A group-connected BD-RIS-assisted near-field communication location prediction method and system based on time-division switching acquisition and dual-channel temporal feature fusion

Abstract

This invention relates to a group-connected BD-RIS-assisted near-field communication (BNFC) location prediction method and system based on time-division switching acquisition and dual-channel temporal feature fusion, comprising the following steps: S1, dividing multiple reconfigurable units of the group-connected BD-RIS-assisted BNFC system into several reconfigurable unit groups; S2, sequentially switching the reconfigurable unit groups, acquiring channel data, and assembling a high-dimensional feature sample set from the multi-dimensional features of the channel data; S3, determining the positioning contribution weight, and performing weighted processing according to the positioning contribution weight to obtain a weighted high-dimensional feature sample set; S4, inputting the weighted high-dimensional feature sample set into a parallel-processed BiGRU neural network and a BiTCN neural network to obtain fused features; S5, training a fully connected layer with the fused features and the real coordinates of the mobile user terminal to obtain a mapping relationship; S6, executing steps S2-S4 to acquire the current fused features of the mobile user terminal, and converting the current fused features into the current coordinates through the mapping relationship.

Description

Technical Field

[0001] This invention relates to the field of location prediction, specifically to a group connection BD-RIS-assisted near-field communication location prediction method and system based on time-division switching acquisition and dual-channel time-series feature fusion. Background Technology

[0002] Reconfigurable Intelligent Surface (RIS) is an electromagnetic environment control technology. RIS can significantly improve network performance by redefining the propagation paradigm of wireless systems and constructing an intelligent and controllable wireless propagation environment through joint optimization of transmitters, receivers, and the propagation environment.

[0003] In the field of wireless positioning, Reflection Arrays (RIS) can be flexibly deployed within the service area of ​​a base station to assist in positioning and improve positioning accuracy. However, as the number of RIS reflector units increases and the positioning range expands, the data dimensionality and computational complexity also gradually increase. Ordinary RIS-assisted positioning algorithms can no longer meet the demands of high-dimensionality and high-intensity computation. Traditional RIS typically employs a diagonal phase-shift matrix architecture, where each reflector unit can only independently control the phase of the signal, failing to achieve mutual coupling between reflector units. Limited by the single-connection topology and diagonal scattering matrix, the channel dataset generated by traditional RIS only contains the phase modulation response applied independently by each element, failing to characterize the mutual coupling effects, power allocation, and joint beamforming behavior between elements. This results in a low-rank channel matrix and a feature space limited to parameters in a single angular domain. This lack of flexibility restricts the RIS's control capabilities and positioning accuracy in complex scenarios.

[0004] The purpose of this invention is to design a group-connected BD-RIS-assisted near-field communication position prediction method and system based on time-division switching acquisition and dual-channel timing feature fusion to address the problems existing in the prior art. Summary of the Invention

[0005] To address the problems existing in the prior art, the present invention provides a group connection BD-RIS-assisted near-field communication position prediction method and system based on time-division switching acquisition and dual-channel timing feature fusion, which can effectively solve at least one of the problems existing in the prior art.

[0006] The technical solution of this invention is: A group-connected BD-RIS-assisted near-field communication location prediction method based on time-division switching acquisition and dual-channel temporal feature fusion includes the following steps: S1, communication between the mobile user terminal and the base station is realized by grouping the BD-RIS assisted near-field communication system, and the multiple reconfigurable units of the grouped BD-RIS assisted near-field communication system are divided into several reconfigurable unit groups; S2, during the communication process between the mobile user terminal and the base station, the group connection BD-RIS assisted near-field communication system is controlled to switch the reconfigurable unit group sequentially according to a preset time interval, and channel data under the action of the reconfigurable unit group within the time window length is collected, and the multi-dimensional features of the channel data are combined into a high-dimensional feature sample set. S3, determine the positioning contribution weight based on the correlation between the high-dimensional feature sample set of the channel data corresponding to each reconfigurable unit group and the real coordinates of the mobile user terminal, and perform weighted processing on multiple high-dimensional feature sample sets according to the positioning contribution weight to obtain a weighted high-dimensional feature sample set. S4, the weighted high-dimensional feature sample set is input into the parallel processing BiGRU neural network and BiTCN neural network. The BiGRU neural network captures the long-term temporal dependency features of the weighted high-dimensional feature sample set, and the BiTCN neural network captures the local temporal features of the weighted high-dimensional feature sample set. The long-term temporal dependency features and the local temporal features are fused to obtain the fused features. S5, train a fully connected layer with the fused features and the real coordinates of the mobile user terminal to obtain the mapping relationship between the fused features and the real coordinates of the mobile user terminal; S6, execute steps S2~S4 to collect the current fusion features of the mobile user terminal, and the current fusion features are transformed into the current coordinates through the mapping relationship.

[0007] Further, in step S1, dividing the multiple reconfigurable units of the BD-RIS assisted near-field communication system into several reconfigurable unit groups includes: The multiple reconfigurable units of the BD-RIS assisted near-field communication system are divided into several reconfigurable unit groups according to their spatial location or electromagnetic coupling relationship. A sub-scattering matrix is ​​constructed to characterize the phase modulation, amplitude modulation, block scattering gain, and intra-group coupling response of the reconfigurable unit group to the incident electromagnetic wave. The sub-scattering matrix satisfies unitary and symmetry constraints, and the sub-scattering matrix arranged diagonally can form the overall scattering matrix of the group connected to the BD-RIS assisted near-field communication system.

[0008] Further, after step S1, execute: The phase and amplitude offset of each of the reconfigurable unit groups are calibrated by sending known pilot signals, and the clock synchronization of the base station and the mobile user terminal is calibrated to obtain the calibrated sub-scattering matrix.

[0009] Further, in step S2, channel data within the time window length is collected under the action of the reconfigurable unit group, and the multi-dimensional features of the channel data are used to form a high-dimensional feature sample set, including: Load the sub-scattering matrix corresponding to the reconfigurable unit group, control the reconfigurable unit group to enter the scattering state corresponding to the sub-scattering matrix, and collect channel data of a preset time window length; After denoising and smoothing the channel data, multi-dimensional features of the channel data are extracted. The multi-dimensional features include at least one of channel state information, time domain features, and angle domain features. The multi-dimensional features are combined to form a high-dimensional feature sample.

[0010] Further, in step S3, determining the positioning contribution weight based on the correlation between the high-dimensional feature sample set of the channel data corresponding to each of the reconfigurable unit groups and the real coordinates of the mobile user terminal includes: The high-dimensional feature sample sets of each reconfigurable unit group under the real coordinates of each mobile user terminal are combined into a multi-location sample set. The intra-group compactness of each multi-location sample set and the inter-group separation degree between different multi-location sample sets are calculated. The ratio of the inter-group separation degree to the intra-group compactness is normalized and used as the positioning contribution weight of the reconfigurable unit group.

[0011] Further, in step S4, the long-term temporal dependency features and the local temporal features are fused to obtain fused features, including: The long-term temporal dependency feature and the local temporal feature are concatenated along the feature dimension to obtain the concatenated feature; The concatenated features are input into a multi-head attention mechanism to generate a query matrix, a key matrix, and a value matrix, which are then divided into multiple attention heads. The attention weights and attention outputs of each attention head are calculated, and the outputs of multiple attention heads are concatenated and projected to obtain the attention enhancement features. The final output of the attention enhancement feature is taken as the fusion feature.

[0012] Further, step S5 includes: Set a fully connected layer As a coordinate regression layer, in For two-dimensional coordinate estimation, As a feature of fusion, and The weights and biases of the fully connected layer are used as the basis for training the fully connected layer with the fused features and the real coordinates of the mobile user terminal. and The optimal value is determined to construct the mapping relationship between the fused feature and the real coordinates of the mobile user terminal.

[0013] Further, after step S2, the following steps are performed: dimensionality reduction of the high-dimensional feature sample set is performed using a compressed sensing algorithm.

[0014] Furthermore, a group-connected BD-RIS-assisted near-field communication position prediction system based on time-division switching acquisition and dual-channel temporal feature fusion is provided. The system is characterized in that it implements the aforementioned group-connected BD-RIS-assisted near-field communication position prediction method based on time-division switching acquisition and dual-channel temporal feature fusion during operation.

[0015] Therefore, the present invention provides the following effects and / or advantages: This application divides a group-connected BD-RIS assisted near-field communication system into multiple groups. By switching between different reconfigurable unit groups, the contribution weights of different reconfigurable unit groups are obtained and weighted summed to obtain sampling data that better reflects user positioning information. Subsequently, a BiGRU-BiTCN hybrid network structure is used to collaboratively extract long-term dependencies and local spatiotemporal features in the channel, and an attention mechanism is introduced to achieve adaptive feature fusion, thereby outputting high-precision position prediction results. This research method can uncover the intrinsic correlation between channel features and mobile user positions in complex millimeter-wave propagation environments, and overcome positioning errors caused by multipath fading, noise interference, and target dynamics in near-field environments, achieving high-precision, low-complexity real-time positioning of mobile users.

[0016] This application divides the multiple reconfigurable units of the group-connected BD-RIS assisted near-field communication system into several reconfigurable unit groups and constructs a sub-scattering matrix to characterize the scattering features. At the same time, the sub-scattering matrix is ​​constrained to satisfy the overall characteristics of the group-connected BD-RIS assisted near-field communication system. Then, by controlling the switching of different reconfigurable unit groups, corresponding data is obtained. When the user moves, the changes in the strength of the received signals of each group will form a correspondence between the spatial position and the response mode of each group. On the one hand, it can reduce the data complexity of the entire group-connected BD-RIS assisted near-field communication system. On the other hand, by weighting and summing the different contributions of each group to the movement position, data that is more suitable for expressing the user's position can be obtained.

[0017] This application uses a perceptual compression method to reduce the dimensionality of the data, which can reduce data complexity and dimensionality.

[0018] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0019] It should be understood that the above summary and the following detailed description of the invention are exemplary and explanatory, and are intended to provide further explanation of the invention as claimed. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating one embodiment of the present invention.

[0021] Figure 2 A schematic diagram of a near-field communication system architecture assisted by BD-RIS for group connection.

[0022] Figure 3 This is a schematic diagram of the framework of this method.

[0023] Figure 4 A schematic diagram illustrating the contribution weights of each high-dimensional feature sample set.

[0024] Figure 5 This diagram illustrates the convergence of the loss function during the training process for different models.

[0025] Figure 6 This diagram illustrates the cumulative distribution function of different models in the mobile user location prediction task.

[0026] Figure 7 This is a schematic diagram illustrating the positioning error statistics for each model.

[0027] Figure 8 This diagram illustrates the impact of the BD-RIS group connection architecture on the near-field localization performance of the CS-BiGT model for mobile users. Detailed Implementation

[0028] To facilitate understanding by those skilled in the art, the present invention will now be described in further detail with reference to the embodiments: refer to Figure 1 , 3 A group-connected BD-RIS-assisted near-field communication location prediction method based on time-division switching acquisition and dual-channel temporal feature fusion includes the following steps: S1, communication between the mobile user terminal and the base station is realized by grouping the BD-RIS assisted near-field communication system, and the multiple reconfigurable units of the grouped BD-RIS assisted near-field communication system are divided into several reconfigurable unit groups; S2, during the communication process between the mobile user terminal and the base station, the group connection BD-RIS assisted near-field communication system is controlled to switch the reconfigurable unit group sequentially according to a preset time interval, and channel data under the action of the reconfigurable unit group within the time window length is collected, and the multi-dimensional features of the channel data are combined into a high-dimensional feature sample set. S3, determine the positioning contribution weight based on the correlation between the high-dimensional feature sample set of the channel data corresponding to each reconfigurable unit group and the real coordinates of the mobile user terminal, and perform weighted processing on multiple high-dimensional feature sample sets according to the positioning contribution weight to obtain a weighted high-dimensional feature sample set. S4, the weighted high-dimensional feature sample set is input into the parallel processing BiGRU neural network and BiTCN neural network. The BiGRU neural network captures the long-term temporal dependency features of the weighted high-dimensional feature sample set, and the BiTCN neural network captures the local temporal features of the weighted high-dimensional feature sample set. The long-term temporal dependency features and the local temporal features are fused to obtain the fused features. S5, train a fully connected layer with the fused features and the real coordinates of the mobile user terminal to obtain the mapping relationship between the fused features and the real coordinates of the mobile user terminal; S6, execute steps S2~S4 to collect the current fusion features of the mobile user terminal, and the current fusion features are transformed into the current coordinates through the mapping relationship.

[0029] In this embodiment, the BD-RIS-assisted near-field communication system is first divided into multiple groups. By switching between different reconfigurable unit groups, the contribution weights of different reconfigurable unit groups are obtained and weighted summed to obtain sampling data that better reflects user positioning information. Subsequently, a BiGRU-BiTCN hybrid network structure is used to collaboratively extract long-term dependencies and local spatiotemporal features in the channel, and an attention mechanism is introduced to achieve adaptive feature fusion, thereby outputting high-precision location prediction results. This research method can uncover the intrinsic correlation between channel features and mobile user location in complex millimeter-wave propagation environments and overcome positioning errors caused by multipath fading, noise interference, and target dynamics in near-field environments, achieving high-precision, low-complexity real-time positioning of mobile users.

[0030] Next, we will explain in detail how each step is implemented.

[0031] S1 enables communication between mobile user terminals and base stations through a group-connected BD-RIS assisted near-field communication system.

[0032] In this step, a group-connected BD-RIS-assisted near-field communication system is constructed for mobile user location positioning, and its architecture is as follows: Figure 2As shown, the system comprises a single-antenna base station (BS), a BD-RIS with a group-connected architecture, a single-antenna fixed user (UF), and a single-antenna mobile user (UM). The system operates in the millimeter-wave band with a carrier frequency of 28 GHz and employs time-division duplex (TDD) mode. Due to building obstructions, the BS and user cannot communicate directly via a line-of-sight (LOS) link; communication between them relies on the reflection path established by the BD-RIS.

[0033] Further, in step S1, dividing the multiple reconfigurable units of the BD-RIS assisted near-field communication system into several reconfigurable unit groups includes: The multiple reconfigurable units of the BD-RIS assisted near-field communication system are divided into several reconfigurable unit groups according to their spatial location or electromagnetic coupling relationship. A sub-scattering matrix is ​​constructed to characterize the phase modulation, amplitude modulation, and intra-group coupling response of the reconfigurable unit group to the incident electromagnetic wave. The sub-scattering matrix satisfies unitary and symmetry constraints, and the sub-scattering matrix arranged diagonally can form the overall scattering matrix of the group connected to the BD-RIS assisted near-field communication system.

[0034] Specifically, the traditional diagonal RIS phase shift matrix is ​​usually expressed as: ,in For traditional diagonal reconfigurable smart surfaces, the phase shift matrix, For diagonalization operators, The phase rotation that occurs after the signal is reflected by the RIS unit. j imaginary unit , For phase angle, M This is the Mth RIS reflector unit. Each reflector unit only supports independent phase modulation, limiting its modulation freedom. In contrast, the group-connected BD-RIS of this embodiment overcomes this limitation by introducing an inter-unit coupling mechanism, achieving more flexible beamforming capabilities. Because in the group-connected BD-RIS, there is no cross-group electromagnetic coupling or joint modulation connection between different reconfigurable unit groups; only intra-group units couple and coordinate, so the overall scattering response can only form coupling relationships within the intra-group block, mathematically manifesting as a block diagonal structure. Its overall scattering matrix... It presents a block diagonal structure, specifically in the following form: ; in Let be the scattering matrix of the entire BD-RIS, with dimensions of . M × M , M Let be the number of rows and columns of the matrix. Let be the sub-scattering matrix of the m-th group, with dimensions Ng×Ng, where Ng is the number of rows and columns of the matrix. This is a block diagonalization operation. Each sub-scattering matrix Strict physical constraints must be followed. While ensuring no energy loss during the reflection process, the principle of electromagnetic reciprocity must also be satisfied, meaning that the matrix must simultaneously possess unitary and symmetric properties.

[0035] , ; in, for The conjugate transpose matrix, for transpose matrix, It is a dimension of N × N The identity matrix. This indicates that the sub-scattering matrix is ​​unitary. This indicates that the sub-scattering matrix has symmetry.

[0036] In addition, the modulus of each element in the submatrix must satisfy the following conditions: , For the first m The first group of scattering matrices i Line number j The magnitudes of the column elements indicate that this structure primarily achieves beamforming by modulating the coupling relationship between the signal phase and the propagation path, rather than simply amplifying the signal. Simultaneously, this enables coordinated control of the units within the group.

[0037] Therefore, we can first generate a fundamental matrix that satisfies a random phase distribution, then use unitary transformation to ensure the conservation of source energy, and then use symmetry transformation to make the matrix satisfy the electromagnetic reciprocity constraint, thus obtaining a sub-scattering matrix that simultaneously possesses unitarity and symmetry. Finally, the... m The sub-scattering matrix is ​​expressed as: ; in The randomly generated unitary matrix is ​​obtained by projecting it through singular value decomposition (SVD). for transpose matrix, N The number of reflection units in the m-th group, thus ensuring It simultaneously satisfies both unitary and symmetric properties.

[0038] The sub-scattering matrix represents information such as phase modulation, amplitude modulation, intra-group coupling response, and block scattering gain of each unit in the group for the incident electromagnetic wave of the signal.

[0039] Further, after step S1, execute: The phase and amplitude offset of each of the reconfigurable unit groups are calibrated by sending known pilot signals, and the clock synchronization of the base station and the mobile user terminal is calibrated to obtain the calibrated sub-scattering matrix.

[0040] In this embodiment, the phase and amplitude offset of each element of BD-RIS is calibrated by sending a known pilot signal to generate an accurate sub-scattering matrix. Then, the clocks of the base station and the user terminal are calibrated by wired synchronization to ensure that the TOF and TDoA measurement errors are less than 1ns.

[0041] S2, the mobile user terminal moves along a predetermined route at a constant speed during the process of communicating with the base station; Load the sub-scattering matrix corresponding to the reconfigurable unit group, control the reconfigurable unit group to enter the scattering state corresponding to the sub-scattering matrix, and collect channel data of a preset time window length; After denoising and smoothing the channel data, multi-dimensional features of the channel data are extracted. The multi-dimensional features include at least one of channel state information, time domain features, and angle domain features. The multi-dimensional features are combined to form a high-dimensional feature sample.

[0042] In this embodiment, the mobile user terminal (UM) is controlled to move at a speed of 1 m / s according to a preset trajectory. Data acquisition is triggered once every 1 second, and each sampling point is paused for 100 milliseconds to ensure the integrity of data acquisition. At each sampling point, the sub-scattering matrix corresponding to different groups of reconfigurable units is switched sequentially. Each group corresponds to a different combination of phase and amplitude. At this time, each sampling time window corresponds to a clear BD-RIS block scattering state, turning the sub-scattering matrix from a simple structural parameter into an active acquisition control quantity, and enhancing the correlation between the sub-scattering matrix and the positioning features.

[0043] Base station signals reach mobile user terminals after passing through near-field channels and BD-RIS modulation. The flexible modulation capabilities of BD-RIS and the rich multipath characteristics of the near-field channel provide a high-dimensional and highly discriminative feature information foundation for mobile user positioning. The positioning system performs synchronous processing and multi-dimensional feature extraction on the received signals to obtain a set of channel features covering multiple dimensions such as signal amplitude, phase, time, and angle. These features comprehensively reflect the complex mapping relationship between the mobile user's location and channel propagation characteristics, thus providing sufficient discriminative basis for high-precision positioning models.

[0044] Channel state information characteristics The expression is as follows: ; in The mean of the amplitude. For amplitude variance, For amplitude kurtosis, The average phase value, For phase variance, The amplitude-phase cross-covariance.

[0045] Time domain features The expression is as follows: ; in For the arrival time of UM (ToA). This represents the time difference of arrival (TDoA) between UM and a certain reference location. The rate of change of time delay, This is the variance of the time delay estimate.

[0046] Angular domain features The array steering vector estimation based on BD-RIS, including the arrival angle and departure angle, can be specifically expressed as: ; In the formula and These represent the azimuth and elevation angles, respectively. and Its change.

[0047] Finally, a kernel function is used to map the above multi-domain features to a high-dimensional space to achieve nonlinear fusion, constructing high-dimensional feature samples for localization: ; In the formula The feature weight matrix, For the feature interaction matrix, For vectorized operations, This is the bias vector. The fusion of multidimensional feature data can effectively utilize the complementarity and structural information between features, suppressing estimation errors caused by multipath fading and noise, thereby significantly improving positioning accuracy and system robustness in complex environments. Simultaneously, it preserves the dynamic changes of mobile users during communication, providing a data foundation for subsequent BiGRU to capture long-term temporal dependencies and BiTCN to capture local temporal changes.

[0048] S3, determining the positioning contribution weight based on the correlation between the high-dimensional feature sample set of the channel data corresponding to each of the reconfigurable unit groups and the real coordinates of the mobile user terminal, including: The high-dimensional feature sample sets of each reconfigurable unit group under the real coordinates of each mobile user terminal are combined into a multi-location sample set. The intra-group compactness of each multi-location sample set and the inter-group separation degree between different multi-location sample sets are calculated. The ratio of the inter-group separation degree to the intra-group compactness is normalized and used as the positioning contribution weight of the reconfigurable unit group.

[0049] In this embodiment, the high-dimensional feature sample sets from multiple locations within the same reconfigurable unit group are combined to form a multi-location sample set. Then, the covariance matrix of these samples in the feature space is calculated, and its trace or determinant is taken as the sample dispersion at that location. The average dispersion of all location points is then calculated to obtain the intra-group compactness index of the corresponding sample set. The smaller the index value, the better the channel feature reproducibility at the same location, and the more stable the positioning fingerprint. Next, the class centers of different multi-location sample sets in the feature space are calculated, and the Euclidean distance between any two class centers is calculated. The average or minimum value is taken as the inter-group separation index. The larger the index value, the more significant the channel feature differences between different locations, and the higher the positioning resolution. Finally, the ratio of the inter-group separation to the intra-group compactness is normalized. The higher the ratio, the more beneficial the data collected under the BD-RIS configuration is for distinguishing different locations, and the higher the weight of its contribution to the positioning task.

[0050] Next, the multiple high-dimensional feature sample sets are weighted according to the positioning contribution weight to obtain a weighted high-dimensional feature sample set; The above steps yielded the contribution weights of each location, which are then weighted and summed.

[0051] refer to Figure 4 , Figure 4 The horizontal axis represents the name of each sample, and the contribution weight of each high-dimensional feature sample set was calculated. By effectively quantifying the feature contribution, the intrinsic relationship between angle, phase features, and location was revealed, providing a reliable basis for feature selection and model optimization. The results show that there is a strong nonlinear mapping relationship between the spatial geometric information between mobile users and BD-RIS and the user's location, thus becoming the core clue for deep learning models to capture localization features. However, BS-related features are greatly affected by multipath fading and noise, and the CS layer has already used reconstruction loss to select and fuse features, resulting in a relative decrease in their original importance in the input layer.

[0052] S4, the weighted high-dimensional feature sample set is input into the parallel-processed BiGRU neural network and BiTCN neural network. The BiGRU neural network captures the long-term temporal dependence features of the weighted high-dimensional feature sample set, and the BiTCN neural network captures the local temporal features of the weighted high-dimensional feature sample set. In this step, the BiGRU-BiTCN hybrid prediction model constitutes the core module of the positioning system. For a weighted high-dimensional feature sample set, this model constructs a hybrid neural network architecture that parallels BiGRU and BiTCN, aiming to collaboratively capture long-term contextual dependencies and local multi-scale feature patterns in the motion trajectory of mobile users. BiGRU excels at modeling long-term memory and bidirectional dependency features of sequence data, while BiTCN effectively extracts local temporal features under different receptive fields through dilated convolution kernels. To further enhance feature representation capabilities, the model introduces an attention enhancement mechanism, dynamically weighting and fusing the outputs of the two branches to highlight the contributions of key temporal segments and important local features, ultimately minimizing positioning error. This hybrid architecture fully leverages the complementary advantages of the two networks, significantly improving the adaptability to near-field channel nonlinearity and time-varying characteristics, thereby enhancing positioning accuracy and system robustness.

[0053] A Bidirectional Gated Recurrent Unit (BiGRU) consists of a forward GRU and a backward GRU. It can simultaneously utilize historical and future information to encode the context information of the current moment, and adaptively adjust the forgetting and updating process of feature information through a gating mechanism. Given the weighted high-dimensional feature sample set... The calculation process of the forward GRU at time step t is as follows: ; in, To enable forward reset, control the degree to which historical information is retained; As a forward update gate, it determines the fusion ratio between historical and current information; It serves as a candidate memory unit, storing candidate feature information at the current moment; For forward hidden state updates, historical hidden states and candidate memory units are merged. for t Input characteristics at time step This is the previous forward-hidden state. The weight matrix is ​​fed forward into the reset gate, update gate, and candidate states. This is the weight matrix from the forward hidden state to the reset gate, update gate, and candidate state. Let be the bias vectors for the forward reset gate, update gate, and candidate states. It is the sigmoid activation function. The hyperbolic tangent activation function is used. This is the Hadamard product. The backward GRU uses a similar structure, reversing the input sequence to obtain the backward hidden state. .

[0054] The hidden state sequence of the forward GRU is The hidden state sequence of the backward GRU is as follows: , L Given the sequence length, the hidden state after bidirectional fusion is: ; in The bidirectional fusion output of the GRU at time t. This is a two-way fusion weight matrix. For the hidden layer dimension, These are the forward and backward hidden states, respectively. This is the fusion bias vector.

[0055] BiTCN consists of a forward TCN and a backward TCN. Through dilated convolutions and residual connections, it expands the receptive field without significantly increasing computational complexity, capturing the local multi-scale correlation characteristics of features. Its core module is the temporal convolutional block (TCB). Each TCB contains two layers of dilated convolutions, Chomp1d pruning, a ReLU activation function, and a Dropout layer. For the input feature sequence... Forward TCN in the first l The output of the layer-time convolutional block is: ; in, This represents the result of the first forward convolution clipping layer. This represents the result of the second forward convolution clipping layer; This represents the output of the first convolutional layer after activation and Dropout. This represents the output of the second convolutional layer after activation and Dropout. This is used as the output of the forward residual connect block to alleviate the gradient vanishing problem. The weights of the convolution kernel, The coefficient of thermal expansion is 1 / 3. For bias, For dropout probability, This is the dimension matching function. Similarly, the backward TCN sequence is reversed to obtain the backward residual connection output. .

[0056] The output sequence of the forward TCN is denoted as The output sequence of the backward TCN is denoted as The TCN output after bidirectional fusion is: ; in, for t Real-time TCN bidirectional fusion output, This is a two-way fusion weight matrix. To output the dimension of the convolutional network, These are the residual connection outputs for the forward and backward directions, respectively. This is the fusion bias vector.

[0057] In this step, the long-term temporal dependency features and the local temporal features are concatenated along the feature dimension to obtain concatenated features; the concatenated features are input into a multi-head attention mechanism to generate query matrix, key matrix, and value matrix, and are divided into multiple attention heads; attention weights and attention outputs are calculated through each attention head, and the outputs of multiple attention heads are concatenated and projected to obtain attention-enhanced features; the final output of the attention-enhanced features is taken as the fusion feature.

[0058] To effectively integrate the output features of BiGRU and BiTCN branches and enhance the representation capability of key temporal information, this study introduces a multi-head attention (MHA) mechanism. First, the outputs of the two branches are concatenated along the feature dimension. , This represents the hybrid features resulting from the splicing of BiGRU and BiTCN. T For time steps, D BiGRU and D BiTCN The two output dimensions constitute the input to the attention mechanism. Therefore, the calculation process for the attention weights is as follows: ; in, These represent the query, key, and value matrices, respectively. For the projection weight matrix of query, key, and value, For the attention head dimension, H To focus on the number of heads, Features resulting from multi-head attention fusion This is the output weight matrix.

[0059] In this step, long-term temporal dependency features mainly reflect the overall trend of mobile user location changes. For example, when a user moves from one area to another, features such as CSI, ToA, AoA, and AoD do not change completely randomly, but rather exhibit continuous trajectory correlations. BiGRU processes sequences in both forward and reverse directions, simultaneously utilizing historical and subsequent information, making it more suitable for extracting such trajectory-level long-term dependencies. Local temporal features mainly reflect local changes within adjacent time windows. For example, in near-field millimeter-wave scenarios, even small changes in a mobile user's location can lead to significant local perturbations in phase, amplitude, multipath components, and angle features. BiTCN, through dilated convolution and residual connections, can extract local multi-scale temporal features across different receptive fields, making it suitable for capturing these short-term channel mutations and detailed patterns. Fusing these two features in parallel enhances the adaptability of the localization model to near-field channel nonlinearity, time-varying characteristics, and dynamic movement features.

[0060] After concatenating the outputs of BiGRU and BiTCN, although two types of features are obtained, not all time segments or feature channels are equally important. For example, in mobile user positioning, the channel features of some time windows are more stable and better reflect the true location; while other time windows may be affected by noise, multipath fading, or transient occlusion, resulting in lower positioning value. The role of the attention mechanism is to allow the model to automatically determine which moments are more critical, which features are more important, and which BiGRU / BiTCN outputs should be assigned higher weights. Therefore, the multi-head attention mechanism is not a simple average fusion, but a dynamic weighted fusion. Different attention heads can focus on features from different perspectives: some focus on long-term trajectory changes, some on local abrupt changes, and some on specific channel dimensions or key time windows.

[0061] S5, set a fully connected layer As a coordinate regression layer, For two-dimensional coordinate estimation, For the transpose of the fusion feature, and The weights and biases of the fully connected layer are used as the basis for training the fully connected layer with the fused features and the real coordinates of the mobile user terminal. and The optimal value is determined to construct the mapping relationship between the fused feature and the real coordinates of the mobile user terminal.

[0062] Specifically, the model ultimately maps the aggregated features to a two-dimensional coordinate estimate of the mobile user through a fully connected layer: ; At this time, it was obtained As a mapping relationship, in subsequent steps, only the fused features need to be input to obtain the corresponding two-dimensional coordinate estimate.

[0063] Further, after step S2, the following step is performed: dimensionality reduction of the high-dimensional feature sample set using a compressed sensing algorithm. This includes: Based on the original high-dimensional feature matrix extracted from the mobile user's received signal, construct the measurement matrix and the sparse transformation matrix; Based on the measurement matrix and the sparse transformation matrix, the original high-dimensional feature matrix is ​​subjected to structured sampling to obtain a compressed low-dimensional observation feature matrix; The low-dimensional observation feature matrix is ​​iteratively reconstructed to obtain a low-dimensional feature sample set for input into the BiGRU-BiTCN hybrid prediction model.

[0064] Specifically, in BD-RIS-assisted wireless communication systems, the original channel features have high dimensionality. Directly inputting them into the prediction model will lead to a sharp increase in computational complexity and easily cause overfitting. To address this issue, this study employs compressed sensing (CS) as the front-end feature preprocessing module, aiming to reduce dimensionality while preserving the nonlinear mapping relationship between mobile user location and channel features through structured sampling.

[0065] CS compresses high-dimensional original features into a low-dimensional observation space by designing a suitable measurement matrix, and then uses a reconstruction algorithm to recover key information. This method not only significantly reduces the computational burden on subsequent neural networks, but also enhances the discriminative power of features through sparsity constraints, thus laying the foundation for achieving accurate localization.

[0066] Let the high-dimensional feature sample set be The compressed high-dimensional feature sample set is The compression process can then be represented as: ; in It is a sparse transformation matrix. For additive measurement noise, For noise power, Let be the identity matrix. To satisfy the constraint isometry (RIP) and enhance feature adaptability, the feature reconstruction problem can be formulated as the following optimization problem: ; C1 ensures that the error between the reconstructed features and the original features is within an acceptable range; C2 ensures that the transformation matrix does not excessively distort the signal energy; C3 ensures that each row of the measurement matrix is ​​normalized; and C4 ensures that the columns of the transformation matrix are orthogonal to enhance compression stability.

[0067] The reconstruction process employs an Iterative Shrink Thresholding (ISTA) algorithm, which optimizes the reconstruction process by incorporating prior information from the near-field channel. The iterative formula is as follows: ;

[0068] in It is a soft thresholding function. , The iteration threshold, As the initial threshold, As the attenuation factor, This represents the number of iterations.

[0069] Furthermore, a group-connected BD-RIS-assisted near-field communication position prediction system based on time-division switching acquisition and dual-channel temporal feature fusion is provided. During operation, it implements the aforementioned group-connected BD-RIS-assisted near-field communication position prediction method based on time-division switching acquisition and dual-channel temporal feature fusion.

[0070] Experimental data refer to Figure 5 The figure illustrates the convergence of the loss function of different models during training. Overall, the loss values ​​of all models decrease with increasing training epochs, but there are significant differences in convergence speed and final stability among the models. Among them, the loss curve of the model in this application (CS-BiGT) converges the fastest and eventually stabilizes at the lowest level. This performance advantage of this application mainly stems from the fact that the CS layer of this model acts as a pre-processing feature module, performing dimensionality reduction and noise filtering on the high-dimensional original data in advance, significantly reducing the learning complexity of the subsequent bidirectional temporal network, allowing the model to quickly capture core features strongly correlated with location.

[0071] refer to Figure 6 The cumulative distribution function (CDF) and 95% confidence intervals of different models in the mobile user location prediction task are compared, as shown in the figure. The results show that the cumulative probability of each model increases with the location error threshold. The CS-BiGT model shows the most significant increase and the narrowest confidence interval. This is attributed to the fact that this model integrates a hybrid architecture of CS and BiGRU-BiTCN, flexibly coordinating feature purification and multimodal temporal modeling to adapt to the dynamic changes in channel characteristics in positioning scenarios, thus achieving more stable generalization.

[0072] refer to Figure 7The figure further illustrates the positioning error statistics for each model. The error distribution shows a significant hierarchical difference in positioning accuracy among the different models. The CS-BiGT model has a median error of only 5 cm, with the most compact upper and lower quartile intervals, and the maximum error is controlled within 10 cm. It significantly outperforms other comparative models in both prediction accuracy and stability.

[0073] refer to Figure 8 The figure illustrates the impact of the BD-RIS group connection architecture on the near-field positioning performance of the CS-BiGT model for mobile users. The results show that as the number of reflective elements on the group-connected BD-RIS assisted near-field communication system increases, the number of elements in each group of BD-RIS also increases. The average positioning error of all group connection architectures and frequency bands continuously decreases, and the standard deviation of the error narrows synchronously, while the positioning accuracy at the 10cm threshold increases synchronously. The G=4 group connection architecture maintains the best performance across all element count ranges. When the number of elements reaches 64, the average positioning error is as low as 4.74cm, and the accuracy at the 10cm threshold reaches 99.15%. This performance advantage is mainly attributed to the G=4 group connection architecture's ability to effectively suppress reflected beam coupling interference and reduce channel estimation errors through independent phase modulation and collaborative optimization of subarrays, thus achieving the optimal balance between positioning accuracy and robustness in dynamic mobile user scenarios.

[0074] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0075] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0076] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0077] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0078] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms should not be construed as necessarily referring to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

Claims

1. A group-connected BD-RIS-assisted near-field communication location prediction method based on time-division switching acquisition and dual-channel temporal feature fusion, characterized in that: Includes the following steps: S1, communication between the mobile user terminal and the base station is realized by grouping the BD-RIS assisted near-field communication system, and the multiple reconfigurable units of the grouped BD-RIS assisted near-field communication system are divided into several reconfigurable unit groups; S2, during the communication process between the mobile user terminal and the base station, the group connection BD-RIS assisted near-field communication system is controlled to switch the reconfigurable unit group sequentially according to a preset time interval, and channel data under the action of the reconfigurable unit group within the time window length is collected, and the multi-dimensional features of the channel data are combined into a high-dimensional feature sample set. S3, determine the positioning contribution weight based on the correlation between the high-dimensional feature sample set of the channel data corresponding to each reconfigurable unit group and the real coordinates of the mobile user terminal, and perform weighted processing on multiple high-dimensional feature sample sets according to the positioning contribution weight to obtain a weighted high-dimensional feature sample set. S4, the weighted high-dimensional feature sample set is input into the parallel processing BiGRU neural network and BiTCN neural network. The BiGRU neural network captures the long-term temporal dependency features of the weighted high-dimensional feature sample set, and the BiTCN neural network captures the local temporal features of the weighted high-dimensional feature sample set. The long-term temporal dependency features and the local temporal features are fused to obtain the fused features. S5, train a fully connected layer with the fused features and the real coordinates of the mobile user terminal to obtain the mapping relationship between the fused features and the real coordinates of the mobile user terminal; S6, execute steps S2~S4 to collect the current fusion features of the mobile user terminal, and the current fusion features are transformed into the current coordinates through the mapping relationship.

2. The group-connected BD-RIS-assisted near-field communication position prediction method based on time-division switching acquisition and dual-channel temporal feature fusion as described in claim 1, characterized in that: In step S1, dividing the multiple reconfigurable units of the BD-RIS assisted near-field communication system into several reconfigurable unit groups includes: The multiple reconfigurable units of the BD-RIS assisted near-field communication system are divided into several reconfigurable unit groups according to their spatial location or electromagnetic coupling relationship. A sub-scattering matrix is ​​constructed to characterize the phase modulation, amplitude modulation, block scattering gain, and intra-group coupling response of the reconfigurable unit group to the incident electromagnetic wave. The sub-scattering matrix satisfies unitary and symmetry constraints, and the sub-scattering matrix arranged diagonally can form the overall scattering matrix of the group connected to the BD-RIS assisted near-field communication system.

3. The group-connected BD-RIS-assisted near-field communication location prediction method based on time-division switching acquisition and dual-channel temporal feature fusion as described in claim 2, characterized in that: After step S1, execute: The phase and amplitude offset of each of the reconfigurable unit groups are calibrated by sending known pilot signals, and the clock synchronization of the base station and the mobile user terminal is calibrated to obtain the calibrated sub-scattering matrix.

4. The group-connected BD-RIS-assisted near-field communication location prediction method based on time-division switching acquisition and dual-channel temporal feature fusion as described in claim 2, characterized in that: In step S2, channel data under the action of the reconfigurable unit group within the time window length is collected, and the multi-dimensional features of the channel data are used to form a high-dimensional feature sample set, including: Load the sub-scattering matrix corresponding to the reconfigurable unit group, control the reconfigurable unit group to enter the scattering state corresponding to the sub-scattering matrix, and collect channel data of a preset time window length; After denoising and smoothing the channel data, multi-dimensional features of the channel data are extracted. The multi-dimensional features include at least one of channel state information, time domain features, and angle domain features. The multi-dimensional features are combined to form a high-dimensional feature sample.

5. The group-connected BD-RIS-assisted near-field communication position prediction method based on time-division switching acquisition and dual-channel temporal feature fusion as described in claim 1, characterized in that: In step S3, determining the positioning contribution weight based on the correlation between the high-dimensional feature sample set of the channel data corresponding to each reconfigurable unit group and the real coordinates of the mobile user terminal includes: The high-dimensional feature sample sets of each reconfigurable unit group under the real coordinates of each mobile user terminal are combined into a multi-location sample set. The intra-group compactness of each multi-location sample set and the inter-group separation degree between different multi-location sample sets are calculated. The ratio of the inter-group separation degree to the intra-group compactness is normalized and used as the positioning contribution weight of the reconfigurable unit group.

6. The group-connected BD-RIS-assisted near-field communication position prediction method based on time-division switching acquisition and dual-channel temporal feature fusion as described in claim 1, characterized in that: In step S4, the long-term temporal dependency features and the local temporal features are fused to obtain fused features, including: The long-term temporal dependency feature and the local temporal feature are concatenated along the feature dimension to obtain the concatenated feature; The concatenated features are input into a multi-head attention mechanism to generate a query matrix, a key matrix, and a value matrix, which are then divided into multiple attention heads. The attention weights and attention outputs of each attention head are calculated, and the outputs of multiple attention heads are concatenated and projected to obtain the attention enhancement features. The final output of the attention enhancement feature is taken as the fusion feature.

7. The group-connected BD-RIS-assisted near-field communication position prediction method based on time-division switching acquisition and dual-channel temporal feature fusion as described in claim 1, characterized in that: Step S5 includes: Set a fully connected layer As a coordinate regression layer, For two-dimensional coordinate estimation, As a feature of fusion, and The weights and biases of the fully connected layer are used as the basis for training the fully connected layer with the fused features and the real coordinates of the mobile user terminal. and The optimal value is determined to construct the mapping relationship between the fused feature and the real coordinates of the mobile user terminal.

8. The group-connected BD-RIS-assisted near-field communication position prediction method based on time-division switching acquisition and dual-channel temporal feature fusion as described in claim 1, characterized in that: After step S2, the following steps are performed: dimensionality reduction of the high-dimensional feature sample set is achieved using a compressed sensing algorithm.

9. A group-connected BD-RIS-assisted near-field communication position prediction system based on time-division switching acquisition and dual-channel temporal feature fusion, characterized in that, During operation, it implements the group connection BD-RIS assisted near-field communication position prediction method based on time-division switching acquisition and dual-channel timing feature fusion as described in any one of claims 1-8.

Images