Multi-uav sensing algorithmic resource collaborative optimization method and device

Abstract

The application provides a multi-unmanned aerial vehicle (UAV) sensing and computing resource collaborative optimization method and device, aiming at the problems of strong coupling of heterogeneous resources and poor data timeliness when the multi-UAV provides sensing and computing integrated services, the application obtains local observation information by each UAV under discrete time slots, and outputs enhanced observation features through a feature extraction network containing an attention mechanism; a multi-agent strategy network independently generates a multi-dimensional joint decision containing dynamic user association, trajectory update, bandwidth and power allocation, and computing power offloading, controls the UAV and the edge node to perform the sensing and computing task, and according to the timeliness of the task, the closed-loop update information age of the decision result is achieved to drive the decision update. The application is used for distributed collaborative scheduling of multi-UAVs, and can effectively improve the data freshness and service reliability.

Description

Technical Field

[0001] This invention relates to the fields of low-altitude intelligent network, unmanned aerial vehicle (UAV) communication and edge computing technology, and in particular to a method and apparatus for collaborative optimization of multi-UAV sensory computing resources for low-altitude intelligent network. Background Technology

[0002] With the rapid development of the low-altitude economy, drones are widely used in traffic monitoring, urban security, and emergency rescue. Building a low-altitude intelligent network to achieve collaborative operation among multiple drones, dynamic ground users, and computing nodes has become an important technological path to improve the efficiency of low-altitude applications. However, the traditional single-sensor monitoring model faces multiple bottlenecks in complex urban environments: independent deployment of radar systems and communication networks leads to high equipment costs, and spectrum resource competition causes interference; with the surge in the number of drones, the requirements for real-time response capabilities in dynamic target tracking have increased dramatically, while the traditional separate architecture struggles to meet millisecond-level response requirements.

[0003] Integrated Sensing and Communication (ISAC) technology enables communication base stations to simultaneously possess environmental sensing capabilities through shared hardware facilities, unified spectrum resources, and collaborative signal processing, achieving a "dual-purpose network" capability and simultaneously improving hardware efficiency and spectrum utilization. Building upon ISAC, edge computing further forms integrated sensing and intelligent computing services. Currently, some existing solutions have begun exploring the multi-UAV resource scheduling problem. Typical technical approaches include: optimizing bandwidth or power separately based on traditional optimization theories (such as convex optimization and matching game theory); or using single-agent reinforcement learning algorithms to plan UAV trajectories. A common feature of these solutions is that they model and solve each dimension (such as trajectory, bandwidth, power, and computational offloading) of heterogeneous integrated sensing and intelligent computing resources separately, or optimize only some of these dimensions.

[0004] However, the aforementioned existing technologies still have the following shortcomings in low-altitude intelligent network scenarios: First, there is a strong coupling relationship between heterogeneous resources in the synergistic computing system (flight trajectory, sensing and communication bandwidth / power, airborne computing power, and edge computing power). It is difficult to obtain the global optimal solution by independently optimizing each dimension of resources, and the traditional static allocation method is also difficult to adapt to the rapid dynamic changes in the number of drones, dynamic user distribution, and link status. Second, existing service quality assessments rely heavily on end-to-end latency indicators, which make it difficult to accurately depict the freshness of perceived data and lack quantitative assessment and feedback mechanisms oriented towards information timeliness, resulting in a lack of timely basis for resource scheduling decisions. Third, in terms of distributed collaborative decision-making, existing solutions usually require a central node to collect global information and schedule it uniformly, which results in high computational overhead and slow response, making it difficult to meet the real-time service requirements in dynamic low-altitude scenarios.

[0005] Therefore, there is an urgent need for a method and device for collaborative optimization of multi-UAV sensory computing resources for low-altitude intelligent networks, in order to solve the problems existing in the current technology. Summary of the Invention

[0006] This invention provides a method and apparatus for collaborative optimization of multi-UAV intelligent computing resources. It addresses the problems of existing technologies, such as the difficulty in achieving low-overhead collaborative scheduling in dynamic environments with strong coupling of heterogeneous resources, and the inability to accurately characterize the freshness of multi-source data by relying on a single end-to-end latency indicator, resulting in a lack of timeliness and adaptability to dynamic environments in resource scheduling decisions.

[0007] The core technology of this invention is to process the local observation information of UAVs through a feature extraction network with an attention mechanism, and to introduce information age and timeliness achievement rate evaluation indicators in the closed-loop feedback, so that each UAV can generate multi-dimensional joint decisions including user association, flight trajectory, wireless resources and computing resources through distributed online reasoning of a multi-agent policy network without relying on a central node.

[0008] In a first aspect, the present invention provides a collaborative optimization method for multi-UAV sensing and computing resources in low-altitude intelligent networks, applied to an edge network architecture consisting of multiple UAVs, multiple ground dynamic users, and at least one edge artificial intelligence processor, executed with discrete time slots as the scheduling cycle, including the following steps: A. System initialization: Set system constraints and initialize age variables; B. Obtain system status information and construct local observation information for each UAV; C. By using a feature extraction network that includes an attention mechanism, local observation information is processed to obtain enhanced observation feature representations; D. Each UAV independently generates multi-dimensional decision-making results for synesthetic computing resources based on enhanced observation feature representations through a multi-agent policy network; E. Execute decision results and provide integrated sensory computing task services; F. Update the information age based on the timeliness of the task to form a closed-loop feedback to drive joint decision-making updates in subsequent time slots.

[0009] Furthermore, the system initialization in step A specifically includes: Establish drone ensembles, ground dynamic user ensembles, and time slot ensembles; Set the location information of ground dynamic users and UAVs within any time slot, where the ground dynamic users are located on the ground plane and the UAVs perform tasks at a fixed flight altitude; Set minimum safe distance constraints between drones, as well as the maximum total bandwidth, maximum total transmit power, and maximum number of ground dynamic users that can be served per time slot for each drone; A dynamic user association decision variable is introduced to represent the service association relationship between the UAV and dynamic ground users; The bandwidth allocation ratio and power allocation ratio are introduced to represent the resource allocation between sensing and communication for UAVs, respectively; We introduce the decision variables of unloading and local processing to represent the strategy of preprocessing perceived data locally or uploading it to an edge AI processor for preprocessing; A computing resource allocation variable is introduced to represent the allocation of local computing power of the drone and computing power of edge AI processors, respectively; Set an initial age value for each dynamic user on the ground.

[0010] Furthermore, the local observation information for each UAV in step B includes: The location coordinates of the UAV within the current time slot; The spatial location relationship information between the drone and each candidate ground dynamic user; The amount of radar data required to perceive each candidate dynamic ground user; Information age values ​​of dynamic users in various locations within the current time slot; The local data processing density of the drone in the current time slot.

[0011] Furthermore, in step C, a feature extraction network containing an attention mechanism is used to perform attention-weighted aggregation of the state features of candidate ground dynamic users associated with the current time slot and the resource state features of UAVs and edge AI processors, mapping the variable-length input caused by the change in the number of dynamic users into a fixed-dimensional enhanced observation feature representation.

[0012] Furthermore, the multi-dimensional decision-making results of the synesthetic computing resources in step D include at least: Ground-based dynamic user-related decision-making; UAV flight trajectory update volume; The bandwidth allocation ratio between sensing and communication; Power allocation ratio between sensing and communication; Local processing and edge offloading decisions for perceived data; Decisions on the allocation of local computing power for drones.

[0013] Furthermore, the execution of the synesthetic computing integrated task service in step E includes: Based on the bandwidth allocation ratio and power allocation ratio, the total bandwidth and total transmit power of each UAV are divided into the sensing part and the communication part, and then allocated to the associated ground dynamic users. Radar sensing is performed on the associated ground dynamic users, and the radar sensing rate is used as the sensing reliability indicator, which is required to be no less than a preset threshold. The number of ground dynamic users associated with each drone in any time slot shall not exceed the preset limit, and each ground dynamic user shall be associated with at most one drone in any time slot. Based on the unloading decision, the perception data is preprocessed locally on the drone side or transmitted to the edge AI processor for preprocessing. The preprocessed perception data is then fused and analyzed by the edge AI processor, and the results are sent to the ground dynamic user. The total service time from data collection to processing completion does not exceed the preset time window threshold.

[0014] Furthermore, in step F: The timeliness achievement determination is based on whether the radar sensing rate meets a preset threshold and whether the end-to-end service time meets a preset time window threshold. The information age is updated as follows: if the timeliness of the ground dynamic user in the current time slot is met and the user is associated with a drone, the information age is updated to the service time of the current time slot; otherwise, the information age is incremented based on the previous time slot. The long-term optimization goal of the system is to maximize the long-term mean of the difference between the timeliness achievement rate and the weighted average information age, while satisfying flight safety constraints, correlation constraints, perception reliability constraints, service time constraints, and energy consumption constraints.

[0015] Furthermore, the execution of the integrated sensing and computing task service also includes energy consumption constraints: the total energy consumption on the UAV side includes local data preprocessing energy consumption and data transmission energy consumption, and the total energy consumption on the edge AI processor side includes edge data processing energy consumption; the total system energy consumption of each UAV and edge AI processor in any time slot does not exceed the preset energy consumption limit.

[0016] Secondly, the present invention provides a multi-UAV sensing and computing resource collaborative optimization device for low-altitude intelligent networks, comprising: The initialization module is used to set system constraints and initialize age variables. The observation construction module is used to acquire system status information and construct local observation information for each UAV. The feature extraction module is used to process local observation information through a feature extraction network containing an attention mechanism to obtain enhanced observation feature representations; The decision generation module is used to independently generate multi-dimensional decision results of synesthetic computing resources based on enhanced observation feature representation through a multi-agent policy network; The task execution module is used to execute decision results and provide integrated sensory computing task services; The feedback update module is used to update the information age based on the timeliness of the task and the judgment result, forming a closed-loop feedback to drive the joint decision update in subsequent time slots.

[0017] Thirdly, the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to execute the above-described method for collaborative optimization of multi-UAV sensory computing resources for low-altitude intelligent networks.

[0018] Fourthly, the present invention provides a readable storage medium storing a computer program, the computer program including program code for controlling a process to execute the process, the process including the above-described method for collaborative optimization of multi-UAV sensory computing resources for low-altitude intelligent networks.

[0019] The main contributions and innovations of this invention are as follows: 1. Under the discrete time slot framework, this invention jointly models and coordinates the dynamic user association, flight trajectory update, bandwidth allocation for sensing and communication, power allocation for sensing and communication, task offloading decision and local computing power allocation. This overcomes the problem of low resource utilization efficiency caused by traditional methods that optimize the above dimensions independently or only partially jointly, and realizes the full-process coordination of sensing and computing resources.

[0020] 2. This invention uses a feature extraction network with an attention mechanism to perform attention-weighted aggregation of variable-length, multi-dynamic user state features, mapping the variable-length input caused by changes in the number of dynamic users to a fixed-dimensional enhanced observation feature representation. This solves the problem that traditional fully connected networks cannot handle variable-length inputs, enabling the policy network to adapt to scenarios with changing numbers of dynamic users.

[0021] 3. This invention constructs a joint optimization index for timeliness achievement rate and information age based on dual criteria of radar sensing rate threshold and end-to-end service time threshold. This forms a closed-loop feedback mechanism from task execution to performance evaluation to policy update, overcoming the shortcomings of traditional single-index evaluation methods that rely solely on end-to-end latency, which cannot reflect the freshness of sensing data, and improving the system's adaptive scheduling capability in dynamic environments.

[0022] 4. This invention adopts a multi-agent policy network architecture based on an attention mechanism. During the training phase, global information is used for centralized training. During the execution phase, each UAV relies only on local observations to independently reason and output decisions, realizing distributed collaboration without a central node. This avoids the high computational overhead caused by global re-optimization in every time slot and meets the real-time requirements of millisecond-level online decision-making in low-altitude intelligent network scenarios.

[0023] Details of one or more embodiments of the present invention are set forth in the following drawings and description, so that other features, objects and advantages of the invention will be more readily understood. Attached Figure Description

[0024] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the system architecture of the multi-UAV sensory computing resource collaborative optimization method for low-altitude intelligent networks according to an embodiment of the present invention; Figure 2 This is a framework diagram of the Transformer-based Multi-agent Proximal Policy Optimization Algorithm (TMAPPO) according to an embodiment of the present invention. Figure 3 This is a comparison curve of the convergence of the TMAPPO algorithm of the present invention and the existing technology algorithm in the simulation embodiment; Figure 4 This is a comparison chart of the average reward, timeliness achievement rate, and average AoI of the TMAPPO algorithm of the present invention and existing technology algorithms under different numbers of dynamic ground users in the simulation embodiment. Figure 5 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0025] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0026] It should be noted that the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification in other embodiments. In some other embodiments, the methods may include more or fewer steps than described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments.

[0027] This invention provides a collaborative optimization method for multi-UAV sensing and computing resources in low-altitude intelligent networks. It is applied to an edge network architecture consisting of multiple UAVs, multiple ground-based dynamic users, and at least one edge AI processor, and is executed on a discrete time slot scheduling cycle. Here, "sensing and computing intelligence" refers to an integrated collaborative service encompassing communication, sensing, computing, and intelligence. Information Age (AoI) is an indicator of data freshness, reflecting the time elapsed since the most recent successful acquisition of sensing data to the current moment; a smaller value indicates fresher data.

[0028] like Figure 1 As shown, the collaborative network architecture includes multiple drones, multiple ground dynamic users, and at least one edge AI processor. The drones serve as optional centralized training nodes during the training phase. Multiple drones move along their flight paths within the service area, performing integrated sensory data acquisition from ground dynamic users and preprocessing the perceived data locally and / or offloading it to the edge AI processor. The edge AI processor receives the perceived data from each drone, performs fusion analysis, and sends the analysis results back to the ground dynamic users. The system operates on a discrete time slot-based scheduling cycle. Within each time slot, each drone independently performs state observation, decision generation, and task execution, completing a full perception-transmission-computation-feedback closed loop.

[0029] Example 1 This embodiment provides a collaborative optimization method for multi-UAV sensing and computing resources in low-altitude intelligent networks, which specifically includes the following steps: Step A: System Initialization Establish drone ensemble Ground dynamic user set K and time slot set Assume the system has a total of A drone, numbered ; Total A ground-based dynamic user, numbered as The scheduling process is divided into: There are discrete time slots, numbered as follows: .

[0030] Let any time slot t∈ Location information of dynamic users k∈K on the inland ground and UAVs Location information. Ground-based dynamic users. The three-dimensional coordinates are:

[0031] in and For ground dynamic users in time slots The horizontal and vertical coordinates are set, with the vertical component set to 0, indicating that the altitude of the dynamic user is ignored. (Drone) The three-dimensional coordinates are:

[0032] in and The drones in time slots The horizontal and vertical coordinates, This refers to the fixed flight altitude of the drone.

[0033] Set minimum safe distance constraints between drones. Any two drones and The Euclidean distance between them must satisfy:

[0034] in This is the minimum safe distance used to prevent collisions between drones.

[0035] Set the total bandwidth limit for each drone With the upper limit of total transmit power And set the maximum number of dynamic ground users that each drone can serve in each time slot to be [number missing]. .

[0036] Introduce the following decision variables: 1. Dynamic user-related decision variables ,when Time indicates time slot Internal drones With ground dynamic users Establish service associations.

[0037] 2. Bandwidth allocation ratio With power distribution ratio ,in Indicates drone In the time slot The proportion of bandwidth used for sensing in the total bandwidth Indicates drone In the time slot The proportion of power used for sensing to the total power.

[0038] 3. Unloading and Local Processing of Decision Variables ,in Indicates drone For ground dynamic users The perceived data is preprocessed locally on the drone side. This means that the perceived data will be uploaded to an edge AI processor for preprocessing.

[0039] 4. Variables in computing resource allocation, including the local computing capabilities of drones. Its value range is:

[0040] and the computing power of edge AI backend processors (In terms of cycles per second), its value range is:

[0041] For each ground dynamic user Set the initial value for Information Age AoI. .

[0042] Step B: Acquisition of System State Information and Construction of Local Observations In any time slot Each drone Obtain the system state information for the current time slot and construct local observation information. Local observation information. Includes the following:

[0043] The meanings of each component are as follows: For drones In the time slot Position coordinates; For time slots Internal drones With ground dynamic users Spatial relationship information between them; For drones In the time slot Sensing ground dynamic users The required amount of radar data (in bits) is a known constant; For ground dynamic users In the time slot Information about age values; For drones In the time slot Local data processing density (in terms of the number of cycles required per bit).

[0044] Step C: Attention feature extraction yields enhanced observation feature representations. Treat each drone as an independent decision-making entity, based on local observations. By employing a feature extraction network incorporating an attention mechanism, weighted modeling of the spatiotemporal features of multiple dynamic users and resource states is achieved, resulting in enhanced observation feature representations. .

[0045] like Figure 2 As shown, the attention feature extraction network encodes multi-dynamic user features and resource status in local observation information. The specific process is as follows.

[0046] First, the components of the local observation information are mapped into high-dimensional feature vectors through an embedding layer, and positional encoding is added to preserve sequence information. For features related to dynamic users (such as location relationships, data volume, information age, etc.), the state information of each candidate dynamic user is encoded into a feature vector; for features related to resource status (such as drone location, local processing density, etc.), they are encoded separately into resource feature vectors.

[0047] Secondly, the encoded features are aggregated using a multi-head self-attention mechanism with attention weighting. The calculation formula for the self-attention mechanism is as follows:

[0048] in , , These are the query matrix, key matrix, and value matrix, respectively. Let be the dimension of the key vector. is an optional attention mask matrix. The multi-head self-attention mechanism captures the spatiotemporal dependencies between the drone and each dynamic user from different subspaces by computing multiple sets of attention weights in parallel.

[0049] A key advantage of this attention-weighted aggregation process is that traditional fixed-dimensional neural networks struggle to directly handle variable-length inputs when the number of dynamic users in the system changes. The attention mechanism can map the variable-length input features caused by changes in the number of dynamic users into a fixed-dimensional augmented observation feature vector. This effectively adapts to the dynamic and continuously changing user density characteristics in low-altitude scenarios.

[0050] Subsequently, the representation of observation features will be enhanced. Input the multi-agent policy network to generate joint actions for each time slot.

[0051] Step D: Multidimensional Joint Decision Output Based on enhanced observation feature representation The multi-agent policy network governs the drone. In the time slot Output the following action vector:

[0052] The meanings of each dimension of decision-making are as follows: For ground-based dynamic user-related decision-making, determine the drone The set of dynamic ground users that need to be served in the current time slot; Determine the division between sensing bandwidth and communication bandwidth based on the bandwidth allocation ratio; To determine the power allocation ratio, the division between sensing power and communication power is established; To update the drone's flight trajectory, determine the direction and distance of flight movement in the current time slot, and the updated position is... ; To determine the processing method for each sensed data, whether it is preprocessed locally or unloaded at the edge, the decision between unloading and local processing is made. Allocate local computing power to determine the computing capacity that the drone can use for local data processing.

[0053] Step E: Execution of the integrated sensory computing task The execution process of step E includes a sensing phase, a communication transmission phase, and a computational processing phase, as detailed below.

[0054] (a) Bandwidth Allocation and Sub-bandwidth Distribution drones Total bandwidth According to bandwidth allocation ratio Divided into sensing bandwidth and communication bandwidth:

[0055] The sensing bandwidth is further allocated to each associated dynamic ground user:

[0056] Similarly, communication bandwidth is allocated to each associated user:

[0057] (II) Power partitioning and sub-power allocation drones Total power According to power allocation ratio Divided into sensing power and communication power:

[0058] The sensing power is further distributed to the associated ground dynamic users:

[0059] Communication power is allocated similarly:

[0060] (III) Radar sensing process and sensing reliability constraints drones Using radar signals to monitor associated ground dynamic users Perform integrated sensor data acquisition. (UAV) For ground dynamic users The power model for the two-way radar sensing channel is as follows:

[0061] in, For radar sensing channel reference gain, For drones With ground dynamic users The Euclidean distance between them.

[0062] Considering the interference effects between multiple drones, time slots Chinese drones For ground dynamic users The radar signal-to-interference-plus-noise ratio (SINR). for:

[0063] in, The noise power is represented by the summation term, which indicates the co-frequency interference generated by other UAVs when sensing the same ground dynamic user.

[0064] Radar sensing rate is used as an indicator of sensing reliability. (UAV) For ground dynamic users Perception rate for:

[0065] To ensure perception quality, the perception rate must not be lower than a preset minimum threshold.

[0066] in, Set a minimum threshold for radar sensing rate.

[0067] (iv) Association Constraints The number of ground-based dynamic users associated with each drone in any given time slot shall not exceed a preset limit. :

[0068] Each ground-based dynamic user can be associated with at most one drone in any given time slot:

[0069] (v) Communication transmission process After completing the perception, the drone Based on uninstallation decision The decision is made regarding whether the perceived data is preprocessed locally or transmitted to an edge AI processor for preprocessing. Considering the transmission link between the drone and the edge AI processor, a logarithmic distance path loss model is used to model the channel gain between them:

[0070] in, This is the reference channel gain. The path loss index is related to flight altitude. And related to environmental parameters, the model is as follows:

[0071] in , and These are parameters related to the propagation environment. and All are zero-mean Gaussian distributed random variables. Characterizing shadow fading properties, Characterizes the range of changes in path loss.

[0072] According to Shannon's theorem, in time slots Inside, drones Transmitting ground dynamic users The communication rate from the sensed data to the edge AI processor is:

[0073] (vi) Calculation and processing procedures when At that time, drones For ground dynamic users The perceived data is preprocessed locally on the drone side. The local processing time is:

[0074] The energy consumption for local data preprocessing is:

[0075] in, The effective capacitance coefficient of the drone processor is determined by the drone's hardware parameters.

[0076] when At that time, the amount of data after local preprocessing was (usually smaller than the original data volume) This requires transmitting the pre-processed data to an edge AI processor for fusion analysis. The amount of data after preprocessing has a different meaning from d(i,k) in step (iii) above, which represents the spatial distance between the UAV and the ground dynamic user. The former is usually smaller than the amount of original data. .when At that time, raw sensory data All data is transmitted to an edge AI processor for preprocessing and fusion analysis. The time required for offloading and transmission is:

[0077] The energy consumption for offloading transmission is:

[0078] UAV-side processing of ground dynamic users The total energy consumption of the data is the sum of the local computing energy consumption and the data transmission energy consumption:

[0079] The processing time and energy consumption of the edge AI processor are modeled as follows:

[0080]

[0081] in, This represents the data processing density of the edge AI processor within time slot t. The effective capacitance coefficient of the edge AI processor (determined by the processor hardware parameters, used to characterize the actual energy consumption per unit of computation of the processor). This indicates the computing power of the edge AI processor.

[0082] (vii) End-to-end service time constraints The total service time for sensing data from acquisition to completion of processing consists of local processing time, data transmission time, and edge processing time.

[0083] The above time quantities are in time slots Calculated at the start time. Maximum end-to-end service time window threshold. Defined as the maximum allowed time from data acquisition to processing completion, this time can cover multiple consecutive time slots. The total service time shall not exceed this threshold.

[0084] (viii) Energy consumption constraints In the preferred embodiment that incorporates an energy consumption model, time slots The total system energy consumption is the sum of the energy consumption of all drones and edge AI processors:

[0085] The system's total energy consumption must meet the preset energy consumption upper limit constraint:

[0086] Step F: Determining the timeliness of task completion, updating information age, and providing closed-loop feedback. To form a closed-loop scheduling mechanism, the following timeliness assessment and information age update process is constructed.

[0087] (a) Timeliness of achieving the target quantity For each ground dynamic user Set time-sensitive target quantity Its value is determined jointly by two conditions: perceived reliability and service timeliness.

[0088] That is, local ground dynamic users When the perceived rate is not lower than the preset threshold and the end-to-end service time does not exceed the maximum service time window threshold, the timeliness achievement indicator value is 1, indicating that the user's task in the current time slot is completed with sufficient perceived quality within the effective time window.

[0089] (ii) Timeliness achievement rate Define time slot Timeliness achievement rate The percentage of ground-based dynamic users who meet the timeliness standard within the current time slot:

[0090] (III) Information Age Update In the time slot No. Information age of ground dynamic users The update rules are as follows. If the timeliness meets the indicated quantity... And the ground dynamic user by drone Association (i.e.) If the information age is updated to the service time of the current time slot, then the information age will be updated to the service time of the current time slot. This reflects that newly acquired sensory data improves data freshness; otherwise, the information age increases by 1 from the previous time slot value, indicating that the data is aging further over time.

[0091] Define time slot The average information age of all ground-based dynamic users is:

[0092] (iv) Long-term optimization goals of the system Based on timeliness achievement rate and average information age The system's long-term optimization objective is to maximize the long-term mean of the difference between the timeliness achievement rate and the weighted average information age across all time slots, while satisfying constraints related to flight safety, correlation, perception reliability, mission offloading decision, computational resources, maximum service time, and energy consumption.

[0093] in, This is a weighting coefficient for information age, used to balance the relative importance between timeliness achievement rate and data freshness. Optimization variables include dynamic user association strategies. Bandwidth allocation Power distribution Flight trajectory Uninstallation strategy and computing resource allocation .

[0094] The constraints include: minimum safe distance constraint between UAVs (the Euclidean distance between any two UAVs i and j satisfies...). Dynamic user association constraints The number of ground dynamic users associated with each drone in any time slot meets the following requirements. Each ground-based dynamic user can be associated with at most one drone in any given time slot. ), bandwidth and power allocation ratio constraints ( Perception reliability constraints (minimum threshold constraint for radar perception rate) ), binary constraints for task unloading decision ( Upper and lower bound constraints on computing resources for drones and edge AI processors , Maximum end-to-end service time constraint ) and the system's total energy consumption upper limit constraint ( ).

[0095] The aforementioned optimization problem is a multi-stage stochastic optimization problem containing discrete and continuous variables, characterized by high dimensionality, non-convexity, and strong coupling. Traditional optimization methods struggle to solve it within acceptable time complexity. Therefore, this embodiment employs the Transformer-based Multi-agent Proximal Policy Optimization (TMAPPO) algorithm for solving the problem.

[0096] In this embodiment, as Figure 2 As shown, the implementation of the TMAPPO algorithm includes the following key steps: (i) Constructing a distributed partially observable Markov decision process model The multi-UAV cooperative scheduling process is modeled as a decentralized partially observable Markov Decision Process (Dec-POMDP), defining the global state. Local observation Joint actions and global rewards The global reward function is designed as follows:

[0097] in, The positive reward weight for meeting the timeliness target, The penalty weight is used for failure to meet timeliness standards. This reward function drives the policy network to learn efficient collaborative scheduling strategies by positively incentivizing timely performance, negatively penalizing failure to meet timeliness standards, and guiding the reduction of information age.

[0098] (II) Transformer State Feature Extraction The encoder, based on the Transformer architecture, embeds and encodes the input state, and uses a multi-head self-attention mechanism to capture the spatiotemporal dependencies between drones, dynamic users, and edge nodes.

[0099] By stacking multiple encoders, a high-dimensional state feature is obtained after encoding, which integrates nonlinear coupling information between multiple UAVs.

[0100] (III) Decoding and Decision Generation The decoder receives the encoded high-dimensional state features and generates joint action strategies for each UAV through an autoregressive approach, including dynamic user association, bandwidth and power allocation, flight trajectory, and computational offload decisions.

[0101] (iv) Centralized training and distributed execution During the training phase, global information (including complete state information of multiple drones, multiple dynamic users, and edge resources) is utilized. (Collected uniformly by centralized training nodes) to optimize the Critic network and Actor network. The Critic network evaluates the value of global state-action pairs, and the Actor network updates the policy parameters of each UAV.

[0102] During the execution phase, each drone relies solely on its own local observation information. Through attention feature extraction and inference by the trained Transformer-Actor network, the optimal action is output independently without the need for a central coordination node, realizing distributed synesthetic computing and intelligent collaborative optimization in low-altitude environments.

[0103] Example 2 This embodiment provides a multi-UAV sensing and computing resource collaborative optimization device for low-altitude intelligent networks, including the following modules: The initialization module is used to set system constraints and initialize age variables. The observation construction module is used to acquire system status information and construct local observation information for each UAV. The feature extraction module is used to process the local observation information through a feature extraction network containing an attention mechanism to obtain an enhanced observation feature representation; The decision generation module is used to independently generate multi-dimensional decision results of synesthetic computing resources based on the enhanced observation feature representation through a multi-agent policy network; The task execution module is used to execute the decision results and provide integrated sensory computing task services. The feedback update module is used to update the information age based on the timeliness of the task and the judgment result, forming a closed-loop feedback to drive the joint decision update in subsequent time slots.

[0104] The specific functional implementation of each module is the same as the corresponding steps described in Embodiment 1, and will not be repeated here.

[0105] This embodiment also includes simulation verification: To verify the effectiveness of the method of this invention, a low-altitude intelligent network simulation environment was built using the PyTorch framework on the Python 3.8 platform. The simulation parameters were set as follows: The simulation area is a 200m × 200m square service area, where three drones are deployed to collaboratively provide integrated sensing, computing, and communication services to dynamic ground users. The edge AI processor is fixed at the center coordinates (100, 100, 0) of the area, and the initial flight altitude of the drones is uniformly set to 50m. The system uses discrete time slots as the scheduling cycle, with each time slot lasting 100ms.

[0106] The key parameters for the simulation environment are set as follows: Radar sensing channel reference gain. for dB, noise power for Radar sensing rate threshold The reference channel gain is 0.3. The total bandwidth of the drone is 40dB. The total power of the drone is 1MHz. The maximum processing service time threshold is 1W. The timeout is 500ms (this threshold covers 5 consecutive time slots, allowing the task to complete end-to-end processing within multiple time slots). The drone's computing power ranges from 5-10GHz, while the edge AI processor's computing power ranges from 50-100GHz, with maximum energy consumption constraints. It is 100J.

[0107] The hyperparameters for the reinforcement learning network are set as follows: Actor learning rate. Critic learning rate Discount factor The batch size is 128, the number of training rounds is 300, and the number of steps in each round is 500.

[0108] To verify the effectiveness of the proposed method, the TMAPPO algorithm is compared with the following baseline algorithms: Multi-Agent Proximal Policy Optimization (MAPPO), REINFORCE (a policy gradient algorithm in reinforcement learning), and Multi-Agent Deep Deterministic Policy Gradient (MADDPG). Table 1 shows the simulation results comparison. Table 1. Performance results of TMAPPO algorithm compared with benchmarks

[0109] As shown in Table 1, the average AoI of the method of the present invention is 3.926 s, the timeliness achievement rate is 78.2%, and the single-step running time is 8.26 ms. Compared with the baseline algorithm, it can improve the timeliness achievement rate while reducing AoI and meet the millisecond-level online decision-making requirements.

[0110] like Figure 3 The convergence of different algorithms during training was compared in terms of average reward, timeliness achievement rate, and average AoI. Figure 3 As can be seen from part a, the method of this invention tends to stabilize within a short number of training sessions, and the reward level in the stable phase is significantly higher than that of existing technologies; Figure 3 (b) It is evident that the method of the present invention can maintain a higher timeliness achievement rate; Figure 3 As can be seen from part c, the method of the present invention maintains a lower average AoI after convergence, indicating that the method has better training stability in complex low-altitude dynamic scenarios.

[0111] like Figure 4 The paper presents a comparison of the algorithms in terms of average reward, timeliness achievement rate, and average AoI under different numbers of ground users. As the number of dynamic users increases, Figure 4 In part a, the average reward decreased. Figure 4 In part b, the timeliness achievement rate generally decreased. Figure 4 In part c, the average AoI increases; however, the method of the present invention still maintains higher effectiveness and lower AoI under different dynamic user scales, demonstrating its adaptability and robustness to changes in dynamic user density.

[0112] Based on the above results, under typical simulation settings, the method of the present invention can achieve a low average AoI (approximately 3.926s), a high timeliness achievement rate (approximately 78.2%), and a single-step decision time in the millisecond range (approximately 8.26 ms), demonstrating feasibility and real-time performance for online deployment in dynamic low-altitude intelligent network scenarios.

[0113] It should be noted that the Transformer architecture and TMAPPO algorithm used in the above embodiments are only one of the preferred implementations of the technical solution of the present invention. Those skilled in the art can use other neural network structures containing attention mechanisms (such as figure attention networks, cross-attention networks, etc.) or multi-agent reinforcement learning algorithms (such as multi-agent Actor-Critic algorithms, QMIX algorithms, etc.) to achieve the same or similar technical effects without departing from the concept of the present invention, and all such implementations should fall within the protection scope of the present invention.

[0114] Example 3 This embodiment also provides an electronic device, see reference. Figure 5It includes a memory 402 and a processor 401, wherein the memory 402 stores a computer program and the processor 401 is configured to run the computer program to perform the steps in any of the above method embodiments.

[0115] Specifically, the processor 401 may include a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of the present invention.

[0116] The memory 402 may include a mass storage device for data or instructions. For example, and not limitingly, the memory 402 may include a hard disk drive (HDD), a floppy disk drive, a solid-state drive (SSD), flash memory, an optical disk drive, a magneto-optical disk drive, magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, the memory 402 may include removable or non-removable (or fixed) media. Where appropriate, the memory 402 may be internal or external to a data processing device. In a particular embodiment, the memory 402 is non-volatile memory. In a particular embodiment, the memory 402 includes read-only memory (ROM) and random access memory (RAM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable read-only memory (PROM), an erasable read-only memory (EPROM), an electrically erasable read-only memory (EEPROM), an electrically alterable read-only memory (EAROM), or flash memory, or a combination of two or more of these. Where appropriate, the RAM can be Static Random-Access Memory (SRAM) or Dynamic Random-Access Memory (DRAM). DRAM can be Fast Page Mode Dynamic Random Access Memory (FPMDRAM), Extended Data Out Dynamic Random Access Memory (EDODRAM), Synchronous Dynamic Random-Access Memory (SDRAM), etc.

[0117] The memory 402 can be used to store or cache various data files that need to be processed and / or communicated, as well as possible computer program instructions executed by the processor 401.

[0118] The processor 401 reads and executes computer program instructions stored in the memory 402 to implement any of the multi-UAV sensory computing resource collaborative optimization methods for low-altitude intelligent networks in the above embodiments.

[0119] Optionally, the electronic device may further include a transmission device 403 and an input / output device 404, wherein the transmission device 403 is connected to the processor 401 and the input / output device 404 is connected to the processor 401.

[0120] The transmission device 403 can be used to receive or send data via a network. Specific examples of the network described above may include wired or wireless networks provided by the communication provider of the electronic device. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 403 may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0121] Input / output device 404 is used for inputting or outputting information. It can be a speaker, microphone, monitor, or keyboard.

[0122] Example 4 This embodiment also provides a readable storage medium storing a computer program, which includes program code for controlling and executing a process, the process including the multi-UAV sensory computing resource collaborative optimization method for low-altitude intelligent networks according to Embodiment 1.

[0123] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0124] Generally, various embodiments can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects of the invention can be implemented in hardware, while others can be implemented by firmware or software executed by a controller, microprocessor, or other computing device, but the invention is not limited thereto. Although various aspects of the invention may be shown and described as block diagrams, flowcharts, or using some other graphical representation, it should be understood that, by way of non-limiting example, these blocks, apparatuses, systems, techniques, or methods described herein can be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0125] Embodiments of the present invention can be implemented by computer software, which may be executable by a data processor of a mobile device, such as a processor entity, or by hardware, or by a combination of software and hardware. Computer software or programs (also referred to as program products), including software routines, applets, and / or macros, can be stored in any device-readable data storage medium, and they include program instructions for performing specific tasks. A computer program product may include one or more computer-executable components configured to perform embodiments when the program is run. One or more computer-executable components may be at least one piece of software code or a portion thereof. Additionally, it should be noted that any block in the logical flow of the figures may represent a program step, or interconnected logical circuitry, blocks and functions, or a combination of program steps and logical circuitry, blocks and functions. The software may be stored on physical media such as memory chips or blocks of storage implemented within a processor, magnetic media such as hard disks or floppy disks, and optical media such as, for example, DVDs and their data variants, CDs, etc. The physical medium is a non-transient medium.

[0126] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0127] The above embodiments are merely illustrative of several implementations of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the appended claims.

Claims

1. A collaborative optimization method for multi-UAV sensory computing resources in low-altitude intelligent networks, applied to an edge network architecture consisting of multiple UAVs, multiple ground dynamic users, and at least one edge artificial intelligence processor, executed with discrete time slots as the scheduling period, characterized in that... Includes the following steps: A. System initialization: Set system constraints and initialize age variables; B. Obtain system status information and construct local observation information for each UAV; C. The local observation information is processed through a feature extraction network with an attention mechanism to obtain an enhanced observation feature representation. The feature extraction network with an attention mechanism is used to perform attention-weighted aggregation on the state features of candidate ground dynamic users associated with the current time slot and the resource state features of UAVs and edge AI processors, mapping the variable-length input caused by the change in the number of dynamic users into a fixed-dimensional enhanced observation feature representation. D. Based on the enhanced observation feature representation, each UAV independently generates multi-dimensional decision-making results for synesthetic computing resources through a multi-agent policy network; E. Execute the decision results to provide integrated sensing and computing task services; according to the bandwidth allocation ratio and power allocation ratio, divide the total bandwidth and total transmission power of each UAV into a sensing part and a communication part, and allocate them to the associated ground dynamic users; perform radar sensing on the associated ground dynamic users, using the radar sensing rate as a sensing reliability indicator, and the sensing rate is not lower than a preset threshold. F. Update the information age based on the timeliness achievement determination result to form a closed-loop feedback to drive the joint decision update of subsequent time slots; the timeliness achievement determination is jointly determined based on whether the radar perception rate meets the preset threshold and whether the end-to-end service time meets the preset time window threshold; if the timeliness of the ground dynamic user in the current time slot is achieved and the user is associated with the UAV, the information age is updated to the service time of the current time slot, otherwise the information age is increased based on the previous time slot.

2. The multi-UAV sensory intelligence resource collaborative optimization method as described in claim 1, characterized in that, The system initialization in step A specifically includes: Establish drone ensembles, ground dynamic user ensembles, and time slot ensembles; Set the location information of ground dynamic users and UAVs within any time slot, where the ground dynamic users are located on the ground plane and the UAVs perform tasks at a fixed flight altitude; Set minimum safe distance constraints between drones, as well as the maximum total bandwidth, maximum total transmit power, and maximum number of ground dynamic users that can be served per time slot for each drone; A dynamic user association decision variable is introduced to represent the service association relationship between the UAV and dynamic ground users; The bandwidth allocation ratio and power allocation ratio are introduced to represent the resource allocation between sensing and communication for UAVs, respectively; We introduce the decision variables of unloading and local processing to represent the strategy of preprocessing perceived data locally or uploading it to an edge AI processor for preprocessing; A computing resource allocation variable is introduced to represent the allocation of local computing power of the drone and computing power of edge AI processors, respectively; Set an initial age value for each dynamic user on the ground.

3. The multi-UAV sensory intelligence resource collaborative optimization method as described in claim 1, characterized in that, The local observation information for each UAV in step B includes: The location coordinates of the UAV within the current time slot; The spatial location relationship information between the drone and each candidate ground dynamic user; The amount of radar data required to perceive each candidate dynamic ground user; Information age values ​​of dynamic users in various locations within the current time slot; The local data processing density of the drone in the current time slot.

4. The multi-UAV sensory intelligence resource collaborative optimization method as described in claim 1, characterized in that, The multi-dimensional decision-making results of the synesthetic computing resources in step D include at least the following: Ground-based dynamic user-related decision-making; UAV flight trajectory update volume; The bandwidth allocation ratio between sensing and communication; Power allocation ratio between sensing and communication; Local processing and edge offloading decisions for perceived data; Decisions on the allocation of local computing power for drones.

5. The multi-UAV sensory intelligence resource collaborative optimization method as described in claim 1, characterized in that, The execution of the integrated sensory computing task service in step E also includes: The number of ground dynamic users associated with each drone in any time slot shall not exceed the preset limit, and each ground dynamic user shall be associated with at most one drone in any time slot. Based on the unloading decision, the perception data is preprocessed locally on the drone side or transmitted to the edge AI processor for preprocessing. The preprocessed perception data is then fused and analyzed by the edge AI processor, and the results are sent to the ground dynamic user. The total service time from data collection to processing completion does not exceed the preset time window threshold.

6. The multi-UAV sensory intelligence resource collaborative optimization method as described in claim 1, characterized in that, In step F: The long-term optimization goal of the system is to maximize the long-term mean of the difference between the timeliness achievement rate and the weighted average information age, while satisfying flight safety constraints, correlation constraints, perception reliability constraints, service time constraints, and energy consumption constraints.

7. The multi-UAV sensory intelligence resource collaborative optimization method as described in claim 5, characterized in that, The execution of the integrated sensing and computing task service also includes energy consumption constraints: the total energy consumption on the UAV side includes local data preprocessing energy consumption and data transmission energy consumption, and the total energy consumption on the edge AI processor side includes edge data processing energy consumption; the total system energy consumption of each UAV and edge AI processor in any time slot does not exceed the preset energy consumption limit.

8. An apparatus for implementing the multi-UAV sensory intelligence resource collaborative optimization method described in any one of claims 1 to 7, characterized in that, include: The initialization module is used to set system constraints and initialize age variables. The observation construction module is used to acquire system status information and construct local observation information for each UAV. The feature extraction module is used to process the local observation information through a feature extraction network containing an attention mechanism to obtain an enhanced observation feature representation; The decision generation module is used to independently generate multi-dimensional decision results of synesthetic computing resources based on the enhanced observation feature representation through a multi-agent policy network; The task execution module is used to execute the decision results and provide integrated sensory computing task services. The feedback update module is used to update the information age based on the timeliness of the task and the judgment result, forming a closed-loop feedback to drive the joint decision update in subsequent time slots.

9. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to execute the multi-UAV sensory computing resource collaborative optimization method according to any one of claims 1 to 7.

Images