A communication and sensing integrated beamforming method and system based on alternating optimization and hybrid multilayer perceptron

Abstract

The application provides a communication and sensing integrated beamforming method and system based on alternating optimization and a hybrid multilayer perceptron, comprising: adding a full-duplex technology in an ISAC system, constructing an augmented beamforming matrix; based on the augmented beamforming matrix, constructing a corresponding multi-objective joint constraint condition; based on the multi-objective joint constraint condition, combining mobile edge computing resource allocation, taking maximum system energy efficiency as the target, constructing a beamforming optimization problem; combining an MLP-Mixer network and an alternating optimization method, constructing an MLP-Mixer-AO framework, and using the MLP-Mixer-AO framework to solve the beamforming optimization problem to obtain an optimal solution, and completing communication and sensing integrated beamforming. The application can effectively solve the FIM matrix singularity problem, accurately model the extended target scattering characteristics, and significantly improve the sensing accuracy, energy efficiency and real-time performance while ensuring communication quality.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, specifically relating to a communication-sensing integrated beamforming method and system based on alternating optimization and hybrid multilayer perceptrons. Background Technology

[0002] With the explosive growth of data traffic in IoT systems and the evolution of fifth-generation mobile communication systems towards 5G-Advanced and sixth-generation mobile communication, the business models carried by wireless networks are undergoing fundamental changes. Emerging applications such as autonomous driving, industrial IoT, and extended reality not only require high-speed communication but also high-precision environmental perception capabilities. However, traditional separate communication and sensing systems, due to resource monopolization and mutual interference issues, struggle to meet the efficiency demands of high-concurrency intelligent applications. This trend has spurred the rapid development of integrated sensing and communication technologies. Their core value lies in achieving deep integration of communication and sensing functions through a unified hardware platform and spectrum resources, providing a new paradigm for solving the problem of scarce spectrum resources.

[0003] ISAC technology, by sharing signal processing links and spectrum resources, significantly improves system spectral efficiency and reduces cost and power consumption, becoming a key enabling technology for 6G networks. Current research mainly focuses on the core issue of beamforming design, achieving a performance balance between communication and sensing by optimizing spatial signal distribution. Existing research has formed technical routes represented by minimizing the Cramer-Rao bound, guaranteeing the signal-to-interference-plus-noise ratio, and multi-objective optimization. Beamforming schemes based on alternating optimization and semi-definite relaxation methods have emerged, providing important support for improving system performance.

[0004] However, with the increasing complexity of application scenarios, existing ISAC systems face severe challenges in collaborative resource allocation and spectral efficiency optimization. Current research focuses primarily on the collaborative allocation of resources and interference management between communication and sensing functions, such as improving sensing accuracy while maintaining communication service quality through optimized beamforming. As the number of users increases, the FIM matrix used for sensing performance estimation becomes singular due to increased dimensionality, leading to deviations between the estimated and actual values. Simultaneously, existing algorithms generally simplify the sensing target to an ideal single-scattering point model, ignoring multipath scattering effects and spatial expansion characteristics, resulting in distorted representation of the target's electromagnetic properties. This leads to scattering mechanism mismatch and deterioration of detection accuracy in complex scenarios. The impact of energy efficiency on ISAC system performance has not been fully considered. Furthermore, most existing studies have failed to adequately address the real-time computational bottleneck of the massive amounts of data generated by high-precision sensing, resulting in delays in sensing information extraction and making it difficult to meet the real-time requirements of low-latency intelligent applications such as autonomous driving and the Industrial Internet of Things. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a beamforming algorithm based on a multilayer perceptron mixer combined with alternating optimization. Unlike existing work, this invention considers the singularity of the Fisher information matrix, multiple scattering effects and spatial distribution characteristics, energy efficiency, and real-time processing requirements of sensing data in a full-duplex communication-sensing integrated system. A joint beamforming framework integrating mobile edge computing and spatially extended targets is proposed. The transmit degrees of freedom are enhanced through orthogonal dedicated detection flows, and the target characteristics are nonlinearly modeled using an MLP-Mixer model, combined with alternating optimization to solve the beamforming problem. To achieve the above objectives, this invention provides the following solution:

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A communication-sensing integrated beamforming method based on alternating optimization and hybrid multilayer perceptrons, comprising:

[0008] Full-duplex technology is incorporated into the ISAC system to construct an augmented beamforming matrix;

[0009] Based on the augmented beamforming matrix, corresponding multi-objective joint constraints are constructed.

[0010] Based on the aforementioned multi-objective joint constraints and combined with the allocation of mobile edge computing resources, a beamforming optimization problem is constructed with the goal of maximizing the energy efficiency of the ISAC system.

[0011] By combining the MLP-Mixer network and the alternating optimization method, an MLP-Mixer-AO framework is constructed. The MLP-Mixer-AO framework is then used to solve the beamforming optimization problem, obtaining the optimal solution and completing the integrated beamforming for communication and sensing.

[0012] Preferably, the method for constructing the augmented beamforming matrix includes:

[0013] Full-duplex technology is incorporated into the ISAC system to construct an FD-ISAC sensing model based on extended objectives;

[0014] Based on the FD-ISAC sensing model, construct the FD-ISAC-MEC system model;

[0015] Based on the FD-ISAC-MEC system model, an orthogonal dedicated detection stream is introduced to design the FD-ISAC base station transmit signal matrix and construct an augmented beamforming matrix; the FD-ISAC base station transmit signal matrix is ​​designed as follows:

[0016]

[0017] This represents the augmented beamforming matrix of the FD-ISAC base station transmitter. Assigning a sub-matrix to the communication beam, To assist in sensing the beamforming matrix, For the augmented signal matrix, N t D represents the number of transmitting antennas, and D represents the number of downlink users per antenna.

[0018] Preferably, the multi-objective joint constraints include minimum signal-to-interference-plus-noise ratio (SINR) constraints for downlink communication users, minimum SINR constraints for uplink communication users, total transmit power constraints for the base station, transmit power range constraints for uplink user terminals, task processing delay constraints, and task offloading ratio constraints.

[0019] Preferred methods for constructing beamforming optimization problems include:

[0020] The total system power is calculated based on the sum of the uplink communication user's transmit power, the device's inherent circuit power consumption, the computing power, and the circuit's static power consumption.

[0021] Energy efficiency is calculated based on the total system power, the minimum signal-to-interference-plus-noise ratio (SINR) constraints for downlink and uplink communication users.

[0022] Based on the system's Cramer-Rao bound, resource allocation weighting factor, and energy efficiency, combined with the multi-objective joint constraints, the beamforming optimization problem is constructed.

[0023] Preferably, the beamforming optimization problem is decomposed into three interrelated sub-problems using an alternating optimization method. Two sets of variables are fixed, and the remaining set is optimized sequentially until the algorithm converges. The three sub-problems include the uplink user transmit power optimization problem, the base station receive beamforming matrix optimization problem, and the base station transmitter augmented beamforming matrix optimization problem. The method for solving each optimization problem includes:

[0024] By fixing the augmenting beamforming matrix and the base station receiving beamforming matrix, the uplink communication user's transmit power optimization problem is modeled as a fractional programming problem. The Dinkelbach algorithm is used to introduce parameters to transform the fractional programming problem into an equivalent parameterized subproblem. The unconstrained optimal power solution is obtained by solving the first derivative.

[0025] With fixed uplink user transmit power and augmented beamforming matrix, the optimization problem of base station receive beamforming matrix is ​​expressed as the uplink user and rate maximization problem. Using the generalized Rayleigh quotient principle, the uplink user and rate maximization problem is transformed into a generalized eigenvalue problem. The optimal solution of the generalized eigenvalue problem is the eigenvector corresponding to the maximum generalized eigenvalue of the preset matrix pair. After normalization by the unit norm, the optimal receive beamforming vector is obtained.

[0026] With fixed uplink user transmit power and base station receive beamforming matrix, the optimization problem of augmented beamforming matrix at the base station transmitter is modeled as a non-convex problem involving Cramer-Rao bounds and system energy efficiency weighting. A semidefinite relaxation technique is used to introduce a covariance variable into the non-convex problem, transforming it into a convex semidefinite programming problem. A successive convex approximation method is then used to linearize the non-convex terms of the objective function of the convex semidefinite programming problem to obtain a convex optimization problem. Solving the convex optimization problem yields the optimal covariance matrix, and finally, the augmented beamforming matrix is ​​recovered.

[0027] Preferably, the MLP-Mixer network is used to provide an initial solution for the alternating optimization method; the structure of the MLP-Mixer network includes:

[0028] The input layer is used to map the raw system state information into a fixed-dimensional feature representation; the raw system state information includes the uplink and downlink user signal-to-interference-plus-noise ratio thresholds and the total power budget.

[0029] Convolutional layers are used to locally model the feature representation to obtain spatial feature information;

[0030] The Mixer layer is used to convert the spatial feature information into two-dimensional data and divide it into patches. The Mixer layer is divided into a token mixing multilayer perceptron and a channel mixing multilayer perceptron. Each multilayer perceptron consists of two fully connected layers and a ReLU activation function.

[0031] The Per-patch Fully-connected layer is used to convert the patch into a corresponding feature embedding sequence; the feature embedding sequence is then embedded into the Mixer layer for feature extraction and mixing.

[0032] The output layer is used to generate an initial solution for the alternating optimization method; the initial solution includes the augmented beamforming matrix, the base station receive beamforming matrix, and the transmit power of the uplink communication user.

[0033] The present invention also provides a communication-sensing integrated beamforming system based on alternating optimization and hybrid multilayer perceptrons, for implementing the method, comprising:

[0034] The matrix construction module is used to add full-duplex technology to the ISAC system and construct the augmented beamforming matrix;

[0035] The constraint construction module is used to construct corresponding multi-objective joint constraint conditions based on the augmented beamforming matrix.

[0036] The optimization problem construction module is used to construct a beamforming optimization problem based on the multi-objective joint constraints and the allocation of mobile edge computing resources, with the goal of maximizing the energy efficiency of the ISAC system.

[0037] The problem-solving module is used to combine the MLP-Mixer network and the alternating optimization method to construct the MLP-Mixer-AO framework, and use the MLP-Mixer-AO framework to solve the beamforming optimization problem, obtain the optimal solution, and complete the integrated beamforming of communication and sensing.

[0038] Preferably, the matrix construction module includes:

[0039] The perception model building unit is used to add full-duplex technology to the ISAC system and build an FD-ISAC perception model based on extended objectives.

[0040] The system model building unit is used to build the FD-ISAC-MEC system model based on the FD-ISAC perception model.

[0041] The matrix construction unit is used to design the FD-ISAC base station transmit signal matrix and construct the augmented beamforming matrix based on the FD-ISAC-MEC system model, by introducing orthogonal dedicated detection streams; wherein, the FD-ISAC base station transmit signal matrix is ​​designed as follows:

[0042]

[0043] This represents the augmented beamforming matrix of the FD-ISAC base station transmitter. Assigning a sub-matrix to the communication beam, To assist in sensing the beamforming matrix, For the augmented signal matrix, N t D represents the number of transmitting antennas, and D represents the number of downlink users per antenna.

[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0045] This invention addresses the problem of insufficient target detection accuracy caused by the singularity of the Fisher information matrix and the simplified target model in the FD-ISAC system. Through efficient and coordinated resource allocation, it improves spectrum utilization while balancing the performance contradictions between communication and sensing, thereby increasing sensing accuracy while ensuring communication speed. Attached Figure Description

[0046] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a schematic diagram of the system model of FD-ISAC-ME according to an embodiment of the present invention;

[0048] Figure 2 This is a diagram of the MLP-Mixer network structure according to an embodiment of the present invention;

[0049] Figure 3 This is a schematic diagram illustrating the change of the algorithm's CRB with the number of training iterations under MLP-Mixer initialization and traditional random initialization conditions in an embodiment of the present invention.

[0050] Figure 4 This is a schematic diagram showing the users and rates of the four algorithms in this embodiment of the invention under different base station transmit powers;

[0051] Figure 5 This is a schematic diagram of the CWD distillation structure according to an embodiment of the present invention;

[0052] Figure 6 This is a flowchart of the communication-sensing integrated beamforming method based on alternating optimization and hybrid multilayer perceptrons, according to an embodiment of the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0055] Example 1:

[0056] like Figure 6 As shown, a communication-sensing integrated beamforming method based on alternating optimization and hybrid multilayer perceptrons includes:

[0057] S1: Incorporate full-duplex technology into the ISAC system to construct an augmented beamforming matrix.

[0058] A further implementation method involves constructing the augmented beamforming matrix, including:

[0059] S11: Full-duplex technology is incorporated into the ISAC system to construct an FD-ISAC sensing model based on extended targets. Specifically, the transmitted signal matrix records discrete signal samples transmitted by each antenna in L consecutive time slots. To simplify the sensing model, it is assumed that the target consists of many reflection points located at the same distance from the antennas. A uniform linear array with half-wavelength antenna spacing is used to further enhance the sensing capability. Furthermore, based on narrowband conditions, only the angular dimension information of the target needs to be analyzed. The echo signal matrix Y received by the FD-ISAC base station is:

[0060]

[0061] in For the target response matrix, This represents the transmit signal matrix of the FD-ISAC base station. This represents the additive white Gaussian noise (AWGN) matrix, where the variance of each term is... N t N represents the number of base station transmit antennas. r This represents the number of receiving antennas at the base station.

[0062] For extended target scenarios (e.g., mobile vehicles located near a base station), this invention focuses on the response characteristics under near-field conditions, where the extended target response matrix G can be expressed as:

[0063]

[0064] Where N s α represents the number of scattering units. For the i-th scattering unit, α i and θ i These correspond to its reflection coefficient and angle, respectively. H represents the steering vector of the transmitting antenna, and H represents the conjugate transpose operation.

[0065] In a full-duplex ISAC system, the Fisher information matrix of the parameter matrix G is:

[0066]

[0067] This represents the noise variance, which is usually assumed to be the variance of additive white Gaussian noise.

[0068] The CRB (Cramer-Rao boundary) of the system can be represented as:

[0069] .

[0070] The representative sample covariance matrix.

[0071] S12: Based on the FD-ISAC sensing model, construct the FD-ISAC-MEC system model; specifically, based on the FD-ISAC sensing model, considering the real-time computing bottleneck of the massive data generated by high-precision sensing, which leads to the delay in sensing information extraction, the FD-ISAC-MEC system model is constructed as follows:

[0072] Consider an FD-ISAC MEC system, which consists of N... t Root transmitting antenna and N r The system consists of an FD-BS with one receiving antenna, M extended radar targets, U single-antenna uplink users, and D single-antenna downlink users. The FD-BS, equipped with a dedicated MEC server, can perform computational tasks offloaded from the user terminals. Each user terminal has a computationally intensive task, which can be executed locally or partially migrated to the BS. The received signal Y of the downlink users... D This includes communication signals from U uplink users, echo signals from M radar targets, and noise signals.

[0073]

[0074] Among them, downlink communication channel At the FD-ISAC base station, it is known that its elements are independently distributed. Here, D represents the number of downlink users, and the matrix... Let represent an additive white Gaussian noise matrix, where the variance of each element is . The signal-to-noise-to-interference ratio (SNR) of the d-th downlink user is:

[0075]

[0076] in, This represents the channel vector from the base station to the d-th downlink user. This represents the beamforming vector used by the base station to send data to the d-th user. This represents the beamforming vector that the base station sends to other user i. This represents the sensing beamforming matrix.

[0077] Therefore, the downlink reachable speed for users is:

[0078]

[0079] B represents the channel bandwidth. Considering that the focus is on the signal-to-interference-plus-noise ratio and the impact on the data rate, the bandwidth will be set to 1 in the following text.

[0080] The received signal Y of the FD-ISAC base station R This includes communication signals from U uplink users, radar echo signals from M targets, and noise signals. Its expression is as follows:

[0081]

[0082] Among them, the uplink communication channel At the FD-ISAC base station, it is known that its elements are independently distributed. This indicates that the uplink user is transmitting a signal. Let represent additive white Gaussian noise, with the variance of each element being . To separate the communication component of a specific user u from the mixed signal, the base station employs a receive beamforming matrix. As a spatial filter, after beamforming processing, the communication signal components of the uplink user... :

[0083]

[0084] The signal-to-noise ratio of the u-th uplink user is

[0085]

[0086] in, This represents the channel vector from the base station to the u-th uplink user. The beamforming vector represents the beamforming vector used by the base station to send data to the u-th user. The uplink channel vector of the p-th user, P u For user u's transmit power, P p This represents the transmit power of other uplink connections. Therefore, the downlink achievable data rate for users is:

[0087]

[0088] To focus on the core process of uplink task offloading, this invention omits the energy consumption and latency of the MEC server transmitting downlink calculation results to device u. The energy consumption and latency analysis in this patent primarily targets uplink transmission and edge computing. Let x... u ∈[0,1] represents the task offloading ratio, that is, the proportion of user u's tasks offloaded to the MEC server, with the remaining portion executed locally. Therefore, the total latency and total energy consumption of device u in executing tasks are expressed as follows:

[0089]

[0090]

[0091] in This represents the computational latency of the u-th device. This represents the task offloading latency of IoT device u; This represents the energy consumption of the u-th device. This represents the energy consumption of task offloading for IoT device u.

[0092] S13: To achieve integrated communication and sensing, the FD-ISAC base station's transmitted signals need to carry user data streams and have their spatial beam characteristics optimized to support high-precision sensing. Addressing the issue of singularity in the FIM matrix of multiple sensing users, based on the FD-ISAC-MEC system model, an orthogonal dedicated detection stream is introduced to design the FD-ISAC base station's transmitted signal matrix and construct an augmented beamforming matrix; such as... Figure 1 As shown. The transmit signal matrix of the FD-ISAC base station is designed as follows:

[0093]

[0094] This represents the augmented beamforming matrix of the FD-ISAC base station transmitter. Assigning a sub-matrix to the communication beam, To assist in sensing the beamforming matrix, For the augmented signal matrix, N t D represents the number of transmitting antennas, and D represents the number of downlink users per antenna.

[0095] Specialized perceptual symbol matrix S C With the communication symbol matrix S A They are mutually orthogonal, satisfying:

[0096]

[0097] The sample covariance matrix R can be obtained from the statistical characteristics of the transmitted signal. X for:

[0098]

[0099] This design ensures that full rank solves the FIM singularity problem, enabling lower bounds on the mean squared error and unbiased estimation of the first D columns W. C Used for transmitting user data, after N t Column W A Used to enhance perception performance.

[0100] S2: Based on the augmented beamforming matrix, to balance communication and sensing performance, corresponding multi-objective joint constraints are constructed.

[0101] A further implementation involves multi-objective joint constraints including minimum signal-to-interference-plus-noise ratio (SINR) constraints for downlink and uplink users, total transmit power constraints for the base station, transmit power range constraints for uplink user terminals, task processing delay constraints, and task offloading ratio constraints. These constraints collectively ensure that the system can simultaneously meet the requirements of communication service quality, sensing performance, energy efficiency, and real-time processing of computational tasks when optimizing beamforming.

[0102] Define uplink and downlink signal-to-interference-plus-noise ratio constraints: , This ensures that the signal quality of each communication link meets an acceptable minimum standard, thereby maintaining reliable data transmission; Represents the downlink user SINR threshold. This represents the SINR threshold for uplink users.

[0103] Define transmit power constraints P T This indicates the maximum total available transmit power of the base station;

[0104] Define user power constraints ;

[0105] Define delay constraints , This represents the total latency of the u-th user task from start to finish. This indicates the maximum allowable delay threshold required by the system.

[0106] In this invention, the noise ratio of the d-th downlink communication user is expressed as:

[0107]

[0108] in, This represents the channel vector from the base station to the d-th downlink user. This represents the beamforming vector used by the base station to send data to the d-th user. This represents the beamforming vector that the base station sends to other user i. This represents the sensing beamforming matrix.

[0109] In this invention, the signal-to-interference-plus-noise ratio (SIR) of the u-th uplink communication user is expressed as:

[0110]

[0111] in, This represents the channel vector from the base station to the u-th uplink user. The beamforming vector represents the beamforming vector used by the base station to send data to the u-th user. The uplink channel vector of the p-th user, P u For user u's transmit power, P p This indicates the transmit power of other uplinks.

[0112] Meanwhile, considering the physical feasibility of task offloading decisions and the rationality of system resource allocation after the introduction of mobile edge servers, this invention must adjust the task offloading ratio x. u Constraints are applied within the range [0,1]. From a physical perspective, this ratio defines the division of labor between the local machine and the server for task execution, x u =0 and x u =1 corresponds to the two extreme cases of completely local execution and complete unloading, respectively; values ​​outside this range have no practical meaning. From a system perspective, this constraint ensures that communication resources and computing resources can be allocated collaboratively within clear boundaries, avoiding resource demand conflicts or distortion of optimization goals caused by uncontrolled ratios.

[0113] S3: Based on multi-objective joint constraints and combined with mobile edge computing resource allocation, a beamforming optimization problem (minimizing the weighted sum of CRB and the inverse of energy efficiency) is constructed with the goal of maximizing the energy efficiency of the ISAC system.

[0114] A further implementation method involves constructing a beamforming optimization problem, including:

[0115] Based on base station downlink transmit power Uplink communication user's transmit power Calculate power With circuit static power consumption The sum of these values ​​is used to calculate the total system power P; the calculation formula is as follows:

[0116]

[0117] Where P u =[p 1, p2,…,p u ] T For the transmit power of uplink users, p u P represents the transmit power of the u-th uplink user. c Calculate the power consumption based on the inherent circuit power consumption of the device. , Indicates the frame duration.

[0118] Energy efficiency is calculated based on the total system power, the minimum signal-to-interference-plus-noise ratio (SIR) constraints for downlink and uplink users; energy efficiency Defined as the number of information bits transmitted per unit of energy, i.e.:

[0119]

[0120] Where R=R U +R D This represents the total speed of the system. This represents the square of the Frobenius norm of the matrix.

[0121] Based on the system's Cramer-Rao bound, resource allocation weighting factor, and energy efficiency, combined with multi-objective joint constraints, a beamforming optimization problem is constructed. The beamforming optimization problem can be modeled as follows:

[0122]

[0123] in Resource allocation weighting factors are used to flexibly adjust the trade-off between sensing performance and energy efficiency, achieving targeted optimization of performance targets. C1 and C2 satisfy uplink user signal-to-noise ratio constraints, C3 is the base station's maximum transmit power constraint, C4 is the transmit power constraint for each uplink user, C5 is the user transmission time constraint, and C6 is the resource allocation weighting factor constraint.

[0124] S4: Combining the MLP-Mixer network and the alternating optimization method, construct the MLP-Mixer-AO framework, and use the MLP-Mixer-AO framework to solve the beamforming optimization problem, obtain the optimal solution, and complete the integrated beamforming of communication and sensing.

[0125] A further implementation method involves, considering the complex coupling relationships among the variables in the proposed optimization problem, decomposing the beamforming optimization problem into three interrelated sub-problems using an alternating optimization method (Algorithm 1). Two sets of variables are fixed, and the remaining set is optimized sequentially until the algorithm converges. The three sub-problems include the uplink user transmit power optimization problem, the base station receive beamforming matrix optimization problem, and the base station transmitter augmented beamforming matrix optimization problem. The method for solving each optimization problem includes:

[0126] By fixing the augmenting beamforming matrix and the base station receiving beamforming matrix, the uplink user transmit power optimization problem is modeled as a fractional programming problem. The Dinkelbach algorithm is used, and parameters are introduced to transform the fractional programming problem into an equivalent parameterized subproblem. The unconstrained optimal power solution is obtained by solving the first derivative; specifically, the uplink user transmit power is optimized. (Fixed augmenting beamforming matrix) and the base station receive beamforming matrix W U First, the uplink user transmit power The optimization problem can be modeled as a fractional programming problem:

[0127]

[0128] Where the definition , , , , To be with P u Irrelevant terms; this problem is a nonconvex partitioning programming problem. The Dinkelbach algorithm is used, introducing parameters... This is transformed into an equivalent parameterized subproblem, and the unconstrained optimal power solution is obtained by solving the first derivative:

[0129]

[0130] Finally, the theoretical solution is mapped to the feasible region that satisfies the signal-to-interference-plus-noise ratio (SINR) and maximum transmit power constraints through projection operations, and the Dinkelbach parameters are iteratively updated until convergence.

[0131] With fixed uplink user transmit power and augmented beamforming matrix, the optimization problem of base station receive beamforming matrix is ​​formulated as the uplink user and rate maximization problem. Using the generalized Rayleigh quotient principle, the uplink user and rate maximization problem is transformed into a generalized eigenvalue problem. The optimal solution of the generalized eigenvalue problem is the eigenvector corresponding to the largest generalized eigenvalue of the preset matrix pair. After normalization by the unit norm, the optimal receive beamforming vector is obtained.

[0132] With fixed uplink user transmit power and base station receive beamforming matrix, the optimization problem of the augmented beamforming matrix at the base station transmitter is modeled as a non-convex problem involving Cramer-Rao bounds and system energy efficiency weighting. A positive semidefinite relaxation technique is used to introduce a covariance variable into the non-convex problem, transforming it into a convex positive semidefinite programming problem. A successive convex approximation method is then used to linearize the non-convex terms of the objective function of the convex positive semidefinite programming problem, resulting in a convex optimization problem. Solving the convex optimization problem yields the optimal covariance matrix, and finally, the augmented beamforming matrix is ​​recovered. Specifically, the augmented beamforming matrix at the base station transmitter is optimized. With fixed uplink user transmit power and base station receive beamforming matrix, the optimization problem of the augmented beamforming matrix is ​​modeled as a non-convex problem involving Cramer-Rao bounds and system energy efficiency weighting.

[0133]

[0134] in ,

[0135] This problem is highly nonconvex due to the presence of matrix inverses, fractions, and logarithmic terms. A positive semidefinite relaxation technique is used to introduce a covariance variable:

[0136]

[0137] in This indicates that the matrix is ​​positive definite.

[0138] Since the objective function contains non-convex fractional terms and matrix inverses, and the constraints are quadratic fractional forms, the original problem is a non-convex optimization problem overall. To transform it into a solvable convex positive semi-definite optimization problem, SDR is used, introducing a covariance variable X. D And relax the rank constraint. Let , , , can be obtained , X D =X A +X C At this point, the SINR of the uplink user u is converted into a semi-positive definite variable X. A Linear inequalities: ,in To represent the SINR of the d-th user in the downlink, let... , achievable , .

[0139] Linearization is performed using a successive convex approximation method, resulting in the final SDP subproblem:

[0140]

[0141] The optimal covariance matrix is ​​obtained by solving this convex optimization problem, and the augmented beamforming matrix is ​​finally recovered.

[0142] A further implementation method considers that the alternating optimization solution is sensitive to initial values, has a slow convergence speed, and is prone to getting trapped in local optima. Therefore, an MLP-Mixer network is proposed to provide a high-performance initial solution. The MLP-Mixer network structure is as follows: Figure 2 As shown. The MLP-Mixer network is used to provide initial solutions for the alternating optimization method; the structure of the MLP-Mixer network includes:

[0143] The input layer is used to map the raw ISAC system state information S into a fixed-dimensional feature representation; the raw system state information includes the uplink and downlink user signal-to-interference-plus-noise ratio thresholds and the total power budget.

[0144] Convolutional layers are used to locally model the feature representation and obtain spatial feature information; rectified linear units (ReLU) are used as the activation function, giving the network robust non-linear learning capabilities. This provides a more discriminative representation for the subsequent mixer module.

[0145] The Mixer layer is used to convert spatial feature information into two-dimensional data and divide it into patches. The Mixer layer is divided into token mixing multilayer perceptron and channel mixing multilayer perceptron. Each multilayer perceptron consists of two fully connected layers and a ReLU activation function. The Per-patch Fully-connected layer is used to convert the patch into the corresponding feature embedding sequence. The feature embedding sequence is then embedded into the Mixer layer for feature extraction and mixing.

[0146] Specifically, the Mixer layer is divided into a Token-mixing Multi-Layer Perceptron (MLP) and a Channel-mixing Multi-Layer Perceptron (MLP). Each MLP consists of two fully connected layers and a ReLU activation function. The Token-mixing MLP captures spatial correlations, enabling the network to integrate features from different tokens, thus better representing multi-user, multi-target joint perception and communication environments. The Channel-mixing MLP performs operations along the feature dimension, mining dependencies between features to improve the model's ability to nonlinearly characterize extended target scattering properties and power allocation patterns. These two types of layers are stacked alternately to facilitate interaction between the two input dimensions. The Token-mixing MLP operates on the columns of the input feature matrix, starting with the transpose of the input feature matrix, and the MLP parameters are shared across all columns. Subsequently, the output is transposed again. In contrast, Channel-mixing MLP operates on the matrix of input features rows, sharing MLP parameters across all rows.

[0147] If the feature matrix input to the Mixer sublayer is represented as H M W represents the number of users. M H represents the feature dimension corresponding to each entity. In this invention, H M =D+U+M,W M =64, then the output of the token-mixing MLP It can be represented as:

[0148]

[0149] in This is the output of the layer, with the same shape as the input. From X MTake the i-th column, which corresponds to the value of the i-th channel across all tokens. W1 and W2 represent the two fully connected matrices, respectively. `<ReLU>` represents the ReLU activation function, and `LayerNorm` represents the normalization process applied to the vector. The essence of this structure lies in mixing information along the token dimension, enabling the same channel to establish connections between different users and targets, thereby improving the model's ability to characterize spatial correlations. To further enhance the model's ability to express feature dimension dependencies, the Mixer layer introduces a Channel-mixing MLP, whose output... As shown below:

[0150]

[0151] Where W3 and W4 represent the weight matrices of the two linear transformations, respectively. This represents the feature vector of the j-th token across all channels.

[0152] The output layer generates the initial solution for the alternating optimization method; the initial solution includes the augmented beamforming matrix. Base station receive beamforming matrix W U and the transmit power P of uplink communication users u The network is trained end-to-end, with a weighted loss function that minimizes CRB and maximizes energy efficiency as the optimization objective, thereby effectively alleviating the singularity problem of FIM and improving the accuracy and convergence speed of beamforming solution.

[0153] In this embodiment, an MLP-Mixer-AO framework is proposed to handle the non-convexity and coupling of the problem and obtain the optimal solution, as shown in Table 1:

[0154] Table 1

[0155]

[0156] The specific simulation parameters for this example are shown in Table 2.

[0157] Table 2

[0158]

[0159] Figure 3 This shows the change in the algorithm's CRB with the number of training iterations under MLP-Mixer initialization and traditional random initialization conditions. Figure 3As can be seen, the CRB values ​​under both initialization methods gradually decrease and tend to stabilize with the increase of the number of iterations, but there are significant differences in their convergence speed and final performance. Specifically, the traditional random initialization iteration speed is slower than that of MLP-Mixer initialization, and it requires more iterations to converge, resulting in a higher convergence value. In contrast, MLP-Mixer initialization requires fewer iterations, the algorithm converges faster, and the convergence value is lower. This is because MLP-Mixer initialization can provide more reasonable parameter initialization at the beginning of training, improve the weight distribution, and make the subsequent AO optimization process closer to the global optimum, thus accelerating the convergence speed.

[0160] Figure 4 A comparison of the sensing performance of various schemes under different base station transmit powers is presented. It can be seen that the sensing performance of all four algorithms improves with increasing base station transmit power. This is because higher transmit power enhances the signal-to-noise ratio of the echo signal, thereby reducing parameter estimation errors and improving target sensing accuracy. From the comparison results of different algorithms, traditional random initialization performs the worst. Convex relaxed boundary improves performance to some extent. The product Riemannian manifold optimization method further reduces the CRB (Credit Ratio Bias). The proposed MMA consistently exhibits the best sensing performance across the entire power range, proving the effectiveness of the proposed algorithm.

[0161] Figure 5 The table shows the user and data rate of four algorithms under different base station transmit powers. It can be seen that as the base station transmit power increases, the user and data rate of all algorithms show a significant upward trend, indicating a positive correlation between transmit power and received signal strength. This is because higher transmit power can improve the received signal strength, thereby improving link quality and data transmission rate. Furthermore, among the four algorithms, the MMA algorithm exhibits excellent performance at all base station transmit powers, performing well not only in the low-power region but also showing a more pronounced advantage in the high-power region. This is because the MLP-Mixer module introduced in the AO framework can provide initialization results closer to the optimal solution for iterative optimization through deep feature extraction and nonlinear mapping of channel state information, enabling the optimization process to converge faster and avoid getting trapped in local optima.

[0162] Example 2

[0163] This invention also provides a communication-sensing integrated beamforming system based on alternating optimization and hybrid multilayer perceptrons, for implementing the method of Embodiment 1, comprising:

[0164] The matrix construction module is used to add full-duplex technology to the ISAC system and construct the augmented beamforming matrix;

[0165] The constraint construction module is used to construct corresponding multi-objective joint constraints based on the augmented beamforming matrix.

[0166] The optimization problem construction module is used to construct beamforming optimization problems based on multi-objective joint constraints and mobile edge computing resource allocation, with the goal of maximizing the energy efficiency of the ISAC system.

[0167] The problem-solving module combines the MLP-Mixer network with the alternating optimization method to construct the MLP-Mixer-AO framework, and uses the MLP-Mixer-AO framework to solve the beamforming optimization problem, obtain the optimal solution, and complete the integrated beamforming for communication and sensing.

[0168] A further implementation method includes a matrix construction module comprising:

[0169] The perception model building unit is used to add full-duplex technology to the ISAC system and build an FD-ISAC perception model based on extended objectives.

[0170] The system model building unit is used to build the FD-ISAC-MEC system model based on the FD-ISAC perception model.

[0171] The matrix construction unit is used to design the FD-ISAC base station transmit signal matrix and construct the augmented beamforming matrix based on the FD-ISAC-MEC system model, by introducing orthogonal dedicated detection streams; wherein, the FD-ISAC base station transmit signal matrix is ​​designed as follows:

[0172]

[0173] This represents the augmented beamforming matrix of the FD-ISAC base station transmitter. Assigning a sub-matrix to the communication beam, To assist in sensing the beamforming matrix, For the augmented signal matrix, N t D represents the number of transmitting antennas, and D represents the number of downlink users per antenna.

[0174] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A communication-sensing integrated beamforming method based on alternating optimization and hybrid multilayer perceptrons, characterized in that, include: Full-duplex technology is incorporated into the ISAC system to construct an augmented beamforming matrix; Based on the augmented beamforming matrix, corresponding multi-objective joint constraints are constructed. Based on the aforementioned multi-objective joint constraints and combined with the allocation of mobile edge computing resources, a beamforming optimization problem is constructed with the goal of maximizing the energy efficiency of the ISAC system. By combining the MLP-Mixer network and the alternating optimization method, an MLP-Mixer-AO framework is constructed. The MLP-Mixer-AO framework is then used to solve the beamforming optimization problem to obtain the optimal solution, thus completing the integrated beamforming for communication and sensing. The beamforming optimization problem is decomposed into three interrelated sub-problems using an alternating optimization method. Two sets of variables are fixed, and the remaining set is optimized sequentially until the algorithm converges. The three sub-problems include the uplink user transmit power optimization problem, the base station receive beamforming matrix optimization problem, and the base station transmitter augmented beamforming matrix optimization problem. The method for solving each optimization problem includes: By fixing the augmenting beamforming matrix and the base station receiving beamforming matrix, the uplink communication user's transmit power optimization problem is modeled as a fractional programming problem. The Dinkelbach algorithm is used to introduce parameters to transform the fractional programming problem into an equivalent parameterized subproblem. The unconstrained optimal power solution is obtained by solving the first derivative. With fixed uplink user transmit power and augmented beamforming matrix, the optimization problem of base station receive beamforming matrix is ​​expressed as the uplink user and rate maximization problem. Using the generalized Rayleigh quotient principle, the uplink user and rate maximization problem is transformed into a generalized eigenvalue problem. The optimal solution of the generalized eigenvalue problem is the eigenvector corresponding to the maximum generalized eigenvalue of the preset matrix pair. After normalization by the unit norm, the optimal receive beamforming vector is obtained. With fixed uplink user transmit power and base station receive beamforming matrix, the optimization problem of augmented beamforming matrix at the base station transmitter is modeled as a non-convex problem involving Cramer-Rao bounds and system energy efficiency weighting. A semidefinite relaxation technique is used to introduce a covariance variable into the non-convex problem, transforming it into a convex semidefinite programming problem. A successive convex approximation method is then used to linearize the non-convex terms of the objective function of the convex semidefinite programming problem to obtain a convex optimization problem. Solving the convex optimization problem yields the optimal covariance matrix, and finally, the augmented beamforming matrix is ​​recovered.

2. The communication-sensing integrated beamforming method based on alternating optimization and hybrid multilayer perceptrons according to claim 1, characterized in that, Methods for constructing augmented beamforming matrices include: Full-duplex technology is incorporated into the ISAC system to construct an FD-ISAC sensing model based on extended objectives; Based on the FD-ISAC sensing model, construct the FD-ISAC-MEC system model; Based on the FD-ISAC-MEC system model, an orthogonal dedicated detection stream is introduced to design the FD-ISAC base station transmit signal matrix and construct an augmented beamforming matrix; the FD-ISAC base station transmit signal matrix is ​​designed as follows: This represents the augmented beamforming matrix of the FD-ISAC base station transmitter. Assigning a sub-matrix to the communication beam, To assist in sensing the beamforming matrix, For augmented signal matrix, N t D represents the number of transmitting antennas, and D represents the number of downlink users per antenna.

3. The communication-sensing integrated beamforming method based on alternating optimization and hybrid multilayer perceptrons according to claim 1, characterized in that, The multi-objective joint constraints include the minimum signal-to-interference-plus-noise ratio (SIR) constraint for downlink communication users, the minimum SIR constraint for uplink communication users, the total transmit power constraint of the base station, the transmit power range constraint for uplink user terminals, the task processing delay constraint, and the task offloading ratio constraint.

4. The communication-sensing integrated beamforming method based on alternating optimization and hybrid multilayer perceptrons according to claim 3, characterized in that, Methods for constructing beamforming optimization problems include: The total system power is calculated based on the sum of the uplink communication user's transmit power, the device's inherent circuit power consumption, the computing power, and the circuit's static power consumption. Energy efficiency is calculated based on the total system power, the minimum signal-to-interference-plus-noise ratio (SINR) constraints for downlink and uplink communication users. Based on the system's Cramer-Rao bound, resource allocation weighting factor, and energy efficiency, combined with the multi-objective joint constraints, the beamforming optimization problem is constructed.

5. The communication-sensing integrated beamforming method based on alternating optimization and hybrid multilayer perceptrons according to claim 1, characterized in that, The MLP-Mixer network is used to provide an initial solution for the alternating optimization method; the structure of the MLP-Mixer network includes: The input layer is used to map the raw system state information into a fixed-dimensional feature representation; the raw system state information includes the uplink and downlink user signal-to-interference-plus-noise ratio thresholds and the total power budget. Convolutional layers are used to locally model the feature representation to obtain spatial feature information; The Mixer layer is used to convert the spatial feature information into two-dimensional data and divide it into patches. The Mixer layer is divided into a token mixing multilayer perceptron and a channel mixing multilayer perceptron. Each multilayer perceptron consists of two fully connected layers and a ReLU activation function. The Per-patch Fully-connected layer is used to convert the patch into a corresponding feature embedding sequence; the feature embedding sequence is then embedded into the Mixer layer for feature extraction and mixing. The output layer is used to generate an initial solution for the alternating optimization method; the initial solution includes the augmented beamforming matrix, the base station receive beamforming matrix, and the transmit power of the uplink communication user.

6. A communication-sensing integrated beamforming system based on alternating optimization and hybrid multilayer perceptrons, used to implement the method described in any one of claims 1-5, characterized in that, include: The matrix construction module is used to add full-duplex technology to the ISAC system and construct the augmented beamforming matrix; The constraint construction module is used to construct corresponding multi-objective joint constraint conditions based on the augmented beamforming matrix. The optimization problem construction module is used to construct a beamforming optimization problem based on the multi-objective joint constraints and the allocation of mobile edge computing resources, with the goal of maximizing the energy efficiency of the ISAC system. The problem-solving module is used to combine the MLP-Mixer network and the alternating optimization method to construct the MLP-Mixer-AO framework, and use the MLP-Mixer-AO framework to solve the beamforming optimization problem, obtain the optimal solution, and complete the integrated beamforming of communication and sensing.

7. The integrated communication and sensing beamforming system based on alternating optimization and hybrid multilayer perceptrons according to claim 6, characterized in that, The matrix construction module includes: The perception model building unit is used to add full-duplex technology to the ISAC system and build an FD-ISAC perception model based on extended objectives. The system model building unit is used to build the FD-ISAC-MEC system model based on the FD-ISAC perception model. The matrix construction unit is used to design the FD-ISAC base station transmit signal matrix and construct the augmented beamforming matrix based on the FD-ISAC-MEC system model, by introducing orthogonal dedicated detection streams; wherein, the FD-ISAC base station transmit signal matrix is ​​designed as follows: This represents the augmented beamforming matrix of the FD-ISAC base station transmitter. Assigning a sub-matrix to the communication beam, To assist in sensing the beamforming matrix, For augmented signal matrix, N t D represents the number of transmitting antennas, and D represents the number of downlink users per antenna.

Images