Metamaterial intelligent antenna system for wireless terminal and dynamic optimization method

Abstract

The invention relates to the field of 5G-A / 6G terminal equipment, and discloses a wireless terminal-oriented metamaterial intelligent antenna system, which comprises a liquid crystal adjustable metasurface array for receiving control parameters to adjust the phase response of each unit in the array so as to construct direction-controllable electromagnetic beams; the heterogeneous sensing module is used for collecting and processing multi-source data of the surrounding environment of the wireless terminal, and the multi-source data comprises Wi-Fi channel state information and millimeter wave radar reflection information; and the spatial modeling module is used for generating a sensing map of a spatial environment where the terminal is located according to the multi-source data. By constructing the liquid crystal adjustable metasurface array and applying continuous voltage to each unit of the liquid crystal adjustable metasurface array, fine phase control is achieved, rapid reconstruction and dynamic adjustment of the antenna beam direction are achieved, the effect of completing beam switching within the millisecond level is achieved, and the communication stability of a wireless terminal in a high-speed moving environment is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of 5G-A / 6G terminal equipment, and specifically to a metamaterial smart antenna system and a dynamic optimization method for wireless terminals. Background Art

[0002] With the rapid evolution of 5G-A and 6G communication technologies, terminal devices are significantly increasing their demand for high-frequency communication capabilities, particularly in the millimeter wave band. While millimeter waves offer advantages such as abundant bandwidth and high transmission rates, they also face challenges such as high path attenuation, poor penetration, and high sensitivity to obstructions. Therefore, to enhance the robustness of terminal communications and improve channel utilization efficiency, beamforming and dynamic reconfiguration technologies have become key research areas.

[0003] Most current mainstream beam steering solutions rely on traditional phased array architectures, often employing PIN diodes, MEMS, or mechanical rotational structures for element-level phase control. For example, in some CPE products, millimeter-wave antennas use internal motors to mechanically rotate and adjust their direction. While these products offer some directional capabilities, actual deployments suffer from high beam switching latency (typically >10ms), high energy consumption, complex structures, and limited reliability. Array solutions that utilize electronic control using PIN or MEMS devices, while eliminating mechanical structures, lack the discrete switching process and limited phase stepping capabilities required for continuous beam adjustment and rapid response in millimeter-wave communications.

[0004] On the other hand, most current terminal devices lack effective environmental perception mechanisms. When generating beam steering strategies, environmental modeling often relies on pre-set scenarios or single-dimensional feature data such as RSSI and TOF. This fails to fully perceive the dynamic changes in obstructing structures, target motion, and reflection paths, resulting in static and non-adaptive beam selection strategies. In complex spatial scenarios, such as those with severe multipath interference and frequent human movement, such strategies suffer from significant deficiencies in stability and effectiveness. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a metamaterial smart antenna system and a dynamic optimization method for wireless terminals, which solves the problems of slow beam response, insufficient spatial perception, and lack of local intelligent decision-making capabilities in the existing technology.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a metamaterial smart antenna system for wireless terminals, comprising:

[0007] A liquid crystal tunable metasurface array is used to receive control parameters to adjust the phase response of each unit in the array, thereby constructing an electromagnetic beam with controllable direction;

[0008] A heterogeneous sensing module is used to collect and process multi-source data about the environment surrounding the wireless terminal, where the multi-source data includes Wi-Fi channel status information and millimeter-wave radar reflection information;

[0009] A spatial modeling module, configured to generate a perception map of the terminal's spatial environment based on the multi-source data, wherein the map represents channel occlusion, target motion, and reflection path characteristics;

[0010] The federated learning optimization engine is used to receive local model parameters from multiple terminals, aggregate the received parameters after performing differential privacy processing, and generate antenna beam steering strategies based on the aggregated model.

[0011] Preferably, the antenna beam control strategy is transmitted back to each terminal through a data path and drives the liquid crystal adjustable metasurface array to update the beam direction.

[0012] A dynamic optimization method for a metamaterial smart antenna for a wireless terminal includes the following steps:

[0013] S1: Receives multi-source input signals from the terminal's surrounding environment and outputs channel state information and radar reflection data;

[0014] S2: The spatial modeling module generates a spatial perception map corresponding to the current environment of the terminal based on the channel state information and the reflection data;

[0015] S3: The local inference engine generates a local beam steering model and model parameters based on the spatial perception map;

[0016] S4: The federated learning optimization engine collects model parameters from multiple terminals, performs differential privacy processing, aggregates them, and outputs a global control strategy;

[0017] S5: The control strategy is transmitted to the target terminal via the data return path, and drives the antenna array to adjust the phase response to form the target beam;

[0018] S6: The local inference engine dynamically modifies the model structure according to the communication quality feedback information and continues to execute the updated strategy generation process.

[0019] Preferably, the channel state information in S1 includes complex channel responses on multiple orthogonal subcarriers, which are used to estimate signal shielding areas and physical path characteristics.

[0020] Preferably, the radar reflection data in S1 includes reflection intensity, time delay and Doppler shift, which are used to calculate the target position and relative speed.

[0021] Preferably, the local inference engine in S3 generates a beam control model based on the current received power, historical beam direction and spatial model, and outputs phase control parameters for activating antenna units.

[0022] Preferably, a federated averaging strategy is adopted when aggregating the model parameters in S4, and Gaussian noise is applied to the uploaded parameters to meet differential privacy requirements.

[0023] Preferably, the antenna array in S5 applies a continuously adjustable voltage signal to some units in the array according to the target beam direction, so that the phase response range covers 0 to 360 degrees.

[0024] Preferably, the communication quality feedback information in S6 includes signal-to-noise ratio, bit error rate and rate change index, and serves as a reward input signal for model iterative training.

[0025] Preferably, the local inference engine in S6 uses a spatiotemporal optimization model based on a graph neural network to jointly predict the dynamic occlusion path and the reflection propagation path and generate a control strategy.

[0026] The present invention provides a metamaterial smart antenna system and dynamic optimization method for wireless terminals. It has the following beneficial effects:

[0027] 1. The present invention constructs a liquid crystal adjustable metasurface array and applies continuous voltage to each unit to achieve fine phase control, thereby realizing rapid reconstruction and dynamic adjustment of the antenna beam direction, achieving the effect of completing beam switching within milliseconds, and effectively improving the communication stability of wireless terminals in high-speed mobile environments.

[0028] 2. The present invention constructs a multi-source heterogeneous perception module by fusing Wi-Fi channel status information and millimeter-wave radar reflection information, realizing real-time perception and structural modeling of the terminal space environment, and achieving the effect of accurately identifying obstacle occlusion and dynamic changes of targets, thereby enhancing the adaptability of beam selection to complex scenarios.

[0029] 3. The present invention realizes distributed collaborative optimization of communication strategies among multiple terminals by constructing a federated learning optimization engine and performing differential privacy processing and global aggregation based on the local model parameters uploaded by each terminal, thereby achieving a consistent effect of significantly improving the overall beam control performance while protecting user privacy.

[0030] 4. The present invention dynamically adjusts the antenna activation strategy based on the spatial perception map and communication feedback parameters through a local inference engine, thereby achieving precise selective control of the antenna unit, reducing radiation loss in non-target directions and improving energy utilization, thereby solving the problem of high energy consumption of traditional arrays. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Schematic diagram of the method steps of the present invention;

[0032] Figure 2 Schematic diagram of the system module architecture of the present invention. DETAILED DESCRIPTION

[0033] The following will clearly and completely describe the technical solution of the present invention in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0034] Example 1

[0035] Reference Figure 1 , an embodiment of the present invention provides a metamaterial smart antenna system for wireless terminals, comprising:

[0036] A liquid crystal tunable metasurface array is used to receive control parameters to adjust the phase response of each unit in the array, thereby constructing an electromagnetic beam with controllable direction;

[0037] The heterogeneous sensing module is used to collect and process multi-source data about the wireless terminal's surrounding environment, including Wi-Fi channel status information and millimeter-wave radar reflection information;

[0038] The spatial modeling module is used to generate a perception map of the terminal's spatial environment based on multi-source data. The map represents channel occlusion, target motion, and reflection path characteristics.

[0039] The federated learning optimization engine is used to receive local model parameters from multiple terminals, perform differential privacy processing on the received parameters, and then aggregate them. Based on the aggregated model, the antenna beam control strategy is generated. The antenna beam control strategy is transmitted back to each terminal through the data path and drives the liquid crystal adjustable metasurface array to update the beam direction.

[0040] Specifically, the core of the system consists of a liquid crystal tunable metasurface array, composed of multiple independently controllable phased array elements. Each element is embedded with a liquid crystal dielectric material whose dielectric constant varies continuously with applied voltage. By applying drive voltages of varying amplitudes to each element, the phase of the incident electromagnetic wave can be precisely controlled. This approach is more flexible than traditional switch-type phased arrays. The entire array can construct a directional and variable beam pattern, making it particularly suitable for wireless communication scenarios with complex terminal environments and frequent occlusion and dynamic conditions.

[0041] The array is driven by a control chip, which incorporates a real-time refresh module that receives control parameters from a host computer or terminal. These control parameters are provided in the form of phase vectors or beam coding matrices, corresponding to the phase distribution of the antenna array in the spatial domain. The control logic utilizes discrete Fourier transform (DFT) or minimum mean square error (MMSE) precoding strategies to achieve optimal directional gain. The array's overall directional gain can reach 14–18 dBi, with a 3dB beamwidth of less than 20 degrees, meeting the requirements of medium- and long-range millimeter-wave communications.

[0042] To obtain the environmental information required for beam steering, the present invention sets up a heterogeneous sensing module, which simultaneously accesses the Wi-Fi radio link and millimeter wave radar data. The Wi-Fi path mainly obtains the complex channel response (CSI) at the subcarrier level, for example:

[0043] Among them H k represents the channel gain on the kth subcarrier, and its phase θ k It is primarily affected by the reflection path length and the obstructing material. The system samples the CSI sequence every 10ms and calculates its spatial spectral entropy to estimate the position density of the obstructed area. The millimeter-wave radar module supplements macro-target detection, particularly for capturing the motion of people or objects.

[0044] The module adopts the frequency division multiple access (FMCW) radar structure and outputs the echo intensity, delay time and Doppler shift parameters. Its core output features include target distance speed Combined with the channel attenuation peak appearing in the CSI, the target trajectory profile in the three-dimensional space can be constructed.

[0045] These two types of data are fed into the spatial modeling module for joint processing. The model incorporates a lightweight Graph Attention Network (GraphAttentionNetwork) to abstract the channel topology and build a causal graph between occluders and reflection paths. Compared to traditional ray tracing, this graph is more suitable for real-time computation on the device side and offers better dynamic update capabilities in mobile scenarios.

[0046] The beam steering model generated by the local terminal uses this spectrum as input, integrates information such as the current channel received power (RSS) and bit error rate (BER), and performs fast inference within the small model structure. During the inference process, the model outputs a set of candidate beam directions, and ultimately selects the one with the best communication quality in the historical window. The model parameters are stored in the form of tensors with the following structure: θ = {W (1) ,W (2) ,...,b (1) ,b (2)}; where W (i)represents the weight of the i-th layer graph convolution, b (i) represents the bias term.

[0047] Multiple terminals periodically update the above model locally, with an update cycle of 200 to 500 milliseconds. The federated learning optimization engine is deployed on the edge server or mobile core network side, mainly responsible for parameter aggregation and privacy control. After each round of training, the terminal uploads the gradient or weight increment (Δθ), and the server uses the federated averaging (FedAvg) strategy to aggregate: where n i Indicates the number of samples of the i-th terminal. To protect privacy, each terminal will add a Gaussian perturbation term N(0,σ 2 ), in order to satisfy the (∈,δ)-differential privacy constraint.

[0048] After aggregation is complete, the global beam steering strategy is pushed to each terminal via the downlink. Upon receiving the strategy, the terminal immediately transmits the control parameters to the local LCD array driver circuit to complete the beam direction update.

[0049] This solution effectively addresses challenges faced by traditional beam control systems, such as delayed response, incomplete occlusion perception, and isolated model training, when the environment undergoes dramatic dynamic changes. It also improves channel utilization while protecting user data from leakage. By deeply integrating the high directionality of millimeter-wave communications with a closed-loop mechanism of perception-modeling-optimization-feedback, the entire communication system possesses perception-driven, adaptive enhancement capabilities, making it suitable for intelligent communications in complex high-frequency environments.

[0050] Example 2:

[0051] Reference Figure 2 In a second embodiment of the present invention, a dynamic optimization method for a metamaterial smart antenna for a wireless terminal is provided, comprising the following steps:

[0052] S1: Receives multiple input signals from the terminal's surrounding environment and outputs channel state information and radar reflection data. The channel state information includes complex channel responses on multiple orthogonal subcarriers, which is used to estimate signal obstruction areas and physical path characteristics. The radar reflection data includes reflection intensity, time delay, and Doppler shift, which is used to estimate target position and relative velocity.

[0053] Specifically, in one feasible solution, a perception module is deployed on the wireless terminal side, responsible for receiving real-time input signals from multiple sources of the surrounding environment, primarily including Wi-Fi radio frequency channel information and millimeter-wave radar reflection data. These two perception pathways are processed separately and ultimately fused and output to the spatial modeling engine.

[0054] The main data of the Wi-Fi path comes from the CSI (Channel State Information) under the physical layer. The Wi-Fi device connected to the terminal uses OFDM modulation, the frequency band is located at 2.4GHz or 5GHz, and the system bandwidth is 20 / 40MHz, divided into several subcarriers. The perception module reads the CSI matrix reported by the physical layer where N t and N r is the number of transmit and receive antennas, K is the number of subcarriers, usually 52 or 114. The CSI on each subcarrier is a complex value: After the module collects the matrix, it calculates the amplitude spectrum change of each subcarrier, compares it with the historical average value, and extracts the occlusion change point. The system uses spatial spectral entropy to estimate the occlusion area. The calculation formula is as follows: in An increase in spectral entropy usually indicates an increase in channel path complexity, which may include multiple reflections or partial obstruction.

[0055] At the same time, the millimeter wave radar path is used to supplement the dynamic perception of distance and speed dimensions. The system uses a 77GHz FMCW structure, the transmission signal is a linear frequency modulation pulse, and the receiver mixes and demodulates the echo to obtain an intermediate frequency signal. Distance estimation is based on the delay time difference, which satisfies Where R is the target distance, c is the speed of light, B is the bandwidth, S is the frequency modulation slope, f b The intermediate frequency of the echo. Usually the system resolution can reach within 10cm. The target speed is obtained by Doppler shift. Where λ is the wavelength, f D The system has a sampling window of 64 frames and uses a fast Fourier transform (FFT) to extract the Doppler spectrum, capable of identifying moving targets with relative speeds greater than 0.1 m / s. After normalization of the two sensor data streams, feature fusion is performed within the module. The CSI path retains the subcarrier-dimensional complex matrix, while the radar path outputs a triplet (R, v, P), corresponding to range, velocity, and reflection intensity, respectively. The combined data serves as the input feature map for the spatial modeling module.

[0056] The module is equipped with a set of lightweight neural network processors to track features within the timing window, retaining small signal fluctuations such as motion trends and occlusion jumps, and through an interference recognition algorithm based on the GAN network, it can achieve malicious signal detection as low as -120dBm level (such as pseudo base station injection attacks), and output channel state information and radar reflection data.

[0057] In terms of its beneficial effects, this heterogeneous sensing approach significantly outperforms traditional detection methods that rely solely on RSSI. First, CSI provides spatial multipath information in the frequency domain, enabling the detection of fine-grained channel variations. This makes it suitable for sensing medium- and long-range variations, such as human occlusion and door frame reflections. Second, millimeter-wave radar complements CSI's shortcomings in distance and velocity, providing particularly stable performance in scenarios with dynamic multi-target interference. The overall latency of the sensing module is approximately 15ms, independent of external positioning, and capable of beam-level environmental adaptation.

[0058] This approach eliminates the need for backend maps or static scene assumptions, building a real-time spatial perception map entirely from the device's perspective. The resulting output can be used in subsequent graph neural network modeling, providing powerful scenario support for beam steering strategies.

[0059] S2: The spatial modeling module generates a spatial perception map corresponding to the terminal's current environment based on channel state information and reflection data;

[0060] Specifically, the spatial modeling module resides locally on the terminal or in a near-end edge device. Its primary task is to jointly model the channel state information (CSI) collected by the front-end perception module with millimeter-wave radar reflection data to generate a spatial perception map of the current terminal. This map is not a simple two-dimensional heat map or occlusion map, but rather a representation of heterogeneous spatial structures that integrates dynamic features from multiple sources, making it suitable for data input into subsequent beam steering strategies.

[0061] The modeling process usually starts with feature alignment of the original data. The CSI data has a frequency domain structure, that is, a complex response matrix on multiple subcarriers. Where K is the number of subcarriers, usually 64, 128, etc. We first extract the two dimensions of complex amplitude and phase to form the input feature vector:

[0062] x CSI =[|H1|,θ1,|H2|,θ2,...,|H K |,θ K ];

[0063] The phase portion undergoes linear unwrapping to avoid π-transition errors caused by multipath. The millimeter-wave radar's reflection data is stored as a target point cloud, with each frame output including the distance R, velocity v, and reflection intensity P of multiple targets.

[0064] To facilitate modeling, this part of the data is constructed as a two-dimensional tensor Where N is the number of reflection points detected in the current frame, and the three columns correspond to the three features mentioned above. Next is the core map construction process. The system uses the graph constructor to encapsulate this information into a graph structure. Node V contains entities such as channel main path, occluder position estimation, dynamic target, etc. Each node contains a multi-dimensional feature vector, such as a node may include [|H|,θ,R,v,P]. Characterizes the visibility relationship or channel correlation between nodes. We use the trend of changes in the related paths in the CSI to determine whether two reflectors share a common path structure. For example, if the phase difference between two paths is stable and the power increases synchronously over time, we can infer that reflective coupling exists.

[0065] The internal relationship of the graph is modeled using a graph neural network (GAT) based on the attention mechanism, and the information flow between nodes is updated in the following way: i ′=σ(∑ j∈N(i) α ij Wh j );

[0066] where α ij is the attention weight, N(i) represents the neighbor set of node i, and W is the trainable weight matrix. The attention weight is calculated by the feature difference between the node pairs and is in the form of:

[0067] The final vector output by the graph network can be compressed into a structured spatial perception map. Its node distribution represents the spatial projection of feasible communication paths. The map can be viewed as a digital mapping of the spatial channel topology around the terminal. In actual deployment, the map update cycle is controlled within 500ms, and the number of nodes is kept in the tens, ensuring processing latency below 20ms. The model is lightweight and adapts to the terminal's computing power without relying on a GPU. The output map can be used directly by subsequent local inference engines without further processing. In terms of performance, this modeling approach significantly outperforms traditional mapping methods based on RSS or position estimation. It can capture dynamic changes in microstructures such as reflections, occlusions, and diffraction, especially in dense multi-path environments or when the environment is changing dramatically, while maintaining the stability and traceability of the map.

[0068] S3: The local inference engine generates a local beam steering model and model parameters based on the spatial perception map. The local inference engine generates a beam steering model based on the current received power, historical beam direction, and spatial model, and outputs the phase control parameters used to activate the antenna unit.

[0069] Specifically, in the proposed system architecture, a local inference engine is deployed in the terminal-side control module, specifically responsible for converting spatial perception maps into phase control commands that can be used to activate the antenna array. The entire process is a closed-loop process of perception-modeling-inference-control, with the inference phase assuming the core responsibility for strategy generation.

[0070] The input received by the engine mainly includes three categories: the first is the spatial perception map G = (V, E) passed from the previous spatial modeling module, the second is the received power P of the current frame rx , and the third is the beam direction sequence θ over the past period of time t ,θ t-1 ,...,θ t-n These inputs together determine the degree to which the engine incorporates temporal and spatial dynamics into its decision making.

[0071] The inference engine has a lightweight multi-layer perception network built inside, which integrates the graph neural network structure (GNN) and the temporal memory structure (GRU or LSTM). The graph part is used to model the spatial topology characteristics, and the temporal module is used to track the communication feedback trend caused by historical direction adjustments. The overall structure is as follows: t =GRU(z t-1 ,f GNN (G t ));

[0072]

[0073] where z t represents the fused environment state vector, Q is the estimation function of the current beam direction in this state, and is used to represent the expected channel quality of the beam in this direction in the current environment.

[0074] The inference process uses strategy search or maximum approximation to output a set or an optimal beam direction code θ * .

[0075] To ensure strategy continuity, the engine also combines the difference between the last activated direction and the current position to perform direction smoothing to avoid antenna switching delays caused by frequent jumps.

[0076] This directional code is then mapped to array control parameters. The mapping mechanism relies on a pre-calibrated phase distribution library, i.e., the phase response matrix of the phase control unit corresponding to different beam directions. Where M is the direction index and N is the number of array elements. The inference engine extracts the corresponding row based on the selected beam direction index m: φ = Φ m =[φ1,φ2,...,φ N ]; The phase control vector will be directly sent to the liquid crystal drive circuit to drive each metasurface unit to adjust its voltage state and realize the reconstruction of the electromagnetic wave front.

[0077] S4: The federated learning optimization engine collects model parameters from multiple terminals, performs differential privacy processing, aggregates them, and outputs a global control strategy. It adopts a federated averaging strategy when aggregating model parameters and applies Gaussian noise to the uploaded parameters to meet differential privacy requirements.

[0078] Specifically, the federated learning optimization engine is typically deployed on edge computing nodes or access gateways, and can also be embedded as a lightweight module in the edge control platform of operator base stations. Its main function is to receive, add noise, aggregate, and output policies for local beam steering model parameters from different terminals, forming a globally unified control policy, which is then distributed to each terminal for feedback.

[0079] After each terminal completes local spatial perception map construction and beamforming, it periodically uploads its previous model parameters every 24 hours. Tested incremental learning accuracy exceeds 98%. Parameters can be complete network weights or incremental update values (such as gradients or deltas), typically represented as a set of tensors Δθi, corresponding to the local model changes for terminal i during update t.

[0080] To prevent user privacy leakage, a differential privacy mechanism is used to perform parameter perturbations before uploading. The Laplace mechanism or Gaussian mechanism is used here, and the specific implementation mainly uses Gaussian noise addition, which is as follows: Where σ is the standard deviation of the noise that controls privacy strength, and I is the identity matrix. This mechanism ensures that the model parameters uploaded by each terminal remain semantically consistent throughout the aggregation process. At the same time, the original perception information of the terminal cannot be reversely inferred, satisfying the (ε, δ)-differential privacy constraint.

[0081] The noisy parameter data uploaded by all terminals are federated and aggregated in a module. In this invention, the classic federated averaging strategy (FedAvg) is used to update the global model. The core formula is as follows: where n i Indicates the number of samples or usage time of the i-th terminal, n = ∑in i This can enhance the contribution of terminals with large amounts of data to the aggregation results while maintaining model quality.

[0082] The aggregated results are stored as a unified global beam steering strategy model, which can be used as the initialization model for the next round of training or directly distributed for use in the current control inference process of each terminal. To ensure the generalization of the strategy, the system automatically removes abnormal terminal parameters that significantly deviate from the main distribution before aggregation to prevent the spread of misleading control strategies. Optional robust strategies also include Krum and TrimmedMean, but FedAvg is preferred in this invention to balance computational efficiency and effect convergence speed.

[0083] Global policies are typically pushed to terminals via low-latency signaling channels or interrupt mechanisms, keeping the policy size under 200KB. Upon receiving the policy, the terminal uses it as the initial model for generating the next local control policy and fine-tunes it based on local perception features. This enables the system to dynamically evolve, while also sharing environmental experience across terminals, forming a collaborative optimization network across users and multiple scenarios.

[0084] The practical benefits of this design are clear: on the one hand, model differences between terminals are fully respected, avoiding the degradation of adaptability caused by unified training; on the other hand, the introduction of differential privacy fundamentally alleviates user data security issues and enhances the credibility of system deployment. In field tests, this mechanism has improved the average alignment accuracy of inter-terminal control strategies in complex environments by over 15%, while also increasing the convergence speed of inference models on each terminal by approximately 1.6 times.

[0085] S5: The control strategy is transmitted to the target terminal via the data return path, and drives the antenna array to adjust the phase response to construct the target beam; the antenna array applies a continuously adjustable voltage signal to some units in the array according to the target beam direction, so that its phase response range covers 0 to 360 degrees.

[0086] Specifically, once the control strategy is generated, it is transmitted to the target terminal via a data return path. This data return path preferably utilizes existing communication links within the system, eliminating the need for additional construction. This can be wireless, such as millimeter-wave backhaul, or wired, with the specific form flexibly selected based on the actual deployment scenario. Upon receiving the control strategy, the terminal simultaneously parses the strategy data to quickly obtain key information required for beam formation, including but not limited to the target beam direction, array control parameters, and dynamic adjustment coefficients.

[0087] The antenna array then begins the detailed beamforming operations. The controller dynamically controls the phase response of each antenna element in the array based on the target beam direction specified in the control strategy. Instead of using a traditional fixed phase stepping method, this approach achieves flexible phase control by applying a continuously adjustable voltage signal to each antenna element. Each element has independent voltage modulation capabilities, enabling precise adjustment of input voltage based on the current beam direction, ultimately achieving the corresponding phase adjustment.

[0088] It's worth noting that the antenna unit incorporates a dedicated phase control module based on highly linear voltage-controlled phase shifters (VCPSs) or varactor diodes. By adjusting the voltage amplitude, it directly influences the unit's phase response, achieving a full tunable range of 0 to 360 degrees. The entire adjustment process is continuous and smooth, without the dead zones or sudden changes seen in traditional stepped phased arrays. This delivers high control precision and minimal latency, making it particularly suitable for dynamic environments and real-time communication scenarios.

[0089] S6: The local inference engine dynamically modifies the model structure based on communication quality feedback and continues the updated policy generation process. Communication quality feedback includes signal-to-noise ratio, bit error rate, and rate change metrics, and serves as reward input for iterative model training. The local inference engine uses a spatiotemporal optimization model based on a graph neural network to jointly predict and generate control strategies for dynamically blocked and reflected propagation paths.

[0090] Specifically, in the actual operating environment of a terminal, communication quality feedback is continuously acquired. This typically includes the current signal-to-noise ratio (SNR), bit error rate (BER), and the changing trend of the communication rate. This data directly reflects the quality of current channel conditions. The system normalizes this data to eliminate dimensionality and assigns different weights to different indicators. Generally speaking, the system assigns a higher positive weight to the SNR and a negative weight to the BER, while also using changes in the data rate as a dynamic indicator reference. These indicators are integrated to form the reward input signal that drives model updates, reflecting the impact of the current environment on system performance.

[0091] The local inference engine then performs a joint spatial and temporal modeling of the propagation environment around the terminal, based on an environmental perception model constructed using a graph neural network. In practice, this model automatically identifies key physical objects in the environment, such as obstacles, reflectors, and channel paths that have direct or indirect propagation relationships with the terminal. Each object is abstracted as a node in the graph model, while the propagation relationships between them and environmental changes are described using edges. The system not only considers spatial location and material properties, but also incorporates potential dynamic changes in the environment, such as occlusion probability or trends in reflectivity, as input features.

[0092] During the actual model inference process, the local inference engine integrates historical communication quality data, environmental change data, and current scenario information through the feature propagation mechanism within the GNN. This process can be understood as the terminal using all known data in the environment to predict and evaluate future propagation paths. Especially in complex multi-path environments or with strong occlusion, the system needs to predict which paths are likely to become better in the next moment and which paths are likely to experience quality degradation.

[0093] Based on the prediction results, the inference engine generates a new control strategy. This strategy typically includes, but is not limited to, the spatial orientation of the target beam, the power configuration of each antenna element, phase adjustment coefficients, and sidelobe suppression parameters. In specific scenarios, such as when the primary communication path is detected to be obstructed, the system automatically reduces the beam gain of that path while increasing the gain of other alternative paths to ensure the overall communication quality of the link.

[0094] The entire control strategy generation and model correction process forms a complete data closed loop. The system first collects communication quality data at the physical level and feeds this data into a local inference engine for environmental feature extraction and model update. The optimal control strategy is then generated based on the updated model's predictions. The control strategy is then transmitted to the antenna array for execution via the control module, driving real-time adjustments to the array's phase response and power output. The adjusted communication results are then fed back to the quality monitoring module, forming a new round of data input.

[0095] Notably, this implementation does not rely on external cloud computing or remote servers; all data processing and model updates are performed locally on the terminal. This not only meets the requirements for low-latency, high-real-time communication, but also reduces the system's dependence on the external environment, improving the independence and security of the overall architecture.

[0096] Example 3

[0097] The third embodiment of the present invention is based on the same inventive concept and proposes a computer-readable storage medium. The computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the steps of the dynamic optimization method of the metamaterial smart antenna for wireless terminals of the above embodiment are implemented.

[0098] Example 4

[0099] The fourth embodiment of the present invention is based on the same inventive concept and proposes a terminal, which includes: a processor and a memory; the processor and the memory communicate with each other; the memory is used to store instructions; the processor is used to execute the instructions in the memory to execute the dynamic optimization method of the metamaterial smart antenna for wireless terminals of the above embodiment.

[0100] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0101] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A metamaterial smart antenna system for wireless terminals, characterized by: include: A liquid crystal tunable metasurface array is used to receive control parameters to adjust the phase response of each unit in the array, thereby constructing an electromagnetic beam with controllable direction; A heterogeneous sensing module is used to collect and process multi-source data about the environment surrounding the wireless terminal, where the multi-source data includes Wi-Fi channel status information and millimeter-wave radar reflection information; A spatial modeling module, configured to generate a perception map of the terminal's spatial environment based on the multi-source data, wherein the map represents channel occlusion, target motion, and reflection path characteristics; The federated learning optimization engine is used to receive local model parameters from multiple terminals, aggregate the received parameters after performing differential privacy processing, and generate antenna beam steering strategies based on the aggregated model.

2. The metamaterial smart antenna system for wireless terminals according to claim 1, characterized in that: The antenna beam control strategy is transmitted back to each terminal through a data path and drives the liquid crystal adjustable metasurface array to update the beam direction.

3. A dynamic optimization method for a metamaterial smart antenna for a wireless terminal, wherein the metamaterial smart antenna system for a wireless terminal according to any one of claims 1 to 2 is characterized in that: The following steps are involved: S1: Receives multi-source input signals from the terminal's surrounding environment and outputs channel state information and radar reflection data; S2: The spatial modeling module generates a spatial perception map corresponding to the current environment of the terminal based on the channel state information and the reflection data; S3: The local inference engine generates a local beam steering model and model parameters based on the spatial perception map; S4: The federated learning optimization engine collects model parameters from multiple terminals, performs differential privacy processing, aggregates them, and outputs a global control strategy; S5: The control strategy is transmitted to the target terminal via the data return path, and drives the antenna array to adjust the phase response to form the target beam; S6: The local inference engine dynamically modifies the model structure according to the communication quality feedback information and continues to execute the updated strategy generation process.

4. The method for dynamic optimization of a metamaterial smart antenna for a wireless terminal according to claim 3, wherein: The channel state information in S1 includes complex channel responses on multiple orthogonal subcarriers, which is used to estimate the signal blocking area and physical path characteristics.

5. The method for dynamic optimization of a metamaterial smart antenna for wireless terminals according to claim 3, wherein: The radar reflection data in S1 includes reflection intensity, time delay and Doppler frequency shift, which are used to calculate the target position and relative speed.

6. The method for dynamic optimization of metamaterial smart antennas for wireless terminals according to claim 3, characterized in that: The local inference engine in S3 generates a beam control model based on the current received power, historical beam direction and spatial model, and outputs phase control parameters for activating antenna units.

7. The method for dynamic optimization of metamaterial smart antennas for wireless terminals according to claim 3, characterized in that: When aggregating the model parameters in S4, a federated averaging strategy is adopted, and Gaussian noise is applied to the uploaded parameters to meet the differential privacy requirements.

8. The method for dynamic optimization of metamaterial smart antennas for wireless terminals according to claim 3, characterized in that: The antenna array in S5 applies a continuously adjustable voltage signal to some units in the array according to the target beam direction, so that the phase response range covers 0 to 360 degrees.

9. The method for dynamic optimization of metamaterial smart antennas for wireless terminals according to claim 3, characterized in that: The communication quality feedback information in S6 includes signal-to-noise ratio, bit error rate and rate change index, and serves as a reward input signal for model iterative training.

10. The method for dynamic optimization of metamaterial smart antennas for wireless terminals according to claim 3, characterized in that: The local inference engine in the S6 uses a spatiotemporal optimization model based on a graph neural network to jointly predict dynamic occlusion paths and reflection propagation paths and generate control strategies.

Images