Intelligent site selection method and system for drone take-off and landing points for low-altitude logistics

Abstract

The present invention discloses a method and system for intelligent site selection of drone take-off and landing points for low-altitude logistics, which relates to the field of logistics management technology. The method comprises: collecting logistics order data, airspace status data and environmental perception data in real time through edge computing terminals, and normalizing the collected data; constructing a five-dimensional real-time state vector including demand intensity factor, drone cluster status, airspace compliance index, terrain resistance coefficient and cost constraint vector based on the processed data; constructing a reinforcement learning decision engine, designing an action space including node operation, task allocation and resource allocation, and filtering the action space based on airspace rules and physical constraints. The present invention significantly improves the real-time response capability of low-altitude logistics, enhances adaptability to complex terrain, optimizes multi-objective decision-making, and ensures the robustness and compliance of the system, providing a reliable solution for the intelligent management of low-altitude logistics.

Description

Technical Field

[0001] The present invention relates to the field of logistics management technology, and in particular to a method and system for intelligent site selection of drone take-off and landing points for low-altitude logistics. Background Art

[0002] With the rapid development of the low-altitude logistics industry, drones, with their efficiency and flexibility, are increasingly being used in areas such as urban express delivery and medical supply transportation. However, significant shortcomings remain in the critical process of selecting drone takeoff and landing sites.

[0003] When responding to sudden surges in low-altitude logistics demand (such as a surge in medical supplies or e-commerce promotional orders), many rely on static clustering analysis methods (such as DBSCAN) based on historical data. This approach results in significant lags in takeoff and landing point adjustments, with response times exceeding four hours, making it difficult to meet the "small-batch, high-frequency, and immediacy" requirements of drone logistics. Furthermore, existing technologies lack effective processing of real-time dynamic data, unable to quickly perceive and adapt to dynamic factors such as changes in logistics demand, airspace status adjustments, and environmental fluctuations. This makes it difficult for site selection to guarantee drone flight safety and logistics efficiency, limiting the further development and large-scale application of the low-altitude logistics industry. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for intelligent site selection of drone take-off and landing points for low-altitude logistics, so as to solve the problems raised in the above background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent site selection of drone take-off and landing points for low-altitude logistics, comprising the following steps:

[0006] Use edge computing terminals to collect logistics order data, airspace status data, and environmental perception data in real time, and normalize the collected data;

[0007] Based on the processed data, a five-dimensional real-time state vector is constructed, which includes the demand intensity factor, the UAV cluster status, the airspace compliance index, the terrain resistance coefficient, and the cost constraint vector.

[0008] Build a reinforcement learning decision engine, design an action space that includes node operations, task allocation, and resource allocation, and filter the action space based on spatial rules and physical constraints;

[0009] Construct a multi-objective reward function that includes timeliness rewards, cost rewards, safety rewards, and compliance rewards, and achieve balanced optimization of multiple objectives through dynamic weight adjustment;

[0010] Use proximal policy optimization algorithms to train deep reinforcement learning models, and combine edge computing with federated learning mechanisms for model deployment.

[0011] Based on the real-time state vector, a site selection plan for the drone take-off and landing points is generated through the reinforcement learning decision engine;

[0012] The site selection plan is output to the drone scheduling system and airspace management system, and the model parameters are optimized through a closed-loop feedback mechanism.

[0013] Preferably, the five-dimensional real-time state vector including the demand intensity factor, the UAV cluster status, the airspace compliance index, the terrain resistance coefficient and the cost constraint vector is constructed based on the processed data, including:

[0014] Calculate the demand intensity factor based on the exponentially weighted moving average, and determine the emergency order weighting factor based on the order type;

[0015] Calculate the drone cluster state vector, which includes the normalized remaining battery power of the drones, the drone cluster load rate, and the number of faulty drones;

[0016] Calculate the airspace compliance index of the candidate area based on the dynamic no-fly zone data provided by the airspace management system;

[0017] Calculate the terrain slope between take-off and landing points based on digital elevation data, and determine the terrain drag coefficient based on the flight duration;

[0018] The cost constraint vector is obtained through Bayesian network training, which includes normalized construction cost, maintenance cost and noise complaint risk value.

[0019] Preferably, the construction of the reinforcement learning decision engine, designing an action space including node operations, task allocation, and resource deployment, and filtering the action space based on spatial rules and physical constraints, includes:

[0020] Define three types of atomic action combinations: node operation, task allocation, and resource allocation;

[0021] Generate temporary take-off and landing points in the compliance area based on the order heat map, and perform perturbation processing on the coordinates of the temporary take-off and landing points;

[0022] The Hungarian algorithm is used to solve the matching problem between missions and take-off and landing points. The objective function includes flight cost and take-off and landing point startup cost.

[0023] Invalid actions that violate airspace rules or physical constraints are blocked in real time.

[0024] Preferably, the multi-objective reward function comprising timeliness reward, cost reward, safety reward and compliance reward is constructed, and the balanced optimization of the multiple objectives is achieved through dynamic weight adjustment, including:

[0025] Build time-efficiency rewards based on decision time and order type;

[0026] Build cost incentives based on flight costs, landing point activation costs, and dynamic cost thresholds;

[0027] Building safety incentives based on obstacle density in a spatial system;

[0028] Establishing compliance incentives based on the approval results of the airspace management system;

[0029] The weights of the multi-objective reward function are adjusted according to the preset time period and cluster load status.

[0030] Preferably, the proximal policy optimization algorithm is used to train the deep reinforcement learning model, and the model is deployed in combination with edge computing and federated learning mechanisms, specifically including:

[0031] Constructing a neural network structure including an input layer, a hidden layer, and an output layer, wherein the hidden layer adopts an activation function and sets an anti-overfitting mechanism;

[0032] Conduct online learning on edge computing terminals and update model parameters using the federated learning mechanism.

[0033] The locally collected order data is processed for privacy protection before participating in model training.

[0034] Preferably, the method further comprises the following steps:

[0035] When abnormal changes in order volume are detected, the emergency response strategy is triggered to adjust the action space and reward function calculation logic;

[0036] When the model output is abnormal, a backup site selection plan is generated through the historical case matching mechanism.

[0037] Preferably, the edge computing terminal integrates a meteorological sensor and a high-precision positioning module for collecting wind speed, wind direction and take-off and landing point location data;

[0038] The airspace management system provides dynamic no-fly zone data;

[0039] The data normalization process includes format conversion and standardization of data of different protocols.

[0040] The present invention also provides an intelligent site selection system for drone take-off and landing points for low-altitude logistics, comprising:

[0041] The data collection module is used to collect logistics order data, airspace status data, and environmental perception data in real time through edge computing terminals, and normalize the collected data;

[0042] A decision engine module, which is used to construct a five-dimensional real-time state vector containing the demand intensity factor, drone cluster status, airspace compliance index, terrain resistance coefficient, and cost constraint vector;

[0043] The decision engine module is also used to design an action space including node operations, task allocation, and resource allocation, and filter the action space based on spatial rules and physical constraints;

[0044] The decision engine module is also used to construct a multi-objective reward function that includes timeliness rewards, cost rewards, safety rewards, and compliance rewards, and achieves balanced optimization of multiple objectives through dynamic weight adjustment;

[0045] The decision engine module is also used to train deep reinforcement learning models using a proximal policy optimization algorithm and deploy the models in combination with edge computing and federated learning mechanisms.

[0046] The decision engine module is also used to generate a site selection plan for the drone’s take-off and landing points based on the real-time state vector;

[0047] An execution feedback module is used to output the site selection plan to the drone scheduling system and airspace management system, and optimize the model parameters through a closed-loop feedback mechanism.

[0048] The present invention further provides an electronic device, wherein the electronic device is a physical device and comprises:

[0049] a processor and a memory, wherein the memory is communicatively connected to the processor;

[0050] The memory is used to store at least one executable instruction executed by the processor, and the processor is used to execute the executable instruction to implement the above-mentioned intelligent site selection method for drone take-off and landing points for low-altitude logistics.

[0051] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the above-mentioned intelligent site selection method for drone take-off and landing points for low-altitude logistics.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] By building a dynamic data acquisition layer to obtain multi-source information in real time, and combining it with a reinforcement learning decision engine for intelligent analysis, efficient site selection for drone take-off and landing points is achieved. By defining a state space with multiple dimensions such as demand intensity and drone cluster status, the complex scenarios of low-altitude logistics are accurately portrayed; designing an efficient action space and combining it with an action mask mechanism improves the accuracy and efficiency of decision-making; constructing a multi-objective reward function and dynamically adjusting the weights to balance multiple factors such as timeliness, cost, and safety; at the same time, adopting a lightweight training and execution process to ensure the stable operation of the system in different scenarios; the present invention significantly improves the real-time response capability of low-altitude logistics, enhances adaptability to complex terrain, optimizes multi-objective decision-making, and ensures the robustness and compliance of the system, providing a reliable solution for the intelligent management of low-altitude logistics. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a main flow chart of a method for intelligent site selection of drone take-off and landing points for low-altitude logistics provided by an embodiment of the present invention;

[0055] Figure 2 A schematic diagram of the structure of an intelligent site selection system for drone take-off and landing points for low-altitude logistics provided by an embodiment of the present invention;

[0056] Figure 3 A schematic structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0057] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0058] The execution subject of the method of this embodiment is a terminal, which can be a mobile phone, tablet computer, PDA, notebook or desktop computer, etc. Of course, it can also be other devices with similar functions, and this embodiment is not limited.

[0059] See also Figure 1 The present invention provides a method for intelligent site selection of take-off and landing points for UAVs for low-altitude logistics, comprising:

[0060] Step 100: Collect logistics order data, airspace status data, and environmental perception data in real time through the edge computing terminal, and normalize the collected data.

[0061] In this embodiment, the edge computing terminal integrates a meteorological sensor and a high-precision positioning module to collect wind speed, wind direction, and take-off and landing point location data;

[0062] The airspace management system provides dynamic no-fly zone data;

[0063] The data normalization process includes format conversion and standardization of data of different protocols.

[0064] Specifically, high-performance edge computing terminals such as the Huawei Atlas 500 are deployed in low-altitude logistics systems. These terminals boast powerful real-time data collection capabilities, enabling comprehensive collection of detailed logistics order information, including order quantity, destination, weight, and volume. They also accurately capture airspace status data, such as available airspace at different altitudes, flight-restricted areas, and the distribution of other aircraft within the current airspace. Various sensors provide precise environmental perception, encompassing meteorological data such as wind speed, direction, temperature, humidity, and visibility, as well as geographic information such as topography and building distribution. The collected multi-source data is transmitted and normalized via the efficient Apache Kafka data bus. Apache Kafka, with its superior performance, ensures a stable data throughput of ≥100,000 records per second, ensuring rapid and orderly data flow and providing a solid foundation for subsequent data analysis and decision-making. As a data bus, Apache Kafka can achieve data normalization processing through its ecosystem components and extension solutions. For example, by integrating with Apache Flink or Spark Streaming, its richer transformation operators (such as window aggregation and regular expression parsing) can be used to implement complex normalization logic.

[0065] Step 200 , based on the processed data, construct a five-dimensional real-time state vector S(t)=[S1, S2, S3, S4, S5] including the demand intensity factor, the UAV cluster status, the airspace compliance index, the terrain resistance coefficient and the cost constraint vector.

[0066] Specifically, step 200 includes:

[0067] Step 210, demand intensity factor (S1):

[0068] Calculate the demand intensity factor based on exponentially weighted moving average ,determine the urgent order weighting factor in combination with the order type;

[0069] in, This is the current 5-minute order volume. This data is obtained in real time using the API provided by the logistics platform. To ensure the timeliness and accuracy of the data, the acquisition frequency is set to ≥ 100 times / second. For example, during peak logistics business hours, order data is frequently updated every second, allowing the system to grasp the latest order dynamics in real time.

[0070] is the historical mean calculated based on the exponentially weighted moving average (EWMA), where the weighting coefficient A value of 0.3 is used. This calculation method can give greater weight to the influence of recent data. When analyzing the trend of order volume changes over time, it allows the system to more sensitively capture recent fluctuations in order volume, rather than simply averaging historical data.

[0071] This is a weighting factor for emergency orders, assigned differently depending on the type of goods involved in the order. For medical supply orders, as they involve urgent needs like life-saving, a value of 1.5 is assigned; for fresh cold chain orders, given the timeliness and freshness requirements of the goods, a value of 1.2 is assigned; and for general orders, as they do not require special timeliness, a value of 1 is assigned.

[0072] Step 220, UAV cluster status (S2):

[0073] Calculate the state vector of the drone cluster , including the normalized remaining battery power of the drone, the load rate of the drone cluster and the number of faulty drones;

[0074] in, is the normalized remaining battery power of the drone, is the UAV cluster load rate, is the number of faulty drones;

[0075] Normalized power value It is used to reflect the relative position of the current power of the drone within its power range. The calculation formula is: ,here, Represents a critical power level of 20%. When the drone's power level drops to this value, it is necessary to consider landing and charging immediately to ensure safe flight. It corresponds to 100% full power state;

[0076] Load ratio Used to measure the overall load of the drone cluster. The calculation formula is: , where Indicates the i The real-time payload of the drone, For the i The maximum payload of a drone, nRepresents the number of drones in the entire swarm. By accumulating and averaging the load ratios of each drone, it can be used to intuitively reflect whether the swarm load is close to saturation.

[0077] also, Represents the number of faulty drones. This value is used to evaluate the health status of the drone cluster. The number of faulty drones directly affects the stability and reliability of cluster mission execution.

[0078] Step 230, airspace compliance index (S3):

[0079] Calculate the airspace compliance index of the candidate area based on the dynamic no-fly zone data provided by the airspace management system ;

[0080] in, Represents the effective area of the candidate area, that is, the area that complies with specific airspace usage rules and can be safely used for drone takeoff and landing operations. is the total area of the candidate region, covering the entire scope of the evaluated area. It is a raster mask matrix converted from the UTM system. The matrix is constructed based on various airspace restrictions (such as no-fly zones, altitude restrictions, etc.). The value of the corresponding airspace restriction position is 0, and the value of the airspace position that can be used normally is 1. It is used to filter out areas that meet the conditions. It is a regional identification matrix, which is used to clarify the target area to which each grid belongs. By multiplying and adding the corresponding elements of the two, the effective area ratio can be accurately obtained to quantify the degree of airspace compliance.

[0081] Step 240, terrain resistance coefficient (S4):

[0082] Calculate the terrain slope between take-off and landing points based on digital elevation data, and determine the terrain drag coefficient based on the flight duration;

[0083] Specifically, there are two take-off and landing points: and The calculation formula of its average slope is ,in and Respectively represent and Altitude, is the distance between two points on the horizontal plane;

[0084] Taking into account the particularity of the night flight environment, compared with daytime, night flight will have additional resistance. After actual testing and research analysis, in this embodiment, when in the night flight state (i.e. =1), an additional 10% drag will be added to more accurately simulate and consider factors such as flight energy consumption, and the corresponding drag correction factor The calculation formula is ;

[0085] The above formula can fully consider the impact of terrain slope on drone flight, and make reasonable corrections for the additional resistance of night flight, thereby providing more accurate data support for drone flight path planning and energy consumption estimation. It ensures that under different terrain and time conditions, the correlation and differences between take-off and landing points can be scientifically evaluated, laying a solid foundation for subsequent decision-making.

[0086] Step 250, cost constraint vector (S5):

[0087] The cost constraint vector is obtained through Bayesian network training, which includes normalized construction cost, maintenance cost and noise complaint risk value;

[0088] During the training process, population density data is taken into consideration. This data reflects the density of human distribution in the area and is of great significance to the potential impact around drone landing and take-off points. Land rent data is also input, which is related to the land cost investment required for the construction of the landing and take-off points. After complex and precise calculations by the Bayesian network, the final output is normalized construction cost, maintenance cost, and noise complaint risk value. These values are all limited to the range of [0,1] to facilitate subsequent comprehensive analysis and application.

[0089] Step 300: Build a reinforcement learning decision engine, design an action space including node operation, task allocation, and resource allocation, and filter the action space based on spatial rules and physical constraints.

[0090] Specifically, step 300 includes:

[0091] Step 310 defines three types of atomic action combinations: node operation, task allocation, and resource allocation. 32 composite strategies are supported, and invalid actions are filtered in real time through the action mask mechanism:

[0092] Node operations cover the activation and deactivation of existing take-off and landing points, as well as the creation of new temporary ones. Task allocation aims to rationally distribute orders to various take-off and landing points for efficient delivery. Resource allocation involves the rational arrangement of drone cluster resources, such as power and payload. These three types of atomic actions, through different combinations, form 32 composite strategies that can flexibly respond to diverse low-altitude logistics scenarios. The action masking mechanism acts like an intelligent filter. During real-time decision-making, it quickly screens out invalid actions that do not conform to the actual situation based on the current system status, greatly improving decision-making efficiency and accuracy.

[0093] Step 320: Generate temporary take-off and landing points in the compliance area based on the order heat map, and perform perturbation processing on the coordinates of the temporary take-off and landing points;

[0094] To more accurately reflect the actual demand distribution, we conducted operations near the peak point where the temperature exceeded 40°C on the order heat map based on the Delaunay triangulation algorithm. Taking into account the rationality and flexibility of site selection, we limited the scope to an area with a radius of 500m around the peak point, and this area must meet the compliance indicators. , this compliance area can ensure that the selected node has a good basis for real-world applications. In this compliance area, a virtual node is generated, and its coordinates are obtained by the peak point coordinates. , that is, the virtual node coordinates are ,in and The values are randomly selected within a range of ±200m. The introduction of this random perturbation avoids the clustering problem that may arise from selecting locations based entirely on peak points, helps to more evenly cover the demand area, and improves the efficiency and rationality of subsequent logistics scheduling.

[0095] Step 330: Use the Hungarian algorithm to solve the matching problem between the mission and the take-off and landing point. The objective function includes the flight cost and the take-off and landing point startup cost. The objective function is:

[0096] ,

[0097] in, Represents the distance from the mission starting point to the take-off and landing point j The comprehensive flight cost is composed of several key factors. The specific formula is:

[0098] ,

[0099] It represents the straight-line flight distance between the mission starting point and the take-off and landing point on the horizontal plane, in kilometers. The coefficient of 0.02 reflects the impact weight of flight distance on cost, that is, the cost increase brought about by each kilometer of flight distance; is the altitude difference between the mission start point and the take-off and landing point, in meters. The coefficient 0.1 reflects the degree of effect of altitude difference on cost, which means the cost change associated with each meter of altitude difference; Represents the load factor of the UAV when performing a mission. It is a proportional value between 0 and 1. The coefficient of 0.05 indicates the impact of the load factor on the cost, that is, the cost change caused by a one-unit change in the load factor. It is specifically used to measure the take-off and landing points j The single startup cost covers the one-time investment costs such as startup energy consumption and equipment loss of the equipment related to the take-off and landing points. As a 0-1 decision variable, when the task i Assigned to take-off and landing points j hour, The value is 1, otherwise it is 0, so as to accurately control the matching relationship between the mission and the take-off and landing points.

[0100] Step 340: Shield invalid actions that violate airspace rules or physical constraints in real time.

[0101] Understandably, drones' maneuvers are screened in real time based on established airspace regulations, such as no-fly zones and restricted altitude zones, as well as physical constraints, including but not limited to topography (mountains, water, and other obstacles) and obstacles (tall buildings, communication towers, etc.). Any maneuvers that violate these regulations are immediately identified as invalid and blocked, ensuring that drones maintain legal and safe flight paths.

[0102] Step 400: Construct a multi-objective reward function that includes time reward, cost reward, safety reward, and compliance reward. And achieve balanced optimization of multiple objectives through dynamic weight adjustment.

[0103] Specifically, step 400 includes:

[0104] Step 410: Create a time-efficiency reward based on decision time and order type ;

[0105] ,

[0106] in, Represents the decision time, which is accurate to milliseconds and is the time interval from the decision start moment to the final determination of the solution. It is an urgency indicator. When a task is judged as urgent, The value is 1, otherwise it is 0. This formula means that the shorter the decision time and the more urgent the task, the higher the timeliness reward.

[0107] Step 420: Build a cost reward based on the flight cost, take-off and landing point startup cost, and dynamic cost threshold. ;

[0108] ,

[0109] Where, It represents the total cost of the drone during flight, including fuel consumption, equipment loss and other expenses directly related to the flight; The cost of setting up the take-off and landing point, including one-time expenses such as site rental and equipment deployment; is the dynamic cost threshold, which is expressed by the formula Calculated, where is the average of past flight and setup costs, is the standard deviation. This formula reflects the relationship between the actual cost and the cost threshold. The lower the cost, the higher the cost reward.

[0110] Step 430: Construct a safety reward based on the obstacle density of the spatial voxels ;

[0111] ,

[0112] This formula is calculated based on 50m³ voxels. The obstacle density within each 50m³ voxel is the proportion of space occupied by obstacles within that voxel. By multiplying all relevant voxels, the lower the obstacle density within a voxel, the higher the safety bonus, which is used to measure the safety of the environment around the take-off and landing point.

[0113] Step 440: Create a compliance reward based on the airspace management system's approval results ;

[0114] Compliance rewards are awarded when the flight plan successfully passes the UTM (Unmanned Aircraft System Traffic Management) approval. The value is 1; if it fails to pass the approval, The value is -0.5. UTM approval mainly considers whether the flight plan complies with a series of regulatory requirements such as airspace management rules and safety standards.

[0115] Step 450 , adjusting the weight of the multi-objective reward function according to a preset time period and cluster load status;

[0116] The system can automatically switch weights according to different time periods (morning peak, off-peak, night) and load rate conditions. For example, during the morning peak period, due to the extremely high demand for logistics timeliness, the system can automatically switch weights according to different time periods (morning peak, off-peak, night) and load rate conditions. At this time, the weight of time-efficiency reward is the largest, highlighting the principle of time-efficiency priority. During off-peak hours, various factors are relatively balanced, and the weight may be adjusted to , , , , during night time, considering flight safety and cost control, the weights will be different, such as , , , ,The load rate will also affect the weight. When the load rate is high, in order to ensure the smooth completion of the task, the weight of the cost reward and safety reward may be appropriately increased.

[0117] Step 500: Use the proximal policy optimization algorithm to train the deep reinforcement learning model, and combine edge computing with the federated learning mechanism to deploy the model.

[0118] Specifically, step 500 includes:

[0119] Step 510: construct a neural network structure including an input layer, a hidden layer, and an output layer, wherein the hidden layer adopts an activation function and sets an anti-overfitting mechanism;

[0120] The PPO algorithm is used, and the neural network consists of a 512-dimensional input layer, two hidden layers with 256 neurons (ReLU activation + dropout = 0.2), and a policy / value network output layer.

[0121] Step 520: Perform online learning on the edge computing terminal and update the model parameters in combination with the federated learning mechanism;

[0122] Perform online learning on the edge computing terminal, set the edge node computing power ≥ 2TOPS, and make the decision delay less than 200ms.

[0123] Among them, in the intelligent site selection process, this embodiment adopts the proximal policy optimization (PPO) algorithm to drive the decision-making process. The constructed neural network input layer is set to 512 dimensions, which can fully absorb various types of complex data information. The middle layer contains two hidden layers with 256 neurons each. Each layer uses the ReLU activation function to enhance the nonlinear expression ability of the model, and is matched with a dropout ratio of 0.2 to effectively prevent overfitting. A policy / value network output layer is provided at the end of the network, which is responsible for outputting the final decision result. In order to ensure the efficient operation of the neural network, clear requirements are put forward for the computing power of the edge node, which must reach a level of ≥2TOPS to ensure a rapid response when processing large-scale data. The decision delay is strictly controlled at <200ms to achieve near real-time decision feedback.

[0124] Step 530: Perform privacy protection processing on the locally collected order data before participating in model training;

[0125] The local node performs differential privacy processing on the order data (ε=0.5), and the central server aggregates the model through the FedAvg algorithm, triggering a parameter update every 500 decisions.

[0126] Understandably, to ensure data privacy and security, this embodiment uses a federated learning mechanism. When processing order data, local nodes introduce differential privacy technology and set the privacy budget parameter ε to 0.5. While ensuring data availability, it minimizes the risk of data leakage. The central server uses the FedAvg algorithm to periodically aggregate the model parameters uploaded by each local node. After every 500 cumulative decisions, a parameter update operation is triggered, prompting the model to be continuously optimized within the entire federated system, thereby improving the accuracy and generalization ability of the global model.

[0127] Step 540: When an abnormal change in order volume is detected, the emergency response strategy is triggered to adjust the action space and reward function calculation logic;

[0128] When the order volume change rate ΔQ is greater than 300%, a Level 3 emergency response is triggered, skipping complex calculations and increasing decision-making speed by another 50%;

[0129] Step 550: When the model output is abnormal, a backup site selection plan is generated through a historical case matching mechanism.

[0130] When the model fails, the historical optimal solution is retrieved as a backup action through cosine similarity Sim>0.8.

[0131] Specifically, when the order demand change rate ΔQ>300%, it means an extreme order peak. At this time, the core area supply guarantee mechanism is immediately triggered. This mechanism will skip the conventional complex calculation process and adopt a more efficient decision-making logic, so that the decision-making speed can be increased by another 50% in this extreme case, fully guaranteeing the supply of materials in the core area.

[0132] In addition, to address model failure scenarios, this embodiment also incorporates a safety margin. Using the cosine similarity algorithm, the system searches for decision solutions in historical data. When the cosine similarity Sim of the search results exceeds 0.8, the historical solution is considered highly similar to the current scenario and the optimal historical solution is directly invoked to maintain basic system operation and avoid business stagnation caused by model anomalies.

[0133] Step 600: Generate a site selection plan for the drone take-off and landing points through the reinforcement learning decision engine based on the real-time state vector.

[0134] Step 700: Output the site selection plan to the drone scheduling system and airspace management system, and optimize the model parameters through a closed-loop feedback mechanism.

[0135] This site selection plan is exported to the drone dispatch system and airspace management system, ensuring that the drone dispatch system can efficiently plan drone flight routes and assign tasks based on the plan, and the airspace management system can use this information to rationally manage airspace resources for low-altitude logistics. Simultaneously, through a closed-loop feedback mechanism, actual drone operation data and system feedback are continuously collected, allowing for dynamic adjustment and optimization of model parameters to improve the model's adaptability and accuracy to different environments and mission requirements.

[0136] In this embodiment, by constructing a dynamic data acquisition layer to obtain multi-source information in real time, and combining it with a reinforcement learning decision engine for intelligent analysis, efficient site selection for drone take-off and landing points is achieved. By defining a state space with multiple dimensions such as demand intensity and drone cluster status, the complex low-altitude logistics scenarios are accurately portrayed; an efficient action space is designed and combined with an action mask mechanism to improve the accuracy and efficiency of decision-making; a multi-objective reward function is constructed and the weights are dynamically adjusted to balance multiple factors such as timeliness, cost, and safety; at the same time, a lightweight training and execution process is adopted to ensure the stable operation of the system in different scenarios; the present invention significantly improves the real-time response capability of low-altitude logistics, enhances adaptability to complex terrain, optimizes multi-objective decision-making, and ensures system robustness and compliance, providing a reliable solution for the intelligent management of low-altitude logistics.

[0137] Based on the above embodiments, Figure 2 As shown, the present invention also provides an intelligent site selection system for drone take-off and landing points for low-altitude logistics, which is used to support the intelligent site selection method for drone take-off and landing points for low-altitude logistics in the above embodiment. The intelligent site selection system for drone take-off and landing points for low-altitude logistics includes:

[0138] The data collection module 11 is used to collect logistics order data, airspace status data, and environmental perception data in real time through the edge computing terminal, and normalize the collected data;

[0139] A decision engine module 12 is used to construct a five-dimensional real-time state vector including a demand intensity factor, a drone cluster state, an airspace compliance index, a terrain resistance coefficient, and a cost constraint vector;

[0140] The decision engine module 12 is further configured to design an action space including node operations, task allocation, and resource deployment, and filter the action space based on spatial rules and physical constraints;

[0141] The decision engine module 12 is also used to construct a multi-objective reward function including timeliness reward, cost reward, safety reward and compliance reward, and achieve balanced optimization of multiple objectives through dynamic weight adjustment;

[0142] The decision engine module 12 is also used to train the deep reinforcement learning model using a proximal policy optimization algorithm and deploy the model in combination with edge computing and federated learning mechanisms;

[0143] The decision engine module 12 is further used to generate a site selection plan for the drone take-off and landing points based on the real-time state vector;

[0144] The execution feedback module 13 is used to output the site selection plan to the drone scheduling system and the airspace management system, and optimize the model parameters through a closed-loop feedback mechanism.

[0145] In an optional embodiment, the decision engine module 12 is further configured to:

[0146] The demand intensity factor is calculated based on the exponentially weighted moving average, and the emergency order weighting factor is determined in combination with the order type; the drone cluster state vector is calculated, including the normalized remaining drone power, drone cluster load rate and the number of faulty drones; the airspace compliance index of the candidate area is calculated based on the dynamic no-fly zone data provided by the airspace management system; the terrain slope between the take-off and landing points is calculated based on the digital elevation data, and the terrain resistance coefficient is determined in combination with the flight period; and the cost constraint vector is obtained through Bayesian network training, including the normalized construction cost, maintenance cost and noise complaint risk value.

[0147] In an optional embodiment, the decision engine module 12 is further configured to:

[0148] Define three types of atomic action combinations: node operation, task allocation, and resource allocation; generate temporary take-off and landing points in the compliant area based on the order heat map, and perturb the coordinates of the temporary take-off and landing points; use the Hungarian algorithm to solve the matching problem between tasks and take-off and landing points, and the objective function includes flight cost and take-off and landing point startup cost; and block invalid actions that violate airspace rules or physical constraints in real time.

[0149] In an optional embodiment, the decision engine module 12 is further configured to:

[0150] Build time rewards based on decision time and order type; build cost rewards based on flight cost, take-off and landing point startup cost and dynamic cost threshold; build safety rewards based on obstacle density of spatial voxels; build compliance rewards based on the approval results of the airspace management system; adjust the weights of the multi-objective reward function according to the preset time period and cluster load status.

[0151] In an optional embodiment, the decision engine module 12 is further configured to:

[0152] Construct a neural network structure consisting of an input layer, a hidden layer, and an output layer. The hidden layer uses an activation function and sets an anti-overfitting mechanism. Perform online learning on the edge computing terminal and update the model parameters in combination with the federated learning mechanism. Perform privacy protection processing on the locally collected order data before participating in model training.

[0153] In an optional embodiment, the decision engine module 12 is further configured to:

[0154] When abnormal changes in order volume are detected, the emergency response strategy is triggered to adjust the action space and reward function calculation logic; when the model output is abnormal, a guaranteed site selection plan is generated through the historical case matching mechanism.

[0155] In this embodiment, by constructing a dynamic data acquisition layer to obtain multi-source information in real time, and combining it with a reinforcement learning decision engine for intelligent analysis, efficient site selection for drone take-off and landing points is achieved. By defining a state space with multiple dimensions including demand intensity and drone cluster status, the complex scenarios of low-altitude logistics are accurately portrayed; an efficient action space is designed and combined with an action mask mechanism to improve the accuracy and efficiency of decision-making; a multi-objective reward function is constructed and the weights are dynamically adjusted to balance multiple factors such as timeliness, cost, and safety; at the same time, a lightweight training and execution process is adopted to ensure the stable operation of the system in different scenarios; the present invention achieves the effect of significantly improving the real-time response capability of low-altitude logistics, enhancing adaptability to complex terrain, optimizing multi-objective decision-making, and ensuring system robustness and compliance, providing a reliable solution for the intelligent management of low-altitude logistics.

[0156] Furthermore, the intelligent site selection system for drone take-off and landing points for low-altitude logistics can run the above-mentioned intelligent site selection method and system for drone take-off and landing points for low-altitude logistics. The specific implementation can be found in the method embodiment, which will not be repeated here.

[0157] Based on the above embodiments, Figure 3 As shown, the present invention also provides an electronic device, the electronic device comprising:

[0158] At least one processor 22, at least one memory 21, a communication interface 23 and a communication bus 24, wherein the processor 22 is communicatively connected to the memory 21;

[0159] In this embodiment, the memory 21 can be implemented in any appropriate manner, for example, the memory 21 can be a read-only memory, a mechanical hard disk, a solid-state drive, or a USB flash drive, etc. The memory 21 is used to store at least one executable instruction executed by the processor;

[0160] In this embodiment, the processor 22 can be implemented in any appropriate manner. For example, the processor 22 can take the form of a microprocessor or a processor and a computer-readable medium storing computer-readable program code (such as software or firmware) that can be executed by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers, etc.; the processor is used to execute the executable instructions to implement the above-mentioned method and system for intelligent site selection of drone take-off and landing points for low-altitude logistics.

[0161] Based on the above embodiments, the present invention also provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, the method and system for intelligent site selection of drone take-off and landing points for low-altitude logistics as described above are implemented.

[0162] Those skilled in the art will appreciate that the modules and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0163] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described devices, equipment and modules can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0164] In the several embodiments provided in this application, it should be understood that the disclosed devices, apparatuses and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the modules is merely a logical function division. In actual implementation, there may be other division methods, such as multiple modules or units can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interface, device or equipment, which can be electrical, mechanical or other forms.

[0165] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed across multiple network elements. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0166] In addition, each functional module in each embodiment of the present invention may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0167] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program instructions, such as a USB flash drive, a mobile hard disk, a read-only storage server, a random access storage server, a magnetic disk, or an optical disk.

[0168] In addition, it should be noted that the combination of the various technical features in this case is not limited to the combination described in the claims of this case or the combination described in the specific embodiments. All technical features recorded in this case can be freely combined or combined in any way unless there is a contradiction between them.

[0169] It should be noted that the above examples are merely specific embodiments of the present invention. Obviously, the present invention is not limited to the above examples, and many similar variations are possible. All variations directly derived from or associating with the present invention by those skilled in the art are intended to fall within the scope of protection of the present invention.

[0170] The above are only preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An intelligent site selection method for drone take-off and landing points for low-altitude logistics, characterized by: The following steps are involved: Use edge computing terminals to collect logistics order data, airspace status data, and environmental perception data in real time, and normalize the collected data; Based on the processed data, a five-dimensional real-time state vector is constructed, which includes the demand intensity factor, the UAV cluster status, the airspace compliance index, the terrain resistance coefficient, and the cost constraint vector. Build a reinforcement learning decision engine, design an action space that includes node operations, task allocation, and resource allocation, and filter the action space based on spatial rules and physical constraints; Construct a multi-objective reward function that includes timeliness rewards, cost rewards, safety rewards, and compliance rewards, and achieve balanced optimization of multiple objectives through dynamic weight adjustment; Use proximal policy optimization algorithms to train deep reinforcement learning models, and combine edge computing with federated learning mechanisms for model deployment. Based on the real-time state vector, a site selection plan for the drone take-off and landing points is generated through the reinforcement learning decision engine; Output the site selection plan to the drone dispatch system and airspace management system, and optimize the model parameters through a closed-loop feedback mechanism; The five-dimensional real-time state vector including the demand intensity factor, the UAV cluster status, the airspace compliance index, the terrain resistance coefficient and the cost constraint vector is constructed based on the processed data, including: Calculate the demand intensity factor based on the exponentially weighted moving average, and determine the emergency order weighting factor based on the order type; Calculate the drone cluster state vector, which includes the normalized remaining battery power of the drones, the drone cluster load rate, and the number of faulty drones; Calculate the airspace compliance index of the candidate area based on the dynamic no-fly zone data provided by the airspace management system; Calculate the terrain slope between take-off and landing points based on digital elevation data, and determine the terrain drag coefficient based on the flight duration; The cost constraint vector is obtained through Bayesian network training, which includes normalized construction cost, maintenance cost and noise complaint risk value; The reinforcement learning decision engine is constructed, and an action space including node operation, task allocation, and resource allocation is designed. The action space is filtered based on spatial rules and physical constraints, including: Define three types of atomic action combinations: node operation, task allocation, and resource allocation; Generate temporary take-off and landing points in the compliance area based on the order heat map, and perform perturbation processing on the coordinates of the temporary take-off and landing points; The Hungarian algorithm is used to solve the matching problem between missions and take-off and landing points. The objective function includes flight cost and take-off and landing point startup cost. Real-time blocking of invalid actions that violate airspace rules or physical constraints; The multi-objective reward function is constructed, which includes timeliness reward, cost reward, safety reward, and compliance reward. The balanced optimization of multiple objectives is achieved through dynamic weight adjustment, including: Build time-efficiency rewards based on decision time and order type; Build cost incentives based on flight costs, landing point activation costs, and dynamic cost thresholds; Building safety incentives based on obstacle density in a spatial system; Establishing compliance incentives based on the approval results of the airspace management system; The weights of the multi-objective reward function are adjusted according to the preset time period and cluster load status.

2. The method according to claim 1, characterized in that The proposed method uses a proximal policy optimization algorithm to train a deep reinforcement learning model and combines edge computing with a federated learning mechanism for model deployment, specifically including: Constructing a neural network structure including an input layer, a hidden layer, and an output layer, wherein the hidden layer adopts an activation function and sets an anti-overfitting mechanism; Conduct online learning on edge computing terminals and update model parameters using the federated learning mechanism. The locally collected order data is processed for privacy protection before participating in model training.

3. The method according to claim 1, characterized in that The following steps are also included: When abnormal changes in order volume are detected, the emergency response strategy is triggered to adjust the action space and reward function calculation logic; When the model output is abnormal, a backup site selection plan is generated through the historical case matching mechanism.

4. The method according to claim 1, wherein The edge computing terminal integrates meteorological sensors and high-precision positioning modules to collect wind speed, wind direction and take-off and landing point location data; The airspace management system provides dynamic no-fly zone data; The data normalization process includes format conversion and standardization of data of different protocols.

5. An intelligent site selection system for drone take-off and landing points for low-altitude logistics, characterized by: include: The data acquisition module is used to collect logistics order data, airspace status data, and environmental perception data in real time through edge computing terminals, and normalize the collected data; A decision engine module, which is used to construct a five-dimensional real-time state vector containing the demand intensity factor, drone cluster status, airspace compliance index, terrain resistance coefficient, and cost constraint vector; The decision engine module is also used to design an action space including node operations, task allocation, and resource allocation, and filter the action space based on spatial rules and physical constraints; The decision engine module is also used to construct a multi-objective reward function that includes timeliness rewards, cost rewards, safety rewards, and compliance rewards, and achieves balanced optimization of multiple objectives through dynamic weight adjustment; The decision engine module is also used to train deep reinforcement learning models using a proximal policy optimization algorithm and deploy the models in combination with edge computing and federated learning mechanisms. The decision engine module is also used to generate a site selection plan for the drone’s take-off and landing points based on the real-time state vector; An execution feedback module is used to output the site selection plan to the drone scheduling system and airspace management system, and optimize model parameters through a closed-loop feedback mechanism; The decision engine module is further configured to calculate a demand intensity factor based on an exponentially weighted moving average and determine an emergency order weighting factor based on the order type; calculate a drone cluster state vector, including the normalized remaining battery power of the drones, the drone cluster load rate, and the number of faulty drones; calculate the airspace compliance index of the candidate area based on the dynamic no-fly zone data provided by the airspace management system; calculate the terrain slope between the take-off and landing points based on digital elevation data and determine the terrain resistance coefficient based on the flight time; and obtain a cost constraint vector through Bayesian network training, including the normalized construction cost, maintenance cost, and noise complaint risk value. The decision engine module is also used to define three types of atomic action combinations: node operation, task allocation, and resource allocation; generate temporary take-off and landing points in the compliance area based on the order heat map, and perform perturbation processing on the coordinates of the temporary take-off and landing points; use the Hungarian algorithm to solve the matching problem between tasks and take-off and landing points, with the objective function including flight cost and take-off and landing point startup cost; and block invalid actions that violate airspace rules or physical constraints in real time. The decision engine module is also used to build a time reward based on the decision time and order type; build a cost reward based on the flight cost, take-off and landing point startup cost and dynamic cost threshold; build a safety reward based on the obstacle density of the spatial system; build a compliance reward based on the approval results of the airspace management system; and adjust the weight of the multi-objective reward function according to the preset time period and cluster load status.

6. An electronic device, characterized in that: The electronic device comprises: a processor and a memory, wherein the memory is communicatively connected to the processor; The memory is used to store at least one executable instruction executed by the processor, and the processor is used to execute the executable instruction to implement the intelligent site selection method for drone take-off and landing points for low-altitude logistics as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the intelligent site selection method for drone take-off and landing points for low-altitude logistics as described in any one of claims 1 to 4.

Images