Artificial intelligence-based entanglement disentangling and reconstruction method, device, equipment and medium

Abstract

The application relates to an entanglement disentangling and reconstruction method and device based on artificial intelligence, equipment and a medium. In the application, configuration information of a terahertz base station and a plurality of entanglement sensing reconfigurable intelligent metasurfaces is acquired, entanglement sensing data corresponding to a downlink pilot signal is collected through an entanglement sensing reconfigurable intelligent metasurface unit, a channel entanglement tensor is constructed based on the configuration information and the entanglement sensing data; a channel matrix corresponding to the channel entanglement tensor is compressed to an intrinsic space to perform quantum entanglement decoupling processing, and intrinsic parameters are obtained; the intrinsic parameters and a distorted terahertz hologram are input into a holographic reconstruction engine carrying a physical enhanced neural network to obtain a channel environment model; according to the channel entanglement relationship of the channel environment model and the channel entanglement tensor, entanglement sensitive beamforming is performed to obtain beamforming parameters; and according to the beamforming parameters, the reflection phase and amplitude of the entanglement sensing reconfigurable intelligent metasurface unit are configured, and the optimization of the terahertz channel is completed.

Description

Technical Field

[0001] This invention relates to the field of data transmission and communication technology, and specifically to a method, apparatus, device, and medium for de-entanglement and reconstruction based on artificial intelligence. Background Technology

[0002] Channel deentanglement and holographic reconstruction are core technologies of the physical layer in 6G (sixth generation mobile communication technology) terahertz massive MIMO (multiple-input multiple-output) communication systems. They aim to solve the challenges of ultra-high-dimensional channel modeling, high-precision environmental perception, and efficient beamforming in the terahertz band. As 6G networks fully evolve towards the terahertz band, massive MIMO systems can achieve Tbps (terabits per second) transmission rates and μs (microseconds) latency by deploying hundreds to thousands of antenna elements. However, the extremely short wavelength, strong directionality, and complex multipath scattering and atmospheric absorption characteristics of terahertz waves result in deep correlations between channel parameters of different users, paths, and frequency components. Simultaneously, electromagnetic wave interference in the environment forms electromagnetic holograms. Channel deentanglement effectively reduces channel dimensionality and estimation errors, while holographic reconstruction obtains high-fidelity spatial environmental information. The synergy of these two technologies provides fundamental support for achieving precise beamforming and efficient integrated communication and perception.

[0003] Existing channel deentanglement techniques are mainly based on traditional statistical channel models, treating the channel as an independent random matrix or tensor, and reducing the channel dimension through matrix factorization, compressed sensing, and other methods. Holographic reconstruction techniques often directly adopt algorithms from the optical field, such as two-dimensional FFT (Fast Fourier Transform) and three-dimensional neural networks, to reconstruct environmental information by detecting the amplitude and phase distribution of electromagnetic waves. Meanwhile, to improve the real-time performance and generalization ability of channel estimation, some schemes introduce federated learning frameworks to achieve distributed training. Reconfigurable smart surfaces are also used to assist in channel estimation and beamforming, reconstructing the electromagnetic propagation environment by adjusting the phase, amplitude, and other parameters of passive reflective units.

[0004] However, the aforementioned existing technologies fail to effectively identify and utilize the inherent physical entanglement relationships between terahertz channel parameters, resulting in the inability to fundamentally solve the channel dimension explosion problem. Pilot overhead and computational complexity remain high. Furthermore, the directly ported optical holographic reconstruction algorithm does not consider the special scattering, diffraction, and atmospheric absorption characteristics of terahertz waves, leading to severe image distortion and phase distortion. The centralized processing architecture and flat federated learning mechanism also struggle to meet the real-time requirements of the extremely short coherence time of terahertz channels, resulting in insufficient cross-scenario generalization capabilities. These shortcomings severely restrict the practical deployment and performance improvement of terahertz large-scale MIMO systems.

[0005] Therefore, how to effectively decouple the physical entanglement of terahertz channels and achieve high-precision, low-delay holographic reconstruction is an urgent problem to be solved. Summary of the Invention

[0006] The technical problem to be solved by this invention is how to effectively decouple the physical entanglement relationship of terahertz channels and achieve high-precision, low-latency holographic reconstruction.

[0007] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: This application provides an artificial intelligence-based method for untangling and reconstructing entanglement, including: The configuration information of a terahertz base station and multiple entangled sensing reconfigurable smart metasurfaces is obtained, the terahertz base station is controlled to transmit downlink pilot signals, and entangled sensing data corresponding to the downlink pilot signals is collected through entangled sensing reconfigurable smart metasurface units. Based on the configuration information and the entangled sensing data, a channel entanglement tensor is constructed. The channel matrix corresponding to the channel entanglement tensor is compressed into the eigenspace to perform quantum entanglement decoupling processing and obtain eigen parameters, wherein the dimension of the channel matrix is ​​higher than the dimension of the eigenspace; The intrinsic parameters and the distorted terahertz holograms acquired by the entangled sensing reconfigurable smart metasurface unit are input into a holographic reconstruction engine equipped with a physically enhanced neural network to perform holographic reconstruction through preset physical constraints and obtain a channel environment model. Based on the channel environment model and the channel entanglement relationship corresponding to the channel entanglement tensor, entanglement-sensitive beamforming is performed to obtain beamforming parameters. Based on the beamforming parameters, the reflection phase and amplitude of the entangled sensing reconfigurable smart metasurface unit are configured to complete the optimization of the terahertz channel.

[0008] The beneficial effects of this invention are as follows: By acquiring the configuration information of the terahertz base station and the entanglement-sensing reconfigurable smart metasurface, the base station is controlled to transmit downlink pilot signals and collect entanglement-sensing data, thereby constructing a channel entanglement tensor, providing basic data for subsequent processing. The high-dimensional channel matrix corresponding to the channel entanglement tensor is compressed into a low-dimensional eigenspace. Through quantum entanglement decoupling processing, the channel dimension is effectively reduced, and mutually independent eigenvalues ​​are extracted, thus decoupling the physical entanglement relationship of the terahertz channel. The eigenvalues ​​and distorted terahertz holograms are input into a holographic reconstruction engine equipped with a physically enhanced neural network. Holographic reconstruction is performed using preset physical constraints. The physically enhanced neural network is designed for terahertz wave characteristics and can recover a high-fidelity environment model from the distorted hologram, achieving high-precision holographic reconstruction. Entanglement-sensitive beamforming is performed based on the channel environment model and channel entanglement relationship. Considering the correlation of channel parameters and combining near-field propagation characteristics, beamforming effect can be effectively improved, meeting real-time requirements and achieving low latency. By configuring the reflection phase and amplitude of the intelligent metasurface unit according to the beamforming parameters, the terahertz channel is optimized, and the physical entanglement relationship is effectively decoupled and high-precision, low-latency holographic reconstruction is achieved.

[0009] Based on the above technical solution, the present invention can be further improved as follows.

[0010] Furthermore, the construction of the channel entanglement tensor based on the configuration information and the entanglement-aware data includes: The configuration information is parsed to obtain four-dimensional parameters of the channel tensor, which correspond to the number of base station antennas, the number of users, the number of propagation paths, and the number of frequency sub-bands, respectively. Based on the four-dimensional parameters, the entangled sensing data is subjected to dimensional feature extraction to obtain the channel parameters corresponding to the four-dimensional parameters. The channel parameters include the signal amplitude and phase corresponding to different antennas, different users, different paths, and different sub-bands. The channel parameters are subjected to tensor mapping to obtain the initial channel tensor; The initial channel tensor is normalized in magnitude and phase to obtain the channel entanglement tensor.

[0011] The advantages of adopting the above-mentioned further scheme are as follows: Dimensional feature extraction ensures the comprehensive and targeted acquisition of channel parameters from different dimensions from entanglement-sensing data, enabling the channel tensor to accurately reflect the multifaceted characteristics of the terahertz channel. Tensor mapping integrates these parameters into tensor form, providing a suitable data structure for subsequent complex processing such as tensor-based quantum entanglement decoupling. Amplitude and phase normalization processing makes the channel tensor data more consistent and comparable, which is beneficial to improving the accuracy of subsequent calculations and analyses.

[0012] Furthermore, the step of compressing the channel matrix corresponding to the channel entanglement tensor into the eigenspace for quantum entanglement decoupling processing to obtain eigenparameters includes: Tensor train decomposition is performed on the channel entanglement tensor to obtain multiple target tensors; The multiple target tensors are iteratively optimized using the alternating least squares algorithm to obtain the optimized target tensors. The termination condition for the iterative optimization is that the reconstruction error satisfies the preset convergence condition. The optimized target tensor is subjected to redundant parameter pruning to obtain the pruned target tensor. The redundant parameters are those whose contribution to the channel representation is lower than a preset contribution threshold. The intrinsic parameters are extracted from the pruned target tensor to obtain the intrinsic parameters, which include the scattering center position, main path delay, and interference subspace information.

[0013] The beneficial effects of adopting the above-mentioned further scheme are as follows: Tensor train decomposition of the channel entanglement tensor transforms the complex high-dimensional channel tensor into a more tractable structure, laying the foundation for reducing the channel dimensionality. Iterative optimization using the alternating least squares algorithm can find the optimal solution among multiple interrelated target tensors, minimizing reconstruction errors and ensuring that the decomposed tensor accurately represents the original channel information, thus guaranteeing accurate extraction of intrinsic parameters. Redundant parameter pruning removes parameters with low contribution to the channel representation, reducing data volume and improving processing efficiency without affecting the expression of key channel features. Finally, the extracted intrinsic parameters are independent of each other, forming a low-dimensional channel representation, effectively reducing the channel dimensionality and achieving quantum entanglement decoupling.

[0014] Furthermore, the physically enhanced neural network includes a convolutional neural network and a transposed convolutional network. The intrinsic parameters and the distorted terahertz hologram acquired by the entanglement-aware reconfigurable intelligent metasurface unit are input into a holographic reconstruction engine equipped with the physically enhanced neural network to perform holographic reconstruction through preset physical constraints, thereby obtaining a channel environment model. This includes: The distorted terahertz hologram is subjected to noise removal and amplitude-phase calibration to obtain a standardized hologram; The standardized hologram and the intrinsic parameters are input into the encoder of the convolutional neural network to perform multi-scale feature extraction and obtain a multi-scale feature map. According to preset physical constraints, constraints are applied to the multi-scale feature map to obtain a constrained feature map; The constrained feature map is input into the decoder of the transposed convolutional network to reconstruct the environment model and obtain the channel environment model.

[0015] The beneficial effects of adopting the above-mentioned further scheme are as follows: noise removal and amplitude and phase calibration of the distorted terahertz hologram improve the quality of the input data, creating favorable conditions for accurate feature extraction and environmental model reconstruction. Multi-scale feature extraction can comprehensively capture different levels of features in the hologram and intrinsic parameters, making the reconstructed environmental model more detailed and accurate. Pre-set physical constraints, combined with the actual physical characteristics of terahertz waves, ensure that the reconstruction result conforms to the propagation law of terahertz waves, effectively reducing the distortion caused by directly transplanting traditional optical holographic algorithms and improving reconstruction accuracy. Environmental model reconstruction uses a transposed convolutional network to accurately convert the processed feature map into a channel environment model, enabling the final channel environment model to reflect the actual channel conditions with high fidelity.

[0016] Furthermore, the step of applying constraints to the multi-scale feature map according to preset physical constraints to obtain a constrained feature map includes: Scattering characteristic matching analysis is performed on the multi-scale feature map to obtain the scattering model consistency loss; Atmospheric attenuation simulation analysis was performed on the multi-scale feature map to obtain the atmospheric absorption model loss; Phase gradient calculation is performed on the multi-scale feature map to obtain the phase continuity loss; The total loss function is obtained by weighted summing of the data reconstruction loss, the scattering model consistency loss, the atmospheric absorption model loss, and the phase continuity loss. Based on the total loss function, the parameters of the encoder of the convolutional neural network are optimized by backpropagation to obtain the optimized encoder, and based on the optimized encoder, the constrained feature map is obtained.

[0017] The beneficial effects of adopting the above-mentioned further scheme are as follows: Scattering characteristic matching analysis, atmospheric attenuation simulation analysis, and phase gradient calculation are performed on the multi-scale feature maps. These constrain the reconstruction process from the actual physical characteristics of terahertz waves on different material surfaces, such as scattering, atmospheric absorption, and phase continuity, yielding corresponding loss values. These loss values, along with the data reconstruction loss, are weighted and summed to form a total loss function, comprehensively considering various factors in terahertz wave propagation and the differences between the reconstructed and ideal conditions. Backpropagation optimization of the convolutional neural network encoder based on the total loss function can adjust the encoder parameters, enabling it to process multi-scale feature maps more accurately. The resulting constrained feature maps better conform to the physical characteristics of terahertz waves and are closer to the ideal reconstruction result.

[0018] Furthermore, the step of performing entanglement-sensitive beamforming based on the channel environment model and the channel entanglement relationship corresponding to the channel entanglement tensor to obtain beamforming parameters includes: The channel entanglement tensor is used to calculate the user link correlation degree to obtain the channel entanglement degree between different user links. The channel entanglement degree is used to indicate the channel entanglement relationship between different user links. The channel entanglement degrees between the different user links are sorted from high to low to obtain a priority sequence of user links; Near-field propagation characteristics of the channel environment model were analyzed to obtain Fresnel diffraction model parameters; Based on the priority sequence of the user links and the Fresnel diffraction model parameters, entanglement-sensitive beamforming is performed to obtain the beamforming parameters corresponding to each user link.

[0019] The beneficial effects of adopting the above-mentioned further scheme are as follows: Calculating the user link correlation degree using the channel entanglement tensor clarifies the channel entanglement relationship between different user links, providing a foundation for the rational allocation of beamforming resources. Ranking the channel entanglement degree yields a priority sequence for user links, enabling priority processing of user links with high entanglement degrees during beamforming, effectively coordinating signal interference between user links and improving communication quality. Analyzing the near-field propagation characteristics of the channel environment model and obtaining Fresnel diffraction model parameters fully considers the special propagation characteristics of terahertz waves in the near field, making beamforming more consistent with the actual propagation environment. Entanglement-sensitive beamforming based on the priority sequence and Fresnel diffraction model parameters can calculate the optimal beamforming parameters for each user link.

[0020] Furthermore, the method also includes: The pairwise correlation degree of the channel entanglement tensors corresponding to multiple entangled sensing reconfigurable smart metasurface nodes is calculated to obtain the channel entanglement degree matrix between nodes. Spectral clustering is performed on the inter-node channel entanglement matrix to obtain multiple node subnets, wherein nodes with channel entanglement degrees higher than a preset entanglement degree threshold belong to the same node subnet. Each node in a subnet trains a model based on its local channel dataset, resulting in multiple local models. Multiple local models within the same node subnet are subjected to subnet federated averaging to obtain the subnet global model; The target global model is obtained by performing inter-subnet federated averaging on multiple subnet global models. The target global model is distributed to each entangled sensing reconfigurable smart metasurface node to update the holographic reconstruction engine.

[0021] The beneficial effects of adopting the above-mentioned further scheme are as follows: The pairwise correlation degree of the channel entanglement tensors of multiple entanglement-aware reconfigurable intelligent metasurface nodes is calculated and a matrix is ​​constructed, clearly demonstrating the channel entanglement relationship between nodes. Spectral clustering divides nodes into subnets based on entanglement degree, ensuring that nodes with high entanglement degrees are grouped into the same subnet. This division method fully utilizes the correlation between nodes, improving the efficiency and accuracy of model training. Each node trains a local model based on its local channel dataset, fully exploring local channel characteristics. The federated averaging within a subnet integrates information from nodes within that subnet, reducing the bias of individual node data and obtaining a more representative global model for the subnet. The federated averaging between subnets further integrates information from each subnet, obtaining a target global model adapted to the channel characteristics of different subnets. The target global model is then distributed to update the holographic reconstruction engine, enabling the engine to utilize a more comprehensive and accurate model, effectively improving the accuracy and adaptability of holographic reconstruction.

[0022] This application also provides an artificial intelligence-based de-entanglement and reconstruction device, the device comprising: The acquisition module is used to acquire configuration information of a terahertz base station and multiple entangled sensing reconfigurable smart metasurfaces, control the terahertz base station to transmit downlink pilot signals, and collect entangled sensing data corresponding to the downlink pilot signals through entangled sensing reconfigurable smart metasurface units, and construct a channel entanglement tensor based on the configuration information and the entangled sensing data; The entanglement decoupling module is used to compress the channel matrix corresponding to the channel entanglement tensor into the eigenspace to perform quantum entanglement decoupling processing and obtain eigen parameters, wherein the dimension of the channel matrix is ​​higher than the dimension of the eigenspace; The holographic reconstruction module is used to input the intrinsic parameters and the distorted terahertz hologram collected by the entangled sensing reconfigurable smart metasurface unit into the holographic reconstruction engine equipped with a physically enhanced neural network, so as to perform holographic reconstruction through preset physical constraints and obtain a channel environment model. The beamforming module is used to perform entanglement-sensitive beamforming based on the channel environment model and the channel entanglement relationship corresponding to the channel entanglement tensor, and to obtain beamforming parameters. The channel optimization module is used to configure the reflection phase and amplitude of the entangled sensing reconfigurable smart metasurface unit according to the beamforming parameters, thereby optimizing the terahertz channel.

[0023] This application also provides an electronic device, including a processor and a memory, wherein the memory stores multiple instructions; the processor loads instructions from the memory to execute steps in any of the artificial intelligence-based de-entanglement and reconstruction methods provided in this application.

[0024] This application also provides a computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to execute steps in any of the artificial intelligence-based de-entanglement and reconstruction methods provided in this application. Attached Figure Description

[0025] Figure 1 A flowchart illustrating the artificial intelligence-based de-entanglement and reconstruction method provided in an embodiment of the present invention; Figure 2 A schematic diagram of the structure of the artificial intelligence-based de-entanglement and reconstruction device provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0026] The principles and features of the present invention are described below. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0027] This application provides an artificial intelligence-based method, apparatus, electronic device, and medium for unentanglement and reconstruction.

[0028] Specifically, the AI-based de-entanglement and reconstruction device can be integrated into an electronic device, such as a terminal or server. The terminal can be a mobile phone, tablet, smart Bluetooth device, laptop, or personal computer (PC); the server can be a single server or a server cluster consisting of multiple servers.

[0029] In some embodiments, the AI-based unentanglement and reconstruction device can also be integrated into multiple electronic devices. For example, the AI-based unentanglement and reconstruction device can be integrated into multiple servers, and the AI-based unentanglement and reconstruction method of this application can be implemented by multiple servers.

[0030] In some embodiments, the terminal can also be used as a server to implement some or all of the functions of a server.

[0031] In the embodiments of this application, the terms "module" or "unit" may refer to a computer program or part of a computer program with a predetermined function, which works together with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) may be used to implement one or more modules or units. Furthermore, each module or unit may be part of an overall module or unit that includes the functions of that module or unit.

[0032] The following sections provide detailed descriptions of each example. It should be noted that the order of the following embodiments is not intended to limit the preferred order of the embodiments.

[0033] This embodiment provides an artificial intelligence-based method for de-entanglement and reconstruction. See also... Figure 1 , Figure 1 This is a flowchart illustrating the artificial intelligence-based de-entanglement and reconstruction method provided in an embodiment of the present invention, as shown below. Figure 1 As shown, the specific process of this AI-based unentanglement and reconstruction method can be summarized as follows: 110. Obtain configuration information of a terahertz base station and multiple entangled sensing reconfigurable smart metasurfaces, control the terahertz base station to transmit downlink pilot signals, and collect entangled sensing data corresponding to the downlink pilot signals through entangled sensing reconfigurable smart metasurface units, and construct a channel entanglement tensor based on the configuration information and the entangled sensing data; 120. Compress the channel matrix corresponding to the channel entanglement tensor into the eigenspace to perform quantum entanglement decoupling processing and obtain eigenparameters, wherein the dimension of the channel matrix is ​​higher than the dimension of the eigenspace; 130. Input the intrinsic parameters and the distorted terahertz hologram acquired by the entangled sensing reconfigurable smart metasurface unit into a holographic reconstruction engine equipped with a physically enhanced neural network, so as to perform holographic reconstruction through preset physical constraints and obtain a channel environment model. 140. Based on the channel environment model and the channel entanglement relationship corresponding to the channel entanglement tensor, perform entanglement-sensitive beamforming to obtain beamforming parameters; 150. Based on the beamforming parameters, configure the reflection phase and amplitude of the entangled sensing reconfigurable smart metasurface unit to complete the optimization of the terahertz channel.

[0034] The terahertz base station is the device responsible for generating and transmitting terahertz communication signals in the entire communication system. It includes a multi-antenna array, supporting large-scale multiple-input multiple-output configurations such as 1024×1024, operating in the 0.1-3 terahertz frequency band, preferably 0.3 terahertz, with a bandwidth of 10-20 gigahertz, and possessing beamforming and spatial multiplexing capabilities. Its baseband processing unit performs signal modulation, encoding, and precoding, while the quantum entanglement analyzer is responsible for analyzing the entanglement relationships in the channel and identifying highly entangled channel parameter pairs.

[0035] The entanglement-aware reconfigurable smart metasurface consists of a large number of programmable reflection units, such as 1024×1024 units. Each unit can independently control the reflection phase, ranging from 0 to 2π with a precision of 3 bits, and the amplitude range is from 0 to 1 with a precision of 2 bits. Each subarray consists of 16×16 units and shares an entanglement-aware module. This module integrates a Schottky diode detector and a correlation calculation unit, which can detect the entanglement characteristics of the local channel, such as phase coupling between paths and interference patterns between users, and can also measure the entanglement degree between different paths and different users in real time.

[0036] The channel entanglement tensor is a multi-order tensor obtained by modeling a terahertz large-scale multiple-input multiple-output (MIMO) channel. Its dimension is determined by the number of base station antennas, users, paths, and frequency sub-bands. For example, in a 1024×1024 MIMO system, considering 8 users, 32 paths, and 64 sub-bands, the tensor dimension would be very high. By introducing the concept of quantum entangled states, the channel tensor is modeled as an entangled state, unifying the representation of all channel parameters. The degree of entanglement is defined as the reciprocal of the tensor rank, thus reflecting the degree of correlation between parameters.

[0037] The intrinsic space is a low-dimensional space obtained by compressing the high-dimensional channel matrix corresponding to the channel entanglement tensor; the dimension of the channel matrix is ​​much higher than the dimension of the intrinsic space. In this space, the intrinsic parameters extracted through quantum entanglement decoupling are independent of each other, and these parameters constitute a low-dimensional channel representation, thereby achieving an effective and simplified description of the channel.

[0038] Among them, the Physically Enhanced Neural Network (PEN) is a neural network architecture designed specifically for the unique scattering characteristics of terahertz waves. It employs a U-shaped network architecture as the holographic reconstruction network, including an encoding path (downsampling) to extract deep features and a decoding path (upsampling) to restore spatial resolution. Simultaneously, a terahertz physical constraint module is introduced into the neural network. Through three key components—scattering model consistency loss, atmospheric absorption model loss, and phase continuity loss—it ensures that the reconstruction result conforms to the propagation characteristics of terahertz waves, recovering a high-fidelity environmental model from the distorted hologram.

[0039] The holographic reconstruction engine is an edge computing unit deployed on the side of the smart metasurface, running a terahertz holographic reconstruction network. The edge computing unit adopts a hybrid architecture of field-programmable gate arrays (FPGAs) and graphics processors (GPUs). The FPGAs are responsible for high-speed data acquisition and preprocessing, while the GPUs are responsible for neural network inference. The inference latency is less than 10 milliseconds, enabling the mapping from distorted holograms to high-fidelity environmental models.

[0040] Among them, the channel environment model is a model obtained after holographic reconstruction, which includes information such as the location of the scattering center, path parameters, and interference relationships. It can reflect the three-dimensional spatial situation of the terahertz channel with high fidelity and provide accurate environmental information for subsequent beamforming.

[0041] Entanglement-sensitive beamforming is a beamforming method based on the channel environment model and channel entanglement relationships. Considering the entanglement relationships between channel parameters, it prioritizes optimizing user links with high entanglement to achieve reasonable resource scheduling and beam allocation, thereby improving the performance of terahertz communication systems.

[0042] Among them, beamforming parameters refer to the parameters required to configure the reflection phase and amplitude of the intelligent metasurface unit in order to achieve the preset beamforming effect. By adjusting the intelligent metasurface unit through these parameters, the beam of the terahertz signal can better adapt to the channel environment and improve communication quality.

[0043] For example, the specific method for obtaining configuration information and collecting data to construct the channel entanglement tensor can be as follows: First, the configuration information of the terahertz base station and multiple entanglement-sensing reconfigurable smart metasurfaces is acquired. This information includes key parameters such as the number of base station antennas, the number of users, the number of propagation paths, and the number of frequency subbands. Then, the terahertz base station is controlled to transmit downlink pilot signals. The pilots employ an orthogonal frequency division multiplexing (OFDM) structure with a subcarrier spacing of 1 MHz and a pilot period of 100 microseconds. Each entanglement-sensing reconfigurable smart metasurface unit acquires the amplitude and phase information corresponding to the downlink pilot signals through its entanglement sensing module, and measures entanglement sensing data such as the degree of entanglement between different paths and different users. Next, the configuration information is dimensionally analyzed to obtain four-dimensional parameters. Based on these parameters, dimensional feature extraction is performed on the entanglement sensing data to obtain channel parameters such as signal amplitude and phase corresponding to different antennas, different users, different paths, and different subbands. Tensor mapping is then performed on the channel parameters to obtain the initial channel tensor. Finally, the initial channel tensor is normalized in amplitude and phase to obtain the channel entanglement tensor.

[0044] For example, the specific method for compressing the channel matrix into the eigenspace and obtaining the eigenparameters can be as follows: The constructed channel entanglement tensor is decomposed using tensor train decomposition to obtain multiple target tensors. These target tensors are then iteratively optimized using an alternating least squares algorithm, continuously adjusting the decomposition parameters to minimize the reconstruction error. Iteration stops when the reconstruction error meets a preset convergence condition, yielding the optimized target tensors. Redundancy pruning is then performed on the optimized target tensors, removing parameters whose contribution to the channel representation is below a preset contribution threshold, resulting in pruned target tensors. Finally, intrinsic parameters such as the scattering center location, main path delay, and interference subspace information are extracted from the pruned target tensors.

[0045] For example, the specific method for obtaining a channel environment model by holographic reconstruction of input data can be as follows: First, noise removal and amplitude-phase calibration are performed on the distorted terahertz holograms acquired by the entanglement-aware reconfigurable intelligent metasurface unit to obtain standardized holograms. The standardized holograms and the previously obtained intrinsic parameters are then input into a convolutional neural network encoder of a physically enhanced neural network. This encoder uses a U-shaped network architecture encoding path to extract multi-scale features, resulting in multi-scale feature maps. Based on preset physical constraints, scattering characteristic matching analysis is performed on the multi-scale feature maps to obtain scattering model consistency loss, atmospheric attenuation simulation analysis to obtain atmospheric absorption model loss, and phase gradient calculation to obtain phase continuity loss. The data reconstruction loss and these three losses are weighted and summed to obtain a total loss function. Based on this total loss function, the parameters of the convolutional neural network encoder are optimized through backpropagation to obtain an optimized encoder, which in turn yields constrained feature maps. Finally, the constrained feature maps are input into a decoder of a transposed convolutional network. Through the residual connections and skip connections of the decoder, an environmental model is reconstructed to obtain a channel environment model.

[0046] For example, the specific method for obtaining beamforming parameters through entanglement-sensitive beamforming based on the channel environment model and entanglement relationship can be as follows: First, the channel entanglement tensor is used to calculate the user link correlation degree, obtaining the channel entanglement degree between different user links. This entanglement degree represents the channel entanglement relationship between different user links. Next, the channel entanglement degrees of different user links are sorted from high to low to obtain a priority sequence of user links. Then, near-field propagation characteristics are analyzed on the channel environment model to obtain Fresnel diffraction model parameters. Finally, based on the user link priority sequence and Fresnel diffraction model parameters, entanglement-sensitive beamforming is performed to obtain the beamforming parameters corresponding to each user link.

[0047] For example, the specific method for obtaining the channel optimization by configuring intelligent metasurface units based on beamforming parameters can be as follows: Based on beamforming parameters, the reflection phase and amplitude of the entanglement-aware reconfigurable smart metasurface unit are configured so that the beam of the terahertz signal can better adapt to the channel environment and optimize the performance of the terahertz channel, such as improving communication quality and reducing interference.

[0048] For example, consider a terahertz communication scenario. The terahertz base station uses a 1024×1024 multiple-input multiple-output configuration, operates in the 0.3 terahertz band with a bandwidth of 15 GHz, serves 8 users, has 32 communication paths, and divides the frequency subbands into 64. The base station transmits downlink pilot signals with a 1 MHz subcarrier spacing and a 100 microsecond pilot period, according to an orthogonal frequency division multiplexing (OFDM) structure. An entangled sensing reconfigurable smart metasurface, composed of 1024×1024 programmable reflector units, collects signal amplitude, phase information, and measures entanglement degree data through an entanglement sensing module in each 16×16 unit subarray. Dimensional parameters are parsed based on the configuration information of the base station and the smart metasurface. Channel parameters are extracted from the collected data by dimension, and a channel entanglement tensor is constructed through tensor mapping and normalization. The channel matrix is ​​compressed to the intrinsic space through operations such as tensor train decomposition and iterative optimization using alternating least squares algorithm, and intrinsic parameters are extracted. Simultaneously, the intelligent metasurface acquires distorted terahertz holograms, which, after processing, are input along with intrinsic parameters into a holographic reconstruction engine equipped with a physically enhanced neural network. Through a series of operations, a channel environment model is obtained. Based on this model and the channel entanglement relationship, the entanglement degree of user links is calculated and ranked. Combined with Fresnel diffraction model parameters obtained from near-field propagation characteristic analysis, entanglement-sensitive beamforming is performed to obtain beamforming parameters. Finally, based on these parameters, the reflection phase and amplitude of the intelligent metasurface units are configured to complete the optimization of the terahertz channel.

[0049] As can be seen, in this embodiment, by acquiring the configuration information of the terahertz base station and the entanglement-sensing reconfigurable smart metasurface, the base station is controlled to transmit downlink pilot signals and collect entanglement-sensing data, thereby constructing a channel entanglement tensor, providing basic data for subsequent processing. The high-dimensional channel matrix corresponding to the channel entanglement tensor is compressed to a low-dimensional eigenspace. Through quantum entanglement decoupling processing, the channel dimension is effectively reduced, and mutually independent eigenvalues ​​are extracted, thus decoupling the physical entanglement relationship of the terahertz channel. The eigenvalues ​​and the distorted terahertz hologram are input into a holographic reconstruction engine equipped with a physically enhanced neural network. Holographic reconstruction is performed using preset physical constraints. The physically enhanced neural network is designed for terahertz wave characteristics and can recover a high-fidelity environment model from the distorted hologram, achieving high-precision holographic reconstruction. Entanglement-sensitive beamforming is performed based on the channel environment model and channel entanglement relationship. Considering the correlation of channel parameters and combining near-field propagation characteristics, beamforming effect can be effectively improved, meeting real-time requirements and achieving low latency. By configuring the reflection phase and amplitude of the intelligent metasurface unit according to the beamforming parameters, the terahertz channel is optimized, and the physical entanglement relationship is effectively decoupled and high-precision, low-latency holographic reconstruction is achieved.

[0050] Preferably, in an embodiment, constructing a channel entanglement tensor based on the configuration information and the entanglement-aware data includes: The configuration information is parsed to obtain four-dimensional parameters of the channel tensor, which correspond to the number of base station antennas, the number of users, the number of propagation paths, and the number of frequency sub-bands, respectively. Based on the four-dimensional parameters, the entangled sensing data is subjected to dimensional feature extraction to obtain the channel parameters corresponding to the four-dimensional parameters. The channel parameters include the signal amplitude and phase corresponding to different antennas, different users, different paths, and different sub-bands. The channel parameters are subjected to tensor mapping to obtain the initial channel tensor; The initial channel tensor is normalized in magnitude and phase to obtain the channel entanglement tensor.

[0051] Among them, dimensional analysis refers to the analysis and extraction of parameters corresponding to four key dimensions related to the channel tensor from the configuration information of terahertz base stations and entangled sensing reconfigurable smart metasurfaces. These parameters are the basic dimensional information for constructing the channel tensor.

[0052] Among them, the dimensional feature extraction is based on the four-dimensional parameters of the obtained channel tensor. The entangled sensing data collected by the entangled sensing reconfigurable smart metasurface unit is subjected to feature mining according to different dimensions. This allows us to obtain channel parameters such as signal amplitude and phase under different antennas, different users, different paths, and different subbands. These parameters will be used to construct the channel tensor in the future.

[0053] Tensor mapping is a process of converting the channel parameters obtained through multidimensional feature extraction into tensor form according to certain rules to form the initial channel tensor. It is a key step in integrating the channel parameters into a tensor structure, providing a suitable data format for subsequent processing.

[0054] The amplitude and phase normalization process involves processing the amplitude and phase information in the initial channel tensor to ensure that the amplitude and phase values ​​are within a preset standard range. The purpose is to eliminate the influence of magnitude differences between different data, making the channel tensor data more consistent and comparable, and ultimately obtaining the channel entanglement tensor.

[0055] For example, the specific method for performing dimensional parsing of configuration information can be as follows: By analyzing the configuration information of terahertz base stations and entanglement-sensing reconfigurable smart metasurfaces, information directly related to the channel tensor dimension is extracted, and four dimensional parameters are identified: the number of base station antennas, the number of users, the number of propagation paths, and the number of frequency subbands. This provides a dimensional basis for subsequent processing of entanglement-sensing data.

[0056] For example, the specific method for extracting dimensional features from entanglement-aware data based on four dimensional parameters can be as follows: Guided by four established parameters, the entangled sensing data collected by the entangled sensing reconfigurable smart metasurface unit is classified and processed. For the dimension of the number of base station antennas, the signal amplitude and phase corresponding to different antennas are extracted; for the dimension of the number of users, the signal amplitude and phase related to different users are obtained; similarly, channel parameters such as signal amplitude and phase of the corresponding paths and sub-bands are extracted from the dimensions of the number of propagation paths and the number of frequency sub-bands, respectively, to prepare data for constructing the initial channel tensor.

[0057] For example, the specific method for performing tensor mapping processing on channel parameters can be as follows: The channel parameters extracted from different dimensions are mapped into a tensor form according to the tensor construction rules, generating the initial channel tensor. This completes the transformation from scattered parameters to a tensor structure, facilitating subsequent unified mathematical operations and processing of the channel data.

[0058] For example, the method for normalizing the amplitude and phase of the initial channel tensor can be as follows: For the amplitude information in the initial channel tensor, all its amplitude values ​​are normalized to the range of 0 to 1 by a preset calculation method, such as dividing by a maximum amplitude value or scaling by a certain ratio; for the phase information, the phase value is adjusted to the standard range of 0 to 2π through corresponding mathematical transformations, thereby obtaining the final channel entanglement tensor, ensuring the consistency and comparability of the data.

[0059] In this embodiment, multi-dimensional feature extraction ensures the comprehensive and targeted acquisition of channel parameters from different dimensions from entanglement-sensing data, enabling the channel tensor to accurately reflect the multifaceted characteristics of the terahertz channel. Tensor mapping integrates these parameters into tensor form, providing a suitable data structure for subsequent complex processing such as tensor-based quantum entanglement decoupling. Amplitude and phase normalization processing further enhances the consistency and comparability of the channel tensor data, improving the accuracy of subsequent calculations and analyses.

[0060] Preferably, in the embodiment, compressing the channel matrix corresponding to the channel entanglement tensor into the eigenspace for quantum entanglement decoupling processing to obtain eigenparameters includes: Tensor train decomposition is performed on the channel entanglement tensor to obtain multiple target tensors; The multiple target tensors are iteratively optimized using the alternating least squares algorithm to obtain the optimized target tensors. The termination condition for the iterative optimization is that the reconstruction error satisfies the preset convergence condition. The optimized target tensor is subjected to redundant parameter pruning to obtain the pruned target tensor. The redundant parameters are those whose contribution to the channel representation is lower than a preset contribution threshold. The intrinsic parameters are extracted from the pruned target tensor to obtain the intrinsic parameters, which include the scattering center position, main path delay, and interference subspace information.

[0061] Tensor train decomposition is a method for handling high-dimensional tensors, decomposing a high-dimensional channel entanglement tensor into a tensor product of multiple relatively low-dimensional target tensors. This decomposition transforms complex high-dimensional tensors into structures that are easier to process and analyze, providing a foundation for subsequent reduction of channel dimensionality and decoupling of quantum entanglement relationships.

[0062] Alternating least squares is an iterative optimization algorithm used to approach the optimal solution by alternately fixing other variables and performing least squares optimization on a single variable when multiple variables are interrelated. In this scenario, it is used to optimize multiple target tensors obtained from tensor train decomposition to minimize reconstruction error and ensure that the decomposed tensors can represent the original channel tensor as accurately as possible.

[0063] The reconstruction error refers to the degree of difference between the reconstructed channel tensor and the original entangled channel tensor after tensor train decomposition and iterative optimization. By calculating this difference, the effectiveness of the decomposition and optimization can be evaluated, providing a basis for determining whether the preset convergence conditions are met.

[0064] The preset convergence condition is a pre-defined standard used to determine whether the iterative optimization process of the alternating least squares algorithm should be stopped. When the reconstruction error reaches or falls below this preset condition, the iterative optimization is considered to have reached an acceptable level of accuracy, and the optimized target tensor obtained at this point can be used for subsequent processing.

[0065] Redundant parameters are those that contribute little to accurately describing channel characteristics in the channel representation. Although these parameters are included in the channel tensor, they have little impact on the expression of key channel features. Removing them will not cause significant loss of the main information of the channel; on the contrary, it can simplify the channel representation and improve processing efficiency.

[0066] The preset contribution threshold is a standard used to define whether a parameter is redundant. When a parameter's contribution to the channel representation is lower than this threshold, it is considered a redundant parameter and can be removed during redundant parameter pruning.

[0067] Redundant parameter pruning involves filtering the optimized target tensor and removing redundant parameters whose contribution to the channel representation is below a preset contribution threshold, resulting in a more concise pruned target tensor. This process reduces data volume, improves the efficiency of subsequent processing, and preserves key channel information.

[0068] For example, the specific method for performing tensor train decomposition on the channel entanglement tensor can be as follows: Taking the channel entanglement tensor as the object, the tensor train decomposition method is used to decompose it into a tensor product of multiple target tensors. This decomposition is based on the mathematical properties of tensors, breaking down high-dimensional channel information into multiple low-dimensional parts according to certain rules, making the complex channel tensor structure simpler and clearer, and facilitating further processing and analysis.

[0069] For example, the iterative optimization of multiple target tensors using the alternating least squares algorithm can be specifically as follows: An alternating least squares algorithm is employed. In each iteration, all target tensors except for one are fixed, and least squares optimization is performed on that target tensor. By continuously adjusting the parameters of the target tensors, the reconstruction error between the reconstructed channel tensor and the original channel entanglement tensor gradually decreases. This process of alternating optimization of each target tensor is repeated until the reconstruction error meets a preset convergence condition, yielding the optimized target tensor. This continuous adjustment of tensor parameters improves the accuracy of the decomposed tensor in representing the original channel tensor.

[0070] For example, the specific method for pruning redundant parameters of the optimized target tensor can be as follows: After obtaining the optimized target tensor, each parameter is evaluated based on a preset contribution threshold to determine its contribution to the channel representation. Parameters with contributions below the threshold are identified as redundant and removed, resulting in the pruned target tensor. This approach preserves key channel information while reducing data volume and improving subsequent processing efficiency.

[0071] For example, the specific method for extracting intrinsic parameters from the pruned target tensor can be as follows: From the pruned target tensor, intrinsic parameters representing the core characteristics of the channel are extracted, such as the scattering center location, main path delay, and interference subspace information. These intrinsic parameters are independent of each other and constitute a low-dimensional channel representation, providing concise and crucial information for subsequent channel processing and analysis.

[0072] In this embodiment, the channel entanglement tensor is decomposed using tensor train decomposition, transforming the complex high-dimensional channel tensor into a more tractable structure, laying the foundation for reducing the channel dimensionality. Iterative optimization using the alternating least squares algorithm finds the optimal solution among multiple interrelated target tensors, minimizing reconstruction error and ensuring that the decomposed tensor accurately represents the original channel information, thus guaranteeing accurate extraction of intrinsic parameters. Redundant parameter pruning removes parameters with low contribution to the channel representation, reducing data volume and improving processing efficiency without affecting the expression of key channel features. The finally extracted intrinsic parameters are independent of each other, forming a low-dimensional channel representation, effectively reducing the channel dimensionality and achieving quantum entanglement decoupling.

[0073] Preferably, in an embodiment, the physically enhanced neural network includes a convolutional neural network and a transposed convolutional network. The step of inputting the intrinsic parameters and the distorted terahertz hologram acquired by the entanglement-aware reconfigurable intelligent metasurface unit into a holographic reconstruction engine equipped with the physically enhanced neural network to perform holographic reconstruction through preset physical constraints to obtain a channel environment model includes: The distorted terahertz hologram is subjected to noise removal and amplitude-phase calibration to obtain a standardized hologram; The standardized hologram and the intrinsic parameters are input into the encoder of the convolutional neural network to perform multi-scale feature extraction and obtain a multi-scale feature map. According to preset physical constraints, constraints are applied to the multi-scale feature map to obtain a constrained feature map; The constrained feature map is input into the decoder of the transposed convolutional network to reconstruct the environment model and obtain the channel environment model.

[0074] Due to the strong scattering characteristics of terahertz waves, distorted terahertz holograms contain a significant amount of noise. Noise removal involves using pre-defined filtering algorithms or signal processing methods to identify and eliminate these noise signals, resulting in a clearer hologram that is more conducive to subsequent processing. This noise may be caused by various factors, such as environmental interference and equipment-related issues; noise removal improves the quality of the hologram.

[0075] Terahertz holograms contain amplitude and phase information of electromagnetic waves, but deviations or inconsistencies may occur in amplitude and phase during acquisition. Amplitude and phase calibration involves adjusting the amplitude and phase of the hologram through measurement or calculation to conform to the actual propagation characteristics of terahertz waves, thereby obtaining a standardized hologram. This provides a reliable data foundation for subsequent accurate feature extraction and environmental model reconstruction.

[0076] In multi-scale feature extraction, a convolutional neural network encoder is used to process the input standardized hologram and intrinsic parameters through convolutional kernels of different sizes and network structures of different levels, extracting the feature information contained therein from different scales. Features at different scales can reflect the characteristics of images or data at different levels of detail. For example, large-scale features can reflect the overall structure, while small-scale features can reflect subtle local changes. These multi-scale features are crucial for accurately reconstructing the channel environment model.

[0077] The preset physical constraints are conditions set for the special physical characteristics of terahertz waves, including scattering model consistency loss, atmospheric absorption model loss, and phase continuity loss. These constraints ensure that the reconstruction results conform to the physical laws of terahertz waves in actual propagation, avoiding unrealistic reconstruction results and thus improving the accuracy of the reconstruction.

[0078] In the environmental model reconstruction process, the feature map, after being processed by preset physical constraints, is input into the decoder of the transposed convolutional network. Through the transposed convolution operation, the low-dimensional feature map is restored into a high-dimensional channel environment model. The transposed convolutional network can learn how to reconstruct channel environment information with actual physical meaning from the feature map, such as the location of the scattering center and path parameters, thus realizing the mapping from features to the environment model.

[0079] For example, the specific method for noise removal and amplitude-phase calibration of distorted terahertz holograms can be as follows: Filtering algorithms, such as Gaussian filtering and median filtering, are used to smooth noise in the distorted terahertz hologram, removing randomly distributed noise points. Then, based on terahertz wave propagation theory and pre-acquired reference data, the amplitude and phase of the hologram are calibrated. For example, the amplitude is adjusted by comparing the difference between the amplitude values ​​at different locations and the theoretical values; the phase is corrected by analyzing the phase variation pattern, thus obtaining a standardized hologram.

[0080] For example, the method of inputting standardized holograms and intrinsic parameters into the encoder of a convolutional neural network for multi-scale feature extraction can be as follows: The standardized hologram and intrinsic parameters are used as input to the convolutional neural network encoder. Internally, the encoder performs convolution operations by sliding convolutional kernels of different sizes across the image and data. For example, larger kernels are used initially to capture overall feature information, and then smaller kernels are gradually used to extract more detailed features. In different network layers, the input is continuously downsampled, reducing data dimensionality while enhancing feature extraction, ultimately resulting in a multi-scale feature map containing feature information at different scales.

[0081] For example, the specific way to apply constraints to multi-scale feature maps according to preset physical constraints can be as follows: For multi-scale feature maps, the scattering model consistency loss, atmospheric absorption model loss, and phase continuity loss are calculated separately. For the scattering model consistency loss, the differences in scattering characteristics of terahertz waves on different material surfaces are utilized to compare the reconstructed environmental model with the prior scattering model, and the difference between the two is calculated as the loss value. For the atmospheric absorption model loss, based on the preset atmospheric absorption coefficient in the terahertz band, the attenuation of terahertz waves during propagation is simulated, and the difference between the reconstructed result and the theoretical result considering atmospheric absorption is calculated as the loss value. For the phase continuity loss, the spatial gradient of the reconstructed wavefront phase is calculated to maintain phase continuity and avoid phase jumps; this gradient value is used as the loss value. Finally, these loss values ​​are summed with the data reconstruction loss according to certain weights to obtain the total loss function. Based on this, the parameters of the convolutional neural network encoder are optimized through backpropagation, thus obtaining the constrained feature map.

[0082] For example, the specific method for reconstructing the environment model by inputting the constrained feature map into the decoder of the transposed convolutional network can be as follows: The constrained feature map is input into the decoder of the transposed convolutional network. The decoder gradually restores the dimensionality of the feature map through transposed convolution operations, mapping low-dimensional feature information back to a high-dimensional space. During this dimensionality restoration process, the relationship between the features learned by the transposed convolutional network and the channel environment is combined to fill in detailed information, ultimately reconstructing a channel environment model containing information such as the scattering center location, path parameters, and interference relationships.

[0083] In this embodiment, noise removal and amplitude-phase calibration are performed on the distorted terahertz hologram, improving the quality of the input data and creating favorable conditions for accurate feature extraction and environmental model reconstruction. Multi-scale feature extraction can comprehensively capture different levels of features in the hologram and intrinsic parameters, making the reconstructed environmental model more detailed and accurate. Pre-set physical constraints, combined with the actual physical characteristics of terahertz waves, ensure that the reconstruction result conforms to the propagation law of terahertz waves, effectively reducing distortion caused by directly transplanting traditional optical holographic algorithms and improving reconstruction accuracy. The environmental model reconstruction uses a transposed convolutional network to accurately convert the processed feature map into a channel environment model, enabling the final channel environment model to reflect the actual channel conditions with high fidelity.

[0084] Preferably, in the embodiment, the step of applying constraints to the multi-scale feature map according to preset physical constraints to obtain a constrained feature map includes: Scattering characteristic matching analysis is performed on the multi-scale feature map to obtain the scattering model consistency loss; Atmospheric attenuation simulation analysis was performed on the multi-scale feature map to obtain the atmospheric absorption model loss; Phase gradient calculation is performed on the multi-scale feature map to obtain the phase continuity loss; The total loss function is obtained by weighted summing of the data reconstruction loss, the scattering model consistency loss, the atmospheric absorption model loss, and the phase continuity loss. Based on the total loss function, the parameters of the encoder of the convolutional neural network are optimized by backpropagation to obtain the optimized encoder, and based on the optimized encoder, the constrained feature map is obtained.

[0085] Among them, scattering characteristic matching analysis: Terahertz waves exhibit different scattering characteristics on different material surfaces. For example, metals have a high reflectivity for terahertz waves, typically greater than 90%, while concrete has a relatively low reflectivity, less than 30%. Scattering characteristic matching analysis compares the scattering of terahertz waves on different material surfaces as reflected in the multi-scale feature map with the known scattering characteristics of terahertz waves on various materials. This comparison measures the degree of agreement between the reconstructed results and the actual physical properties, thereby obtaining the consistency loss of the scattering model.

[0086] Terahertz waves are affected by atmospheric absorption during propagation. Different frequencies of terahertz waves are absorbed to varying degrees in the atmosphere; for example, a water vapor absorption peak exists in the 0.3 terahertz band. Atmospheric attenuation simulation analysis simulates the attenuation of terahertz waves during propagation in the atmosphere based on a pre-defined atmospheric absorption coefficient for the terahertz band and factors such as propagation distance. The simulation results are then compared with the terahertz wave propagation data presented in the multi-scale feature map to determine the atmospheric absorption model loss.

[0087] In the propagation of terahertz waves, the wavefront phase should remain spatially continuous to avoid phase jumps. Phase gradient calculation involves calculating the spatial gradient of the reconstructed wavefront phase in a multi-scale feature map. By calculating the rate of change of phase at adjacent positions, the continuity of the phase is evaluated. The resulting gradient value is used as the phase continuity loss to constrain the reconstruction result and ensure it conforms to the actual propagation characteristics of terahertz waves.

[0088] Data reconstruction loss is an indicator used to measure the difference between the reconstructed model and the desired ideal model during the process of reconstructing a channel environment model using a neural network. It reflects the degree of deviation from reality when reconstructing an environment model based on multi-scale feature maps. By minimizing the data reconstruction loss, the reconstructed model can be made closer to the real channel environment.

[0089] In the weighted summation, several different types of losses—data reconstruction loss, scattering model consistency loss, atmospheric absorption model loss, and phase continuity loss—are added together according to certain weight ratios. Different losses have varying degrees of impact on the reconstruction result. By appropriately setting the weights, the influence of various factors on the reconstruction result can be comprehensively considered, thus obtaining a total loss function that fully reflects the reconstruction error.

[0090] Backpropagation optimization is an algorithm used to train neural networks. Based on the total loss function, it calculates the gradient of the loss function with respect to each parameter in the convolutional neural network encoder, and then updates the parameters in the opposite direction of the gradient, gradually reducing the value of the loss function. By continuously repeating this process, the encoder parameters are adjusted to better process multi-scale feature maps, thereby obtaining more accurate constrained feature maps and optimizing the reconstruction effect of the entire channel environment model.

[0091] For example, the specific method for obtaining the consistency loss of the scattering model by performing scattering characteristic matching analysis on multi-scale feature maps can be as follows: Feature information related to terahertz wave scattering on different material surfaces is extracted from multi-scale feature maps, such as identifying different material regions through pre-defined convolution operations. These features are then compared with known terahertz wave scattering characteristics on various materials. For each material region in the reconstructed environmental model, the difference between its actual scattering response and the corresponding material scattering response in the prior scattering model is calculated, and the sum of the squared differences across all material regions is used as the scattering model consistency loss. For example, if the reconstructed model includes metal and concrete regions, their differences from the known scattering characteristics of metal and concrete are calculated separately, and the sum is used to obtain the scattering model consistency loss.

[0092] For example, the specific method for obtaining the atmospheric absorption model loss by performing atmospheric attenuation simulation analysis on multi-scale feature maps can be as follows: An atmospheric attenuation model is constructed based on the pre-defined atmospheric absorption coefficient and propagation distance in the terahertz band. Propagation information of terahertz waves at different frequencies and locations is obtained from multi-scale feature maps, and this information is substituted into the atmospheric attenuation model to simulate the propagation of terahertz waves in the atmosphere. The simulation results are compared with the actual terahertz wave propagation observed in the multi-scale feature maps, and the differences between the two are calculated. For example, the sum of squared differences between the reconstructed results of different frequency components and the theoretical results considering atmospheric absorption is calculated to obtain the atmospheric absorption model loss.

[0093] For example, the specific method for calculating the phase gradient of a multi-scale feature map to obtain the phase continuity loss can be as follows: For the wavefront phase information reconstructed from the multi-scale feature map, gradient calculations are performed point-by-point in space. For the phase value at each location, the rate of change of its phase value with that of its neighboring locations is calculated, and the phase gradient is obtained through a certain mathematical formula (such as the finite difference method). The sum of the squares of the phase gradients at all locations is used as the phase continuity loss to constrain phase changes, ensuring that the wavefront phase is spatially continuous and avoiding unreasonable phase jumps.

[0094] For example, the total loss function can be obtained by weighted summation of the data reconstruction loss, scattering model consistency loss, atmospheric absorption model loss, and phase continuity loss as follows: Based on the varying importance of different losses to the reconstruction results, weights are pre-defined for data reconstruction loss, scattering model consistency loss, atmospheric absorption model loss, and phase continuity loss. For example, considering the data reconstruction loss to have a significant impact on the overall result, its weight is set to 0.5; the weights for scattering model consistency loss, atmospheric absorption model loss, and phase continuity loss are 0.2 and 0.1 respectively. These four loss values ​​are multiplied by their respective weights and then summed to obtain a comprehensive total loss function that reflects the reconstruction error. This function fully considers various physical characteristics during terahertz wave propagation and the differences between the reconstructed and ideal models.

[0095] For example, the specific method for obtaining the constrained feature map can be as follows: Using the total loss function, the gradient of the loss function with respect to each parameter in the convolutional neural network encoder is calculated via backpropagation. The gradient represents the effect of small changes in the parameters on the loss function. Updating the parameters in the opposite direction of the gradient reduces the value of the loss function. For example, for a specific convolutional kernel parameter in the encoder, its value is adjusted based on the calculated gradient. After multiple iterations of parameter updates, the encoder can better handle multi-scale feature maps, resulting in an optimized encoder. This optimized encoder is then used to process the multi-scale feature maps, yielding constrained feature maps that are optimized in terms of conforming to the physical properties of terahertz waves and their closeness to the ideal reconstruction result.

[0096] In this embodiment, scattering characteristic matching analysis, atmospheric attenuation simulation analysis, and phase gradient calculation are performed on the multi-scale feature maps. The reconstruction process is constrained by the actual physical characteristics of terahertz waves on different material surfaces, such as scattering, atmospheric absorption, and phase continuity, resulting in corresponding loss values. These loss values ​​are weighted and summed with the data reconstruction loss to form a total loss function, comprehensively considering various factors in terahertz wave propagation and the differences between the reconstructed and ideal conditions. Backpropagation optimization of the convolutional neural network encoder based on the total loss function adjusts the encoder parameters, enabling it to process multi-scale feature maps more accurately. The resulting constrained feature maps better conform to the physical characteristics of terahertz waves and are closer to the ideal reconstruction result.

[0097] Preferably, in the embodiment, the step of performing entanglement-sensitive beamforming based on the channel environment model and the channel entanglement relationship corresponding to the channel entanglement tensor to obtain beamforming parameters includes: The channel entanglement tensor is used to calculate the user link correlation degree to obtain the channel entanglement degree between different user links. The channel entanglement degree is used to indicate the channel entanglement relationship between different user links. The channel entanglement degrees between the different user links are sorted from high to low to obtain a priority sequence of user links; Near-field propagation characteristics of the channel environment model were analyzed to obtain Fresnel diffraction model parameters; Based on the priority sequence of the user links and the Fresnel diffraction model parameters, entanglement-sensitive beamforming is performed to obtain the beamforming parameters corresponding to each user link.

[0098] The calculation of user link correlation involves analyzing the relationships between channel parameters related to different user links in the channel entanglement tensor to determine the degree of correlation. In terahertz channels, signals from different user links influence each other, and this influence is reflected in the correlation of channel parameters. By using a pre-defined calculation method, this correlation is quantified to obtain the channel entanglement degree between different user links, thus providing a clear understanding of the channel entanglement relationships between them.

[0099] Channel entanglement is a quantitative indicator used to accurately indicate the degree of channel entanglement between different user links. It reflects the strength of signal mutual influence between different user links in a terahertz channel. The higher the value, the stronger the entanglement between user links, which means that the correlation between them, such as signal interference, is closer, and more closely coordinated processing is required during beamforming.

[0100] The user link priority sequence is formed by arranging the channel entanglement degrees of different user links in descending order. This sequence determines the processing priority of different user links during entanglement-sensitive beamforming. User links with high entanglement degrees will be optimized first during beamforming because they have greater mutual influence, and prioritizing them can better coordinate signals and improve communication quality.

[0101] In the near-field propagation characteristic analysis, the propagation characteristics of terahertz waves in the near-field region are studied based on the channel environment model. In the near-field region, terahertz waves exhibit spherical wave characteristics, and the beam focus shifts nonlinearly with distance, which is significantly different from the planar wave characteristics in the far field. By analyzing the relevant parameters in the channel environment model and combining them with the physical principles of terahertz wave propagation in the near field, Fresnel diffraction model parameters describing these propagation characteristics are derived, providing a basis for accurate beamforming.

[0102] Among them, the Fresnel diffraction model parameters describe the relevant parameters involved in Fresnel diffraction theory when terahertz waves propagate in the near-field region. These parameters include propagation distance, wavelength, aperture size, etc., and they can accurately characterize the propagation characteristics of terahertz waves in the near field, such as beam focusing and focal point shift. By obtaining these parameters, near-field effects can be fully considered during beamforming, allowing the beam to better adapt to the near-field environment and improve communication performance.

[0103] For example, the specific method for obtaining the channel entanglement degree between different user links can be as follows: Channel parameters related to different user links are extracted from the channel entanglement tensor, such as signal amplitude and phase information of different user links. Then, correlation analysis and other methods are used to calculate the degree of correlation between these parameters. For example, by calculating the cross-correlation of signal amplitude and phase of different user links, a numerical value is obtained to quantify their correlation. This numerical value is the degree of channel entanglement between different user links, thus clearly showing the channel entanglement relationship between different user links.

[0104] For example, the specific method for obtaining the priority sequence of user links can be: The calculated channel entanglement degrees between different user links are collected and sorted using algorithms such as bubble sort and quicksort, in descending order of entanglement degree values. Each entanglement degree value corresponds to a user link, and as the entanglement degrees are sorted, the corresponding user links form a sequence. This sequence is the priority sequence of user links, providing a processing order for subsequent beamforming.

[0105] For example, the specific method for obtaining Fresnel diffraction model parameters can be as follows: This study delves into various information within the channel environment model, including the locational relationship between the terahertz base station and the user, and the characteristics of the propagation medium. Based on the Fresnel diffraction theory governing terahertz wave propagation in the near field, and combining relevant parameters in the model, such as propagation distance, frequency (which can be converted to wavelength), and the aperture size of the antenna array or reflector, the Fresnel diffraction model parameters describing the propagation characteristics of terahertz waves in this near-field environment are derived through pre-defined calculation formulas or simulation methods. These parameters include the beam focusing position and focus offset.

[0106] For example, the specific method for obtaining beamforming parameters can be: Based on the priority sequence of user links, those with high entanglement are processed first. During processing, the propagation characteristics of terahertz waves in the near field are considered in conjunction with Fresnel diffraction model parameters. For example, the beam emission angle and shape are adjusted according to the focus offset and focusing position. Through optimization algorithms, such as iterative optimization, the optimal beamforming parameters for each user link, considering near-field effects and entanglement relationships, are calculated. These parameters will be used to configure entanglement-aware reconfigurable smart metasurface units to achieve more efficient communication.

[0107] In this embodiment, the correlation degree of user links is calculated using the channel entanglement tensor, clarifying the channel entanglement relationship between different user links and providing a foundation for the rational allocation of beamforming resources. The priority sequence of user links is obtained by ranking the channel entanglement degree, enabling priority processing of user links with high entanglement degrees during beamforming, effectively coordinating signal interference between user links and improving communication quality. Near-field propagation characteristics analysis of the channel environment model is performed, and Fresnel diffraction model parameters are obtained, fully considering the special propagation characteristics of terahertz waves in the near field, making beamforming more consistent with the actual propagation environment. Entanglement-sensitive beamforming based on the priority sequence and Fresnel diffraction model parameters can calculate the optimal beamforming parameters for each user link.

[0108] Preferably, in an embodiment, the method further includes: The pairwise correlation degree of the channel entanglement tensors corresponding to multiple entangled sensing reconfigurable smart metasurface nodes is calculated to obtain the channel entanglement degree matrix between nodes. Spectral clustering is performed on the inter-node channel entanglement matrix to obtain multiple node subnets, wherein nodes with channel entanglement degrees higher than a preset entanglement degree threshold belong to the same node subnet. Each node in a subnet trains a model based on its local channel dataset, resulting in multiple local models. Multiple local models within the same node subnet are subjected to subnet federated averaging to obtain the subnet global model; The target global model is obtained by performing inter-subnet federated averaging on multiple subnet global models. The target global model is distributed to each entangled sensing reconfigurable smart metasurface node to update the holographic reconstruction engine.

[0109] In the pairwise correlation calculation, for each of the multiple entangled sensing reconfigurable smart metasurface nodes, the correlation between the channel entanglement tensors of any two nodes is calculated one by one. This calculation reveals the closeness of the interrelationships between channel parameters of different nodes, providing a data foundation for subsequent clustering analysis.

[0110] The inter-node channel entanglement matrix arranges the pairwise channel entanglement degrees between multiple entanglement-aware reconfigurable smart metasurface nodes in matrix form. The rows and columns of the matrix correspond to different nodes, and the element values ​​represent the channel entanglement degrees between the corresponding two nodes. This matrix comprehensively and intuitively displays the channel entanglement relationships between all nodes.

[0111] Spectral clustering is a graph-based clustering algorithm that analyzes the channel entanglement matrix between nodes and groups them according to the similarity of their channel entanglement. It utilizes information such as the eigenvalues ​​and eigenvectors of the matrix to divide nodes into different categories, resulting in higher channel entanglement between nodes within the same category and relatively lower channel entanglement between different categories.

[0112] The preset entanglement threshold is a pre-defined standard value for determining whether nodes belong to the same subnet. When the channel entanglement between two nodes is higher than this threshold, they are considered to have a strong correlation and should be classified into the same node subnet; otherwise, they may belong to different subnets. Setting this threshold is crucial for the reasonable division of node subnets.

[0113] The local channel dataset is a collection of data related to the local channel collected and stored by each entanglement-aware reconfigurable smart metasurface node. This data contains various information about the channels surrounding the node, such as signal strength, phase, and multipath propagation characteristics, and serves as the foundational data for model training.

[0114] The subnet federated averaging process involves averaging the parameters of multiple local models trained on local channel datasets for each node within the same subnet. This method integrates the training results of all nodes within the subnet to obtain a global subnet model that represents the overall situation of the subnet, reducing the limitations of data from individual nodes.

[0115] The inter-subnet federated averaging process involves averaging the global models obtained from each of the multiple subnets. By integrating and averaging the global model parameters from different subnets, a target global model that encompasses all network node information is obtained. This allows the model to adapt to the channel characteristics of different subnets and has broader applicability.

[0116] For example, the specific method for obtaining the inter-node channel entanglement matrix can be as follows: For each entanglement-aware reconfigurable smart metasurface node's channel entanglement tensor, the correlation degree is calculated with the channel entanglement tensors of all other nodes. For example, by calculating the correlation of corresponding elements in two channel entanglement tensors, a numerical value representing the channel entanglement degree between the two nodes is obtained. Arranging these calculation results pairwise between all nodes in matrix form yields the inter-node channel entanglement degree matrix. For example, with 5 nodes, a 5×5 matrix is ​​calculated, where the element in the i-th row and j-th column represents the channel entanglement degree between the i-th node and the j-th node.

[0117] For example, the specific method for obtaining multiple node subnets can be as follows: A spectral clustering algorithm is used to process the channel entanglement matrix between nodes. First, an undirected weighted graph is constructed based on the matrix, with nodes as vertices and the channel entanglement between nodes as edge weights. Then, by calculating the eigenvalues ​​and eigenvectors of the graph's Laplacian matrix, appropriate eigenvectors are selected for clustering according to certain rules. For example, nodes with similar element values ​​in their eigenvectors are grouped into one class, resulting in multiple node subnets. During this process, nodes with channel entanglement degrees higher than a preset entanglement degree threshold are assigned to the same node subnet.

[0118] For example, the specific methods for obtaining multiple local models can be as follows: Within each node subnet, each node uses its own collected local channel dataset and employs appropriate machine learning algorithms, such as neural network training algorithms, to train the model. Each node adjusts the model's parameters based on the characteristics of its local data to better fit the features of the local channel, thus obtaining multiple local models that reflect the local channel conditions. For example, in a node subnet with 10 nodes, each node trains a local model based on its own local channel dataset.

[0119] For example, the specific method for obtaining the global model of a subnet can be: The parameters of multiple local models trained on various nodes within the same subnet are aggregated. For each model parameter, the average value of that parameter across all local models is calculated. For example, assuming a local model has 10 parameters, for the first parameter, the sum of the first parameters from all 10 local models is added together and divided by 10 to obtain the average value. This process is repeated to obtain the average value of all parameters. The model composed of these average values ​​is the subnet's global model. In this way, the training information from each node within the subnet is integrated to obtain a more representative model.

[0120] For example, the specific method for obtaining the target global model can be: The subnet global models obtained from multiple subnets are aggregated and averaged. For each parameter in the subnet global model, the average value of that parameter across all subnet global models is calculated, resulting in a target global model that integrates information from all subnets. For example, assuming there are 3 subnets, each with 20 parameters in its global model, the first parameter is summed from the first parameters of the 3 subnet global models, and the result is divided by 3 to obtain the average value. This process is repeated to obtain the average value of all parameters, forming the target global model.

[0121] For example, updating the holographic reconstruction engine can be specifically done as follows: The obtained target global model is transmitted to each entangled-aware reconfigurable smart metasurface node. After receiving the model, the node replaces the original model parameters in the holographic reconstruction engine with the parameters of the target global model, enabling the holographic reconstruction engine to work with this more comprehensive and accurate model, thereby improving the performance and adaptability of holographic reconstruction.

[0122] In this embodiment, pairwise correlations are calculated for the channel entanglement tensors of multiple entanglement-aware reconfigurable smart metasurface nodes, and a matrix is ​​constructed to clearly demonstrate the channel entanglement relationships between nodes. Spectral clustering divides nodes into subnets based on entanglement degree, grouping nodes with high entanglement degrees into the same subnet. This division method fully utilizes the correlations between nodes, improving the efficiency and accuracy of model training. Each node trains a local model based on its local channel dataset, fully exploring local channel characteristics. Federated averaging within a subnet integrates information from nodes within that subnet, reducing the bias of individual node data and resulting in a more representative global subnet model. Federated averaging between subnets further integrates information from each subnet, yielding a target global model adapted to the channel characteristics of different subnets. The target global model is then distributed to update the holographic reconstruction engine, enabling the engine to utilize a more comprehensive and accurate model, effectively improving the accuracy and adaptability of holographic reconstruction.

[0123] To better implement the above methods, this application also provides an artificial intelligence-based unentanglement and reconstruction device, which can be integrated into electronic devices, such as servers or terminals. The terminal may include tablet computers, laptop computers, and / or personal computers.

[0124] For example, see Figure 2 , Figure 2 This is a schematic diagram of the structure of the artificial intelligence-based deentanglement and reconstruction device provided in an embodiment of the present invention, as shown below. Figure 2 As shown, the AI-based de-entanglement and reconstruction device may include an acquisition module 201, an entanglement decoupling module 202, a holographic reconstruction module 203, a beamforming module 204, and a channel optimization module 205, as follows: The acquisition module 201 is used to acquire configuration information of a terahertz base station and multiple entangled sensing reconfigurable smart metasurfaces, control the terahertz base station to transmit downlink pilot signals, and collect entangled sensing data corresponding to the downlink pilot signals through the entangled sensing reconfigurable smart metasurface unit, and construct a channel entanglement tensor based on the configuration information and the entangled sensing data; Entanglement decoupling module 202 is used to compress the channel matrix corresponding to the channel entanglement tensor to the eigenspace for quantum entanglement decoupling processing to obtain eigen parameters, wherein the dimension of the channel matrix is ​​higher than the dimension of the eigenspace; The holographic reconstruction module 203 is used to input the intrinsic parameters and the distorted terahertz hologram collected by the entangled sensing reconfigurable smart metasurface unit into a holographic reconstruction engine equipped with a physically enhanced neural network, so as to perform holographic reconstruction through preset physical constraints and obtain a channel environment model. Beamforming module 204 is used to perform entanglement-sensitive beamforming based on the channel environment model and the channel entanglement relationship corresponding to the channel entanglement tensor to obtain beamforming parameters; The channel optimization module 205 is used to configure the reflection phase and amplitude of the entangled sensing reconfigurable smart metasurface unit according to the beamforming parameters, so as to complete the optimization of the terahertz channel.

[0125] In practice, the above modules can be implemented as independent entities or combined in any way to be implemented as the same or several entities. For the specific implementation of the above modules, please refer to the previous method implementation examples, which will not be repeated here.

[0126] As described above, in this embodiment, by acquiring the configuration information of the terahertz base station and the entanglement-sensing reconfigurable smart metasurface, the base station is controlled to transmit downlink pilot signals and collect entanglement-sensing data, thereby constructing a channel entanglement tensor, providing basic data for subsequent processing. The high-dimensional channel matrix corresponding to the channel entanglement tensor is compressed into a low-dimensional eigenspace. Through quantum entanglement decoupling, the channel dimension is effectively reduced, and mutually independent eigenvalues ​​are extracted, thus decoupling the physical entanglement relationship of the terahertz channel. The eigenvalues ​​and the distorted terahertz hologram are input into a holographic reconstruction engine equipped with a physically enhanced neural network. Holographic reconstruction is performed using preset physical constraints. The physically enhanced neural network is designed for terahertz wave characteristics and can recover a high-fidelity environment model from the distorted hologram, achieving high-precision holographic reconstruction. Entanglement-sensitive beamforming is performed based on the channel environment model and channel entanglement relationship. Considering the correlation of channel parameters and combining near-field propagation characteristics, beamforming effect can be effectively improved, meeting real-time requirements and achieving low latency. By configuring the reflection phase and amplitude of the intelligent metasurface unit according to the beamforming parameters, the terahertz channel is optimized, and the physical entanglement relationship is effectively decoupled and high-precision, low-latency holographic reconstruction is achieved.

[0127] In some embodiments, the AI-based de-entanglement and reconstruction apparatus of the present invention can be implemented in a combination of hardware and software. As an example, the AI-based de-entanglement and reconstruction apparatus of the present invention can be a processor in the form of a hardware decoding processor, which is programmed to execute the AI-based de-entanglement and reconstruction method of the present invention. For example, the processor in the form of a hardware decoding processor can be one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0128] The modules described in the embodiments of this invention can be implemented in software or hardware. The names of the modules are not, in some cases, limiting the scope of the module itself.

[0129] An electronic device according to an embodiment of the present invention includes a processor and a memory, wherein the memory stores multiple instructions; the processor loads instructions from the memory to execute the steps in the artificial intelligence-based de-entanglement and reconstruction method described above. That is, an electronic device according to an embodiment of the present invention may include, but is not limited to: a processor and a memory; the memory is used to store a computer program; the processor is used to execute the artificial intelligence-based de-entanglement and reconstruction method shown in any embodiment of the present invention by calling the computer program.

[0130] In one alternative embodiment, an electronic device is provided, such as Figure 3 As shown, Figure 3 The illustrated electronic device 3000 includes a processor 3001 and a memory 3003. The processor 3001 and the memory 3003 are connected, for example, via a bus 3002. Optionally, the electronic device 3000 may further include a transceiver 3004, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver 3004 is not limited to one type, and the structure of the electronic device 3000 does not constitute a limitation on the embodiments of the present invention.

[0131] Processor 3001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this invention. Processor 3001 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0132] Bus 3002 may include a path for transmitting information between the aforementioned components. Bus 3002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 3002 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 3The bus 3002 is represented by only one thick line, but this does not mean that there is only one bus or one type of bus.

[0133] The memory 3003 may be ROM (Read Only Memory) or other types of static storage devices capable of storing static information and instructions, RAM (Random Access Memory) or other types of dynamic storage devices capable of storing information and instructions, or EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0134] The memory 3003 stores application code (computer program) for executing the present invention, and its execution is controlled by the processor 3001. The processor 3001 executes the application code stored in the memory 3003 to implement the content shown in the foregoing method embodiments.

[0135] Among them, electronic devices can also be terminal devices, which can be any device that can install applications, including at least one of smartphones, tablets, laptops, desktop computers, smart speakers, smartwatches, smart TVs, and smart in-vehicle devices.

[0136] It should be noted that, Figure 3 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0137] An embodiment of the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to execute the steps in any of the above-described artificial intelligence-based de-entanglement and reconstruction methods.

[0138] Alternatively, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, a floppy disk, and an optical data storage device, etc.

[0139] In an exemplary embodiment, a computer program product or computer program is also provided, which includes computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the electronic device to perform the steps in the artificial intelligence-based de-entanglement and reconstruction method described above.

[0140] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0141] It should be understood that the flowcharts and block diagrams in the accompanying drawings illustrate the implementable architecture, functions, and operations of methods and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code, which includes one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0142] The computer-readable storage medium provided in this invention can be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EEPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium that includes or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0143] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the method shown in the above embodiments.

[0144] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0145] It should be noted that the terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and represent a limitation on a specific order or sequence. Where appropriate, the order of use for similar objects can be interchanged so that the embodiments of this application described herein can be implemented in an order other than that shown or described.

[0146] Those skilled in the art will recognize that this invention can be implemented as a system, method, or computer program product. Therefore, this invention can be specifically implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, this invention can also be implemented as a computer program product in one or more computer-readable media, the computer-readable media including computer-readable program code.

[0147] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. An artificial intelligence-based method for unentanglement and reconstruction, characterized in that, include: The configuration information of a terahertz base station and multiple entangled sensing reconfigurable smart metasurfaces is obtained, the terahertz base station is controlled to transmit downlink pilot signals, and entangled sensing data corresponding to the downlink pilot signals is collected through entangled sensing reconfigurable smart metasurface units. Based on the configuration information and the entangled sensing data, a channel entanglement tensor is constructed. The channel matrix corresponding to the channel entanglement tensor is compressed into the eigenspace to perform quantum entanglement decoupling processing and obtain eigen parameters, wherein the dimension of the channel matrix is ​​higher than the dimension of the eigenspace; The intrinsic parameters and the distorted terahertz holograms acquired by the entangled sensing reconfigurable smart metasurface unit are input into a holographic reconstruction engine equipped with a physically enhanced neural network to perform holographic reconstruction through preset physical constraints and obtain a channel environment model. Based on the channel environment model and the channel entanglement relationship corresponding to the channel entanglement tensor, entanglement-sensitive beamforming is performed to obtain beamforming parameters. Based on the beamforming parameters, the reflection phase and amplitude of the entangled sensing reconfigurable smart metasurface unit are configured to complete the optimization of the terahertz channel.

2. The artificial intelligence-based de-entanglement and reconstruction method according to claim 1, characterized in that, The construction of the channel entanglement tensor based on the configuration information and the entanglement-aware data includes: The configuration information is parsed to obtain four-dimensional parameters of the channel tensor, which correspond to the number of base station antennas, the number of users, the number of propagation paths, and the number of frequency sub-bands, respectively. Based on the four-dimensional parameters, the entangled sensing data is subjected to dimensional feature extraction to obtain the channel parameters corresponding to the four-dimensional parameters. The channel parameters include the signal amplitude and phase corresponding to different antennas, different users, different paths, and different sub-bands. The channel parameters are subjected to tensor mapping to obtain the initial channel tensor; The initial channel tensor is normalized in magnitude and phase to obtain the channel entanglement tensor.

3. The artificial intelligence-based de-entanglement and reconstruction method according to claim 1, characterized in that, The step of compressing the channel matrix corresponding to the channel entanglement tensor into the eigenspace for quantum entanglement decoupling to obtain eigenparameters includes: Tensor train decomposition is performed on the channel entanglement tensor to obtain multiple target tensors; The multiple target tensors are iteratively optimized using the alternating least squares algorithm to obtain the optimized target tensors. The termination condition for the iterative optimization is that the reconstruction error satisfies the preset convergence condition. The optimized target tensor is subjected to redundant parameter pruning to obtain the pruned target tensor. The redundant parameters are those whose contribution to the channel representation is lower than a preset contribution threshold. The intrinsic parameters are extracted from the pruned target tensor to obtain the intrinsic parameters, which include the scattering center position, main path delay, and interference subspace information.

4. The artificial intelligence-based de-entanglement and reconstruction method according to claim 1, characterized in that, The physically enhanced neural network includes a convolutional neural network and a transposed convolutional network. The intrinsic parameters and the distorted terahertz hologram acquired by the entanglement-aware reconfigurable intelligent metasurface unit are input into a holographic reconstruction engine equipped with the physically enhanced neural network to perform holographic reconstruction under preset physical constraints, thereby obtaining a channel environment model, including: The distorted terahertz hologram is subjected to noise removal and amplitude-phase calibration to obtain a standardized hologram; The standardized hologram and the intrinsic parameters are input into the encoder of the convolutional neural network to perform multi-scale feature extraction and obtain a multi-scale feature map. According to preset physical constraints, constraints are applied to the multi-scale feature map to obtain a constrained feature map; The constrained feature map is input into the decoder of the transposed convolutional network to reconstruct the environment model and obtain the channel environment model.

5. The artificial intelligence-based de-entanglement and reconstruction method according to claim 4, characterized in that, The step of applying constraints to the multi-scale feature map according to preset physical constraints to obtain a constrained feature map includes: Scattering characteristic matching analysis is performed on the multi-scale feature map to obtain the scattering model consistency loss; Atmospheric attenuation simulation analysis was performed on the multi-scale feature map to obtain the atmospheric absorption model loss; Phase gradient calculation is performed on the multi-scale feature map to obtain the phase continuity loss; The total loss function is obtained by weighted summing of the data reconstruction loss, the scattering model consistency loss, the atmospheric absorption model loss, and the phase continuity loss. Based on the total loss function, the parameters of the encoder of the convolutional neural network are optimized by backpropagation to obtain the optimized encoder, and based on the optimized encoder, the constrained feature map is obtained.

6. The artificial intelligence-based de-entanglement and reconstruction method according to claim 1, characterized in that, The step of performing entanglement-sensitive beamforming based on the channel environment model and the channel entanglement relationship corresponding to the channel entanglement tensor to obtain beamforming parameters includes: The channel entanglement tensor is used to calculate the user link correlation degree to obtain the channel entanglement degree between different user links. The channel entanglement degree is used to indicate the channel entanglement relationship between different user links. The channel entanglement degrees between the different user links are sorted from high to low to obtain a priority sequence of user links; Near-field propagation characteristics of the channel environment model were analyzed to obtain Fresnel diffraction model parameters; Based on the priority sequence of the user links and the Fresnel diffraction model parameters, entanglement-sensitive beamforming is performed to obtain the beamforming parameters corresponding to each user link.

7. The artificial intelligence-based de-entanglement and reconstruction method according to claim 1, characterized in that, The method further includes: The pairwise correlation degree of the channel entanglement tensors corresponding to multiple entangled sensing reconfigurable smart metasurface nodes is calculated to obtain the channel entanglement degree matrix between nodes. Spectral clustering is performed on the inter-node channel entanglement matrix to obtain multiple node subnets, wherein nodes with channel entanglement degrees higher than a preset entanglement degree threshold belong to the same node subnet. Each node in a subnet trains a model based on its local channel dataset, resulting in multiple local models. Multiple local models within the same node subnet are subjected to subnet federated averaging to obtain the subnet global model; The target global model is obtained by performing inter-subnet federated averaging on multiple subnet global models. The target global model is distributed to each entangled sensing reconfigurable smart metasurface node to update the holographic reconstruction engine.

8. An artificial intelligence-based de-entanglement and reconstruction device, characterized in that, include: The acquisition module is used to acquire configuration information of a terahertz base station and multiple entangled sensing reconfigurable smart metasurfaces, control the terahertz base station to transmit downlink pilot signals, and collect entangled sensing data corresponding to the downlink pilot signals through entangled sensing reconfigurable smart metasurface units, and construct a channel entanglement tensor based on the configuration information and the entangled sensing data; The entanglement decoupling module is used to compress the channel matrix corresponding to the channel entanglement tensor into the eigenspace to perform quantum entanglement decoupling processing and obtain eigen parameters, wherein the dimension of the channel matrix is ​​higher than the dimension of the eigenspace; The holographic reconstruction module is used to input the intrinsic parameters and the distorted terahertz hologram collected by the entangled sensing reconfigurable smart metasurface unit into the holographic reconstruction engine equipped with a physically enhanced neural network, so as to perform holographic reconstruction through preset physical constraints and obtain a channel environment model. The beamforming module is used to perform entanglement-sensitive beamforming based on the channel environment model and the channel entanglement relationship corresponding to the channel entanglement tensor, and to obtain beamforming parameters. The channel optimization module is used to configure the reflection phase and amplitude of the entangled sensing reconfigurable smart metasurface unit according to the beamforming parameters, thereby optimizing the terahertz channel.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing multiple instructions; the processor loads instructions from the memory to execute the steps in the artificial intelligence-based unentanglement and reconstruction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions adapted for loading by a processor to execute the steps of the artificial intelligence-based unentanglement and reconstruction method according to any one of claims 1 to 7.

Images