Hybrid RIS-assisted communication and induction integrated system and safety prediction beam forming method thereof

Abstract

The invention provides a hybrid RIS-assisted communication and sensing integrated system and a safety prediction beam forming method thereof, and the method comprises the steps: an offline stage: generating a training data set which comprises a plurality of training samples; constructing a deep expansion prediction network, and constructing a target function of the joint optimization problem into a loss function of the deep expansion prediction network; performing unsupervised end-to-end training on the deep expansion prediction network by using the training data set and minimizing a loss function; the online stage comprises the following steps: collecting a historical channel state information sequence, inputting the historical channel state information sequence into an encoder module of the deep expansion prediction network, and generating a feature vector containing a next time slot; inputting the feature vector into a decoder, wherein the decoder simulates an iteration process of an alternating direction multiplier method through a plurality of cascaded network layers in the decoder; outputting a joint optimization strategy of the next time slot through one-time forward propagation; and at the beginning of the next time slot, signal emission and reflection are carried out according to a joint optimization strategy.

Description

Technical Field

[0001] This invention relates to the field of communications, and more particularly to a hybrid RIS-assisted integrated sensing system and its secure predictive beamforming method. Background Technology

[0002] Integrated Sensing and Communication (ISAC), as a key direction for the evolution of future sixth-generation (6G) wireless networks, achieves deep integration of communication and sensing functions by sharing hardware platforms and spectrum resources, thereby significantly improving system efficiency and spectrum utilization. However, in actual deployment, ISAC systems still face severe challenges. Especially in sensing, signals experience severe attenuation during propagation. When the signal passes through the reflection link of "base station-sensing target-base station," especially after passing through multi-hop links such as "base station-reconfigurable intelligent surface (RIS)-sensing target-RIS-base station," the signal strength drops sharply, forming a serious "double path loss" problem. This greatly limits the system's sensing accuracy and detection range for distant or small reflective cross-section targets.

[0003] To address this challenge, reconfigurable intelligent metasurface (RIS) technology has been introduced into ISAC systems. As an artificial electromagnetic surface composed of numerous low-cost passive components, RIS can actively shape the propagation environment by intelligently controlling the phase, amplitude, and other electromagnetic properties of the incident signal. However, traditional passive RIS, which can only passively reflect signals and cannot provide signal gain, still suffers from inherent "double path loss" when dealing with the aforementioned "four-hop" sensing link, making it difficult to fundamentally meet the signal strength requirements of high-precision sensing. To overcome the inherent shortcomings of passive RIS, hybrid RIS has emerged. Hybrid RIS integrates a small number of active amplification units into a large number of passive reflection units, enabling the reflective surface to amplify the reflected signal while maintaining low power consumption. This structure achieves a more promising balance between system performance gain and energy consumption cost, bringing new opportunities to enhance the performance of ISAC systems, especially their sensing links.

[0004] However, the introduction of hybrid RIS has made system resource management unprecedentedly complex. The system needs to perform rapid and accurate joint optimization of base station transmit beamforming, active / passive mode selection of the hybrid RIS, and its reflection coefficients in highly dynamic environments to simultaneously ensure the security of multi-user communication and the accuracy of sensing high-speed moving targets. This joint optimization problem is typically modeled as a mixed-integer non-convex programming problem, with extremely high computational complexity. Existing solutions based on traditional iterative algorithms are insufficient to meet the millisecond-level real-time decision-making requirements in highly dynamic scenarios, constituting a core bottleneck restricting the practical application of hybrid RIS-assisted ISAC systems.

[0005] Therefore, there is an urgent need in this field for an innovative method that can achieve low-complexity, high-real-time joint resource allocation in order to fully unleash the potential of hybrid RIS in sensory integration systems. Summary of the Invention

[0006] Purpose of the invention: This invention provides a hybrid RIS-assisted inductive integrated system and its safe predictive beamforming method to solve the problems mentioned in the background art.

[0007] Technical solution:

[0008] This invention provides a hybrid RIS-assisted sensing integrated system, including a central base station, a hybrid RIS model, multiple users, and an eavesdropper as the sensing target, wherein the central base station, the hybrid RIS model, the multiple legitimate users, and the eavesdropper are connected in pairs for communication.

[0009] The central base station is configured to transmit communication signals to multiple users and receive signals from legitimate users and eavesdroppers who are the targets of perception.

[0010] The hybrid RIS model is used to enhance the channel.

[0011] As an improvement of the present invention, the communication signal transmitted by the central base station in the nth time slot Modeled as a linear weighted sum of multiple user data streams:

[0012] ;

[0013] in, It is the beamforming matrix transmitted by the central base station. It is the beam vector assigned to the k-th user. It is a normalized communication symbol sent to the k-th user in the n-th time slot. This refers to the number of antennas at the central base station.

[0014] As an improvement to the present invention, the hybrid RIS model comprises M RIS units, the operating states of which are determined by a mode selection vector. ( and reflection coefficient diagonal matrix This indicates that each RIS unit can independently switch between passive reflection mode and active amplification mode; wherein, the reflection amplitude of each RIS unit... Affected by its working mode constraint:

[0015] like Then the RIS unit is in passive reflection mode. ;

[0016] like Then the RIS unit is in active amplification mode. ;

[0017] in, This represents the operating state of the m-th RISC unit in the n-th time slot, where T is the transpose of the matrix. This represents the maximum amplification gain.

[0018] As an improvement to the present invention, the equivalent concatenated channel from the central base station to the k-th user and the eavesdropper... and It is composed of the direct path and the reflection path via the hybrid RIS model, respectively:

[0019] ;

[0020] in, For direct channels, For the central base station-hybrid RIS model channel, For the hybrid RIS model - user / eavesdropper channel;

[0021] The received signals of the kth legitimate user and the eavesdropper and They are respectively:

[0022] ;

[0023] ;

[0024] in, and It is additive white Gaussian noise.

[0025] As an improvement of the present invention, a secure predictive beamforming method for a hybrid RIS-assisted sensory integration system is also provided, which is applied to the hybrid RIS-assisted sensory integration system as described above, including an offline training phase and an online prediction and execution phase.

[0026] The offline training phase includes:

[0027] (a) Generate a training dataset, which contains multiple training samples, each training sample including a sequence of historical channel state information and a corresponding sequence of future channel state information;

[0028] (b) Construct a deep unfolding prediction network, and construct the objective function of the joint optimization problem as the loss function of the deep unfolding prediction network;

[0029] (c) Using the training dataset, the depth unfolding prediction network is trained in an unsupervised end-to-end manner by minimizing the loss function;

[0030] The online prediction and execution phase includes:

[0031] (1) Collect the historical channel state information sequence of multiple consecutive time slots in the past and input it into the encoder module of the deep unfolded prediction network to generate a feature vector containing the channel state prediction information of the next time slot.

[0032] (2) The feature vector is input into the decoder of the deep unfolded prediction network. The decoder simulates the iterative process of the alternating direction multiplier method through multiple cascaded network layers inside it to update and optimize the variables.

[0033] (3) Through one forward propagation of the decoder, the joint optimization strategy for the next time slot is output, wherein the joint optimization strategy includes the transmit beamforming matrix of the central base station. Hybrid RIS mode selection vector and the reflection coefficient matrix of RIS ;

[0034] (4) At the start of the next time slot, control the central base station and the hybrid RIS model to transmit and reflect signals according to the joint optimization strategy.

[0035] As an improvement to the present invention, the joint optimization problem is:

[0036] ;

[0037] in, For the expected value of channel prediction error The worst-case secrecy rate is given by ρ, where ρ is the weighting factor. For the k-th legitimate user, the signal-to-interference-plus-noise ratio (SIR) is... The minimum data rate threshold for users. For user sets Any user in , This represents the current actual total power consumption of the hybrid RIS. The maximum allowable power consumption of the hybrid RIS. Transmit beamforming matrix for central base station The conjugate transpose of . This represents the maximum total transmit power of the central base station. For the RIS unit set any unit in C1 is the Cramer-Rao lower bound for estimating from the eavesdropper's perspective; C2 is the minimum user rate constraint; C3 is the total power consumption constraint of the RIS; C4 is the total transmit power constraint of the central base station; C5 is the RIS mode selection constraint; and C5 is the RIS reflection amplitude constraint.

[0038] As an improvement to the present invention, the worst-case security rate The calculation formula is:

[0039] ;

[0040] in, Let S be the signal-to-interference-plus-noise ratio at the k-th legitimate user. The signal-to-interference-plus-noise ratio (SIR) is the signal-to-interference-plus-noise ratio (SINR) when an eavesdropper is eavesdropping on the signal of the k-th user.

[0041] As an improvement of the present invention, the signal-to-interference-plus-noise ratio at the legitimate user location The calculation formula is:

[0042] ;

[0043] in, Indicates the first In the time slot, the central base station to the first Equivalent concatenated channel for one user It is a transformation on the conjugate transpose of a matrix. For the transmit beamforming matrix The kth column, For the transmit beamforming matrix The j-th column, Noise power for legitimate users;

[0044] And, the signal-to-interference-plus-noise ratio when the eavesdropper eavesdrops on the k-th user signal. The calculation formula is:

[0045] ;

[0046] in, Indicates the first In each time slot, the equivalent concatenated channel from the central base station to the eavesdropper (as the sensing target) is... The noise power of the eavesdropper.

[0047] As an improvement of the present invention, the Cramer-Lao lower bound The calculation process includes:

[0048] According to the transmit beamforming matrix of the central base station Calculate the covariance matrix of the transmitted signal. :

[0049] ;

[0050] Fisher Information Matrix The (i,j)th element The covariance matrix of the transmitted signal and round-trip sensing channel matrix For angle parameters The partial derivatives together determine:

[0051] ;

[0052] in, To perceive the number of snapshots, The noise power of the central base station radar receiver. It is the round-trip sensing channel matrix for the i-th angle parameter The partial derivative matrix;

[0053] The From the Fisher information matrix The trace of the inverse matrix is ​​given:

[0054] .

[0055] Beneficial effects:

[0056] 1. The core effect of this invention is that by transforming the time-consuming online iterative process of the traditional Alternating Direction Multiplier Method (ADMM) into a single forward propagation of a neural network of fixed depth, the decision-making time is reduced from the millisecond level of the traditional iterative algorithm to the sub-millisecond level. This millisecond-level real-time decision-making capability enables the system to effectively cope with the rapid channel changes caused by high-speed moving eavesdropping targets and meets the stringent real-time requirements of high-dynamic ISAC scenarios.

[0057] 2. Unlike traditional passive response mechanisms, this invention uses an encoder based on an attention mechanism to predict channel evolution trends and can generate beamforming strategies for future time slots in advance based on historical channel state information. This enables the system to preemptively form beam nulls in the direction of the eavesdropper, thereby effectively suppressing the quality of the received signal before the eavesdropping occurs and proactively protecting the communication security of all legitimate users.

[0058] 3. This invention improves sensing performance through two levels of synergy: First, it directly introduces the Cramer-Rao lower bound (CRLB) into the optimization target, optimizing the theoretical lower bound of sensing accuracy from an information theory perspective; Second, by jointly optimizing the hybrid RIS, it intelligently utilizes the amplification function of its active units to effectively compensate for the severe signal attenuation in the four-hop sensing link of "base station-RIS-target-RIS-base station", thereby significantly improving the detection accuracy and effective range for distant or weakly reflective targets.

[0059] 4. This invention, through an intelligent hybrid RIS mode selection mechanism, can dynamically decide the working mode (active or passive) of each RIS unit according to real-time communication and sensing needs, and activate the minimum number of active amplification units as needed. This refined energy control strategy significantly reduces the overall power consumption of the hybrid RIS while ensuring system performance, and achieves high energy efficiency operation.

[0060] 5. Traditional methods require the design of complex alternating optimization algorithms to handle coupled variables such as base station beamforming, RIS mode selection, and reflection coefficient separately. This invention solves this complex joint optimization problem in a one-time, end-to-end manner through a deep unfolded network, transforming complex online computation into a single, efficient forward propagation, which greatly simplifies the system's control flow and reduces implementation complexity.

[0061] 6. The deep unfolded network structure of this invention is derived from the iterative steps of the classic ADMM optimization algorithm. Each layer of the network has a clear mathematical and physical meaning (corresponding to one algorithm iteration). Compared with the pure "black box" deep learning model, it has better interpretability and engineering reliability. At the same time, the data-driven training method can help the algorithm escape the local optimal solution that traditional iteration may get stuck in, and obtain better convergence performance and final solution quality than the original algorithm.

[0062] 7. First, the optimization objective of this invention is performed under the expectation of channel prediction error, which enables the system to maintain robust performance even when faced with unavoidable channel estimation errors and prediction uncertainties. Second, the training process adopts unsupervised learning, directly using the augmented Lagrangian function of the original optimization problem as the loss function, without relying on a large amount of hard-to-obtain "optimal solution" label data, significantly reducing the data threshold and implementation difficulty of model training;

[0063] 8. By adjusting the weight factor ρ in the optimization objective, the system can flexibly adjust the allocation strategy of system resources between communication and sensing functions according to the priority of the actual application scenario (e.g., prioritizing communication security or prioritizing improving sensing and tracking accuracy). This inherent flexibility enables the present invention to adapt to diverse business needs and realize the adaptive and configurable integration of sensing and communication. Attached Figure Description

[0064] Figure 1 This is a schematic diagram of the system structure of the present invention;

[0065] Figure 2 This is a schematic flowchart of the method of the present invention;

[0066] Figure 3 This is a schematic diagram of the structure of the depth unfolded prediction network (ADMM-Net) of this invention. Detailed Implementation

[0067] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The specific implementation methods of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0068] This invention discloses a hybrid RIS-assisted sensing-interactive (ISAC) system and its secure predictive beamforming method. The core of this method lies in mapping a complex mixed-integer non-convex optimization problem—aiming to balance communication security and sensing accuracy—to a predictive deep neural network (ADMM-Net) using deep unfolding technology. This network can directly generate joint optimization strategies for base station transmit beamforming, hybrid RIS mode selection, and RIS reflection coefficients for future time slots through a single fast forward propagation based on historical channel information. This achieves a solution to the real-time and security challenges in highly dynamic scenarios with extremely low computational complexity.

[0069] like Figure 1 As shown, the hybrid RIS-assisted sensing integration (ISAC) system provided by the present invention includes a central base station (BS), K legitimate communication users, an airborne mobile eavesdropper as a sensing target, and a hybrid reconfigurable smart metasurface (Hybrid RIS) deployed to enhance the channel.

[0070] (1) Transmitted signal model

[0071] In the nth time slot, the central base station transmits the ISAC communication signal. Modeled as a linear weighted sum of multiple user data streams:

[0072] ;

[0073] in, It is the normalized communication symbol vector of all k users in the nth time slot. , It is the beam vector assigned to the k-th user. It is a normalized communication symbol sent to the k-th user in the n-th time slot. This refers to the number of antennas at the central base station.

[0074] (2) Hybrid RIS model

[0075] The hybrid RIS model comprises M RIS units, whose operating states are determined by a mode selection vector. ( (T is a transformation of the matrix by transpose) and the diagonal matrix of reflection coefficients By common definition, each RIS unit can independently switch between passive reflection mode and active amplification mode. The reflection amplitude of each RIS unit... Affected by its working mode constraint:

[0076] like (Passive reflection mode), then .

[0077] like (Active Amplification Mode) ,in This represents the maximum amplification gain.

[0078] in, This indicates the operating state of the m-th RISC unit in the n-th time slot. This indicates that the m-th RIS unit is in passive reflection mode in the n-th time slot. At this time, the RIS unit has no signal amplification capability, and the reflection coefficient is... When the value is 1, the beam direction can only be adjusted by changing the phase of the signal; This indicates that the m-th RIS unit is in active amplification mode in the n-th time slot. At this time, the RIS unit has signal amplification capability and a reflection coefficient. The amplitude can be Continuous adjustment within the range can both amplify the signal and change its phase.

[0079] (3) Equivalent channel and received signal model

[0080] Equivalent concatenated channel from the central base station to the k-th user and eavesdropper (E). and It is composed of the direct path and the reflection path via the hybrid RIS model, respectively:

[0081] ;

[0082] in, For direct channels, For the central base station-hybrid RIS model channel, This is a hybrid RIS model – user / eavesdropper channel. Therefore, the received signals of the k-th user and the eavesdropper... and They are respectively:

[0083] ;

[0084] ;

[0085] in, and It is additive white Gaussian noise.

[0086] Based on the established system model, in order to quantify the communication security and sensing performance of this invention, the following two core indicators are defined:

[0087] (1) Worst-Case Secrecy Rate (WSR)

[0088] To ensure the security of communications for all users in the system, WSR is defined as the minimum confidentiality rate between all users and the eavesdropper:

[0089] ;

[0090] in, The worst-case security rate for the nth time slot. It is the signal-to-interference-plus-noise ratio at the k-th user. It is the signal-to-interference-plus-noise ratio when the eavesdropper is eavesdropping on the k-th user's signal.

[0091] (2) Perception accuracy (Cramer-Rao lower bound (CRLB))

[0092] To characterize the angle (azimuth) of the eavesdropping target Pitch angle To estimate the theoretically highest achievable accuracy, this invention uses CRLB as a sensing performance metric. CRLB is the lower bound of the estimated variance related to the Fisher Information Matrix (FIM) of the radar echo signal; the smaller the value, the higher the achievable sensing accuracy. This metric can be expressed as... It is the transmission beamforming and RIS reflection coefficient The function.

[0093] Based on the performance metrics defined above, the core technical problem of this invention is constructed as a joint optimization problem aiming to maximize the weighted sum of communication security and sensing accuracy. Considering the existence of channel prediction errors, the optimization focuses on the expected value of the prediction error. Next step:

[0094] ;

[0095] in, , For the expected value of channel prediction error The worst-case secrecy rate is given by ρ, where ρ is the weighting factor. For the k-th legitimate user, the signal-to-interference-plus-noise ratio (SIR) is... The minimum data rate threshold for users. For user sets Any user in , This represents the current actual total power consumption of the hybrid RIS. The maximum allowable power consumption of the hybrid RIS. Transmit beamforming matrix for central base station The conjugate transpose of . This represents the maximum total transmit power of the central base station. For the RIS unit set any unit in , C1 is the Cramer-Rao lower bound for estimating from the eavesdropper's perspective; C2 is the minimum user rate constraint; C3 is the total power consumption constraint of the RIS; C4 is the total transmit power constraint of the central base station; C5 is the RIS mode selection constraint; and C5 is the RIS reflection amplitude constraint.

[0096] Since the objective function is non-convex, and the variables Given binary integers, this problem is a typical mixed integer nonconvex programming (MINLP) problem, which traditional iterative algorithms (such as alternating optimization, SDR, etc.) find difficult to solve in real time in highly dynamic scenarios.

[0097] To address the high computational complexity and real-time bottlenecks caused by the aforementioned MINLP problem, a hybrid RIS-assisted sensory integration system, as described above, is proposed. Figure 2 As shown, this invention proposes a secure predictive beamforming method for a hybrid RIS-assisted inductive and sensory integration system, the method comprising an offline training phase and an online prediction and execution phase.

[0098] The online prediction and execution phase includes the following steps:

[0099] (1) Historical channel information acquisition and spatiotemporal feature extraction

[0100] In each time slot At the end, the system collects past data. A sequence of Channel State Information (CSI) for each consecutive time slot is generated. This CSI sequence is then fed into an encoder module of a Deep Unfolded Prediction Network (ADMM-Net). This encoder employs a multi-head self-attention mechanism, designed to capture the dynamic evolution patterns of the channel in the temporal dimension and its spatial correlations, thereby generating a sequence containing information about the next time slot (the...). The feature vector of channel state prediction information (time slot).

[0101] (2) Neural network expansion of ADMM iterative algorithm

[0102] This invention does not directly solve the aforementioned MINLP problem. Instead, it first decomposes it into multiple alternating subproblems using the Alternating Directional Multiplier Method (ADMM) framework. The core innovation of this invention lies in the fact that it does not actually perform online iteration of ADMM, but rather unfolds its iterative process into a neural network with a specific hierarchical structure, namely the decoder module of ADMM-Net.

[0103] Decomposition of ADMM iteration: First, the original problem is decomposed into iterations of the main variables within the ADMM framework. and dual variables The alternating update steps.

[0104] Mapping between network layers and algorithm steps: The decoder of the ADMM-Net consists of L cascaded layers (blocks), each layer structurally corresponding strictly to one complete iteration of the ADMM algorithm. Specifically, the first... Each network layer receives the optimization variables output from the previous layer. And through its internal learnable parameters (neural network weights), the updated variables are directly calculated. .

[0105] Networked replacement of subproblems: The complex operations (such as inversion or projection) for solving variable updates in the original ADMM algorithm are replaced by efficient, specialized neural network submodules. For example, updating variables... The process of solving subproblems is handled by a learnable neural network module. Replaced. This module is the first The layer's output is used as input to directly calculate the updated value. Similarly, for variables... and dual variables The updates are also handled by their respective dedicated neural network modules. and Complete. This structure preserves the physical interpretability of the optimization algorithm while leveraging the powerful nonlinear fitting capabilities of neural networks.

[0106] (3) Generation of joint optimization strategy

[0107] The ADMM-Net decoder consists of L cascaded layers (blocks), each structurally simulating a complete iteration of the ADMM algorithm. Specifically, the... Layered networks receive optimization variables from the output of the previous layer. And through its internal learnable parameters (neural network weights), the updated variables are directly calculated. The complex inversion or projection operations in the original ADMM algorithm have been replaced by efficient neural network forward computation.

[0108] The feature vector output by the encoder in step (1) is used as the initial input of the decoder, and then... After the forward propagation of the layer, the network directly outputs the final joint optimization strategy, i.e., the transmit beamforming matrix of the central base station in the next time slot, within milliseconds. Hybrid RIS mode selection vector and the reflection coefficient matrix of RIS .

[0109] (4) Deployment and execution of predictive joint optimization strategies

[0110] In the At the moment the time slot begins, the central base station and the hybrid RIS model immediately apply the predictive joint optimization strategy generated by the network in step (3) to transmit and reflect signals, thereby effectively offsetting the time delay caused by channel feedback and computational processing, and achieving accurate and real-time response to highly dynamic environments.

[0111] The method of this invention also includes an offline training phase for a Deep Expanded Prediction Network (ADMM-Net). The powerful predictive ability of the ADMM-Net model stems from sufficient offline training, and the training process of ADMM-Net employs an unsupervised learning approach. The offline training phase includes the following steps:

[0112] (a) Generation of training dataset

[0113] A large number of training samples are generated through computer simulation. In each simulation, the locations of legitimate users and eavesdroppers are randomly generated, resulting in different channel implementations. By simulating the snapshot-like changes of the channel over time, an input sequence containing s consecutive historical CSIs and their corresponding future CSI sequences for the next time slot are constructed (used to calculate the loss during training).

[0114] (b) Construction of loss function

[0115] Its loss function is not based on data labels, but is directly constructed from the augmented Lagrangian function of the optimization problem mentioned above. This unsupervised training method does not require pre-calculation and labeling of the optimal beamforming strategy, thus avoiding the problem of obtaining a large amount of labeled training data, because these "optimal" labels are themselves difficult to solve in real time.

[0116] (c) Training process

[0117] The generated (historical CSI sequence, future CSI sequence) data pairs are input into a system such as... Figure 3 In the ADMM-Net shown, the network outputs a joint optimization policy based on the historical CSI sequence. Then, the loss function is calculated using this policy and the corresponding future CSIs. Finally, the loss function is optimized using backpropagation and a gradient descent optimizer (such as Adam) for all learnable parameters (i.e., each submodule) in the network. The weights and biases of the network are updated end-to-end. By minimizing this loss function, the network is driven during offline training to learn how to directly map from historical CSIs to solutions that optimize the original objective function and satisfy the constraints, thus embedding the intelligence for solving this complex optimization problem into the network parameters.

[0118] To illustrate the invention more specifically, a specific application scenario and simulation parameters will be used as examples to elaborate on the hybrid RIS-assisted inductive integrated system and its safe predictive beamforming method proposed in this invention.

[0119] I. Scene and Parameter Settings

[0120] This embodiment simulates a simple scenario of urban low-altitude monitoring and communication support.

[0121] The system components include a central base station (BS), equipped with Uniform linear array of root antennas (ULA); A single-antenna ground-based legitimate user; a single-antenna unmanned aerial vehicle (UAV) moving at a high speed of 20 m / s as a sensing and eavesdropping target; a Hybrid RIS of the unit.

[0122] Power and Noise: Maximum Transmit Power of Central Base Station RIS maximum power consumption Noise power .

[0123] Channel Model: All channels consider large-scale path loss and small-scale fading. The channel from the central base station to the ground user follows the Rayleigh fading model; the channel from the central base station / RIS to the UAV in the air follows the Rice fading model, with a Rice factor of 10.

[0124] II. Mathematical Model Underlying the Implementation

[0125] The SINR of the k-th user is

[0126] ;

[0127] in, Indicates the first In the time slot, the central base station to the first Equivalent concatenated channel for one user It is a transformation on the conjugate transpose of a matrix. For the transmit beamforming matrix The kth column, For the transmit beamforming matrix The j-th column, Noise power for legitimate users.

[0128] The SINR of the eavesdropper listening to the k-th user signal is:

[0129] ;

[0130] in, for, The noise power of the eavesdropper.

[0131] Cramer-Lao lower bound for perception accuracy (CRLB)

[0132] CRLB is the angle parameter for the UAV. The estimated variance provides a lower bound. The perceptual precision metric used in this embodiment is the trace of the CRLB matrix:

[0133] ;

[0134] in, It is the Fisher Information Matrix (FIM) and the transmitted signal covariance. and round-trip sensing channel It relates to the derivative with respect to angle.

[0135] Fisher Information Matrix (FIM) definition:

[0136] FIM describes the amount of information contained in the observed data about the parameter η to be estimated. For deterministic parameter estimation problems in the context of additive white Gaussian noise, the (i,j)th element of FIM... It can be derived from the expectation of the Hessian matrix of the log-likelihood function, and its calculation formula is as follows:

[0137] ;

[0138] in:

[0139] It is the number of snapshots used for perception;

[0140] It is the noise power of the central base station radar receiver;

[0141] It is the covariance matrix of the transmitted signal, which is the direct bridge connecting transmitted beamforming and sensing accuracy;

[0142] It is a round-trip sensing channel matrix that depends on the angle parameter η;

[0143] It is the round-trip sensing channel matrix for the i-th angle parameter (e.g. The partial derivative matrix of ). This derivative is directly related to the array geometry (i.e., the steering vector) of the central base station and the RIS.

[0144] III. Construction of the ADMM Algorithm Framework

[0145] The core of this invention is solving the Mixed Integer Nonconvex Programming (MINLP) problem. Since this problem cannot be solved directly in a closed form, and traditional iterative algorithms suffer from high computational complexity and difficulty meeting real-time requirements, this invention proposes an innovative solution framework based on deep learning. The framework is constructed in two logical steps: first, an iterative solution blueprint is designed for the original problem using the Alternating Direction Multiplier Method (ADMM); then, this iterative blueprint is expanded into an efficient feedforward deep neural network (ADMM-Net).

[0146] Problem refactoring and variable separation:

[0147] In order to couple the variables together in the original problem Separate the complex constraints (power, integers, physical characteristics) of each main variable and introduce a simpler auxiliary variable for each main variable. And add consistency constraints, for example Thus, the original problem is equivalently transformed into:

[0148] ;

[0149] The augmented Lagrangian function of this optimization problem can be written as:

[0150] ;

[0151] in, It is a dual variable. It is the objective function of the original problem (i.e. ,and This represents the set of constraints for all original problems (such as power constraints, integer constraints, etc.), which have now been completely transferred to auxiliary variables with simpler structures.

[0152] ADMM Iteration Process:

[0153] In the In this iteration, the traditional ADMM algorithm needs to alternately update the following variables:

[0154] Update main variable ( ):

[0155] ;

[0156] This is a coupled and difficult subproblem.

[0157] Update auxiliary variables ( ):

[0158] ;

[0159] This step corresponds to a constraint set. Projection operation on, Because the constraints are separated, the projection can be decomposed into independent, simple operations on each variable (such as power normalization, binarization, etc.), making the computation very efficient.

[0160] Update dual variables :

[0161] ;

[0162] This step is a simple algebraic update used to adjust the penalty intensity, and its computational complexity is negligible.

[0163] The core innovation of this invention lies in the fact that instead of performing the time-consuming iterative process online, it unfolds the process into ADMM-Net and outputs the solution directly through a single forward propagation.

[0164] IV. From ADMM Iteration to Deep Unfolded Networks

[0165] While the aforementioned ADMM framework is theoretically feasible, its online iterative solution mode cannot meet the real-time requirements of highly dynamic scenarios. Therefore, this invention proposes the core idea of ​​expanding this iterative framework into a feedforward neural network.

[0166] The essence of "unfolding" is to map the time iteration dimension of the ADMM algorithm to the spatial hierarchy dimension of the neural network.

[0167] One iteration of ADMM can be considered as one layer of ADMM-Net. The entire ADMM-Net consists of... The cascaded layers directly simulate the iteration of the ADMM algorithm. This process takes several steps.

[0168] ADMM variables are considered as features passed between ADMM-Net layers. Optimization variables Become the first in the network Layer and first Tensors passed between +1 layers.

[0169] For the most complex master variable updates in traditional ADMM, the time-consuming arg min solution is no longer performed. Instead, this invention designs one or more learnable neural network modules (e.g., This complex optimization mapping is approximated by a single fast forward propagation. For example:

[0170] ;

[0171] in, For iteration or network layer index, To augment the Lagrange function, As an auxiliary variable, As dual variables, The context vector output by the encoder. These are the learnable parameters of this neural network module.

[0172] In this way, a traditional iterative algorithm that relies on loops and time-consuming solutions is completely transformed into an end-to-end, fully differentiable, fixed-depth feedforward neural network. This network, trained offline on large datasets, embeds the "intelligence" for solving the original optimization problem into its weight parameters, thereby achieving millisecond-level rapid decision-making.

[0173] V. Specific Implementation Steps

[0174] (1) Offline training phase

[0175] Network Construction: A Deep Unfolded Prediction Network (ADMM-Net) is constructed. Its encoder employs a multi-head self-attention mechanism. Its decoder contains L=5 layers, each structurally simulating and unfolding one iteration of the ADMM described in the previous section. Specifically, the complex arg min operations for solving each variable in traditional ADMM are replaced by dedicated, learnable small neural network modules (such as...). ).

[0176] Training: Tens of thousands of samples containing historical CSI sequences were generated through simulation. An unsupervised learning approach was used to augment the Lagrangian function of ADMM. ADMM-Net was trained end-to-end offline on a GPU platform as the loss function of the network.

[0177] (2) Online prediction and execution phase

[0178] Deployment and Execution: The trained ADMM-Net model is deployed at the central base station. At the end of each time slot n−1, the processor collects CSI data from the past s=5 time slots and sends it to the network.

[0179] Policy Generation: The network performs one forward propagation and, within approximately 0.5 milliseconds (ms), directly outputs the complete joint policy for the next time slot n. .

[0180] VI. Effect Verification and Comparison

[0181] The method of this invention is compared with the baseline scheme that uses the traditional iterative ADMM algorithm (which iteratively solves the problem after acquiring the current channel in each time slot).

[0182] Under the simulation parameters set in this embodiment, the ADMM-Net method proposed in this invention is quantitatively compared with the benchmark scheme in three key indicators: communication security performance, sensing accuracy, and computational real-time performance. The weighting factor ρ is set to 0.5. The results are as follows:

[0183] The ADMM-Net method of this invention achieves a worst-case security rate (WSR) of 1.1306 bps / Hz on average, which is more than 95.20% of the average performance of the traditional iterative ADMM method (1.1876 bps / Hz), proving the effectiveness of this invention in ensuring communication security.

[0184] The average CRLB value of the method of this invention is 1.1111, which is very close to the 1.0460 of the traditional iterative ADMM method, indicating that the perception performance of the proposed method can approach the theoretical optimal benchmark.

[0185] The method of this invention has an average single decision time of 0.14 milliseconds, while the traditional ADMM method takes an average of 1.45 milliseconds to reach convergence. This invention improves decision-making speed by approximately 10.6 times, and the advantage in decision-making speed becomes even greater as the scenario becomes more complex, fully demonstrating its significant advantage in real-time decision-making in highly dynamic scenarios.

[0186] Furthermore, by adjusting the weighting factor ρ in the optimization objective, this invention can flexibly balance communication and sensing performance. When ρ=0.2, the system prioritizes sensing, resulting in better CRLB performance; when ρ=0.8, the system prioritizes communication security, resulting in better WSR performance. Simulation results verify that the system can be flexibly configured according to different service requirements.

[0187] This embodiment clearly demonstrates the implementation process of the present invention through a detailed mathematical model, the construction of the ADMM algorithm framework, and specific parameter settings. Through quantitative performance comparison, it powerfully proves the significant beneficial effects of the present invention in terms of real-time performance, security, and perception accuracy.

[0188] The above embodiments are for illustrative purposes only and are not intended to limit the scope of this invention. Although this invention has been described in detail with reference to the embodiments, those skilled in the art should understand that various combinations, modifications, or equivalent substitutions of the technical solutions of this invention do not depart from the spirit and scope of the technical solutions of this invention and should be covered within the scope of the claims of this invention.

Claims

1. A hybrid RIS-assisted sensing integrated system, comprising a central base station, a hybrid RIS model, multiple users, and an eavesdropper acting as the sensing target, wherein, The central base station, the hybrid RIS model, multiple legitimate users, and the eavesdropper communicate with each other in pairs. The central base station is configured to transmit communication signals to multiple users and receive signals from legitimate users and eavesdroppers who are the targets of perception. The hybrid RIS model is used to enhance the channel.

2. The hybrid RIS-assisted sensory integration system according to claim 1, characterized in that, The communication signal transmitted by the central base station in the nth time slot Modeled as a linear weighted sum of multiple user data streams: ； in, It is the beamforming matrix transmitted by the central base station. It is the beam vector assigned to the k-th user. It is a normalized communication symbol sent to the k-th user in the n-th time slot. This refers to the number of antennas at the central base station.

3. The hybrid RIS-assisted sensory integration system according to claim 1, characterized in that, The hybrid RIS model comprises M RIS units, whose operating states are determined by a mode selection vector. ( and reflection coefficient diagonal matrix This indicates that each RIS unit can independently switch between passive reflection mode and active amplification mode; wherein, the reflection amplitude of each RIS unit... Affected by its working mode constraint: like Then the RIS unit is in passive reflection mode. ; like Then the RIS unit is in active amplification mode. ; in, This represents the operating state of the m-th RISC unit in the n-th time slot, where T is the transpose of the matrix. This represents the maximum amplification gain.

4. The hybrid RIS-assisted sensory integration system according to claim 2 or 3, characterized in that... The equivalent concatenated channel from the central base station to the k-th user and the eavesdropper. and It is composed of the direct path and the reflection path via the hybrid RIS model, respectively: ； in, For direct channels, For the central base station-hybrid RIS model channel, For the hybrid RIS model - user / eavesdropper channel; The received signals of the kth legitimate user and the eavesdropper and They are respectively: ； ； in, and It is additive white Gaussian noise.

5. A secure predictive beamforming method for a hybrid RIS-assisted sensing system, applied to a hybrid RIS-assisted sensing system as described in any one of claims 1 to 4, characterized in that, It includes an offline training phase and an online prediction and execution phase; The offline training phase includes: (a) Generate a training dataset, which contains multiple training samples, each training sample including a sequence of historical channel state information and a corresponding sequence of future channel state information; (b) Construct a deep unfolding prediction network, and construct the objective function of the joint optimization problem as the loss function of the deep unfolding prediction network; (c) Using the training dataset, the depth unfolding prediction network is trained in an unsupervised end-to-end manner by minimizing the loss function; The online prediction and execution phase includes: (1) Collect the historical channel state information sequence of multiple consecutive time slots in the past and input it into the encoder module of the deep unfolded prediction network to generate a feature vector containing the channel state prediction information of the next time slot. (2) The feature vector is input into the decoder of the deep unfolded prediction network. The decoder simulates the iterative process of the alternating direction multiplier method through multiple cascaded network layers inside it to update and optimize the variables. (3) Through one forward propagation of the decoder, the joint optimization strategy for the next time slot is output, wherein the joint optimization strategy includes the transmit beamforming matrix of the central base station. Hybrid RIS mode selection vector and the reflection coefficient matrix of RIS ; (4) At the start of the next time slot, control the central base station and the hybrid RIS model to transmit and reflect signals according to the joint optimization strategy.

6. A secure predictive beamforming method for a hybrid RIS-assisted inductive integrated system according to claim 5, characterized in that, The joint optimization problem is: ； in, For the expected value of channel prediction error The worst-case secrecy rate is given by ρ, where ρ is the weighting factor. For the k-th legitimate user, the signal-to-interference-plus-noise ratio (SIR) is... The minimum data rate threshold for users. For user sets Any user in , This represents the current actual total power consumption of the hybrid RIS. The maximum allowable power consumption of the hybrid RIS. Transmit beamforming matrix for central base station The conjugate transpose of . This represents the maximum total transmit power of the central base station. For the RIS unit set any unit in C1 is the Cramer-Rao lower bound for estimating from the eavesdropper's perspective; C2 is the minimum user rate constraint; C3 is the total power consumption constraint of the RIS; C4 is the total transmit power constraint of the central base station; C5 is the RIS mode selection constraint; and C5 is the RIS reflection amplitude constraint.

7. The method according to claim 6, characterized in that, The worst-case security rate The calculation formula is: ； in, Let S be the signal-to-interference-plus-noise ratio at the k-th legitimate user. The signal-to-interference-plus-noise ratio (SIR) is the signal-to-interference-plus-noise ratio (SINR) when an eavesdropper is eavesdropping on the signal of the k-th user.

8. The method according to claim 7, characterized in that, The signal-to-interference-plus-noise ratio at the legitimate user location The calculation formula is: ； in, Indicates the first In the time slot, the central base station to the first Equivalent concatenated channel for one user It is a transformation on the conjugate transpose of a matrix. For the transmit beamforming matrix The kth column, For the transmit beamforming matrix The j-th column, Noise power for legitimate users; And, the signal-to-interference-plus-noise ratio when the eavesdropper eavesdrops on the k-th user signal. The calculation formula is: ； in, Indicates the first In each time slot, the equivalent concatenated channel from the central base station to the eavesdropper (as the sensing target) is... The noise power of the eavesdropper.

9. The method according to claim 6, characterized in that, The Cramer-Rao lower bound The calculation process includes: According to the transmit beamforming matrix of the central base station Calculate the covariance matrix of the transmitted signal. : ； Fisher Information Matrix The (i,j)th element The covariance matrix of the transmitted signal and round-trip sensing channel matrix For angle parameters The partial derivatives together determine: ； in, To perceive the number of snapshots, The noise power of the central base station radar receiver. It is the round-trip sensing channel matrix for the i-th angle parameter The partial derivative matrix; The From the Fisher information matrix The trace of the inverse matrix is ​​given: 。

Images