization intensity, which theoretically avoids local optima traps and finds the sparsest solution globally. However, when dealing with high-density angular grids, the perceptual dictionary contains a set of highly correlated predictor variables, such as adjacent beams in the dense grid. LASSO tends to randomly select one variable from this set of related variables and assign it a non-zero coefficient, while forcing the others to zero, making path angle estimation susceptible to measurement noise. To overcome this... To address the instability of regularization in the selection of relevant variables, this invention further introduces... Regularization – Ridge Regression and Together, they form an elastic net (EN) (Zou H, Hastie T. Regularization and variable selection via the elastic net[J]. Journal of the Royal Statistical Society, 2005, 67(5):768-768):

[0145]

[0146] in, Used to control the overall regularization strength. Used for control and The proportion. When multiple grid directions are highly correlated. The item will retain these column vectors simultaneously, unlike... Random elimination preserves the true path and its neighborhood completely in the early stages of iteration, avoiding the risk of the true path being erroneously eliminated in the early stages. This provides the necessary numerical stability for multipath recovery in high-coherence environments.

[0147] While Encoding (EN) protects the true path, it also retains a large number of spurious sidelobes, resulting in insufficient sparsity of the solution. To achieve a balance between preserving the true path and eliminating spurious sidelobes, this invention improves upon EN with dynamic weighting, proposing a Coherence-Aware Reweighted Elastic Net (CARE-Net). This model transforms the channel sparse parameter reconstruction process into a series of iteratively solved convex optimization subproblems, specifically including three core stages: preprocessing normalization, coherence-aware iterative weighted optimization, and non-negative least squares debiasing. This enables high-precision parameter reconstruction under ill-conditioned measurement matrices.

[0148] The first stage is preprocessing. To address the significant difference in gain between the antenna beam's main lobe and side lobes, a normalized measurement space is constructed to eliminate search bias caused by uneven antenna beam gain. This is then applied to the measurement vectors. and joint perception dictionary matrix Perform norm normalization to construct a standardized measurement model:

[0149]

[0150]

[0151] in, For a normalized dictionary matrix, satisfying This step ensures that all beam directions have equal energy weights in the initial optimization phase, eliminates scale bias in beam gain, and ensures that the algorithm searches all angular directions fairly.

[0152] Next, iterative optimization is performed to construct a dynamic closed loop of "probe-feedback-coherence suppression". It is important to note that, unlike the atom-by-atom greedy search (adding one path in each iteration) used in the existing WNOMP technique, the iterative process of this invention does not aim to determine non-zero elements one by one, but rather to gradually adjust the regularization penalty weights. In the algorithm's... The next iteration ( In this process, the coefficients of all angle grids are evaluated simultaneously, and dynamic weights are constructed based on the estimation results of each round. This globally balances the data fit and the sparsity of the solution, avoiding the local optima of greedy algorithms. Each iteration solves the following weighted non-negative elastic network subproblem to obtain the distribution of intermediate solutions under the current weights:

[0153]

[0154] in, yes Regularization coefficients are used to control sparsity. When multiple features are highly correlated, the algorithm tends to randomly select one feature and suppress the others; therefore, regularization coefficients are introduced. Regularization coefficients ensure that highly correlated adjacent beams acquire similar non-zero coefficients in the early stages of iteration, avoiding simplistic approaches. Regularization mitigates the risk of randomly discarding highly coherent paths while ensuring the strict convexity and uniqueness of the solution under underdetermined conditions. yes The regularization coefficient is used to control numerical stability and prevent noise amplification caused by the normalization operation in the preprocessing stage.

[0155] This invention uses an iterative weighting mechanism based on a dynamic coherence repulsion field, which differs from traditional weighting methods. The algorithm only weights the signals based on their amplitude, and introduces coherence-aware repulsion weights in the weight construction. By combining sparsity promotion and coherence repulsion, it actively identifies strong paths and imposes penalties on their neighborhoods, effectively suppressing false sidelobes and achieving accurate single-path resolution in ultra-high resolution grids. Defined by the following formula:

[0156]

[0157] in, To prevent small constants with a denominator of zero (e.g.) ), This is the coherence penalty coefficient. Let be the normalization factor, ensuring the weight mean is 1 to maintain the stability of the regularization strength. For the first The set of strong paths identified by round iteration, if grid If an element does not belong to the set of strong paths, then the maximum cross-correlation between it and atoms in the set of strong paths is calculated as the repulsion value. .

[0158]

[0159] If grid If the beam direction is similar to the currently discovered strong paths, then the penalty weight is increased, making... Increase in size, thus optimizing Suppression is zero, thus inhibiting the generation of false splitting paths.

[0160] because Regularization term This will cause a contraction effect on the signal amplitude, and the normalization in the preprocessing stage changes the physical dimensions of the signal, resulting in a smaller reconstructed RSRP power. Therefore, after iterative convergence, a debiasing operation needs to be performed to restore the true physical power.

[0161] Using the non-zero support set obtained in the final iteration Solving the nonnegative least squares problem based on the unnormalized original measurement model:

[0162]

[0163] Supporting sub-matrix Reference from the original perception dictionary In the middle, only the support set is extracted. The column-full rank submatrix reconstructed from the column vectors corresponding to each index. When ,in To determine the total number of identified paths, then: The dimension of this matrix is... ,in To measure the number of beams, The number of paths, usually At the found angular position, unbiased estimation is performed using the least squares method, which solves the amplitude contraction problem caused by sparse regularization, accurately recovers the power gain of the multipath signal, and ensures the physical accuracy of the reconstructed APS.

[0164] In summary, the present invention, CARE-Net, utilizes normalization and coherence weights to accurately find the path angle under the premise of column equality, overcoming the problems of main lobe masking side lobes and high coherence in dense grids. After finding independent and accurate propagation paths in the angle domain, it uses the original matrix to backfill power energy, thus eliminating [the problem of] [other issues]. The amplitude deviation introduced by regularization ensures the physical accuracy of the RSRP prediction value.

[0165] CARE-Net algorithm pseudocode

[0166] Input: Measurement vector Joint perception dictionary matrix Maximum number of iterations Relevance penalty factor Regularization parameters .

[0167] 1) Initialization and preprocessing: Calculate the normalized dictionary and normalized measurement values ,initialization Weight Supports collection .

[0168] 2) Sub-iteration atom selection and update:

[0169] 3) Solving subproblems: Solving the weighted non-negative elastic network optimization problem using the coordinate descent method. To obtain the current estimate .

[0170] 4) Identify strong paths: based on The amplitude distribution determines the set of strong path indices. .

[0171] 5) For all non-strong path meshes :

[0172] 6) Calculate the coherence repulsion force: Calculate its relationship with... The maximum cross-correlation value of atoms in the middle is used as a repulsive term. .

[0173] 7) Update weights: Calculate the weights for the next round. .

[0174] 8) Weight normalization processing: ...

[0175] 9) Non-negative least squares debiased reconstruction: Extracting the final support set .

[0176] 10) Using the original matrix In support set Solve the nonnegative least squares problem above. To obtain the final physical amplitude .

[0177] Output: Sparse APS estimation vector

[0178] To verify the performance advantages of the proposed CARE-Net algorithm in complex multipath environments, five significant propagation paths were established between the transmitter and the user, each with specific AoD and power gain. The observation data consisted of RSRP measurement vectors generated from 32 DFT beams. The algorithm of this invention was compared with the existing WNOMP algorithm, and the path prediction results are shown in Table 1:

[0179] Table 1

[0180]

[0181] Based on 32 codebook beams, RSRP measurement statistics were calculated for 5 known paths and sparse recovery verification was performed. CARE-Net demonstrates superior path recovery capabilities compared to WNOMP, especially in high-coherence angle grid environments, where it can still accurately reconstruct the path APS. In terms of quantitative evaluation, by comparing the reconstruction accuracy of 32-beam RSRP, CARE-Net's mean absolute error (MAE) prediction was compared. The path recovery efficiency is only 1.2909 dB, significantly better than WNOMP's 21.2935 dB. This result strongly demonstrates the rationality and effectiveness of the CARE-Net algorithm in complex channel parameter reconstruction. The CARE-Net algorithm of this invention exhibits superior path recovery capabilities in complex channel parameter reconstruction.

[0182] Synthetic dataset validation and RT prior algorithm model:

[0183] To verify the performance of the algorithm of this invention in a complex scenario that more closely resembles the actual communication environment, this embodiment uses deterministic ray tracing as the dominant method and statistical small-scale fading as the superposition, and weighted hybrid clustering to construct high-fidelity channel data as verification data.

[0184] A deterministic set of multipath components was calculated using RT simulation in the target scene shown in Figure 4. It contains information such as the precise angles of the direct path, reflection path, and diffraction path, characterizing the large-scale spatial structure of the channel.

[0185] Furthermore, based on the 3GPP TR 38.901 standard channel model, random small-scale fading components following a specific statistical distribution are generated. To simulate random scattering and disturbances in the environment.

[0186] The deterministic large-scale components and the statistical small-scale components are weighted, fused, and clustered to generate the final synthetic channel data.

[0187]

[0188] Among them, the weighting coefficient Setting it to 0.9 means that 90% of the energy structure in the synthetic data is determined by the physical environment, and 10% by random perturbations, thus preserving both the geometric authenticity of the physical environment and the random fluctuations of real measurements.

[0189] Based on the channel data construction process described above, it is evident that the energy distribution and spatial multipath structure of actual wireless channels are highly dependent on the geometric and physical characteristics of the environment, and are highly positively correlated with ray tracing results. Existing sparse recovery algorithms typically treat the channel as a completely unknown black box, relying solely on measurement data for blind estimation. This invention proposes a physical prior information-guided grid search. When the number of measurements is limited or the signal-to-noise ratio is low, the reconstruction accuracy is limited. Coarse-grained multipath distribution information of the target scene can be obtained through RT simulation. By reasonably incorporating physical environment information and using physical prior information to guide channel reconstruction, the algorithm is guided to prioritize physically probable angle regions when searching for paths.

[0190] Specifically, discrete RT simulation data Transform it into a continuous probability weight vector covering all angle grids. . No. Prior probability weights of each angle grid Defined as:

[0191]

[0192] in, It is the first APS parameters for each angle grid. It is the first APS parameters for each angle grid. It is a distance metric. It is the tunable diffusion coefficient in the Gaussian kernel.

[0193] Add a priori guidance module to WNOMP and CARE-Net. WNOMP incorporates prior probability weights into its atom selection formula. A priori guiding influence factor:

[0194]

[0195] CARE-Net, on the other hand, is... The regularization coefficients take into account prior probability weights. If the first... If a grid at a given angle has prior support, then the angle of the grid will decrease. Regularization coefficients penalize weights, increasing the probability of selection. The construction formula that combines sparsity and prior knowledge is as follows:

[0196]

[0197] The evaluation method involved RSRP verification with different antenna orientations, due to the multipath distribution in the environment. The environment remains constant, determined by physical entities such as buildings and trees. When the base station antenna rotates, the observation method changes, specifically the beam pointing. This invention evaluates the physical consistency and generalization capability of the channel multipath parameter reconstruction method through antenna array rotation experiments, achieving cross-configuration beam coverage prediction based on environment-hardware decoupling. RSRP measurement data collected by the antenna array under a first preset orientation is utilized. and the corresponding perception dictionary Inversion and reconstruction of spatial APS Subsequently, the antenna array is physically rotated to a second preset orientation to construct a new perception dictionary. Reconstruction using sparse recovery Directly predict the theoretical RSRP value under this orientation and with actual measurements A comparison is then made. Since the multipath propagation structure of the channel in the physical environment does not change with the mechanical rotation of the base station antenna, this evaluation method can verify whether the reconstructed APS vector truly represents the inherent physical properties of the environment, ensuring that the established channel model has portability under different antenna configurations, and providing a reliable predictive basis for subsequent optimization of base station antenna parameters (such as adjusting downtilt and azimuth). The flowchart for verifying the generalization capability of the sparse recovery method for channel multipath parameter reconstruction is shown in Figure 5.

[0198] Based on the dataset and evaluation method described above, beam RSRP was collected under the antenna orientation shown in Figure 3 for sparse reconstruction of multipath parameters, and the RSRP prediction accuracy was observed under the antenna orientation shown in Figure 6.

[0199] The performance of WNOMP and CARE-Net methods with and without prior guidance was compared with that of the benchmark methods NNOMP, LASSO and EN.

[0200] Figure 7 illustrates the single-point beam RSRP reconstruction performance of different algorithms under antenna orientation 1. This experiment, based on RSRP measurement data acquired under antenna orientation 1, used the sparse recovery algorithm to invert the APS, remapped it back to the measurement space, and calculated the theoretical RSRP value under this configuration to verify the algorithm's ability to fit the observed data. Comparison with the actual RSRP reveals that the LASSO algorithm cannot effectively identify sparse paths under dense grid conditions and fails to effectively recover the physical multipath structure, resulting in a large fitting deviation. In contrast, other comparative algorithms all exhibit excellent data fitting capabilities, with the reconstructed curves closely matching the actual measurements.

[0201] Figure 8 further illustrates the beam RSRP prediction performance of each algorithm across antenna configurations. The experiment uses the APS reconstructed under antenna orientation 1 to directly predict the beam RSRP distribution when the antenna is rotated to orientation 2, in order to evaluate the generalization ability of the model.

[0202] The results show that the proposed CARE-Net algorithm and WNOMP algorithm both outperform traditional sparse recovery algorithms. CARE-Net, with its unique global joint optimization and coherence awareness mechanism, can accurately pinpoint the real physical path, thus significantly outperforming WNOMP in cross-configuration prediction. Furthermore, after fusing RT physical prior information, the prediction accuracy of both WNOMP and CARE-Net algorithms is significantly improved, while the prediction error of the proposed method is significantly reduced, with a MAE of only 2.21 dB, demonstrating that its reconstruction parameter APS possesses physical authenticity with "environment-hardware decoupling".

[0203] To avoid random errors in single-point testing, 30 spatial grid points were randomly selected for sparse recovery and RSRP prediction, and the cumulative distribution function (CDF) of the prediction error was calculated, as shown in Figure 9.

[0204] As can be seen from the CDF curves in Figure 9, the CARE-Net curve is steeper than that of WNOMP, with the main errors concentrated below 10dB. After introducing the RT physical prior, both CARE-Net and WNOMP curves show a significant leftward shift compared to recovery without prior knowledge, indicating a further decrease in the overall error level. In particular, the RT-guided CARE-Net curve is closest to the upper left corner, confirming that the introduction of physical prior information provides global navigation for the optimization problem, effectively avoiding abnormal bias samples caused by blind estimation algorithms without prior knowledge getting trapped in local optima, and ensuring the reliability of the overall channel reconstruction.

[0205] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for multipath channel parameter reconstruction and beam coverage prediction, characterized in that... Includes the following steps: 1) Configure antenna beam and orientation parameters and transmit reference signals: Configure the antenna array parameters of the transmitting end and construct a discrete Fourier transform beamforming codebook to control the transmitting end to transmit reference signals in the corresponding beam directions in sequence; 2) The receiving end collects reference signal received strength under different beam conditions: The user terminal receives reference signals from different beam directions and measures their reference signal received strength RSRP to obtain the macroscopic energy statistics characteristics of the location under a specific beam configuration; 3) Deterministic channel modeling and simulation generate channel multipath parameters: The electromagnetic propagation environment of the target area is simulated using deterministic ray tracing technology to obtain coarse-grained multipath distribution information, and a model reflecting the prior probability of channel multipath parameters is constructed to indicate the probability of paths existing in each angular direction; 4) Establish a mapping model between reference signal received strength statistics and channel multipath parameters: Based on the antenna pattern of the antenna array, discrete Fourier transform codebook and other configuration parameters, the theoretical power gain of each beam on each transmission angle grid is calculated, and a linear mapping relationship from the macroscopic low-dimensional measurement domain to the high-dimensional angular domain is established, based on which R... 5) Deterministic channel modeling simulation prior-guided sparse recovery of channel multipath parameters: Guided by the prior probability of channel multipath parameters, the sparse recovery algorithm is used to solve the underdetermined equations and invert the high-dimensional sparse angular power spectrum from the low-dimensional RSRP measurement vector; 6) Obtaining the actual environmental channel propagation model from the channel multipath parameters: Non-zero elements are extracted from the reconstructed sparse angular power spectrum to obtain the real environmental multipath parameters, which constitute the actual environmental channel propagation model independent of the antenna configuration, reflecting the inherent physical propagation properties of the environment; 7) Evaluating the environmental channel map under different antenna configurations: Based on the reconstructed actual environmental channel propagation model, the channel coverage performance of the base station antenna under different physical orientations or beam configurations is predicted and evaluated to verify the accuracy, physical authenticity and generalization ability of the channel model.

2. The multipath channel parameter reconstruction and beam coverage prediction method as described in claim 1, characterized in that... In step 1), a uniform planar array is used as the base station transmitter model, and the antenna array consists of... It consists of several array elements, among which The number of elements in the horizontal direction. The number of array elements in the vertical direction is set to the element spacing. The working center frequency is The corresponding free space wavenumber is The base station is simulated based on the Discrete Fourier Transform (DFT) codebook. Each beam, generated by adjusting the phase shifter, can cover a specific spatial area; No. beams ( The shaped vector of ) Shaped vectors of the horizontal dimension and the shaping vector in the vertical dimension Decide: in, For the shape vector of the horizontal dimension, The frequency is the spatial frequency in the horizontal dimension, j is the imaginary unit, and T represents the transpose operation. It is the base of the natural index; The shaping vector refers to the vertical dimension. Spatial frequency in the vertical dimension; It is the first The beamforming vector, This represents the total number of elements in the antenna array. ， It is the horizontal dimension shaping vector corresponding to the m-th beam. It is the vertical dimension shaping vector corresponding to the m-th beam. This represents the Kronecker product operation; at the transmitter, the continuous AoD spatial angular domain is transformed into a finite mesh; Among them, tilt angle With azimuth angle They are uniformly discretized in free space, and respectively obtained An angle of inclination and Azimuth angle, The vertical angle sampling interval, The horizontal angle sampling interval is π, where π is the mathematical constant pi, and the total number of angle grids is π. 。 3. The multipath channel parameter reconstruction and beam coverage prediction method as described in claim 1, characterized in that... In step 4), a linear mapping relationship is established from the macroscopic low-dimensional measurement domain to the high-dimensional angular domain: in, Number of beam directions This represents the total number of angle grids, which is also the candidate direction of the path; Characterized the first The DFT beam in the first Power response in each spatial angular direction, Indicates the first The path power along each discrete angular grid direction corresponds to the complex gain of that path in the physical channel impulse response. The square of the modulus, that is ； Introducing measured antenna pattern data, ,in, It is the antenna element in Total power gain in the direction, It is the antenna element in the direction Vertical polarization complex gain, It is the antenna element in the direction Horizontal polarization complex gain; It is the tilt angle corresponding to the nth angular direction; It is the azimuth angle corresponding to the nth angular direction; The modulus square operation represents the operation of complex numbers; the array beam gain is obtained by the combined effect of the codebook beam and the array response factor, where the array response factor is: in, The array response factor vector in the horizontal dimension, where j is the imaginary unit, π is pi, and λ is the signal wavelength. It is the spacing between antenna elements. It is the azimuth angle corresponding to the j-th beam. It is the tilt angle corresponding to the i-th beam. The number of elements in the horizontal direction. The number of elements in the vertical direction. It is the base of the natural exponent, and T represents the transpose of a vector; This is the array response factor vector along the vertical dimension. It is the total array response factor vector corresponding to the nth angular direction. It is the horizontal dimension array response factor vector corresponding to the nth angular direction. It is the vertical dimension array response factor vector corresponding to the nth angular direction. This represents the Kronecker product operation.

4. The multipath channel parameter reconstruction and beam coverage prediction method as described in claim 1, characterized in that... In step 4), the construction of the joint sensing dictionary matrix of RSRP measurements and sparse angular power spectrum APS is first achieved by transforming the channel parameter reconstruction into a sparse linear inverse problem through a linear mapping model between beam RSRP and APS. For RSRP measurement vectors, For the APS to be reconstructed, this vector is sparse. ,Right now The vast majority of elements are 0, only Each position is a non-zero positive value. ； For the joint perception dictionary matrix.

5. The multipath channel parameter reconstruction and beam coverage prediction method as described in claim 1, characterized in that... In step 5), the deterministic channel modeling simulation prior guides the sparse recovery of channel multipath parameters, and models the channel sparse parameter reconstruction process as a series of iteratively solved convex optimization subproblems, specifically including three core stages: preprocessing normalization, coherence-aware iterative weighted optimization, and non-negative least squares debiasing, to achieve high-precision parameter reconstruction under ill-conditioned measurement matrices. 5.1) In the preprocessing normalization stage, to address the problem of huge differences in the gain between the main lobe and side lobe of the antenna beam, a normalized measurement space is constructed to eliminate the search bias caused by the uneven gain of the antenna beam. For measurement vector and joint perception dictionary matrix Perform norm normalization to construct a standardized measurement model: in, For a normalized dictionary matrix, satisfying 5.2) In the iterative optimization phase, a dynamic closed loop of "probe-feedback-coherence suppression" is constructed; the regularization penalty weights are gradually modified during the iteration process; each iteration solves the following weighted non-negative elastic network subproblem to obtain the distribution of intermediate solutions under the current weights: in, yes Regularization coefficients are introduced to control sparsity. Regularization coefficients ensure that highly correlated adjacent beams acquire similar non-zero coefficients in the early stages of iteration, avoiding simplistic approaches. Regularization mitigates the risk of randomly discarding highly coherent paths while ensuring the strict convexity and uniqueness of the solution under underdetermined conditions. yes Regularization coefficients; an iterative weighting mechanism based on a dynamic coherence repulsion field is used to introduce coherence-aware repulsion weights into the weight construction. By combining sparsity promotion and coherence repulsion, strong paths are actively identified and their neighborhoods are penalized to suppress false sidelobes, achieving accurate single-path resolution in ultra-high resolution grids. Defined by the following formula: in, To prevent small constants with a denominator of zero (e.g.) ）， This is the coherence penalty coefficient. Let be the normalization factor, ensuring the weight mean is 1 to maintain the stability of the regularization strength; let For the first The set of strong paths identified by round iteration, if grid If an element does not belong to the set of strong paths, then the maximum cross-correlation between it and atoms in the set of strong paths is calculated as the repulsion value. ； If grid If the beam direction is similar to the currently discovered strong paths, then the penalty weight is increased, making... Increase in size, thus optimizing Suppression is set to zero to inhibit the generation of false splitting paths; 5.3) In the non-negative least squares debiasing stage, a debiasing operation is performed to recover the true physical power; using the non-zero support set obtained in the final iteration... Solving the nonnegative least squares problem based on the unnormalized original measurement model: Supporting sub-matrix Reference from the original perception dictionary In the middle, only the support set is extracted. The column-full rank submatrix reconstructed from the column vectors corresponding to each index; when ,in To determine the total number of identified paths, then: The dimension of this matrix is ,in To measure the number of beams, This represents the number of paths.

6. The multipath channel parameter reconstruction and beam coverage prediction method as described in claim 1, characterized in that... In step 5), guided by the prior probability of channel multipath parameters, a physical prior information-guided grid search is proposed. Coarse-grained multipath distribution information of the target scene is obtained through RT simulation, physical environment information is introduced, and physical prior information is used to guide channel reconstruction, guiding the algorithm to prioritize physically probable angle regions when searching for paths. Specifically: discrete RT simulation data... Transform it into a continuous probability weight vector covering all angle grids. ; the Prior probability weights of each angle grid Defined as: in, It is the first APS parameters for each angle grid. It is the first APS parameters for each angle grid. It is a distance metric. It is the tunable diffusion coefficient in the Gaussian kernel; in The regularization coefficients take into account prior probability weights. If the first... If a grid at a given angle has prior support, then the angle of the grid will decrease. Regularization coefficients penalize weights, increasing the probability of selection; weights The construction formula that combines sparsity and prior knowledge is as follows: The physical consistency and generalization ability of the channel multipath parameter reconstruction method will be evaluated through antenna array rotation experiments, so as to achieve cross-configuration beam coverage prediction based on environment-hardware decoupling.

7. The multipath channel parameter reconstruction and beam coverage prediction method as described in claim 1, characterized in that... In step 7), the accuracy, physical authenticity, and generalization ability of the verified channel model are assessed to achieve beam coverage prediction under unmeasured antenna orientation; wherein, the error formula for evaluating the generalization ability is: in, The total number of samples involved in the calculation; i is the index of the sample, ranging from 1 to n; This represents the true value corresponding to the i-th sample; Let be the predicted value corresponding to the i-th sample.

Images