Method and System for UAV Cargo Delivery Management Based on Policy Optimization

Abstract

The present invention discloses a drone cargo delivery management method and system based on strategy optimization, specifically relating to the technical field of cargo delivery management. An improved incremental K-means clustering algorithm is combined with KD-Tree index to optimize nearest neighbor search, thereby improving the efficiency of order area grouping. Based on the order cluster set O, a mixed integer programming + Hungarian algorithm is used to match drone tasks, optimize the task allocation matrix M, and dynamically adjust the delivery path to cope with real-time weather and airspace restrictions. During the delivery process, drone sensors collect data such as order liquidity index, battery consumption rate, and task balance, and use fuzzy logic reasoning to judge scheduling delays, ensuring that the system can still achieve efficient, balanced, and intelligent drone scheduling in high-load scenarios, significantly reducing delivery delays, lowering the empty load rate, and improving overall logistics efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of cargo delivery management, and in particular to a drone cargo delivery management method and system based on strategy optimization. Background Art

[0002] Policy-based optimization management for drone cargo delivery involves using optimization strategies and algorithms to improve drone delivery efficiency, reduce costs, and enhance overall operational management. This management model typically involves multiple aspects, including path planning, task scheduling, energy management, and obstacle avoidance strategies, to ensure drones can efficiently and safely complete delivery missions in complex environments.

[0003] The existing technology has the following shortcomings:

[0004] Existing technologies use the K-means clustering algorithm to group orders from similar areas and reduce duplicate routes. However, traditional K-means requires calculating the distance from each data point to all cluster centers. The computational complexity of each iteration is H(n×k×d) (n = number of orders, k = number of clusters, d = dimension). When the order volume is large, the computation time increases significantly, making it impossible for the system to allocate tasks in real time. For example, during peak periods of large-scale logistics, when a large number of orders are poured in, the K-means computation time can cause delays of several minutes or even hours in the delivery system, affecting overall logistics efficiency. Furthermore, delays in order scheduling can cause drones to wait for extended periods, increasing idle time. Furthermore, if a drone automatically executes a default route due to a timeout, this can lead to wasted resources or even delivery failures. Summary of the Invention

[0005] The purpose of the present invention is to provide a drone cargo delivery management method and system based on strategy optimization to address the shortcomings of the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a drone cargo delivery management method based on policy optimization, comprising:

[0007] S100, obtaining basic information of the order to be delivered, including the order's geographic location, cargo weight, expected delivery time, and drone's load capacity;

[0008] S200, based on the geographic location information of the orders, an improved incremental K-means clustering algorithm is used to group the orders by region, wherein a KD-Tree index structure is used to optimize the nearest neighbor search, and an order cluster set O is generated;

[0009] S300, based on the order cluster set O and the load capacity of the UAV, tasks are matched and the optimal allocation scheme between order clusters and UAVs is calculated to obtain the task allocation matrix M;

[0010] S400 optimizes the delivery path of the UAV based on the task allocation matrix M, generates the initial flight path R, and dynamically adjusts the path based on real-time weather and airspace restrictions;

[0011] S500, execute the delivery task, and collect order liquidity data and drone task load balance data through the drone's sensors to determine whether there is a delay in the current scheduling. If so, update the order cluster set O, adjust the task allocation matrix M, and return to S200 to re-optimize the scheduling; otherwise, complete the delivery task.

[0012] Preferably, the S200 adopts an improved incremental K-means clustering algorithm, wherein when a new order arrives, a KD-Tree is used to perform a nearest neighbor search, and the new order is dynamically assigned to the optimal order cluster, specifically including: using the order geographic location information as the input data of the KD-Tree; constructing the KD-Tree using a bisection hyperplane partitioning method: selecting the dimension with the largest variance in latitude or longitude as the partitioning axis, using the median partitioning method to perform a binary partition on the data, recursively constructing the KD-Tree, and finally forming an efficient index structure; after the KD-Tree is established, the time complexity of the nearest neighbor search is reduced to H(log n), where H(log n) represents the growth rate of the algorithm execution time as the amount of input data increases, that is, the efficiency of the KD-Tree nearest neighbor search.

[0013] Preferably, a dynamic K value optimization strategy is used to determine the optimal number of clusters: calculate the intra-cluster sum of squares under different K values , the expression is: ; Where K is the number of clusters, is the set of data points of the vth cluster, x is the coordinate of an order belonging to the vth cluster, The center of the vth cluster is calculated by standard K-means based on historical order data. When a new order is added, the nearest neighbor search is performed using KD-Tree to quickly match the nearest order cluster. If the distance between the new order and the nearest cluster center is less than the threshold dthresh, the new order is directly added to the cluster. If the distance is greater than the threshold, a new order cluster is created and the K value is updated. The cluster center is updated using the sliding mean update method: ; Where N is the original order number, is the new order coordinate, is the cluster center before updating, is the updated cluster center.

[0014] Preferably, the S300 uses mixed integer programming and Hungarian algorithm to perform drone task matching, specifically: calculating the task cost value of drone i to order cluster j , and construct the task cost matrix , task cost value The calculation expression is: ; Among them: α is the weight of the path length, β is the weight of the cargo weight, γ is the weight of the electricity, is the total weight of the order cluster, The remaining battery power of the drone is in the range of [0,1]. Represents the distance between drone i and order cluster j, and uses the Hungarian algorithm to solve the optimal matching to minimize the total task cost; from the cost matrix Find the minimum value and match it; iteratively update the unmatched order clusters until all tasks are assigned; generate the final task assignment matrix M.

[0015] Preferably, the S400 optimizes the delivery path, including: for each drone i, the order cluster is assigned , n is the total number of drones, and a traveling salesman problem model is constructed. Suppose the drone starts from the current starting position and needs to visit the order clusters in sequence , and finally return to the starting point or the nearest charging station; define the objective function to minimize the total delivery path as , the calculation expression is: ;in, Order cluster and The Euclidean distance between them; calculate the initial path Rinit, randomly select two points in the path , swap the order, calculate the length of the new path; if the new path is shorter, replace the original path; iterate repeatedly until the path can no longer be optimized.

[0016] Preferably, the S500 monitors order liquidity in real time through the drone's GPS, IMU sensor and RFID / NFC identification device, and calculates the order liquidity index, as well as the battery consumption rate, payload ratio and mission balance index.

[0017] Preferably, the order liquidity index, battery consumption rate, effective load ratio and task balance index are used as input items of the fuzzy logic, and the judgment result of whether there is a delay in the current scheduling is used as the output item of the fuzzy logic;

[0018] Formulate fuzzy rules to describe the impact of order liquidity index, battery consumption rate, effective load ratio and task balance index on the current scheduling delay;

[0019] Determine the risk of current scheduling delay based on the fuzzy output results and take corresponding measures.

[0020] The present invention also provides a drone cargo delivery management system based on strategy optimization, which includes an order management module, an order partitioning and clustering module, a task allocation and drone matching module, a drone path adjustment module, and a scheduling optimization module;

[0021] The order management module obtains basic information about the order to be delivered, including the order's geographic location, cargo weight, expected delivery time, and the drone's load capacity;

[0022] The order partition clustering module uses an improved incremental K-means clustering algorithm to group orders by region based on their geographic location information. The KD-Tree index structure is used to optimize the nearest neighbor search and generate an order cluster set O.

[0023] The task allocation and drone matching module matches tasks based on the order cluster set O and the load capacity of the drone, and calculates the optimal allocation plan between order clusters and drones to obtain the task allocation matrix M;

[0024] The UAV path adjustment module optimizes the delivery path of the UAV based on the task allocation matrix M, generates the initial flight path R, and dynamically adjusts the path based on real-time weather and airspace restrictions;

[0025] The scheduling optimization module executes the delivery task and collects order liquidity data and drone task load balance data through the drone's sensors to determine whether there is a delay in the current scheduling. If so, it updates the order cluster set O and adjusts the task allocation matrix M, returning to the order partitioning and clustering module to re-optimize the scheduling; otherwise, the delivery task is completed.

[0026] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0027] 1. This invention uses an improved incremental K-means clustering algorithm, combined with a KD-Tree index to optimize nearest neighbor search, effectively reducing the clustering computational complexity to H(logn), ensuring efficient order region grouping even in large-scale order scenarios. Furthermore, this invention employs a dynamic K-value optimization strategy to dynamically adjust the number of clusters based on order distribution, making order clusters more rational, reducing unnecessary route duplication, and improving drone delivery efficiency.

[0028] 2. The present invention optimizes drone task matching through mixed integer programming + Hungarian algorithm, combines task cost matrix calculation, takes into account path length, load capacity and power constraints, ensures balanced task allocation, and improves drone resource utilization. In addition, during the task execution phase, the present invention uses GPS, IMU sensors, RFID / NFC devices to monitor key data such as order liquidity index, battery consumption rate, task balance index, etc. in real time, and uses fuzzy logic reasoning to judge the risk of scheduling delay. If an abnormality is found, the task allocation plan is dynamically adjusted to ensure the real-time and stability of drone scheduling. Overall, the present invention realizes accurate order grouping, intelligent task matching, efficient path optimization and real-time scheduling adjustment, significantly improving the efficiency of drone logistics distribution, reducing energy consumption, and reducing delivery failure rate in large-scale order scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments described in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0030] Figure 1 Flow chart of the method of the present invention.

[0031] Figure 2 It is a system module diagram of the present invention. DETAILED DESCRIPTION

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0033] Example 1, please refer to Figure 1 As shown, the drone cargo delivery management method based on policy optimization described in this embodiment includes:

[0034] S100, obtaining basic information of the order to be delivered, including the order's geographic location, cargo weight, expected delivery time, and drone's load capacity;

[0035] S200, based on the geographic location information of the orders, an improved incremental K-means clustering algorithm is used to group the orders by region, wherein a KD-Tree index structure is used to optimize the nearest neighbor search, and an order cluster set O is generated;

[0036] S300, based on the order cluster set O and the load capacity of the UAV, tasks are matched and the optimal allocation scheme between order clusters and UAVs is calculated to obtain the task allocation matrix M;

[0037] S400 optimizes the delivery path of the UAV based on the task allocation matrix M, generates the initial flight path R, and dynamically adjusts the path based on real-time weather and airspace restrictions;

[0038] S500, execute the delivery task, and collect order liquidity data and drone task load balance data through the drone's sensors to determine whether there is a delay in the current scheduling. If so, update the order cluster set O, adjust the task allocation matrix M, and return to S200 to re-optimize the scheduling; otherwise, complete the delivery task.

[0039] The drone delivery system collects order-related information from multiple data sources, including:

[0040] Order ID: Uniquely identifies each order and is used to track delivery status. Creation Time: Timestamp of the order creation, used to determine priority and scheduling strategies. Shipping Point Coordinates: The longitude and latitude coordinates of the shipment's departure point. Receipt Point Coordinates: The longitude and latitude coordinates of the order's delivery destination. Delivery Distance: The shortest distance between the shipping point and the receipt point. Cargo Weight: Measured in kg, used to determine whether the drone can carry the order. Cargo Volume: Measured in m³, used to calculate drone cargo hold occupancy.

[0041] Expected delivery time: The latest delivery time requested by the customer, which affects the scheduling priority. Delivery time requirement: The urgency of the order is determined by subtracting the order creation time from the expected delivery time.

[0042] Drone: Used to uniquely identify each drone. Drone Current Location: The drone's current GPS coordinates. Battery Status: % used to calculate delivery feasibility. Maximum Load Capacity: The drone's maximum load capacity (kg) to avoid overloading. Range: km to ensure that the delivery can be completed within the battery's capacity.

[0043] After acquiring the data, preprocessing is required to ensure data integrity, remove outliers, and format the input data so that it is suitable for subsequent algorithms.

[0044] Standardize the format of order information, such as using JSON or SQL database storage. Use the WGS84 coordinate system for longitude and latitude to ensure the accuracy of GPS calculations.

[0045] Geographic coordinate anomaly detection: Filter out orders that fall outside the delivery area, such as those whose longitude and latitude are outside the operational range. Use the DBSCAN algorithm to detect outliers and identify unusual coordinates. If the order's destination is outside the drone's flight range, it will be marked as "requires manual review" or "requires transit delivery." If order data is missing (e.g., weight = NULL), the average value is imputed or the order is rejected. If the expected delivery time for an order is earlier than the current time (i.e., it has expired), the order will be discarded.

[0046] Checks whether all orders contain required fields, automatically completing or removing any missing fields. Uses anomaly detection based on the Mahalanobis distance to identify possible anomalous order data; if the Mahalanobis distance exceeds the 99% confidence interval, the order data is considered anomalous.

[0047] Order data is stored in a cloud database (such as AWS DynamoDB) and synchronized to edge computing devices to reduce scheduling delays. The real-time status of drones is stored in a Kafka or MQTT server to ensure that the latest drone data is available when order matching is performed.

[0048] After order data preprocessing ( S100 ) is completed, the order data to be delivered is extracted from the database. This data primarily includes the order number, order shipping point coordinates, order pickup point coordinates, order weight, and drone mission requirements. This data is used as input for the incremental K-means clustering algorithm.

[0049] Traditional K-means requires traversing all data points to calculate the Euclidean distance to the cluster center when calculating the cluster center. This results in a time complexity of H(n×k×d) (n = number of orders, k = number of clusters, d = dimension). This computational time is prohibitively long for large orders. Therefore, this paper uses a KD-Tree index structure to optimize the nearest neighbor search and reduce computation time.

[0050] The KD-Tree construction process includes:

[0051] Use the order location information as the input data of KD-Tree;

[0052] Use the bisection hyperplane partitioning method to construct the KD-Tree: select the dimension with the largest variance in latitude or longitude as the partition axis, use the median split method to perform binary partitioning on the data to reduce the search depth, and recursively construct the KD-Tree to ultimately form an efficient index structure;

[0053] After the KD-Tree is established, the time complexity of the nearest neighbor search is reduced to H(log n). H(log n) represents the growth rate of the algorithm execution time as the amount of input data increases, that is, the efficiency of the KD-Tree nearest neighbor search.

[0054] The present invention adopts a dynamic K value optimization strategy, combined with the elbow method and Gap Statistic to determine the optimal number of clusters. Gap Statistic is a gap statistic, which is a method to measure the quality of clustering results.

[0055] Calculate the intra-cluster sum of squares under different K values , the expression is: ; Where K is the number of clusters, that is, the number of different areas the orders are divided into, is the set of data points of the vth cluster (order cluster), including the coordinates of all orders assigned to the cluster, x is the coordinate of an order belonging to the vth cluster, usually a two-dimensional or three-dimensional geographic coordinate, is the center point of the vth cluster (mean coordinate), and the K value that makes the WCSS decrease gradually is selected to avoid over-clustering or under-clustering.

[0056] Based on the traditional K-means, this paper adopts incremental K-means so that there is no need to recalculate all clusters when orders change dynamically. The specific steps are as follows:

[0057] Initial clustering: Execute standard K-means to calculate the initial cluster centers based on historical order data.

[0058] Incremental clustering of new orders: When a new order is added, a nearest neighbor search is performed using KD-Tree to quickly match the new order to the nearest order cluster. If the distance between the new order and the nearest cluster center is less than the threshold dthresh, the new order is directly added to the cluster. If the distance is greater than the threshold, a new order cluster is created and the K value is updated.

[0059] Dynamic adjustment of cluster centers: cluster centers are updated using the sliding mean update method: ; Where N is the original order number, is the coordinate of the new order. This avoids recalculating the entire cluster due to new orders and improves clustering efficiency. is the cluster center before updating, is the updated cluster center. After the clustering calculation is completed, the system outputs the order cluster set O for subsequent drone scheduling.

[0060] After the order area grouping is completed (S200), the order cluster set O and the available drone status are extracted from the database, including: the data of the order cluster set O; the cluster number: a unique identifier for each order cluster; the order set: a list of order numbers belonging to the cluster; the cluster center coordinates: the average geographical location of the cluster orders; the cluster total weight: the total weight of all orders in the cluster; the cluster estimated delivery time: the optimal delivery time calculated based on the timeliness of the orders.

[0061] Drone data: Drone ID: uniquely identifies each drone; Current Location: the drone's current coordinates; Maximum Payload: the maximum weight the drone can carry; Current Battery: the remaining battery level (%); Maximum Range: the maximum distance the drone can fly in a single flight. This data is used as input by the task matching algorithm to calculate the optimal allocation.

[0062] This paper uses mixed integer programming to perform optimal task allocation, with the goal of minimizing the total delivery cost C while ensuring that all order clusters can be reasonably assigned to suitable drones. Objective function: Minimize delivery cost: ;in: If drone i is assigned to order cluster j, it takes the value 1, otherwise it takes the value 0; Represents the distance that drone i flies to order cluster j; N is the total number of available drones; M is the total number of order clusters. Load constraint (drone load capacity cannot be overloaded): ensure that the total weight of the assigned order clusters does not exceed the maximum load of the drone. Endurance constraint (drone power is sufficient to complete the delivery task): ensure that the drone can complete the delivery and return safely under the current power. Order uniqueness constraint (each order cluster can only be assigned to one drone): any order cluster can only be handled by one drone. Task balance constraint (try to evenly distribute tasks among all drones): to reduce the situation where some drones are overloaded while other drones are idle.

[0063] The present invention uses the Hungarian algorithm combined with MIP to perform task matching and generate a task allocation matrix M. Calculate the task cost matrix of the drone-order cluster:

[0064] First, calculate the mission cost value of drone i to order cluster j , and construct the task cost matrix , task cost value The calculation expression is: ; Among them: α is the weight of path length (optimizing flight cost); β is the weight of cargo weight (optimizing load balancing); γ is the power weight (avoiding low-power drones from being assigned long-distance missions). is the total weight of the order cluster. For example, if an order cluster set O contains 3 orders with weights of 2kg, 3kg, and 1kg respectively, then: . The remaining battery power of the drone is in the range of [0,1], that is: Indicates full charge (100%); Indicates that the power level is 50%; Indicates that only 20% of the battery remains.

[0065] Use the Hungarian algorithm to find the optimal matching to minimize the total task cost;

[0066] From the cost matrix Find the minimum value and match it;

[0067] Iteratively update the unmatched order clusters until all tasks are assigned;

[0068] Generate the final task allocation matrix M.

[0069] To further optimize task allocation, this invention provides a dynamic adjustment mechanism to cope with real-time changes in orders and drone status. When a new order is received, incremental K-means is used to reallocate the order to the nearest order cluster, and drone task assignments are recalculated to minimize adjustments. When a drone fails a task (e.g., due to low battery or malfunction), a neighboring drone takes over the task, finding the nearest available drone to reassign the task. Task balancing optimization: The task load of each drone is calculated and local task swaps are performed to achieve a more balanced load.

[0070] After task matching is complete in S300, the task allocation matrix M is obtained. This matrix describes the optimal allocation relationship between drones and order clusters. Furthermore, the initial state information of the drone is required: current location, maximum range, remaining battery charge, current wind speed, altitude limit, and other weather information.

[0071] For each drone i, the order cluster is assigned , n is the total number of drones, and it is necessary to determine the optimal delivery sequence, that is, the shortest path, so that the delivery time is shortest and the energy consumption is lowest.

[0072] Construct a traveling salesman problem (TSP) model. Suppose the drone starts from Pdrone, Pdrone is the current starting position, and needs to visit the order clusters in sequence. , and eventually return to the starting point or the nearest charging station.

[0073] The objective function is defined as minimizing the total delivery path: , the calculation expression is: ;in, Order cluster and The Euclidean distance between them.

[0074] Use 2-opt to optimize the TSP path: calculate the initial path Rinit (which can be constructed using a greedy algorithm).

[0075] Use 2-opt algorithm for local optimization: randomly select two points in the path , swap the order, calculate the length of the new path; if the new path is shorter, replace the original path; iterate repeatedly until the path can no longer be optimized.

[0076] After optimization, the initial flight route R of the UAV is obtained, which includes the visiting sequence and navigation mode of the delivery points.

[0077] When drones are performing delivery tasks, they need to dynamically adjust the path based on real-time environmental factors to ensure safe and efficient delivery.

[0078] Wind speed effect: If the wind speed is higher than the safety threshold vmax, an alternative route with lower wind speed will be automatically selected. Calculate the navigation deviation of the drone under the influence of wind speed v ;in is the normal flight speed of the drone, For the planned voyage. If it exceeds the safety range, the path is recalculated.

[0079] Rain and Thunderstorm Avoidance: Real-time weather radar data is collected. If the route passes through a thunderstorm area, a detour is automatically recalculated. The A* search algorithm is used to plan a new route within a safe area. If the drone's route passes through a restricted area (such as an airport or military area), the system automatically recalculates the route and adheres to flight regulations for controlled airspace. Dijkstra's shortest path algorithm is used to regenerate a feasible route, ensuring the drone's legal flight.

[0080] Order liquidity data is mainly used to measure the efficiency of order flow from warehousing, sorting, distribution, and delivery. To achieve accurate monitoring, this paper uses a method that combines drone sensors with a cloud data analysis system. The specific data collected is as follows:

[0081] Order Status Tags: Pending: The order has not yet been loaded onto the drone. In Transit: The order has been loaded onto the drone and is being delivered. Delivered: The order has been successfully delivered. Failed: The order could not be delivered due to weather conditions, no-fly zones, insufficient battery life, etc.

[0082] Order flow time data: The time the order is picked up from the warehouse, the time the drone leaves the distribution center, the time the drone arrives at the destination, the time the order is delivered, and the total time from loading to delivery. GPS trajectory data: The drone's flight path ensures that the order's flow location can be tracked. RFID / NFC identification: RFID / NFC is used to read the shipment status to confirm whether the order has been delivered.

[0083] Calculate the order liquidity index, the expression is: ; If OFI>1, it means that the order delivery is delayed; if OFI<1, it means that the delivery efficiency is high.

[0084] The task load balance of drones is used to measure whether the task allocation of each drone is reasonable, so as to reduce the overload of individual drones and improve the overall scheduling efficiency.

[0085] Drone Flight Load Monitoring: Calculating Battery Drain Rate , the expression is: ;in, The battery capacity at takeoff. If a drone's battery drains too quickly, it may indicate an unreasonable mission allocation.

[0086] Calculating the effective load ratio , the expression is: ;like If it is too low, it means that the drone’s mission load is insufficient, affecting the overall resource utilization.

[0087] Task balance calculation: UAV task allocation matrix (extract data from the task allocation matrix M), calculate the task balance index: ;in: is the number of missions for each UAV, is the standard deviation of the number of tasks, is the average number of tasks. If TBI is close to 1, it means that the tasks are evenly distributed; if it is close to 0, it means that the tasks are unevenly distributed.

[0088] The order liquidity index, battery consumption rate, effective load ratio and task balance index are used as input items of fuzzy logic, and the judgment result of whether there is a delay in the current scheduling is used as the output item of fuzzy logic;

[0089] In drone delivery systems, scheduling delays are often affected by multiple factors, the most important of which include:

[0090] The Order Liquidity Index measures whether the actual delivery speed of an order meets expectations. If the actual delivery time of an order significantly exceeds the estimated delivery time, it indicates poor liquidity, which may lead to scheduling delays. This index is categorized into three fuzzy levels: high (fast order flow), medium (normal order flow), and low (slow order flow).

[0091] Battery consumption rate measures the rate at which a drone's battery drains while performing a mission. Rapid battery consumption may indicate complex flight conditions (such as headwinds) or inappropriate mission assignments (such as overloaded payloads), which can affect delivery efficiency. This metric is categorized into three fuzzy levels: fast (rapid consumption), medium (normal consumption), and slow (stable consumption).

[0092] Effective load ratio measures the ratio of a drone's actual payload to its maximum payload capacity. A payload that is too low (long-term inefficient flight) or too high (close to overloading) can affect delivery efficiency and lead to unstable scheduling. This metric is categorized into three fuzzy levels: high (close to overloading), medium (balanced load), and low (underload).

[0093] The mission balance index measures whether drone missions are evenly distributed. A low mission balance indicates that some drones are overloaded, while others are under-tasked. This situation may cause some orders to be delayed. This index is categorized into three fuzzy levels: high (mission balance), medium (partial imbalance), and low (significant imbalance).

[0094] Whether the current scheduling system is experiencing delays. This output is used to assess whether the current scheduling system is at risk of delays. Results are categorized into three fuzzy levels: High, Medium, and Low. High: Severe delays exist, requiring scheduling optimization. Medium: Some delay exists, but is within an acceptable range. Low: Scheduling is normal and no adjustments are required.

[0095] Fuzzy logic reasoning defines a series of fuzzy rules based on the combination of input items to determine whether there is a delay in the current schedule. For example:

[0096] Rule 1: If the order liquidity index is low (delivery time is too long) and the task balance index is low (dispatching is unbalanced), the risk of scheduling delay is high.

[0097] Rule 2: If the order liquidity index is medium and the battery consumption rate is fast (energy consumption is too high), the scheduling delay risk is medium.

[0098] Rule 3: If the task balance index is high (task allocation is reasonable) and the battery consumption rate is slow (energy consumption is normal), the scheduling delay risk is low.

[0099] Rule 4: If the order liquidity index is low and the payload ratio is high (drone is close to being overloaded), the risk of dispatch delay is high.

[0100] Rule 5: If the order liquidity index is high and the task balance index is high, the scheduling delay risk is low.

[0101] Rule 6: If the battery drain rate is fast and the payload ratio is high, the risk of scheduling delay is medium.

[0102] Rule 7: If the order liquidity index is low and the battery consumption rate is fast, the risk of scheduling delays is high.

[0103] Rule 8: If the order liquidity index is medium and the task balance index is low, the scheduling delay risk is medium.

[0104] These rules output the delay risk level of the current schedule based on different input combinations, thereby assisting in optimizing scheduling decisions.

[0105] The fuzzy logic execution process involves first converting the input data into fuzzy sets. For example: Order Liquidity Index = 0.6 is Medium. Battery Drain Rate = 0.8 is Fast. Payload Ratio = 0.3 is Low. Task Balance Index = 0.7 is Medium.

[0106] The output delay risk corresponding to the input combination is determined through fuzzy inference rules: Rule 2 applies to the current input situation: "If the order liquidity index is medium and the battery consumption rate is fast, the scheduling delay risk is medium."

[0107] Based on the delay risk levels with different weights calculated from multiple rules, the fuzzy results are converted into a specific delay risk score (e.g., a value between 0 and 1). For example, if the calculated delay risk is 0.7, the system may determine that the current schedule has a certain delay risk and that task allocation needs to be adjusted.

[0108] If the delay risk is high (High): Reassign tasks, reduce the delivery load of high-load drones, or prioritize drones with sufficient battery power. Update the order cluster set O and adjust the task allocation matrix M. Return to S200 to re-optimize the schedule.

[0109] If the delay risk is medium: Make some mission adjustments, such as optimizing path planning and reducing the mission burden on overloaded drones.

[0110] If the latency risk is low, maintain the current scheduling policy without adjustment.

[0111] For situations where the risk of delay is medium or low, continue to complete the delivery task.

[0112] Example 2, please refer to Figure 2 As shown, the drone cargo delivery management system based on policy optimization described in this embodiment includes an order management module, an order partitioning and clustering module, a task allocation and drone matching module, a drone path adjustment module, and a scheduling optimization module;

[0113] The order management module obtains basic information about the order to be delivered, including the order's geographic location, cargo weight, expected delivery time, and the drone's load capacity;

[0114] The order partition clustering module uses an improved incremental K-means clustering algorithm to group orders by region based on their geographic location information. The KD-Tree index structure is used to optimize the nearest neighbor search and generate an order cluster set O.

[0115] The task allocation and drone matching module matches tasks based on the order cluster set O and the load capacity of the drone, and calculates the optimal allocation plan between order clusters and drones to obtain the task allocation matrix M;

[0116] The UAV path adjustment module optimizes the delivery path of the UAV based on the task allocation matrix M, generates the initial flight path R, and dynamically adjusts the path based on real-time weather and airspace restrictions;

[0117] The scheduling optimization module executes the delivery task and collects order liquidity data and drone task load balance data through the drone's sensors to determine whether there is a delay in the current scheduling. If so, it updates the order cluster set O and adjusts the task allocation matrix M, returning to the order partitioning and clustering module to re-optimize the scheduling; otherwise, the delivery task is completed.

[0118] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. A drone cargo delivery management method based on policy optimization, characterized by: include: S100, obtaining basic information of the order to be delivered, including the order's geographic location, cargo weight, expected delivery time, and drone's load capacity; S200, based on the geographic location information of the orders, uses an improved incremental K-means clustering algorithm to group the orders by region, wherein a KD-Tree index structure is used to optimize the nearest neighbor search and generate a set of order clusters O. Specifically, the process includes: using the geographic location information of the orders as the input data of the KD-Tree; constructing the KD-Tree using a bisection hyperplane partitioning method: selecting the dimension with the largest variance in latitude or longitude as the partitioning axis, performing a binary partitioning on the data using the median partitioning method, and recursively constructing a KD-Tree to ultimately form an efficient index structure; after the KD-Tree is established, the time complexity of the nearest neighbor search is reduced to H(log n); H(log n) represents the rate of increase in the algorithm execution time as the amount of input data increases, i.e., the efficiency of the KD-Tree nearest neighbor search; S300, based on the order cluster set O, combined with the load capacity of the UAV, the task matching is performed, and the optimal allocation scheme between the order cluster and the UAV is calculated to obtain the task allocation matrix M; specifically: calculate the task cost matrix of UAV i to order cluster j ; Among them: α is the weight of the path length, β is the weight of the cargo weight, γ is the weight of the electricity, is the total weight of the order cluster, The remaining battery power of the drone is in the range of [0,1]. Denotes the distance between drone i and order cluster j. The Hungarian algorithm is used to find the optimal match to minimize the total mission cost: Find the minimum value and match it; iteratively update the unmatched order clusters until all tasks are assigned; generate the final task assignment matrix M; S400 optimizes the delivery path of the UAV based on the task allocation matrix M, generates the initial flight path R, and dynamically adjusts the path based on real-time weather and airspace restrictions; S500: Execute the delivery task and collect order liquidity data and drone task load balance data through the drone's sensors to determine whether there is a delay in the current scheduling. If so, update the order cluster set O and adjust the task allocation matrix M, then return to S200 to re-optimize the scheduling; otherwise, complete the delivery task; The S500 monitors order liquidity in real time through the drone’s GPS, IMU sensors, and RFID / NFC identification devices, and calculates the order liquidity index, as well as the battery consumption rate, payload ratio, and mission balance index; The order liquidity index, battery consumption rate, effective load ratio and task balance index are used as input items of fuzzy logic, and the judgment result of whether there is a delay in the current scheduling is used as the output item of fuzzy logic; Formulate fuzzy rules to describe the impact of order liquidity index, battery consumption rate, effective load ratio and task balance index on the current scheduling delay; Determine the risk of current scheduling delay based on the fuzzy output results and take corresponding measures.

2. The UAV cargo delivery management method based on policy optimization according to claim 1 is characterized by: Use dynamic K value optimization strategy to determine the optimal number of clusters: calculate the intra-cluster sum of squares under different K values , the expression is: ; Where K is the number of clusters, is the set of data points of the vth cluster, x is the coordinate of an order belonging to the vth cluster, The center point of the vth cluster is calculated by performing standard K-means based on historical order data. When a new order is added, the nearest neighbor search is performed using KD-Tree to quickly match the nearest order cluster. If the distance between the new order and the nearest cluster center is less than the threshold dthresh, it is directly added to the cluster; if the distance is greater than the threshold, a new order cluster is created and the K value is updated; the cluster center is updated using the sliding mean update method: ; Where N is the original order number, is the new order coordinate, is the cluster center before updating, is the updated cluster center.

3. The UAV cargo delivery management method based on policy optimization according to claim 1 is characterized by: The S400 optimizes the delivery path, including: for each drone i, the order cluster is assigned , n is the total number of drones, and a traveling salesman problem model is constructed. Suppose the drone starts from the starting point Pdrone and needs to visit the order clusters in sequence , and finally return to the starting point or the nearest charging station; define the objective function to minimize the total delivery path : ;in, Order cluster and The Euclidean distance between them; calculate the initial path Rinit, randomly select two points in the path , swap the order, calculate the length of the new path; if the new path is shorter, replace the original path; iterate repeatedly until the path can no longer be optimized.

4. A drone cargo delivery management system based on policy optimization, used to implement the drone cargo delivery management method based on policy optimization according to any one of claims 1 to 3, characterized in that: It includes order management module, order partitioning and clustering module, task allocation and drone matching module, drone path adjustment module and scheduling optimization module; The order management module obtains basic information about the order to be delivered, including the order's geographic location, cargo weight, expected delivery time, and the drone's load capacity; The order partition clustering module uses an improved incremental K-means clustering algorithm to group orders by region based on their geographic location information. The KD-Tree index structure is used to optimize the nearest neighbor search and generate an order cluster set O. The task allocation and drone matching module matches tasks based on the order cluster set O and the load capacity of the drone, and calculates the optimal allocation plan between order clusters and drones to obtain the task allocation matrix M; The UAV path adjustment module optimizes the delivery path of the UAV based on the task allocation matrix M, generates the initial flight path R, and dynamically adjusts the path based on real-time weather and airspace restrictions; The scheduling optimization module executes the delivery task and collects order liquidity data and drone task load balance data through the drone's sensors to determine whether there is a delay in the current scheduling. If so, it updates the order cluster set O and adjusts the task allocation matrix M, returning to the order partitioning and clustering module to re-optimize the scheduling; otherwise, the delivery task is completed.

Images