A multi-agent collaborative control method and system based on the integration of common sense and artificial intelligence

Abstract

The application discloses a kind of multi-agent collaborative control method and system based on the integration of general sense algorithm wisdom, obtain dynamics, channel, radar, task queue and energy state, construct by Euclidean space, symmetric positive definite matrix manifold and non-negative real number space Generalized state representation;Predict channel state, target trajectory and uncertainty covariance using conditional variational inference network;The prediction result and Lyapunov drift penalty term are introduced into local mean field stochastic state equation and transient stage penalty, and candidate resource strategy and resource marginal technical utility are generated by deep Galerkin method;Under the constraint of total resource, obtain communication, sensing and computing resource allocation vector, according to which adjust event trigger threshold, update parameters and model predictive control solving parameters;Control input is generated under the constraint of control barrier function, and safety fallback quadratic programming is executed when high uncertainty, safety check or solving fails, to realize closed-loop cooperation of resource scheduling, environment prediction and safety control.

Description

Technical Field

[0001] This invention belongs to the technical field of intelligent agent collaborative control, specifically relating to a multi-agent collaborative control method and system based on the integration of sensing, computing and intelligence. Background Technology

[0002] With the development of intelligent decision-making, the Internet of Things, and fifth- and sixth-generation mobile communication technologies, multi-agent systems are gradually moving from theoretical research to practical applications, undertaking complex collaborative tasks in scenarios such as inspection and monitoring, emergency rescue, traffic control, and logistics distribution. Multiple agents need to complete spatiotemporal collaborative tasks that are difficult for a single agent to accomplish independently through continuous information interaction and behavioral coordination.

[0003] Existing multi-agent cooperative control technologies still face the following challenges: First, communication, sensing, computing, and control are typically designed and optimized independently. The communication side prioritizes throughput or bit error rate, the sensing side prioritizes detection accuracy, the computing side prioritizes task processing latency, and the control side often treats changes in communication and sensing as external disturbances, making it difficult to achieve overall optimization on resource-constrained platforms. Second, in urban low-altitude environments, complex electromagnetic environments, or scenarios with rapid task switching, channel quality, interference levels, and task loads may change rapidly. Fixed-cycle control and communication methods are difficult to adapt in a timely manner, easily leading to link congestion, latency jitter, and degraded control performance. Third, as the number of agents increases, the number of system state variables and control variables rises rapidly, making centralized resource scheduling and control solution difficult to complete within a finite decision-making cycle. Fourth, data-driven control methods may output control commands that violate physical safety constraints under extreme conditions, lacking a correction mechanism directly coupled with verifiable safety constraints.

[0004] Therefore, in view of the above-mentioned shortcomings of the existing technology, the present invention discloses a multi-agent collaborative control method and system based on the integration of sensing, computing and intelligence. Summary of the Invention

[0005] This invention discloses a multi-agent collaborative control method and system based on the integration of sensing, computing and intelligence, to solve the problems of resource-separated scheduling, failure of environmental prediction results to enter control decisions, high burden of large-scale collaborative solution, and lack of verifiable safety constraints in data-driven control.

[0006] This invention is achieved through the following technical solution: A multi-agent cooperative control method based on the integration of sensing, computing, and intelligence is implemented using a distributed cooperative system containing agent kinematic models. The distributed cooperative system includes several agents, with at least one agent configured as a resource coordination node. The method includes the following steps: Step 1: Obtain the state data of the agent, and construct a generalized state representation parameter based on the state data, consisting of Euclidean space feature vectors, symmetric positive definite matrices, and non-negative real space vectors; Step 2: Input the generalized state representation parameters into the conditional variational inference network to obtain prediction results including predicted channel state, predicted target trajectory, and uncertainty covariance; Step 3: Construct a Lyapunov drift plus penalty term based on the energy consumption parameters in the task queue data and energy state data. Introduce the prediction results into the local mean field stochastic state equation. Use the Lyapunov drift plus penalty term as the instantaneous cost of the local mean field stochastic state equation. Introduce the neighbor weighted empirical distribution as the local mean field state of the local mean field stochastic state equation. Then, use the deep Galerkin method to solve the local mean field stochastic state equation to obtain the agent's resource marginal technical utility and candidate resource strategy. Step 4: Limit the scope of the agent's resource requirements with candidate resource strategies, construct a technology utility function with the marginal technical utility of resources as the local slope, and maximize the sum of the agent's technical utility under the constraint of the total amount of communication, perception and computing resources. Solve the agent's resource allocation vector and resource sufficiency based on the technology utility function. Step 5: Calculate the communication update parameters and perception update parameters of the agent based on the resource allocation vector, adjust the predictive control solution parameters of the agent's kinematic model based on the resource allocation vector, and calculate the dynamic event triggering threshold of the agent based on the resource sufficiency. Step 6: Establish a safety constraint function based on the prediction results and introduce a control barrier function constraint. Introduce the safety constraint function and the control barrier function constraint into the agent's kinematic model and solve for the agent's control input parameters. Execute the agent based on the control input parameters, obtain the state data after the agent's execution, and update the generalized state representation parameters of the agent in the next control cycle.

[0007] To better realize the present invention, step 2 further includes: Step 2.1: Extract the historical window Euclidean feature sequence of the agent from the generalized state representation parameters, and input the historical window Euclidean feature sequence into the conditional variational inference network; Step 2.2: The latent variable mean and variance parameters are output by the long short-term memory encoder in the conditional variational inference network, and the conditional distribution is output by the decoder in the conditional variational inference network. Step 2.3: Based on the latent variable mean and distribution parameters, the conditional distribution function is obtained by representing the conditional distribution using Gaussian likelihood. The decoder obtains the lower triangular covariance factor by parameterizing the variance parameter through the lower triangular matrix. Step 2.4: Establish a reconstruction loss training objective function based on the conditional distribution function and the lower triangular covariance factor, and train the conditional variational inference network based on the reconstruction loss training objective function to obtain the prediction results.

[0008] To better realize the present invention, the reconstruction loss training objective function is further defined as follows: ; in: [...] represents the reconstruction loss training objective function; Represents the Euclidean feature sequence of the historical window; This represents the channel state and target trajectory sequence for the future window. This represents the mean of the latent variable; Represents the conditional distribution function; This represents an approximate posterior distribution; Represents the conditional prior distribution; Indicates KL divergence; Represents the relative entropy regularization weight that is greater than or equal to zero.

[0009] To better realize the present invention, further, in step 3, the approximate value function of the agent is calculated using the deep Galerkin method, the Euclidean eigenvectors of neighboring agents are aggregated using a graph neural network, the non-negative similarity weights of each component in the Euclidean eigenvectors are normalized to a sum equal to one, and the local mean field state is formed by the weighted empirical distribution of the states of neighboring agents; the local mean field state is input into the approximate value function, and candidate resource strategies are obtained according to the gradient of the approximate value function and the extreme value condition of the Hamiltonian; the marginal technical utility of resources is obtained according to the decrease in the optimal Hamiltonian before and after adding continuous resource components or discrete minimum resource units.

[0010] To better realize the present invention, step 4 further includes: Step 4.1: Extract total resource constraints and agent resource demand constraints from candidate resource strategies; Step 4.2: Using the marginal technical utility of resources as the local slope, construct the technical utility function. Under the constraints of total resource amount and agent resource demand, with the goal of maximizing the sum of agent technical utility, solve the agent's resource allocation vector through the technical utility function. Step 4.3: Set a preset non-negative weight for each component in the resource allocation vector, and the sum of the preset non-negative weights of all components is equal to one. Sum the components and truncate them to the value range of zero to one to obtain the resource sufficiency.

[0011] To better realize the present invention, further, in step 5, a dynamic event triggering threshold function is established based on resource sufficiency and uncertainty covariance in the prediction results, a dynamic event triggering threshold of the agent is calculated based on the dynamic event triggering threshold function, and a dynamic event triggering condition function of the agent is established based on the dynamic event triggering threshold. The dynamic event triggering threshold function is: ; in: This represents a threshold function for dynamic event triggering; Indicates resource sufficiency; Indicates the uncertainty covariance; This indicates the lower limit of the threshold for triggering dynamic events; This represents the threshold adjustment amplitude coefficient; Indicates the stabilization parameter; The scale factor represents the uncertainty; tr(...) represents the trace of the matrix, which is the sum of the elements on the main diagonal of the square matrix; The dynamic event triggering condition function is: ; in: This represents a threshold function for dynamic event triggering; Indicates the measured state value of the event; Indicates the predicted state value of the event; This represents the prediction error between the measured state value of an event and the predicted state value of the event.

[0012] To better realize the present invention, the control barrier function constraint in step 6 further includes: For a safety constraint function of relative order one, apply: ; in: Represents the safety constraint function; This represents the transpose of the gradient vector of the safety constraint function with respect to the system state x; Represents the drift vector field; Represents the input vector field; Indicates the control input parameters; Represents the extended class K function for safety constraint functions; For safety constraint functions with a relative order greater than one, the m-th order recursive function of the safety constraint function is established as follows: , 0≤k≤m, ; in: This represents the k-th order recursive function of the safety constraint function; This represents the (k-1)th order recursive function of the safety constraint function; express The time derivative; (...) represents the extended class K function.

[0013] To better realize the present invention, the candidate resource strategy further includes communication resources, sensing resources, and computing resources; the communication resources include transmit power, orthogonal frequency division multiplexing subcarriers, time slots, and beamforming parameters; the sensing resources include radar beam pointing, beamwidth, pulse repetition frequency, and coherent accumulation time; the computing resources include processor clock frequency, edge offload ratio, and number of solution iterations.

[0014] To better realize the present invention, further, the state data of the intelligent agent in step 1 includes dynamic state data, channel data, radar data, task queue data, and energy state data. The dynamic state data is represented as a Euclidean space vector, the covariance in the channel data and radar data is represented as a symmetric positive definite matrix, and the task queue data and energy state data are represented as non-negative real space vectors. At the selected reference point, logarithmic mapping and vectorization are performed on the symmetric positive definite matrix, and concatenation and normalization processing is performed on the Euclidean space vector and the non-negative real space vector to form a Euclidean feature vector.

[0015] A multi-agent cooperative control system based on integrated sensing, sensing, and computing technologies is implemented using a multi-agent cooperative control method. It includes an integrated sensing radio frequency front-end, sensor interfaces, agent communication interfaces, several agent computing units, a memory, and an actuator interface. At least one agent computing unit is configured as a resource coordination node. The memory stores a state construction module, a joint prediction module, a policy generation module, a resource allocation module, a security control module, and a feedback update module. The state construction module constructs a generalized state representation based on data collected from each interface. The joint prediction module outputs the future channel state, target trajectory, and uncertainty covariance. The policy generation module generates candidate resource policies and resource marginal utility. The resource allocation module, executed by the resource coordination node, outputs communication, sensing, and computing resource allocation vectors. The security control module generates control inputs that satisfy security constraints based on the resource allocation vectors and uncertainty covariance. The feedback update module writes the measured state, task service volume, and energy consumption after execution into the generalized state representation for the next control cycle. The multiple agent computing units execute the modules and output control inputs via the actuator interface.

[0016] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) This invention places generalized state representation, joint environment prediction, resource strategy generation, resource optimization allocation and security control correction in the same control cycle closed loop, so that the prediction mean, uncertainty covariance and resource allocation result are no longer independent post-processing quantities, but are respectively entered into the stochastic state equation, technical utility, event triggering threshold, robust constraint and security backoff criterion. (2) This invention improves the consistency of heterogeneous state representation by product manifold state modeling, reduces the input dimension of large-scale multi-agent policies by local mean field estimation, avoids high-dimensional grid discretization by deep Galerkin approximation, and enhances the adaptability under different environmental conditions by finite element learning. (3) This invention applies safety constraints to the control input under the conditions that the optimization problem is feasible and the model error is bounded by model predictive control and control barrier function safety correction, and switches to safe backoff quadratic programming when there is high uncertainty or the solution fails, thereby improving the robustness in system engineering applications. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a multi-agent cooperative control method. Figure 2 This is a schematic diagram of the structure of a multi-agent cooperative control system; Figure 3 A flowchart for generating resource strategies based on deep Galerkin method and meta-learning; Figure 4 A flowchart illustrating the interaction between the outer loop of resource allocation and the inner loop of security control; Figure 5 This is a schematic diagram of the joint environment prediction network structure based on variational inference. Detailed Implementation

[0018] Example 1: This embodiment discloses a multi-agent cooperative control method based on the integration of sensing, computing, and intelligence. It is implemented using a distributed cooperative system containing agent kinematic models. The distributed cooperative system includes several agents, and at least one agent is configured as a resource coordination node, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain the state data of the agent, and construct a generalized state representation parameter based on the state data, consisting of Euclidean space feature vectors, symmetric positive definite matrices, and non-negative real space vectors; Step 2: Input the generalized state representation parameters into the conditional variational inference network to obtain prediction results including predicted channel state, predicted target trajectory, and uncertainty covariance; Step 3: Construct a Lyapunov drift plus penalty term based on the energy consumption parameters in the task queue data and energy state data. Introduce the prediction results into the local mean field stochastic state equation. Use the Lyapunov drift plus penalty term as the instantaneous cost of the local mean field stochastic state equation. Introduce the neighbor weighted empirical distribution as the local mean field state of the local mean field stochastic state equation. Then, use the deep Galerkin method to solve the local mean field stochastic state equation to obtain the agent's resource marginal technical utility and candidate resource strategy. Step 4: Limit the scope of the agent's resource requirements with candidate resource strategies, construct a technology utility function with the marginal technical utility of resources as the local slope, and maximize the sum of the agent's technical utility under the constraint of the total amount of communication, perception and computing resources. Solve the agent's resource allocation vector and resource sufficiency based on the technology utility function. Step 5: Calculate the communication update parameters and perception update parameters of the agent based on the resource allocation vector, adjust the predictive control solution parameters of the agent's kinematic model based on the resource allocation vector, and calculate the dynamic event triggering threshold of the agent based on the resource sufficiency. Step 6: Establish a safety constraint function based on the prediction results and introduce a control barrier function constraint. Introduce the safety constraint function and the control barrier function constraint into the agent's kinematic model and solve for the agent's control input parameters. Execute the agent based on the control input parameters, obtain the state data after the agent's execution, and update the generalized state representation parameters of the agent in the next control cycle.

[0019] Specifically: The system acquires the dynamic states, channel measurement data, radar sensing data, local and edge task queue states, and energy states of each agent. The dynamic states are represented as vectors in Euclidean space. The channel or radar observation covariance matrix is ​​symmetricized and added to a positive multiple of the identity matrix to form a positive definite matrix in a symmetric positive definite matrix manifold. The task queue states and energy states are represented as vectors in non-negative real space, and the discrete resource index is independently encoded as a category feature.

[0020] Based on the geometric properties of each state, a generalized state representation on the product manifold is constructed. Logarithmic mapping and vectorization are performed on the positive definite matrix. After adding a positive bias to the continuous non-negative states, a logarithmic transformation is performed. The transformation result, the dynamic state vector, and the category features are concatenated into a Euclidean eigenvector. A Lyapunov drift penalty term is constructed based on the task queue and energy consumption.

[0021] The Euclidean eigenvectors of multiple consecutive control cycles are input into the conditional variational inference network to parameterize the conditional probability distribution of the future sequence, thereby obtaining the future channel state prediction, target trajectory prediction, and the corresponding uncertainty covariance.

[0022] The predicted future channel state and target trajectory are used to construct the drift term of the stochastic state equation for each agent. The uncertainty covariance is transformed by a preset mapping matrix to construct the diffusion covariance or risk penalty term, and the Lyapunov drift penalty term is used as the instantaneous cost term. Using the neighbor-weighted empirical distribution as the local mean field state, the approximate value function of each agent is solved using the deep Galerkin method. Candidate resource strategies are obtained based on the gradient of the approximate value function and the Hamiltonian extremum condition. The marginal technical utility of resources is obtained based on the decrease in the optimal Hamiltonian before and after adding continuous resource components or discrete minimum resource units. Meta-learning is used to update the policy parameters within a preset update step and parameter boundary based on environmental samples. The deep residual network used for the approximate value function includes fully connected layers and skip connections, and employs a hyperbolic tangent activation function with continuous second derivatives to support the automatic differentiation calculation of the first-order gradient and second-order derivative matrix of the approximate value function.

[0023] The resource requirements of each agent are limited by the candidate resource strategy, and the marginal technical utility of the resources is used as the local slope of the technical utility function. A technical utility function that decreases with the increase of trajectory tracking error, packet loss rate, perception uncertainty, task queue backlog and energy consumption is constructed. Under the constraint of total communication, perception and computing resources, the sum of the technical utility of each agent is maximized to obtain the dimensionless resource allocation vector and resource sufficiency of each agent.

[0024] The event trigger threshold, communication and perception update parameters, and solution time limit or iteration number of model predictive control are adjusted according to the resource allocation vector. State prediction and robust safety constraints are established based on the target trajectory prediction and uncertainty covariance. When the trace of uncertainty covariance does not exceed the preset entry threshold and the event triggering condition is met, the model predictive control problem with control barrier function constraints is solved. When the event triggering condition is not met, a feasibility check of the control barrier function is performed on the open-loop predictive control input or the previous control input. When the check fails, the trace of uncertainty covariance exceeds the preset entry threshold, or the solver fails to obtain a feasible solution within the specified time limit, a safe quadratic programming problem is solved with the backoff input as the nominal input and satisfying the input amplitude and minimum safe distance constraints. The obtained control input is sent to the agent actuator, and the dynamic measured state, task queue state, task service quantity, and energy state after execution are obtained to update the generalized state representation of the next control cycle.

[0025] Furthermore, in step 1, an atomic resource pool is established, which includes time-frequency resource blocks, radar beam dwell time, calculation cycle frequency, and battery energy budget; the dynamic state is represented as a vector in Euclidean space, the channel or radar observation covariance is represented as a matrix in a symmetric positive definite matrix manifold, and the task queue and energy state are represented as vectors in non-negative real space; at a selected reference point, the symmetric positive definite matrix is ​​logarithmically mapped and vectorized, then concatenated with the remaining state components and normalized to form an Euclidean eigenvector for processing in steps 2 and 3.

[0026] Symmetric positive definite matrix manifolds employ affine invariant metric The product manifold distance obtained based on the metric is used to determine the adjacency weights of the graph neural network or the similarity weights of the joint prediction network; wherein, It is a symmetric positive definite matrix of the same dimension; The order of the matrix; for The inverse square root; The logarithm of the principal matrix of a symmetric positive definite matrix; It is the Frobenius norm; Let be the geodesic distance under the affine invariant Riemannian metric, and hour .

[0027] Let the task queue set be The number of agents is The Lyapunov function is Conditional drift is Minimize drift plus penalty in each control cycle. The upper boundary.

[0028] in: Indicates the first A task queue at time The backlog; This represents the weighted energy consumption for communication, sensing, and computing. The weights are non-negative.

[0029] The upper bound of the Lyapunov drift penalty term is used as the instantaneous cost term in the stochastic differential game of step three; the neighbor state similarity weights are determined based on the distance of the product manifold; wherein... For controlling the periodic index; Index for intelligent agents or task queues; The total number of agents; For the first An intelligent agent at time The backlog of tasks, expressed in bits, data packets, or number of tasks; A collection of task queues; It is a quadratic Lyapunov function; For a given current set of queues, perform a single-step conditional Lyapunov drift; This represents the conditional expectation regarding the current queue state; Weighted energy consumption for communication, sensing, and computing; This is a parameter that balances resource consumption and queue stability. The larger the value, the more emphasis is placed on reducing resource and energy consumption.

[0030] Furthermore, in step 2, as Figure 5 As shown, the historical generalized state sequence The input is a conditional variational inference network consisting of a long short-term memory encoder, a spatiotemporal attention module, a reparameterized sampling layer, and a decoder; the encoder outputs the mean and variance parameters of the latent variables; the decoder does not directly output the future sequence with deterministic values, but instead outputs the conditional distribution. The distribution parameters, where This includes the future channel state sequence and the target trajectory sequence; when using Gaussian likelihood, the conditional distribution is expressed as... ,in To predict the mean, For uncertain covariance, (...) represents a multivariate Gaussian distribution; the decoder uses a lower triangular matrix. Parameterized covariance.

[0031] The decoder includes a channel state prediction head and a target state distribution prediction head. The channel state prediction head outputs future channel quality, interference power, or channel matrix parameters, while the target state distribution prediction head simultaneously outputs the target trajectory mean and uncertainty covariance.

[0032] make ,and The diagonal elements are positiveized using the softplus or exponential function. The network is trained using the predicted negative log-likelihood and relative entropy regularization term, and the reconstruction loss training objective function is: ; in: [...] represents the reconstruction loss training objective function; Represents the Euclidean feature sequence of the historical window; This represents the channel state and target trajectory sequence for the future window. This represents the mean of the latent variable; Represents the conditional distribution function; This represents an approximate posterior distribution; Represents the conditional prior distribution; Indicates KL divergence; Represents the relative entropy regularization weight that is greater than or equal to zero.

[0033] Furthermore, in step 3, a drift term is constructed from the future channel state prediction and target trajectory prediction to form a stochastic state equation; a diffusion term or risk penalty term is constructed from the uncertainty covariance; and an instantaneous cost is constructed from the drift penalty term. A deep residual network is used to approximate the approximate value function. The deep residual network is trained by sampling in the state domain and minimizing the Hamilton-Jacobi-Bellman equation residuals, terminal condition residuals, and boundary condition residuals. Candidate resource strategies are determined based on the gradient of the approximate value function and the extreme value conditions of the Hamiltonian. The decrease in the optimal Hamiltonian before and after adding continuous resource components or discrete minimum resource units is used as the marginal technical utility of the resources. The inner loop of meta-learning updates the strategy parameters using new environment samples, and the outer loop updates the initialization parameters. The number of update steps, learning rate, and parameter projection interval in the online phase are all limited by preset boundaries.

[0034] Furthermore, such as Figure 3 As shown, in step 3, the approximate value function of the agent is calculated using the deep Galerkin method. A graph neural network is used to aggregate the Euclidean eigenvectors of neighboring agents. The non-negative similarity weights of each component in the Euclidean eigenvectors are normalized to a sum equal to one. A local mean-field state is formed by the weighted empirical distribution of the states of neighboring agents. The local mean-field state is input into the approximate value function to reduce the policy input dimension that increases with the number of agents. When the instantaneous cost or drift term change caused by a single neighbor's state change exceeds a preset weak coupling upper bound, the neighbor set is reduced or a centralized resource coordination method is switched. Candidate resource strategies are obtained based on the gradient of the approximate value function and the Hamiltonian extremum condition. The marginal technical utility of the resource is obtained based on the decrease in the optimal Hamiltonian before and after adding continuous resource components or discrete minimum resource units.

[0035] Furthermore, such as Figure 4 As shown, step 4 specifically includes: Step 4.1: Extract total resource constraints and agent resource demand constraints from candidate resource strategies; Step 4.2: Using the marginal technical utility of resources as the local slope, construct the technical utility function. Under the constraints of total resource amount and agent resource demand, with the goal of maximizing the sum of agent technical utility, solve the agent's resource allocation vector through the technical utility function. Step 4.3: Set a preset non-negative weight for each component in the resource allocation vector, and the sum of the preset non-negative weights of all components is equal to one. Sum the components and truncate them to the value range of zero to one to obtain the resource sufficiency.

[0036] The communication, sensing, and computing resources are normalized according to their respective upper and lower limits. Based on the marginal technical utility of resources output in step three, a technical utility function for each agent is constructed. A resource optimization problem is solved with the objective of maximizing the sum of the technical utilities of all agents, constrained by the total amount of each type of resource and the upper limit of the resource demand of each agent, to obtain a dimensionless resource allocation vector. The resource sufficiency is obtained by summing the components of the dimensionless resource allocation vector according to a preset non-negative weight that sums to one and truncating the result to a value range of zero to one. ;in, For the first A dimensionless resource allocation vector for each agent; The total number of resource types, Represents the set of positive integers; For the first Classified and normalized resource components; For resource sufficiency; each resource sufficiency weight is non-negative and the sum of the weights is . Its physical meaning is the degree to which the current allocation of resources comprehensively meets the demand for candidate resources.

[0037] The resource optimization problem is solved using a distributed primal-dual algorithm. Each agent exchanges dimensionless resource demand and dual variables with the resource coordination node. The stopping condition is that neither the primal residual nor the dual residual is higher than a preset threshold, or the preset maximum number of iterations is reached. The feasible resource allocation result at the time of stopping is used in the current control cycle.

[0038] Furthermore, in step 5, a dynamic event triggering threshold function is established based on resource sufficiency and uncertainty covariance in the prediction results; the dynamic event triggering threshold of the agent is calculated based on the dynamic event triggering threshold function; and the dynamic event triggering condition function of the agent is established based on the dynamic event triggering threshold.

[0039] The dynamic event triggering threshold function is:

[0040] ; in: This represents a threshold function for dynamic event triggering; Indicates resource sufficiency; Indicates the uncertainty covariance; This indicates the lower limit of the threshold for triggering dynamic events; This represents the threshold adjustment amplitude coefficient; Indicates the stabilization parameter; The scale factor represents the uncertainty; tr(...) represents the trace of the matrix, which is the sum of the elements on the main diagonal of the square matrix.

[0041] The dynamic event triggering condition function is:

[0042] ; in: This represents a threshold function for dynamic event triggering; Indicates the measured state value of the event; Indicates the predicted state value of the event; This represents the prediction error between the measured state value of an event and the predicted state value of the event.

[0043] When the trace of the uncertainty covariance exceeds a preset threshold, backoff control is executed directly; otherwise, if the event triggering condition is not met, a feasibility check of the control barrier function is performed on the open-loop predictive input or the previous control input. If the check fails, backoff control is executed. and These are the measured state and the predicted state, respectively. Represents an n-dimensional real vector space; This refers to the state prediction error; Threshold for triggering dynamic events; This is the lower limit of the threshold. The threshold adjustment range; To prevent stabilization parameters with a denominator of zero; The scaling factor for the uncertainty trace; To predict the total variance of the covariance; For resource sufficiency. The aforementioned... Used for positive covariance definiteness These are mutually independent parameters.

[0044] Furthermore, the control barrier function constraint in step 6 includes: For a safety constraint function of relative order one, apply:

[0045] ; in: Represents the safety constraint function; This represents the transpose of the gradient vector of the safety constraint function with respect to the system state x; Represents the drift vector field; Represents the input vector field; Indicates the control input parameters; Represents the extended class K function for safety constraint functions; For safety constraint functions with a relative order greater than one, the m-th order recursive function of the safety constraint function is established as follows:

[0046] , 0≤k≤m, ; in: This represents the k-th order recursive function of the safety constraint function; This represents the (k-1)th order recursive function of the safety constraint function; express The time derivative; (...) represents the extended class K function.

[0047] Furthermore, the candidate resource strategy includes communication resources, sensing resources, and computing resources; the communication resources include transmit power, orthogonal frequency division multiplexing subcarriers, time slots, and beamforming parameters; the sensing resources include radar beam pointing, beamwidth, pulse repetition frequency, and coherent accumulation time; and the computing resources include processor clock frequency, edge offload ratio, and number of solution iterations.

[0048] Furthermore, the agent's state data in step 1 includes dynamic state data, channel data, radar data, task queue data, and energy state data. The dynamic state data is represented as a Euclidean space vector, the covariance of the channel and radar data is represented as a symmetric positive definite matrix, and the task queue and energy state data are represented as non-negative real space vectors. At a selected reference point, a logarithmic mapping and vectorization are performed on the symmetric positive definite matrix, and the Euclidean space vector and non-negative real space vector are concatenated and normalized to form a Euclidean eigenvector. The channel data includes path loss, shadowing fading, multipath fading parameters, and interference noise power spectral density; the radar data includes range, angle, Doppler velocity, radar cross section, and observation covariance; the task queue data includes the amount of local tasks to be processed, associated deadline parameters, and the amount of tasks to be unloaded at the edge.

[0049] Example 2: This embodiment discloses a multi-agent collaborative control system based on the integration of sensing, computing, and intelligence, which is an optimization based on Embodiment 1, such as... Figure 2As shown, the system includes an integrated sensing and communication RF front-end, a sensor interface, an agent communication interface, several agent computing units, a memory, and an actuator interface. At least one agent computing unit is configured as a resource coordination node. The memory stores a state construction module, a joint prediction module, a policy generation module, a resource allocation module, a security control module, and a feedback update module. The state construction module constructs a generalized state representation based on data collected from each interface. The joint prediction module outputs the future channel state, target trajectory, and uncertainty covariance. The policy generation module generates candidate resource policies and resource marginal utility. The resource allocation module, executed by the resource coordination node, outputs communication, sensing, and computing resource allocation vectors. The security control module generates control inputs that satisfy security constraints based on the resource allocation vectors and uncertainty covariance. The feedback update module writes the measured state, task service volume, and energy consumption after execution into the generalized state representation of the next control cycle. The multiple agent computing units execute the modules and output control inputs via the actuator interface.

[0050] Furthermore, the integrated sensing RF front-end adopts a multiple-input multiple-output array, sharing the RF link and antenna aperture through time-division or frequency-division; the intelligent agent computing unit includes a programmable logic preprocessing unit and a graphics processing computing unit, which communicate through a high-speed interconnect bus; the programmable logic preprocessing unit is used to perform channel estimation, pulse compression and timestamp synchronization, and the graphics processing computing unit is used to perform joint prediction network inference, value function approximation calculation and model predictive control or secure quadratic programming solution.

[0051] The rest of this embodiment is the same as that of Embodiment 1, so it will not be described again.

[0052] Example 3: This embodiment discloses a multi-agent cooperative control method based on integrated sensing, computing, and intelligence, applied to collaborative inspection of a drone swarm. It optimizes upon Embodiment 1, enabling multiple drones to perform collaborative inspection and target area coverage tasks in complex electromagnetic interference environments. A joint environment prediction network predicts the interference intensity distribution and available communication windows within future time windows. A resource optimization module prioritizes the allocation of communication and computing resources to nodes undertaking critical observation or relay tasks. Model predictive control and control barrier functions are used to maintain track tracking and safe inter-drone distances. Validation metrics include communication interruption rate, task completion rate, minimum inter-drone distance, and energy consumption per unit task.

[0053] In this embodiment, the dynamic state of the i-th UAV includes at least position, velocity, and attitude, and the control input includes at least thrust and attitude control parameters. For any two UAVs i and j, an inter-UAV safety function is constructed by subtracting the square of the minimum safe distance from the square norm of their position difference, and an electronic fence safety function is constructed based on the mission area boundary and the no-fly zone boundary. The resource coordination node outputs bandwidth, radar beam dwell time, and available solution iterations in each control cycle, which are used to limit the communication update frequency, perception update frequency, and computational budget for model predictive control, respectively.

[0054] Model predictive control targets the weighted sum of trajectory tracking error, control energy, and control rate of change, and tightens position, velocity, and safety distance constraints using confidence boundaries of the uncertainty covariance; control barrier function constraints are applied at each prediction step. If the trace of the uncertainty covariance exceeds a preset entry threshold, the barrier function feasibility check fails, or the solver times out, safety quadratic programming finds the feasible control input within the actuator input boundary that minimizes the nominal input distance to hovering, deceleration, or maintaining a safety distance.

[0055] The entry threshold, exit threshold, minimum safe distance, prediction time domain, solution time limit, and weights of each technical utility are calibrated before deployment based on the platform dynamic boundary, sensor error upper bound, and control cycle. The exit threshold is set to be less than the entry threshold, and the feasibility recovery condition is required to be met for multiple consecutive control cycles to avoid frequent switching between normal control and backoff control.

[0056] The rest of this embodiment is the same as that of Embodiment 1, so it will not be described again.

[0057] Example 4: This embodiment discloses a multi-agent collaborative control method based on integrated sensing, computing, and intelligence, applied to UAV urban low-altitude logistics delivery. It optimizes upon Embodiment 1, enabling multiple logistics UAVs to operate at high density in urban low-altitude airspace. The system adjusts the group's position distribution based on local mean field distribution to avoid local congestion and utilizes control barrier functions to construct electronic fences and collision avoidance constraints. Resource coordination nodes adjust scheduling priorities based on task deadlines, queue backlog, and marginal resource utility. Verification metrics include the number of agents that can be accommodated per unit airspace, average task completion time, minimum safe distance, and the number of times control is infeasible.

[0058] The rest of this embodiment is the same as that of Embodiment 1, so it will not be described again.

[0059] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A multi-agent cooperative control method based on the integration of sensing, computing, and intelligence, implemented using a distributed cooperative system containing agent kinematic models, wherein the distributed cooperative system comprises several agents, and at least one agent is configured as a resource coordination node, characterized in that... Includes the following steps: Step 1: Obtain the state data of the agent, and construct a generalized state representation parameter based on the state data, consisting of Euclidean space feature vectors, symmetric positive definite matrices, and non-negative real space vectors; Step 2: Input the generalized state representation parameters into the conditional variational inference network to obtain prediction results including predicted channel state, predicted target trajectory, and uncertainty covariance; Step 3: Construct a Lyapunov drift plus penalty term based on the energy consumption parameters in the task queue data and energy state data. Introduce the prediction results into the local mean field stochastic state equation. Use the Lyapunov drift plus penalty term as the instantaneous cost of the local mean field stochastic state equation. Introduce the neighbor weighted empirical distribution as the local mean field state of the local mean field stochastic state equation. Then, use the deep Galerkin method to solve the local mean field stochastic state equation to obtain the agent's resource marginal technical utility and candidate resource strategy. Step 4: Limit the scope of the agent's resource requirements with candidate resource strategies, construct a technology utility function with the marginal technical utility of resources as the local slope, and maximize the sum of the agent's technical utility under the constraint of the total amount of communication, perception and computing resources. Solve the agent's resource allocation vector and resource sufficiency based on the technology utility function. Step 5: Calculate the communication update parameters and perception update parameters of the agent based on the resource allocation vector, adjust the predictive control solution parameters of the agent's kinematic model based on the resource allocation vector, and calculate the dynamic event triggering threshold of the agent based on the resource sufficiency. Step 6: Establish a safety constraint function based on the prediction results and introduce a control barrier function constraint. Introduce the safety constraint function and the control barrier function constraint into the agent's kinematic model and solve for the agent's control input parameters. Execute the agent based on the control input parameters, obtain the state data after the agent's execution, and update the generalized state representation parameters of the agent in the next control cycle.

2. The multi-agent cooperative control method based on the integration of sensing, computing, and intelligence as described in claim 1, characterized in that, Step 2 specifically includes: Step 2.1: Extract the historical window Euclidean feature sequence of the agent from the generalized state representation parameters, and input the historical window Euclidean feature sequence into the conditional variational inference network; Step 2.2: The latent variable mean and variance parameters are output by the long short-term memory encoder in the conditional variational inference network, and the conditional distribution is output by the decoder in the conditional variational inference network. Step 2.3: Based on the latent variable mean and distribution parameters, the conditional distribution function is obtained by representing the conditional distribution using Gaussian likelihood. The decoder obtains the lower triangular covariance factor by parameterizing the variance parameter through the lower triangular matrix. Step 2.4: Establish a reconstruction loss training objective function based on the conditional distribution function and the lower triangular covariance factor, and train the conditional variational inference network based on the reconstruction loss training objective function to obtain the prediction results.

3. The multi-agent cooperative control method based on the integration of sensing, computing, and intelligence as described in claim 2, characterized in that, The objective function for training the reconstruction loss is: ； in: [...] represents the reconstruction loss training objective function; Represents the Euclidean feature sequence of the historical window; This represents the channel state and target trajectory sequence for the future window. This represents the mean of the latent variable; Represents the conditional distribution function; This represents an approximate posterior distribution; Represents the conditional prior distribution; Indicates KL divergence; Represents the relative entropy regularization weight that is greater than or equal to zero.

4. A multi-agent cooperative control method based on the integration of sensing, computing, and intelligence as described in any one of claims 1-3, characterized in that, In step 3, the approximate value function of the agent is calculated using the deep Galerkin method. The Euclidean eigenvectors of neighboring agents are aggregated using a graph neural network. The non-negative similarity weights of each component in the Euclidean eigenvectors are normalized to a sum of weights equal to one. The local mean field state is formed by the weighted empirical distribution of the states of neighboring agents. The local mean field state is input into the approximate value function. Candidate resource strategies are obtained based on the gradient of the approximate value function and the extreme value condition of the Hamiltonian. The marginal technical utility of the resource is obtained based on the decrease in the optimal Hamiltonian before and after adding continuous resource components or discrete minimum resource units.

5. A multi-agent cooperative control method based on the integration of sensing, computing, and intelligence as described in any one of claims 1-3, characterized in that, Step 4 specifically includes: Step 4.1: Extract total resource constraints and agent resource demand constraints from candidate resource strategies; Step 4.2: Using the marginal technical utility of resources as the local slope, construct the technical utility function. Under the constraints of total resource constraints and agent resource demand constraints, with the goal of maximizing the sum of agent technical utility, solve the agent's resource allocation vector through the technical utility function. Step 4.3: Set a preset non-negative weight for each component in the resource allocation vector, and the sum of the preset non-negative weights of all components is equal to one. Sum the components and truncate them to the value range of zero to one to obtain the resource sufficiency.

6. A multi-agent cooperative control method based on the integration of sensing, computing, and intelligence as described in any one of claims 1-3, characterized in that, In step 5, a dynamic event triggering threshold function is established based on resource sufficiency and uncertainty covariance in the prediction results. The dynamic event triggering threshold of the agent is calculated based on the dynamic event triggering threshold function. The dynamic event triggering condition function of the agent is established based on the dynamic event triggering threshold. The dynamic event triggering threshold function is: ； in: This represents a threshold function for dynamic event triggering; Indicates resource sufficiency; Indicates uncertainty covariance; This indicates the lower limit of the threshold for triggering dynamic events; This represents the threshold adjustment amplitude coefficient; Indicates the stabilization parameter; The scale factor represents the uncertainty; tr(...) represents the trace of the matrix, which is the sum of the elements on the main diagonal of the square matrix; The dynamic event triggering condition function is: ； in: This represents a threshold function for dynamic event triggering; Indicates the measured state value of the event; Indicates the predicted state value of the event; This represents the prediction error between the measured state value of an event and the predicted state value of the event.

7. A multi-agent cooperative control method based on the integration of sensing, computing, and intelligence as described in any one of claims 1-3, characterized in that, The control barrier function constraints in step 6 include: For a safety constraint function of relative order one, apply: ； in: Represents the safety constraint function; This represents the transpose of the gradient vector of the safety constraint function with respect to the system state x; Represents the drift vector field; Represents the input vector field; Indicates the control input parameters; Represents the extended class K function for safety constraint functions; For safety constraint functions with a relative order greater than one, the m-th order recursive function of the safety constraint function is established as follows: ,0≤k≤m, ; in: This represents the k-th order recursive function of the safety constraint function; This represents the (k-1)th order recursive function of the safety constraint function; express The time derivative; (...) represents the extended class K function.

8. A multi-agent cooperative control method based on the integration of sensing, computing, and intelligence as described in any one of claims 1-3, characterized in that, The candidate resource strategy includes communication resources, sensing resources, and computing resources; the communication resources include transmit power, orthogonal frequency division multiplexing subcarriers, time slots, and beamforming parameters; the sensing resources include radar beam pointing, beamwidth, pulse repetition frequency, and coherent accumulation time; and the computing resources include processor clock frequency, edge offload ratio, and number of solution iterations.

9. A multi-agent cooperative control method based on the integration of sensing, computing, and intelligence as described in any one of claims 1-3, characterized in that, The agent's state data in step 1 includes dynamic state data, channel data, radar data, task queue data, and energy state data. The dynamic state data is represented as a Euclidean space vector, the covariance in the channel data and radar data is represented as a symmetric positive definite matrix, and the task queue data and energy state data are represented as non-negative real space vectors. At a selected reference point, logarithmic mapping and vectorization are performed on the symmetric positive definite matrix, and concatenation and normalization processing is performed on the Euclidean space vector and the non-negative real space vector to form Euclidean feature vectors.

10. A multi-agent cooperative control system based on the integration of sensing, computing, and intelligence, implemented based on the multi-agent cooperative control method according to any one of claims 1-9, characterized in that, The system includes an integrated sensing and communication RF front-end, sensor interfaces, agent communication interfaces, several agent computing units, a memory, and an actuator interface, wherein at least one agent computing unit is configured as a resource coordination node. The memory stores a state construction module, a joint prediction module, a policy generation module, a resource allocation module, a security control module, and a feedback update module. The state construction module constructs a generalized state representation based on data collected from each interface. The joint prediction module outputs the future channel state, target trajectory, and uncertainty covariance. The policy generation module generates candidate resource policies and resource marginal utility. The resource allocation module, executed by the resource coordination node, outputs communication, sensing, and computing resource allocation vectors. The security control module generates control inputs that satisfy security constraints based on the resource allocation vectors and uncertainty covariance. The feedback update module writes the measured state, task service volume, and energy consumption after execution into the generalized state representation of the next control cycle. The multiple agent computing units execute the modules and output control inputs via the actuator interface.

Images