A method for unmanned aerial vehicle task scheduling oriented to non-correct information age optimization

Abstract

The application belongs to the field of unmanned aerial vehicle communication network, and discloses a method for unmanned aerial vehicle task scheduling facing non-correct information age optimization, which comprises the following steps: establishing a multi-unmanned aerial vehicle information collection system model; dividing the system running process into multiple fixed-length dynamic decision windows, and extracting window-level state representation at the beginning of each window; inputting the state representation into a dynamic weight decision model to output a five-dimensional weight vector, constructing a comprehensive cost matrix, and solving the task matching result by using the Hungarian algorithm; further performing perception configuration and communication resource allocation; updating the system state according to the window execution result, and optimizing the dynamic weight decision model in the offline stage. The application can adaptively adjust the task matching weight according to the system state, and reduce the long-term average non-correct information age of the system while ensuring the online real-time performance.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) communication networks, specifically relating to a method for UAV task scheduling and resource allocation that optimizes for incorrect information age. Background Technology

[0002] With the development of applications such as low-altitude economy, smart cities, emergency response, and air-to-ground collaborative monitoring, multi-UAV collaborative sensing networks are playing an increasingly important role in wide-area status acquisition, dynamic target monitoring, and real-time information updates. UAVs possess advantages such as maneuverability, rapid deployment, and wide field of view, but they are also constrained by factors such as flight speed, remaining energy, sensing success rate, and wireless backhaul capability. In multi-tasking dynamic environments, how to rationally allocate UAV tasks and coordinate sensing, encoding, and transmission resources is a key issue in improving system performance.

[0003] Existing research often uses throughput, latency, or information age as system performance evaluation metrics. However, information age only describes the time elapsed since the most recent successful update and cannot reflect whether the received information is still correct. When there are perception errors, link latency, link interruptions, or rapid changes in the target state, even if the received information is relatively new, it may be inconsistent with the true state. Incorrect information age, which measures the duration of information errors, can reflect both the timeliness and accuracy of the information and is more suitable for dynamic state update scenarios involving multiple UAVs.

[0004] Existing task scheduling and resource allocation methods generally suffer from the following shortcomings: First, the scheduling cost is constructed using static rules, making it difficult to adaptively adjust to changes in task urgency, spatial location, queue pressure, energy state, and link conditions. Second, task allocation, perception control, and communication resource allocation are interdependent, and direct joint solutions often have high complexity, making it difficult to meet online real-time requirements. Third, some learning-based methods directly output discrete scheduling actions, resulting in insufficient interpretability and difficulty in stably satisfying constraints. Fourth, most methods lack the utilization of historical states and execution feedback within continuous time slots, making it difficult to maintain scheduling stability in dynamic environments.

[0005] Therefore, there is an urgent need for a low-complexity online scheduling method that can adaptively adjust matching weights to reduce the long-term average age of incorrect information and improve the real-time performance and engineering practicality of the system. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide a method for UAV task scheduling and resource allocation based on dynamic window weight learning that minimizes the age of incorrect information.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for UAV mission scheduling optimized for incorrect information age includes the following steps:

[0009] S1: Establish a multi-UAV information acquisition system model, which includes a task state evolution model, UAV motion model, perception and observation model, air-to-ground communication model, queue and energy model, delay state estimation model, and incorrect information age index;

[0010] S2: Dynamic decision window partitioning and state extraction: Divide the system operation process into multiple fixed-length dynamic decision windows, and extract fixed-dimensional window-level state representations at the beginning of each dynamic decision window.

[0011] S3: Dynamic weight generation and task matching: Input the window-level state representation into the dynamic weight decision model, output a five-dimensional dynamic weight vector, construct a comprehensive cost matrix based on the five-dimensional dynamic weight vector and the five-dimensional feature vector of each UAV-task pair, and use the Hungarian algorithm to solve for the optimal task matching result within the current dynamic decision window.

[0012] S4: Perception Configuration and Communication Resource Allocation: Based on the task matching results, the number of repeated perceptions and uplink communication resources are jointly allocated. The uplink communication resource allocation uses a fixed priority parameter vector and the current system state to jointly construct an upload priority score, and determines the uplink resource allocation result for each UAV accordingly.

[0013] S5: State Update and Result Output: Update the task state estimate, queue state, remaining energy, and age of incorrect information based on the window execution result. Optimize the dynamic weight decision model during the offline training phase. During the online deployment phase, output the dynamic weight vector in real time based on the current dynamic decision window state for scheduling decisions. Finally, output the system's long-term average age of incorrect information and energy consumption.

[0014] Optionally, step S2 specifically includes the following steps:

[0015] S21: The system operation process is divided using a fixed-length window mechanism, where the fixed window length is set to 1. Then the first A dynamic decision window is defined as , For the first The starting time slot of each dynamic decision window, with adjacent windows connected end-to-end, satisfies... ;

[0016] S22: Extract fixed-dimensional window-level state representations at the start of each dynamic decision window. The window-level state representations include statistical features of incorrect information age, statistical features of queue load, features of remaining energy state, spatial relationship features between UAV and task points and base stations, system scale features, and time progress features, which are used as inputs to the dynamic weighted decision model.

[0017] Optionally, step S3 specifically includes the following steps:

[0018] S31: Input the window-level state representation into the dynamic weighted decision model, and output the five-dimensional original control vector. , They represent the first Within a time window, the dynamic weighted decision model outputs raw control variables for five state dimensions, which are used to characterize the model's unnormalized assessment of the relative importance of each dimension. The five raw control variables correspond to AOII urgency, spatial reachability, backhaul queue pressure, remaining energy risk, and backhaul link quality, respectively.

[0019] S32: Generating a five-dimensional dynamic weight vector using softmax mapping. ,Right now The five components of the five-dimensional dynamic weight vector are respectively AOII urgency weight, spatial reachability weight, backhaul queue pressure weight, remaining energy risk weight, and backhaul link quality weight.

[0020] S33: For each drone With the task Constructing a five-dimensional eigenvector The five-dimensional feature vectors correspond to the urgency features of task AOII, spatial reachability features, backhaul queue pressure features, remaining energy risk features, and backhaul link quality features, respectively. The five-dimensional dynamic weight vector and the five-dimensional feature vectors are then weighted and combined to obtain the comprehensive cost. Thus, the comprehensive cost matrix is ​​constructed. ;

[0021] S35: Transform the task assignment problem into a linear assignment problem by defining a binary matching variable. The optimization objective is The constraints are satisfied:

[0022]

[0023]

[0024] That is, each drone can serve at most one task, and each task can be served by at most one drone; the Hungarian algorithm is used to find the optimal task matching result within the current dynamic decision window.

[0025] Optionally, step S4 specifically includes the following steps:

[0026] S41: Optimization of Repeated Sensing Count: Assuming a drone For the task The number of repeated perceptions is The payoff function is defined as follows:

[0027]

[0028] in To successfully perceive the probability. To sense energy consumption, To generate data volume, , and The coefficient represents the trade-off between benefits and costs; the optimal number of iterations is... ;

[0029] S42: Construction of Communication Resource Allocation Priority: Let the fixed priority parameter vector be... ,in , , , These represent the queue backlog weight, task urgency weight, energy penalty weight, and channel quality weight, respectively; UAV The upload priority score is defined as ,in, For queue backlog normalized values, The normalized value for the urgency of the team's first mission. For energy risk measurement, This is a normalized value for link quality.

[0030] S43: Bandwidth and Power Allocation: Uplink bandwidth and transmit power are allocated according to the score ratio. The allocated bandwidth and transmit power are as follows:

[0031]

[0032]

[0033] in, Total bandwidth Total transmission power, It is a very small positive number.

[0034] Furthermore, the fixed priority parameter vector is determined through an offline alternating optimization process and remains unchanged during the online deployment phase; the offline alternating optimization process includes: an outer layer iteratively searching for a better fixed priority parameter vector, and an inner layer iteratively training a dynamic weight decision model under given parameters; updating parameters through grid search, random search, or Bayesian optimization to minimize the periodic comprehensive performance index; and repeating the iteration until the stopping condition is met.

[0035] Optionally, step S5 specifically includes:

[0036] S51: Update system status. After the window execution is completed, based on the perception and communication results, update the state estimates of each task (using delayed Kalman filtering), the queue backlog and remaining energy of each UAV, and calculate the age of incorrect information for each task in the current time slot.

[0037] S52: Computational performance metrics: Weighted total age of incorrect information in the computing system when the task is completed. ,in For the task The weight, For the task In the time slot The age of incorrect information; the average weighted age of incorrect information over the entire operating cycle is ;

[0038] S53: Output results: During the offline training phase, according to The system constructs reward functions based on energy consumption to optimize the dynamic weighted decision model; during the online deployment phase, it outputs the scheduling decision for the current window, and after the task is completed, it outputs the long-term average age of incorrect information and the total energy consumption of all drones.

[0039] The dynamic weighted decision model is trained offline using a proximal policy optimization algorithm. Each dynamic decision window is modeled as a decision step in a Markov decision process, where the state is represented by a window-level state, the action is a five-dimensional dynamic weight vector, and the reward is the negative of the window's comprehensive cost. The policy network output is mapped to a five-dimensional dynamic weight vector via softmax, and the value network evaluates the state values. The policy network and value network are jointly optimized by shearing the objective function and the value function loss, and a policy entropy term is introduced to enhance exploration.

[0040] The beneficial effects of this invention are as follows: Addressing the problem of poor information timeliness and low resource utilization in data collection scenarios involving multiple UAVs in the Internet of Things (IoT) and minimizing the age of incorrect information, this invention provides a UAV task scheduling and resource allocation method based on dynamic window weight learning. First, by using a fixed-length dynamic decision window and window-level state summary input, it can characterize multi-task, multi-UAV scenarios with a fixed dimension, balancing online real-time performance and cross-scale adaptability. Second, by outputting dynamic weight vectors through a learning model instead of directly outputting discrete scheduling actions, it combines interpretability, constraint tractability, and online execution efficiency. Third, it explicitly identifies five core features—task information timeliness, spatial reachability, backhaul queue pressure, remaining energy risk, and backhaul link quality—as comprehensive costs, and adaptively adjusts their importance based on system state changes, improving the scheduling strategy's adaptability to complex environments. Simultaneously, it employs a hierarchical mechanism that combines dynamic matching weights with fixed priority parameters, where dynamic weights are used for task matching and fixed parameters are used for communication resource allocation, improving scheduling flexibility and reducing the complexity of online joint optimization. Finally, fixed priority parameters in communication resource allocation are determined through offline alternating optimization, avoiding complete reliance on experience-based settings and improving the stability of the resource allocation module.

[0041] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0042] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0043] Figure 1 This is a diagram illustrating the overall architecture of the multi-UAV collaborative sensing and air-to-ground communication system of the present invention.

[0044] Figure 2 This is a flowchart of the task scheduling and resource allocation method for minimizing the age of incorrect information according to the present invention;

[0045] Figure 3 This is a schematic diagram of the fixed-length dynamic decision window mechanism of the present invention;

[0046] Figure 4 This is a schematic diagram of the decision-making process and status update within a single window according to the present invention;

[0047] Figure 5 This is a diagram illustrating the offline training and online deployment framework of the dynamic weight model of this invention. Detailed Implementation

[0048] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0049] Please see Figure 1 Figure 4 The UAV task scheduling and resource allocation method based on dynamic window weight learning for minimizing the age of incorrect information, as described in this invention, specifically includes the following steps:

[0050] Step 1: Establish a multi-UAV information acquisition system model. In this invention, it is considered that there are [unmanned aerial vehicles] within the monitoring area. There are ground task nodes at fixed locations, and the task set is denoted as . System deployment A fleet of drones, the group of drones is denoted as Specifically, it includes the following steps:

[0051] Step 1.1: Task State Evolution Model. Each task node in the time slot The true state is recorded as The model is based on a first-order autoregressive AR(1) process, i.e.: ,in This is the state-time correlation coefficient. The long-term mean of the state. The noise is the process noise, which follows a mean of 0 and a variance of . The Gaussian distribution.

[0052] Step 1.2: Drone motion model. A drone in a time slot The horizontal position is , Representing a two-dimensional real space, the flight altitude of all drones is fixed at . The maximum flight speed of the drone is The time slot length is If the maximum flight distance is seconds, then the maximum flight distance is The UAV flies in a straight line pointing towards the target mission point, satisfying the following movement constraints: .

[0053] Step 1.3: Perception and Observation Model. Let the first step be... drones and mission nodes In the time slot The Euclidean distance is The probability of a single successful perception is modeled as an exponential decay model. ,in This is the distance attenuation coefficient. Represents a node Location, This represents the probability of a single successful perception. If the drone executes... If the perception is successful the first time, the probability of successful perception is: Once successfully detected, the drone acquires noisy observations. A linear observation model is adopted. Measure noise It follows a Gaussian distribution with zero mean and its variance is... Related to sensing distance and quantization encoding mode. Assume the system supports... A type of quantization encoding mode, the mode selection variable is denoted as Then the measurement noise variance is modeled as follows: ,in Based on the perceived noise variance This is the distance growth factor. For pattern The corresponding additional distortion, and satisfying In other words, the high-precision mode corresponds to less additional distortion.

[0054] Step 1.4: Air-to-Ground Communication Model. Assume the base station is located in a fixed position. The height is drones The horizontal distance from the base station is The three-dimensional distance is The link elevation angle is... The line-of-sight link probability is determined using an empirical model. .in , These are environmentally relevant parameters. The additional losses at line-of-sight and non-line-of-sight are respectively... and The average path loss is .in For free space path loss, , For carrier frequency, The speed of light. Converted to a linear scale as... Setting up drones The transmission power is The received signal-to-noise ratio is ,in This refers to the noise power at the receiving end. If bandwidth is allocated... The uplink transmission rate is The number of bits that can be uploaded within a time slot is .

[0055] Step 1.5: Queue and Energy Model. Each drone maintains a first-in, first-out queue to cache task data packets to be uploaded. Let the... A drone in a time slot The number of queue backlog bits is Its evolution satisfies ,in For time slots The number of newly generated data bits added to the queue.

[0056] The energy consumption of a drone can be divided into three parts: communication energy consumption, propulsion energy consumption, and perception and hovering energy consumption. Communication energy consumption is... The propulsion energy consumption is modeled using a rotary-wing unmanned aerial vehicle (UAV).

[0057]

[0058] in For velocity vectors, , To advance the parameters, It is the acceleration due to gravity. This refers to hovering power. The energy consumption for sensing hovering is... ,in To sense power. Assume the initial energy of the drone is... Then the remaining energy satisfy: , ,and For all Established.

[0059] Step 1.6: Delay State Estimation Model. The base station maintains state estimates for each task. Let the task... In the time slot The estimated value and the error variance are respectively and Before receiving new observations, predictions are performed according to the state evolution model. , . This represents the predicted state estimate for the next time slot. The state autocorrelation coefficient represents the degree of influence of the current state on the state at the next time step. Represents the long-term mean of the state. Let V be the variance of the prediction error for the next time slot. Suppose an observation occurs in time slot V is... Generation, in time slot The transmission delay to the base station is To utilize this delayed observation to update the current state, it is transformed into the current time slot. Equivalent observation. Define the attenuation factor. , Representing transmission delay, the equivalent observation value is , Indicates task The value measured at the time of observation generation has an equivalent observation noise variance of: ,in This represents the variance of the observation noise. The variance of process noise accumulated during the delay period:

[0060]

[0061] Therefore, the Kalman gain is The update of state and error variance is as follows , . This represents the variance of the error corresponding to the prior estimate. Indicates the task In the time slot Prior state estimation, Indicates task In the time slot Posterior state estimation, Indicates time slot The posterior error variance.

[0062] Step 1.7: Incorrect Information Age Assessment Model. Define the task. In the time slot The estimation error is Let the error tolerance threshold be... ,like If the information is correct, then the age is considered correct; otherwise, it is incorrect. ;otherwise ,in For the task The most recently correctly estimated time slot. The system's weighted total age of incorrect information is... ,in For the task The weighting of the average weighted age of incorrect information over the entire operating cycle is... , This represents the total number of time slots in the system.

[0063] Step 2: Dynamic Decision Window Partitioning and State Extraction: Since the above optimization objective involves multiple coupled variables such as task matching, perception configuration, and communication resource allocation, directly solving them in time slots would result in high computational complexity and would be difficult to meet online real-time requirements. Therefore, this invention adopts a fixed-length dynamic decision window mechanism, specifically including the following steps:

[0064] Step 2.1: Dynamic Decision Window Division: Assume a fixed window length of... Then the first A dynamic decision window is defined as , For the first The starting time slot of each dynamic decision window, with adjacent windows connected end-to-end, satisfies... At the start of each window, the system makes a decision on task matching weights only once; within the window, continuous... The matching weight remains unchanged throughout the time slot, thereby reducing the overhead caused by frequent re-solving.

[0065] Step 2.2: Window-level state representation extraction: In order to enable the dynamic decision model to adapt to different task scales and UAV scales, this invention does not directly use the full original state vector as input, but performs a statistical summary of the current system state at the beginning of each window to construct a fixed-dimensional window-level state representation.

[0066] Step 3: Dynamic Weight Generation and Task Matching: At the beginning of each dynamic decision window, the system generates a five-dimensional dynamic weight vector for task matching based on the current window-level state representation. This vector reflects the system's dynamic decision preferences regarding information inaccuracy, information age, spatial cost, queue pressure, energy risk, and link quality. Specifically, this includes the following steps:

[0067] Step 3.1: Input the window-level state representation into the dynamic weighted decision model, and output a five-dimensional original action vector:

[0068]

[0069] They represent the first Within a time window, the dynamic weight decision model outputs raw control variables for five state dimensions, representing the unnormalized assessment of the model's relative importance to each dimension. The five raw control variables correspond to AOII urgency, spatial reachability, backhaul queue pressure, remaining energy risk, and backhaul link quality, respectively. To ensure the output satisfies the constraint that each component is non-negative and its sum is 1, a softmax normalization mapping is performed on the five-dimensional raw control vector to obtain the five-dimensional dynamic weight vector corresponding to the window. ,Right now: The five dynamic weights correspond to AOII urgency, spatial reachability, backhaul queue pressure, remaining energy risk, and backhaul link quality, respectively. Through this design, the system can adaptively adjust its focus on different decision factors during UAV mission matching based on state changes under different dynamic decision windows.

[0070] Step 3.2: Construction of five-dimensional feature vectors. For any UAV... With the task Construct a five-dimensional feature vector:

[0071]

[0072] Each dimension of the feature corresponds to the urgency of the AOII task, spatial accessibility, queue pressure, energy risk, and link quality, respectively, and is used to characterize the comprehensive cost of the UAV performing this task from different perspectives. AOII urgency feature ,in For the task Incorrect age information This is the set urgency saturation value. This feature represents the current urgency of the task; a higher value indicates a more urgent task. Spatial reachability feature. ,in, For drones With the task distance, The entire problem modeling in this invention is in Within the square area, This represents the length of the square region. This feature is used to characterize the spatial cost of the UAV performing the corresponding task; and the backhaul queue pressure feature. ,in, For drones Queue backlog, This is the backlog trigger threshold. This feature reflects the existing cache pressure on the drone; the larger the queue, the higher the cost of continuing to undertake new tasks; remaining energy risk feature. ,in, For drones The remaining energy, This represents the drone's maximum energy. This feature characterizes energy risk; the lower the remaining energy, the higher the value. (Backhaul link quality characteristics) ,in, For drones Distance to base station, This indicates the maximum communication distance in the system scenario. This represents the trade-off coefficient between the distance term and the upload time term. This indicates the time required for the drone to transmit data back to the base station via the backhaul link. This indicates the maximum tolerable latency for the upload time. , For the corresponding link capacity, This indicates the estimated size of the data packet that the drone needs to upload; this feature is used to characterize the impact of backhaul link quality on task performance. To avoid uneven influence from different units on the construction of the overall cost, the above feature is normalized before participating in the weighted calculation.

[0073] Step 3.3: Construction of the Comprehensive Cost Matrix. After obtaining the five-dimensional dynamic weight vector and the five-dimensional feature vector of each UAV-task pair, the UAV is defined... With the task The combined cost between them is Therefore, the comprehensive cost matrix corresponding to the current dynamic decision window can be constructed. Each element in the comprehensive cost matrix corresponds to a candidate UAV-task pair; a smaller value indicates that it is more advantageous for the corresponding UAV to perform the corresponding task in the current window state. Since the weight vector is dynamically adjusted as the window state changes, the comprehensive cost matrix also has the characteristic of adaptively updating with changes in system state.

[0074] Step 3.4: Task Matching Optimization Solution. After constructing the comprehensive cost matrix, the UAV task allocation problem is represented as a linear assignment optimization problem. Define a binary matching variable. When drones Tasks are assigned and executed within the current dynamic decision window. hour, The value is 1 if the condition is met, and 0 otherwise. Therefore, the task matching optimization objective can be expressed as:

[0075]

[0076] The following constraints must be met: , , That is, each drone can execute at most one task within the same dynamic decision window; each task can be assigned to at most one drone within the same dynamic decision window; the matching variable is a binary discrete variable.

[0077] The Hungarian algorithm is used to solve the above task matching problem. When the number of UAVs and tasks is unequal, virtual nodes are introduced to expand the cost matrix into a square matrix. After obtaining the solution from the Hungarian algorithm, the UAV task matching scheme within the current dynamic decision window is obtained. This matching scheme will serve as the input basis for subsequent optimization of repeated sensing times, coding mode selection, and communication resource allocation, and will be applied throughout the entire window length. The corresponding continuous time slots remain unchanged, thereby avoiding the computational overhead and frequent task switching problems caused by rematching every time slot, and improving the continuity and stability of system operation.

[0078] Step 4: Perception Configuration and Communication Resource Allocation. After obtaining the UAV mission matching scheme under the current dynamic decision window, this invention further performs perception configuration and communication resource allocation on the matched UAV-mission pairs to coordinate the relationship between perception success rate, observation accuracy, queue load, link backhaul capability, and remaining energy. Specifically, this includes the following steps:

[0079] Step 4.1: Optimize the number of repeated sensing attempts. Assume a drone... In the Within a dynamic decision window, the task... The number of repeated perceptions is The corresponding return function is defined as follows: ,in, To successfully perceive the probability. To sense energy consumption, To generate data volume, , , This represents the trade-off coefficient between benefits and costs. In this invention, the system can enumerate and calculate the above-mentioned benefit function from a preset set of selectable repeated sensing counts, and select the repeated sensing count that maximizes the benefit function value as the optimal configuration. .

[0080] Step 4.2: Encoding Mode Selection. Assume the system supports... There are several encoding modes, each corresponding to different data compression ratios and additional observation distortions. Generally, higher encoding compression results in smaller data volumes, but may lead to greater additional distortion; higher encoding precision yields higher quality observations, but increases queue backlog and communication load. When the task is urgent and the link quality is good, a higher precision encoding mode can be preferred to improve the accuracy of status updates; when the UAV queue backlog is large or the backhaul link is poor, a higher compression ratio encoding mode can be preferred to reduce the uplink data pressure within the current window.

[0081] Step 4.3: Construction of Communication Resource Allocation Priority. In this invention, a fixed priority parameter vector is introduced to achieve coordinated allocation of communication resources among multiple UAVs. ,in , , and These correspond to the influence coefficients of queue backlog, task urgency, energy risk, and link quality on the upload priority score, respectively. For any drone in any time slot... Its upload priority score is defined as ,in, For queue backlog normalized values, The normalized value for the urgency of the team's first mission. For energy risk measurement, This is the normalized value for link quality.

[0082] Step 4.4: Bandwidth and Power Allocation. After obtaining the upload priority score for each UAV, the total uplink bandwidth and total transmit power of the system are allocated proportionally. Let the UAVs... In the time slot The obtained bandwidth and transmit power are respectively and Then it can be expressed as , .in, The total system bandwidth This represents the total transmit power of the system. To prevent extremely small positive numbers with a denominator of zero.

[0083] Step 4.5: Offline alternating optimization with fixed priority parameter vector. Because the fixed priority parameter vector... Directly affecting bandwidth and transmit power allocation, and consequently the long-term average age of incorrect information in the system, this invention employs an offline alternating optimization method to determine this. The parameter vector optimization and dynamic weight decision model training are divided into two iterative layers: the outer iteration searches for a better parameter vector, while the inner iteration trains the dynamic weight decision model under given conditions.

[0084] First, initialize the parameter vector, setting each component to non-negative values ​​and normalizing it. Then, in the... In the outermost iteration, the current parameter vector is fixed, and offline reinforcement learning training of the dynamic weight decision model is performed until convergence. Then, the model parameters are fixed, and the long-term performance of the current parameter vector is evaluated in a validation environment. The parameter vector is updated using grid search, random search, or Bayesian optimization. Taking grid search as an example, the parameters are adjusted within a preset range. , , and The value of is chosen to minimize the objective function and then used as the input for the next round.

[0085] Repeat the two-stage iterations described above until the stopping condition is met: the improvement of the objective function in two adjacent rounds falls below a preset threshold, or the optimal parameter vector no longer changes for several consecutive rounds. After the alternating optimization is completed, the parameter vector with the best performance is taken as the final parameters and remains unchanged during the online deployment phase.

[0086] Step 5: State Update and Result Output: Update the task state estimate, queue state, remaining energy, and age of incorrect information based on the window execution results, and generate feedback information for optimizing the dynamic weight decision model during the offline training phase. This includes the following steps:

[0087] Step 5.1: Update the state and calculate performance metrics. After the window execution is complete, based on the perception and communication results, update the task's state estimate using delayed Kalman filtering, and simultaneously update the queue backlog and remaining energy of each UAV.

[0088] Step 5.2: Offline training and online deployment, such as... Figure 5 As shown, the dynamic weight decision model is trained offline using the Proximal Policy Optimization (PPO) algorithm. Each window is modeled as an MDP step, and the state is represented as a window-level state. The action is a five-dimensional primitive control vector. Window reward set ,in, This represents the expectation of random factors during system operation, including task state evolution, channel fading, sensing success probability, and observation noise, etc. The average weighted age of incorrect information over the period. As a penalty for average energy consumption, Penalty for collision risk, For the sake of fair punishment, Penalty for mobile costs, These represent the weight coefficients for each penalty term. The policy network output is mapped using softmax to obtain a five-dimensional dynamic weight vector. The value network evaluates the state values. After collecting trajectories, the network parameters are updated until convergence. During online deployment, the pre-trained model is directly loaded, and step S2 is executed sequentially for each window. S5 requires no online optimization and meets real-time requirements.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A UAV mission scheduling method for optimizing age based on incorrect information, characterized in that, Includes the following steps: S1: Establish a multi-UAV information acquisition system model, which includes a task state evolution model, UAV motion model, perception and observation model, air-to-ground communication model, queue and energy model, delay state estimation model, and incorrect information age index; S2: Dynamic decision window partitioning and state extraction: Divide the system operation process into multiple fixed-length dynamic decision windows, and extract fixed-dimensional window-level state representations at the beginning of each dynamic decision window. S3: Dynamic weight generation and task matching: Input the window-level state representation into the dynamic weight decision model, output a five-dimensional dynamic weight vector, construct a comprehensive cost matrix based on the five-dimensional dynamic weight vector and the five-dimensional feature vector of each UAV-task pair, and use the Hungarian algorithm to solve for the optimal task matching result within the current dynamic decision window. S4: Perception Configuration and Communication Resource Allocation: Based on the task matching results, the number of repeated perceptions and uplink communication resources are jointly allocated. The uplink communication resource allocation uses a fixed priority parameter vector and the current system state to jointly construct an upload priority score, and determines the uplink resource allocation result for each UAV accordingly. S5: State Update and Result Output: Update the task state estimate, queue state, remaining energy, and age of incorrect information based on the window execution result. Optimize the dynamic weight decision model during the offline training phase. During the online deployment phase, output the dynamic weight vector in real time based on the current dynamic decision window state for scheduling decisions. Finally, output the system's long-term average age of incorrect information and energy consumption.

2. The UAV mission scheduling method for age optimization based on incorrect information as described in claim 1, characterized in that: The task state evolution model satisfies the following: , Indicates the first Each task node in the time slot The true state This is the state-time correlation coefficient. The long-term mean of the state. This is process noise; The UAV motion model states that the UAV flies in a straight line along the direction pointing to the target mission point, satisfying the following motion constraints: , and They represent the first A drone in a time slot and Horizontal position, The maximum flight speed of the drone, with a time slot length of [value missing]. Second; The perception and observation model: executed by the UAV If the perception is successful the first time, the probability of successful perception is: , This represents the probability of a single successful perception; after a successful perception, the drone obtains noisy observations. , Measure noise It follows a Gaussian distribution with zero mean and its variance is... Related to sensing distance and quantization coding mode, ,in Based on the perceived noise variance This is the distance growth factor. For the first drones and mission nodes In the time slot Euclidean distance, Indicates the additional distortion corresponding to the encoding mode; The air-to-ground communication model: assuming a drone The transmission power is The received signal-to-noise ratio is ,in For receiver noise power, The average path loss is measured linearly, if bandwidth is allocated. The uplink transmission rate is The number of bits that can be uploaded within a time slot is ; The queue and energy model: Let the first... A drone in a time slot The number of queue backlog bits is Its evolution satisfies ,in For time slots The number of newly generated data bits added to the queue; the remaining energy of the drone satisfies: , This indicates the initial energy of the drone. For communication energy consumption, To promote energy consumption, To sense hovering energy consumption; The update of state and error variance in the delayed state estimation model satisfies , , For Kalman gain, Equivalent observations; This represents the variance of the error corresponding to the prior estimate. Indicates the task In the time slot Prior state estimation, Indicates task In the time slot Posterior state estimation, Indicates time slot The posterior error variance; The incorrect information age index defines the task. In the time slot The estimation error is Let the error tolerance threshold be... ,like If the information is correct, then the age is considered correct; otherwise, it is incorrect. ;otherwise ,in For the task The most recently correctly estimated time slot; the system's weighted total age of incorrect information is... ,in For the task The weights, and the average weighted age of incorrect information over the entire operating cycle are: , This represents the total number of time slots in the system.

3. The UAV mission scheduling method for age optimization based on incorrect information as described in claim 1, characterized in that: Step S2 specifically includes the following steps: S21: The system operation process is divided using a fixed-length window mechanism, where the fixed window length is set to 1. Then the first A dynamic decision window is defined as , For the first The starting time slot of each dynamic decision window, with adjacent windows connected end-to-end, satisfies... ; S22: Extract fixed-dimensional window-level state representations at the start of each dynamic decision window. The window-level state representations include statistical features of incorrect information age, statistical features of queue load, features of remaining energy state, spatial relationship features between UAV and task points and base stations, system scale features, and time progress features, which are used as inputs to the dynamic weighted decision model.

4. The UAV mission scheduling method for age optimization based on incorrect information as described in claim 1, characterized in that: Step S3 specifically includes the following steps: S31: Input the window-level state representation into the dynamic weighted decision model, and output the five-dimensional original control vector. , They represent the first Within a time window, the dynamic weighted decision model outputs raw control quantities for five state dimensions, which are used to characterize the model's unnormalized assessment of the relative importance of each dimension; the five raw control quantities correspond to AOII urgency, spatial reachability, backhaul queue pressure, remaining energy risk, and backhaul link quality, respectively. S32: Generating a five-dimensional dynamic weight vector using softmax mapping. ,Right now The five components of the five-dimensional dynamic weight vector are respectively AOII urgency weight, spatial reachability weight, backhaul queue pressure weight, remaining energy risk weight, and backhaul link quality weight. S33: For each drone With the task Constructing a five-dimensional eigenvector The five-dimensional feature vectors correspond to the urgency features of task AOII, spatial reachability features, backhaul queue pressure features, remaining energy risk features, and backhaul link quality features, respectively. The five-dimensional dynamic weight vector and the five-dimensional feature vectors are then weighted and combined to obtain the comprehensive cost. Thus, the comprehensive cost matrix is ​​constructed. ; S35: Transform the task assignment problem into a linear assignment problem by defining a binary matching variable. The optimization objective is The constraints are satisfied: That is, each drone can serve at most one task, and each task can be served by at most one drone; the Hungarian algorithm is used to find the optimal task matching result within the current dynamic decision window.

5. The UAV mission scheduling method for age optimization based on incorrect information according to claim 4, characterized in that: The urgency characteristics of AOII ,in For the task Incorrect information about age, The set saturation value for urgency; The spatial reachability feature ,in, For drones With the task distance, , express The length of the square area; The backhaul queue pressure characteristics ,in, For drones Queue backlog, The backlog trigger threshold; The remaining energy risk characteristics ,in, For drones The remaining energy, This represents the maximum energy output of the drone. The backhaul link quality characteristics ,in, For drones Distance to base station This indicates the maximum communication distance in the system scenario. This represents the trade-off coefficient between the distance term and the upload time term. This indicates the time required for the drone to transmit data back to the base station via the backhaul link. This represents the maximum tolerable latency for uploading.

6. The UAV mission scheduling method for age optimization based on incorrect information according to claim 1, characterized in that: Step S4 specifically includes the following steps: S41: Optimization of Repeated Sensing Count: Assuming a drone For the task The number of repeated perceptions is The payoff function is defined as follows: in To successfully perceive the probability. To sense energy consumption, To generate data volume, , and The coefficient represents the trade-off between benefits and costs; the optimal number of iterations is... ; S42: Construction of Communication Resource Allocation Priority: Let the fixed priority parameter vector be... ,in , , , These represent the queue backlog weight, task urgency weight, energy penalty weight, and channel quality weight, respectively; UAV The upload priority score is defined as ,in, For queue backlog normalized values, The normalized value for the urgency of the team's first mission. For energy risk measurement, This is a normalized value for link quality. S43: Bandwidth and Power Allocation: Uplink bandwidth and transmit power are allocated according to the score ratio. The allocated bandwidth and transmit power are as follows: in, Total bandwidth Total transmission power, It is a very small positive number.

7. The UAV mission scheduling method for age optimization based on incorrect information as described in claim 6, characterized in that: The fixed priority parameter vector is determined through an offline alternating optimization process and remains unchanged during the online deployment phase. The offline alternating optimization process includes: an outer layer iteratively searching for a better fixed priority parameter vector, and an inner layer iteratively training a dynamic weight decision model under given parameters; Update parameters using grid search, random search, or Bayesian optimization to minimize the periodic overall performance index; repeat the iteration until the stopping condition is met.

8. The UAV mission scheduling method for age optimization based on incorrect information according to claim 1, characterized in that: Step S5 specifically includes: S51: Based on the perception and communication results, use delayed Kalman filtering to update the state estimates of each task, the queue backlog and remaining energy of each UAV, and calculate the age of incorrect information for each task in the current time slot. S52: When the task is completed, calculate the weighted total age of incorrect information in the system. ,in For the task The weight, For the task In the time slot The age of incorrect information; the average weighted age of incorrect information over the entire operating cycle is ; S53: During the offline training phase, according to The system also constructs a reward function based on energy consumption to optimize the dynamic weighted decision model; during the online deployment phase, it outputs the scheduling decision for the current window, and after the mission ends, it outputs the long-term average age of incorrect information and the total energy consumption of all drones.

9. A UAV mission scheduling method for age optimization based on incorrect information as described in claim 1 or 4, characterized in that: The dynamic weight decision model is trained offline using a near-end policy optimization algorithm. Each dynamic decision window is modeled as a decision step in a Markov decision process, where the state is a window-level state representation, the action is a five-dimensional dynamic weight vector, and the reward is the negative of the window's comprehensive cost. The policy network output is mapped to a five-dimensional dynamic weight vector via softmax, and the value network evaluates the state value.

Images