Unmanned aerial vehicle cooperative control method and system based on Internet of Things

Abstract

The invention relates to the technical field of unmanned aerial vehicle control, in particular to an unmanned aerial vehicle cooperative control method and system based on the Internet of Things. The method comprises the following steps: collecting scenic spot sensing data, and carrying out multi-modal integrated perception to obtain a scenic spot environment integrated situation map and an unmanned aerial vehicle cluster situation map; performing event-driven dynamic scheduling based on the scenic spot environment integration situation map and the unmanned aerial vehicle cluster situation map to obtain a dynamic task allocation table; obtaining tourist shooting order data, and carrying out intelligent agent track strategy iterative optimization to obtain a multi-intelligent agent strategy set; performing unmanned aerial vehicle cluster global flight path collaborative optimization based on the multi-agent strategy set to obtain an unmanned aerial vehicle individualized execution instruction set; and acquiring a real-time unmanned aerial vehicle shooting image set, performing angle offset adaptability correction to obtain an unmanned aerial vehicle optimized shooting parameter set, and issuing the unmanned aerial vehicle optimized shooting parameter set to each unmanned aerial vehicle terminal. According to the invention, the dynamic scheduling and flight path optimization capability of the multi-unmanned aerial vehicle system can be improved.

Description

Technical Field

[0001] The present invention relates to the field of unmanned aerial vehicle (UAV) control technology, and in particular to a UAV collaborative control method and system based on the Internet of Things. Background Art

[0002] As highly maneuverable, low-cost, intelligent aerial devices, drones are widely used in logistics, disaster monitoring, environmental monitoring, farmland inspection, and other fields. However, in the actual deployment and operation of multi-UAV systems, achieving efficient collaboration and intelligent management and control among multiple drones has become a key challenge that needs to be addressed. Traditional multi-UAV control methods primarily rely on centralized task allocation mechanisms or static, preset task planning. Common approaches include centralized scheduling based on task decomposition, master-slave control based on a hierarchical control structure, and trajectory tracking based on preset flight paths. These methods typically require pre-setting the UAV's operating area, flight trajectory, and task allocation strategy, lacking the ability to adapt to dynamic environments and real-time mission requirements. Furthermore, traditional methods often rely on a single flight control center to provide unified command for all UAVs, resulting in communication bottlenecks and control delays. This results in low efficiency in UAV collaborative operations in complex environments and is susceptible to failure due to unexpected events (such as communication link interruptions and single-unit failures), severely limiting the level of intelligent UAV collaborative operations. Summary of the Invention

[0003] Based on this, it is necessary for the present invention to provide a UAV collaborative control method and system based on the Internet of Things to solve at least one of the above technical problems.

[0004] To achieve the above objectives, a UAV collaborative control method based on the Internet of Things includes the following steps:

[0005] Step S1: collecting scenic area sensor data through the Internet of Things, and performing multimodal integrated perception based on the scenic area sensor data to obtain a scenic area environment integrated situation map and a drone cluster situation map;

[0006] Step S2: Based on the integrated situation map of the scenic area environment and the situation map of the drone cluster, the drone-scene dynamic evolution is performed to obtain a twin interaction situation map; the twin interaction situation map is used to perform event-driven dynamic scheduling to obtain a dynamic task allocation table;

[0007] Step S3: Obtain tourist photo-taking order data through the Internet of Things, and iteratively optimize the agent trajectory strategy of the dynamic task allocation table based on the tourist photo-taking order data to obtain a multi-agent strategy set;

[0008] Step S4: Based on the multi-agent strategy set, the global flight path collaborative optimization of the drone cluster is performed to obtain the drone individual execution instruction set, and the drone individual execution instruction set is distributed to each drone terminal using the edge node deployed by the Internet of Things;

[0009] Step S5: Collect a real-time drone image set, and perform order shooting angle offset evaluation on the real-time drone image set based on the tourist order data to obtain shooting order error data; use the shooting order error data to perform angle offset adaptive correction to obtain the drone optimized shooting parameter set, and send the drone optimized shooting parameter set to each drone terminal.

[0010] The present invention utilizes the high-reliability distributed architecture of the Internet of Things to connect the high-resolution cameras of each drone to the edge node of the Internet of Things, thereby realizing the efficient collection and classification labeling of real-time image data, and ensuring the rapid integration of image data in multi-task scheduling; through the label alignment mapping of tourists' shooting orders and real-time image data, it can accurately reflect the personalized shooting needs of tourists and ensure the dynamic adaptability of drone shooting tasks; based on the angle difference data calculated based on the order mapping data, an angle error feedback mechanism is introduced in the dynamic track optimization, so that the track and attitude adjustment are more in line with the needs of tourists, avoiding the lack of accuracy that may be caused by static planning; through the comprehensive analysis of multi-dimensional parameters such as drone attitude changes and flight speed through the dynamic parameter matrix, an adaptive angle correction model is constructed, which realizes the flexible optimization of real-time track and camera attitude, and effectively improves the dynamic adaptability of drones in complex environments; in the distributed optimization of shooting parameter sets During the distribution process, the MQTT protocol and ROS nodes are combined to ensure reliable data transmission and real-time execution. At the same time, the QoS of the edge node is set to 2, which ensures that the control instructions are not lost in the network, are traceable, and achieve closed-loop monitoring, greatly improving the safety and operation efficiency of multi-machine collaboration of UAVs. In addition, through the dynamic generation of the correction coefficient matrix in the step, not only real-time track attitude correction is achieved, but also through the weight allocation of parameters such as speed, acceleration, and attitude change rate (for example, speed weight 0.3, acceleration weight 0.2, attitude change rate weight 0.25, gimbal response rate weight 0.25), it ensures that the model is still robust under multi-parameter interference, effectively balances local track conflicts and global collaboration requirements in multi-target tasks, avoids resource waste and safety hazards caused by overly concentrated or dispersed task allocation, and thus improves the dynamic scheduling and multi-target track optimization capabilities of the UAV multi-machine system in complex scenarios as a whole.

[0011] Optionally, this specification also provides an Internet of Things-based UAV collaborative control system, which is used to execute the Internet of Things-based UAV collaborative control method described above. The Internet of Things-based UAV collaborative control system includes:

[0012] The multimodal integrated perception module is used to collect scenic area sensor data through the Internet of Things, and perform multimodal integrated perception based on the scenic area sensor data to obtain an integrated situation map of the scenic area environment and a situation map of the drone cluster;

[0013] The task allocation module is used to dynamically evolve the drone-scene based on the integrated situation map of the scenic area environment and the drone cluster situation map to obtain a twin interaction situation map; the twin interaction situation map is used to perform event-driven dynamic scheduling to obtain a dynamic task allocation table;

[0014] The trajectory strategy optimization module is used to obtain tourist photo-taking order data through the Internet of Things, and iteratively optimize the agent trajectory strategy of the dynamic task allocation table based on the tourist photo-taking order data to obtain a multi-agent strategy set;

[0015] The global flight path optimization module is used to collaboratively optimize the global flight path of the drone cluster based on the multi-agent strategy set, obtain the drone's individual execution instruction set, and use the edge nodes deployed by the Internet of Things to distribute the drone's individual execution instruction set to each drone terminal;

[0016] The shooting angle correction module is used to collect real-time drone shooting image sets, and evaluate the order shooting angle offset of the real-time drone shooting image sets based on tourist order data to obtain shooting order error data; use the shooting order error data to perform angle offset adaptive correction to obtain the drone optimized shooting parameter set, and then send the drone optimized shooting parameter set to each drone terminal.

[0017] The Internet of Things-based UAV collaborative control system of the present invention can implement any one of the Internet of Things-based UAV collaborative control methods of the present invention, and is used to combine the operation and signal transmission media between various modules to complete the Internet of Things-based UAV collaborative control method. The internal modules of the system cooperate with each other, thereby improving the dynamic scheduling and multi-target trajectory optimization capabilities of the UAV multi-machine system in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments thereof made with reference to the following drawings:

[0019] Figure 1 This is a schematic flow chart of the steps of the collaborative control method of UAVs based on the Internet of Things of the present invention;

[0020] Figure 2 Detailed step flow diagram of step S1 in the present invention;

[0021] Figure 3 Detailed step flow diagram of step S15 in the present invention;

[0022] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0023] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. It is obvious that the embodiments described are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts are within the scope of protection of the present invention.

[0024] In addition, the accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor and / or microcontroller approaches.

[0025] It should be understood that although the terms "first," "second," and the like may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used solely to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element, without departing from the scope of the exemplary embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the listed associated items.

[0026] To achieve this, please refer to Figures 1 to 3 The present invention provides a method for cooperative control of unmanned aerial vehicles based on the Internet of Things, the method comprising the following steps:

[0027] Step S1: collecting scenic area sensor data through the Internet of Things, and performing multimodal integrated perception based on the scenic area sensor data to obtain a scenic area environment integrated situation map and a drone cluster situation map;

[0028] In this embodiment, the edge nodes deployed by the Internet of Things collect multi-dimensional environmental data such as temperature, humidity, wind speed, PM2.5, tourist flow, road traffic conditions, tree crown density, and drone operation conditions, and are uniformly converted into CSV format through a dedicated interface to form a multi-source heterogeneous original data set (data dimensions include: sensor ID, acquisition time, data type, data value, etc., with a total of 20 columns). On this basis, multi-modal data preprocessing is performed on the multi-source heterogeneous original data set, and the standardized normalization method is used to scale each data to the interval of 0 to 1, and missing value interpolation is performed (setting a threshold of 0.05) to obtain a standardized multimodal feature data set (matrix form is N×20, N is the number of time steps). Subsequently, the standardized multimodal feature data is synchronized in time series using spatial indexing, and after aligning the timestamps, multimodal feature fusion is performed using multidimensional feature selection technology, and a multimodal spatial correlation matrix (in the form of a 20×20 symmetric matrix) is constructed. The clustering threshold is set to 0.6 to divide multiple spatial node sub-graphs. Furthermore, the multimodal spatial association matrix was partitioned and hierarchically analyzed using multi-layer graph convolution to extract community structure data. Finally, based on the community structure and multimodal temporal feature embedding, the community risk value of each node was calculated as a continuous value between 0.3 and 0.9. This completed the generation of an integrated scenic environment situation map and a drone cluster situation map, each stored in JSON format. The graph structure includes node (location ID, attributes, risk value) and edge (connection ID, association weight) information.

[0029] Step S2: Based on the integrated situation map of the scenic area environment and the situation map of the drone cluster, the drone-scene dynamic evolution is performed to obtain a twin interaction situation map; the twin interaction situation map is used to perform event-driven dynamic scheduling to obtain a dynamic task allocation table;

[0030] In this embodiment, based on the integrated situation map of the scenic area environment and the situation map of the drone cluster, each node in the environmental map (scenic spot number, geographic coordinates, risk value) is bidirectionally matched with each node in the drone cluster map (drone ID, current location, remaining power, heading angle). Through real-time mapping of the twin scene, environmental elements are extracted and a twin scene environmental element set is constructed (in data table form, fields include: scene ID, risk level, environmental element type, drone ID). The drone dynamic flight behavior data (including speed, acceleration, angle change, remaining power, etc.) is collected in real time through the edge node, and the feature set is aligned with the twin scene environmental element set to form a drone dynamic behavior feature set (dimensions: drone ID, timestamp, feature sequence). Based on time series association analysis, we further modeled the temporal evolution of scenarios using graph relationships (graph structure: nodes are scene IDs and drone IDs, and edges are time evolution weights). We then extracted event triggering conditions from this graph, such as a sudden change in flight speed exceeding ±20%, battery power falling below 30%, and wind speeds exceeding 5 m / s. These conditions were combined with real-time features from the drone's dynamic behavior feature set (acceleration rate, track offset, etc.) to construct a multidimensional dynamic event triggering mechanism (data structure: event ID, triggering condition, node ID, time window). Finally, we simulated the drone-scenario dynamic relationship graph through the event triggering mechanism, generating a twin interaction situation diagram (stored in a Neo4j graph database, containing node and edge attributes). Event-driven dynamic scheduling was then implemented, outputting a dynamic task allocation table (tabular format: task ID, drone ID, time window, priority, and task content).

[0031] Step S3: Obtain tourist photo-taking order data through the Internet of Things, and iteratively optimize the agent trajectory strategy of the dynamic task allocation table based on the tourist photo-taking order data to obtain a multi-agent strategy set;

[0032] In this embodiment, in terms of acquiring tourist photography order data, tourist order information (including order number, tourist ID, desired attractions, shooting angle, time requirements, special needs, etc.) is collected in real time through the IoT edge node interface and parsed into a structured data table (fields: order ID, timestamp, photography requirements, special tags). Combined with the dynamic task allocation table (fields: task ID, drone ID, start and end time, task description), the tourist photography order data and the dynamic task allocation table are labeled and encoded (label format: tourist demand label set, task label set), and a tourist task label set (JSON format) is generated. The tourist task label set and the dynamic task allocation table are used for feature fusion to construct a multi-objective task feature matrix (matrix form: M×P, M is the number of tasks, P is the fusion label dimension), and the shooting requirement labels in the matrix (such as high-altitude panorama, low-altitude details) are matched with the task feature labels (such as path complexity, task priority) to obtain the initial task allocation strategy (fields include: drone ID, task ID, expected path, time window). Subsequently, combining scenic area environmental data with the UAV's dynamic flight behavior data, the path planning engine (path vector includes: starting point coordinates, target coordinates, track key points, speed, and heading angle) is used to generate an initial track planning dataset (JSON format). Furthermore, the initial task allocation strategy and the initial track planning dataset are subjected to multi-objective coupling optimization to generate a task track optimization dataset (fields: task ID, optimized track point sequence, energy consumption prediction, priority score). Finally, the optimized dataset is subjected to joint modeling and dynamic scheduling simulation to obtain a multi-agent strategy set (stored in GraphML format) for subsequent global flight path collaborative optimization.

[0033] Step S4: Based on the multi-agent strategy set, the global flight path collaborative optimization of the drone cluster is performed to obtain the drone individual execution instruction set, and the drone individual execution instruction set is distributed to each drone terminal using the edge node deployed by the Internet of Things;

[0034] In this embodiment, based on the multi-agent strategy set, the dynamic strategy of each drone is first deconstructed, and the features such as the key points of the flight track, dynamic posture changes (Euler angle three-axis data), and the weight of interaction with the scenic environment are extracted to form a multi-agent feature matrix (in the form of N×Q, where N is the number of drones and Q is the feature dimension, for example, Q=10). Combined with the scenic environment data (including terrain elevation, obstacle coordinates, meteorological conditions, etc.), a global track feasibility analysis is performed, and the constraint parameters set the flight safety distance to 5 meters, the minimum flight altitude to 20 meters, and the maximum track inclination angle to 35 degrees. The above data is used to perform global track constraint modeling and generate a global track constraint set (JSON format). On this basis, the initial global track matrix (matrix form: drone ID, track point number, X coordinate, Y coordinate, Z coordinate, speed) is generated through the continuity detection of the track key points and speed curve fitting. Subsequently, combined with the multi-agent strategy set and the scenic environment data, multi-objective optimization is performed using weighted feature coupling (drone ID, task priority, flight risk index) to obtain global track optimization data (JSON format). Finally, the global trajectory optimization data is broken down by drone ID, and the key point sequence and dynamic attitude change characteristics of each trajectory are extracted. Flight instructions are generated based on the flight safety area (a predefined polygonal area set), and the drone's individual execution instruction set is output (CSV file, containing drone ID, track point coordinates, attitude instructions, execution duration, etc.). The instruction set is distributed to each drone terminal through the IoT edge node interface to complete flight control.

[0035] Step S5: Collect a real-time drone image set, and perform order shooting angle offset evaluation on the real-time drone image set based on the tourist order data to obtain shooting order error data; use the shooting order error data to perform angle offset adaptive correction to obtain the drone optimized shooting parameter set, and send the drone optimized shooting parameter set to each drone terminal.

[0036] In this embodiment, in the acquisition of drone-photographed image data, the drone's high-resolution camera module is called through the Internet of Things interface to collect image data every 2 seconds, and preliminary classification and labeling are performed in real time (labels include scenic spot number, shooting angle, lighting conditions, etc.) to generate a real-time drone-photographed image set (image set ID, drone ID, timestamp, label, image URL). Subsequently, the tourist's photography order data (including the expected shooting angle and scenic spot requirements) is labeled aligned and mapped with the real-time drone-photographed image set (fields: order ID, drone ID, shooting angle, target angle) to obtain the tourist order mapping data (JSON format). Based on the tourist order mapping data, the difference between the tourist's expected shooting angle and the actual image angle is calculated (unit: degree). If the deviation is greater than 3 degrees, it is included in the deviation assessment matrix (matrix form: order ID, drone ID, deviation angle). Based on the shooting order error data and combined with the drone's real-time dynamic parameters (such as flight altitude and track attitude), an adaptive angle offset correction model was constructed. (Model input: deviation angle, drone attitude, environmental factors; output: corrected shooting angle. The model structure uses a multi-layer fully connected network with an input dimension of 6, hidden layer dimensions of 32, 16, and 8, and an output dimension of 1.) This model is used to infer shooting parameters based on the drone's dynamic parameters and tourist order mapping data, resulting in an optimized drone shooting parameter set (fields: drone ID, corrected shooting angle, exposure time, and shutter speed). This optimized shooting parameter set is then distributed to each drone terminal via an IoT interface, enabling dynamic adjustment of the shooting angle.

[0037] Optionally, step S1 specifically includes:

[0038] Step S11: collecting scenic area sensor data through the IoT edge node set deployed in the scenic area, and converting the scenic area sensor data into a format to obtain a multi-source heterogeneous original data set;

[0039] In this embodiment, multiple types of IoT edge nodes are deployed in the main entrances and exits of the scenic area, walking trails, squares and parking lots, including temperature and humidity sensors, wind speed and direction meters, visitor counters, PM2.5 sensors, CO2 concentration detectors, etc., totaling 20 types of devices. Each device uploads data to the edge node every 2 seconds. For drone data collection, a flight status monitoring device is deployed on each drone, including an IMU (inertial measurement unit), a GPS positioning module, a battery power monitoring module, an image acquisition module and a task execution status module. These data are sent to the edge node in real time at a frequency of 1Hz. The drone data fields include: drone ID, timestamp, GPS coordinates, speed (X, Y, Z three axes), heading angle, flight altitude, remaining power (unit: %), current shooting status (Boolean value), etc., totaling 12 columns. Data from all devices and drones is connected to edge nodes via the MQTT protocol. The edge nodes perform a centralized initial data cleansing process, including removing null values, filtering data with signal quality below 80% (using an RSSI threshold of -90dBm), and removing speed outliers (greater than 20m / s) from drone data to prevent sudden changes from interfering with subsequent analysis. After this initial cleansing, the data is uniformly converted to JSON format, with fields including device ID, device type (ground or drone), timestamp, geographic coordinates, data value, and data quality indicator. The data is managed in multiple tables in a PostgreSQL database, one for each device. Drone data is maintained independently but shares a unified timestamp index with ground sensor data, ensuring time alignment during subsequent multimodal data fusion. The resulting multi-source, heterogeneous raw dataset includes not only ground sensor data but also drone flight status and mission execution status data, generating approximately 200,000 data entries per hour. This facilitates subsequent multimodal feature extraction and construction of scenic area situation maps.

[0040] Step S12: performing multimodal data preprocessing and feature standardization on the multi-source heterogeneous original data sets to obtain a standardized multimodal feature data set;

[0041] In this embodiment, for the obtained multi-source heterogeneous original data set, a view is first established in the PostgreSQL database for data connection to ensure that the ground equipment and drone equipment data are associated through a unified timestamp index. For drone data, four feature fields are extracted, namely IMU three-axis speed (unit: m / s), GPS heading angle (unit: °), flight altitude (unit: m) and remaining power (unit: %), and the Z-score normalization method is used to normalize each feature to a standardized interval with a mean of 0 and a standard deviation of 1; for ground sensor data, such as temperature and humidity, PM2.5, CO2, etc., the Min-Max normalization method is used to map the data to the range of 0 to 1. To ensure data processing consistency, all standardization operations are batch processed using Pandas DataFrame objects in the Python environment to generate a standardized multimodal feature data set. The final output data set is labeled by device type (ground or drone) and saved in Parquet format to facilitate subsequent large-scale distributed data analysis and feature alignment.

[0042] Step S13: performing feature alignment and time series synchronization on the standardized multimodal feature dataset to obtain multimodal time series fusion data;

[0043] In this embodiment, in the obtained standardized multimodal feature data set, the ground equipment and drone equipment data are first aligned for multimodal features according to a unified timestamp. Using the window function of PostgreSQL, the window is slid for each time point, the time window length is set to 5 seconds, the step length is 1 second, the data in the window is aggregated, and a multimodal feature sequence of the corresponding time step is generated. Special attention is paid to the flight status and heading angle changes in the drone data. The dynamic characteristics of the drone are captured by calculating the time difference of the heading angle (unit: ° / s), and this dynamic characteristic is included in the multimodal time series fusion data together with the ground sensor data (such as PM2.5, temperature and humidity), ensuring that the time series fusion data contains not only environmental perception information, but also the flight dynamic information of the drone. The final multimodal time series fusion data is stored in Tensor format (shape: N×T×F, N is the number of devices, T is the time step, and F is the feature dimension), where the feature dimension of the drone data is 6 and the feature dimension of the ground equipment data is 10.

[0044] Step S14: performing multimodal feature fusion on the multimodal time series fusion data, and performing spatial correlation analysis based on the feature fusion results to obtain a multimodal spatial correlation matrix;

[0045] In this embodiment, the obtained multimodal time series fusion data is subjected to feature integration processing through the Numpy and Scipy libraries in the Python environment, and the dynamic features of the drone (including speed, heading angle, flight altitude, remaining power, etc.) are fused with the ground environmental features (temperature and humidity, PM2.5, CO2, etc.) for multimodal feature fusion, and a single fusion feature vector is formed in the form of feature splicing. Then, based on the geographic coordinate information (latitude and longitude), the Haversine formula is used to calculate the spatial distance matrix (unit: m) between the devices, and the distance threshold is set to 50m. The relationship between the device pairs exceeding the threshold is set to 0, otherwise it is set to 1, forming a preliminary spatial correlation matrix. In order to reflect the interactive characteristics of the drone and the ground equipment, the flight state influence coefficient of the drone is further introduced into the spatial correlation matrix, and the speed weighting factor of the drone is defined (between 0.8 and 1.2) to correct the correlation strength between the drone and the ground equipment in the spatial correlation matrix. The final multimodal spatial association matrix is ​​a two-dimensional matrix with a matrix shape of (N+M)×(N+M), where N is the number of ground devices and M is the number of drones. The value of each element represents the association strength of the device pair.

[0046] Step S15: Construct a scenic area situation map based on the multimodal spatial association matrix classification, thereby obtaining a scenic area environment integrated situation map and a drone cluster situation map.

[0047] In this example, the obtained multimodal spatial correlation matrix is ​​partitioned by device type (ground or drone), extracting ground device areas, drone areas, and the interaction areas between the two. For the drone area, the focus is on the changing trends of flight altitude and heading angle. By calculating the heading angle change rate matrix (unit: ° / s), the drone dynamic situation is integrated into the overall spatial correlation analysis. For the ground device area, the spatial distribution of environmental parameters such as PM2.5, temperature and humidity is combined with the heading angle change rate and remaining battery distribution in the drone area to comprehensively generate an integrated scenic area environment situation map and a drone cluster situation map. The output scenic area environment integrated situation map is stored in GeoTIFF format, with each pixel containing multi-channel data on PM2.5 concentration, temperature and humidity, and CO2 concentration, with a resolution of 5m. The drone cluster situation map is output in GraphML format, where each node represents a drone, and node attributes include flight altitude, heading angle, and remaining battery. Edge attributes represent the flight state coupling relationship between drones. Ultimately, these two situation maps achieve a unified perception of the overall scenic area situation and drone cluster dynamics through a shared timestamp index.

[0048] Optionally, step S15 is specifically as follows:

[0049] Step S151: performing spatial node partitioning on the multimodal spatial association matrix, using the node partitioning results as independent graph subgraphs, and calculating the edge weights between spatial nodes to obtain a multimodal regional subgraph;

[0050] In this embodiment, in the multimodal spatial association matrix generated in the previous step, the Python NetworkX library is first used to perform spatial node partitioning on the matrix, and the nodes are divided into two categories: one is ground equipment (such as temperature and humidity sensors, CO2 monitoring points, etc.), and the other is drone nodes (including flight speed, remaining power, heading angle and other attributes). In the specific operation, the edge weight threshold is set to 0.5, and all edges below the threshold will be removed to ensure data sparsification in subsequent processing. The multimodal regional subgraph obtained by division is composed of several subgraphs, each of which contains several spatial nodes and edge weight data between them. The output results are saved in GraphML format, and each node in the subgraph is accompanied by an attribute dictionary, such as ground nodes including coordinates, PM2.5 concentration, and humidity values; drone nodes include heading angle, flight altitude and remaining power, and edge weights are recorded in the edge attribute dictionary, with values ​​between 0 and 1.

[0051] Step S152: Perform community segmentation and hierarchical analysis based on the multimodal regional sub-graph to obtain multimodal community structure data;

[0052] In this embodiment, in the generated multimodal regional sub-graph, a community division method based on edge weight threshold stratification is adopted for processing. The edge weight threshold stratification step is set to 0.1, and the sub-graphs with edge weights in different intervals are used as inputs to gradually generate community structures. The node set and edge weight information of each community are extracted to form multimodal community structure data. The heading angle and flight speed of the drone node are recorded separately in the community structure for subsequent drone community risk calculation. Finally, the community structure data is stored in JSON format. The JSON object of each community contains the community number, node set, edge weight distribution and the flight dynamic information of the drone, ensuring the data integrity of the drone and ground nodes in the community division.

[0053] Step S153: Calculate the community comprehensive risk using the node and edge weights in each community structure in the multimodal community structure data to obtain the scenic area environment community risk matrix;

[0054] In this embodiment, for the obtained multimodal community structure data, the community comprehensive risk calculation is performed in the Python environment based on the attribute distribution of the nodes in each community. The flight altitude fluctuation (unit: m) and heading angle deviation (unit: °) of the drone node are used as drone risk factors, and the PM2.5 concentration (unit: μg / m 3) and CO2 concentration (unit: ppm) as ground risk factors. The risk value of each community is obtained through weighted summation, with the weight determined by the proportion of drone nodes (the proportion of drone nodes is between 0.3 and 0.5). The generated scenic area environment community risk matrix is ​​a two-dimensional matrix, with the number of communities in each row and column, and each element represents the risk coupling strength of community i to community j (between 0 and 1). It is ultimately exported in CSV format for subsequent multi-layer scenic area environment modeling.

[0055] Step S154: combining the risk values ​​in the scenic area environment community risk matrix and the community division results, performing multi-layer graph modeling on each community to obtain a multi-layer scenic area environment graph;

[0056] In this embodiment, based on the scenic area environment community risk matrix, the scenic area community is first divided into multiple layers according to the community number, each layer represents a risk level interval (low, medium, relatively high, high), and the comprehensive mean of the ground node environmental risk and the UAV node flight status risk is used as the stratification standard. The heading angle offset range of the UAV node (within ±10°) is used to refine the stratification to ensure that the risk division matches the UAV flight dynamics. The multi-layer scenic area environment map is saved in the form of a multi-layer graph structure, and each layer of the graph is independently recorded in the form of a GraphML file, including node attributes, edge weights and hierarchical risk levels. Each node object is labeled with a risk label to facilitate subsequent visual rendering and UAV dynamic embedding.

[0057] Step S155: extracting the real-time dynamic behavior features of the drones based on the multimodal time series fusion data, embedding the real-time dynamic behavior features of the drones into the multi-layer scenic area environment map, modeling the dynamic interaction behavior between drones, and obtaining the dynamic behavior data of the drone cluster;

[0058] In this embodiment, the real-time dynamic behavior characteristics of the drone are extracted for the obtained multimodal time series fusion data, and the characteristics include flight speed (unit: m / s), flight altitude (unit: m), remaining power (unit: %), heading angle (unit: °) and heading angle change rate (unit: ° / s). These features are uniformly encapsulated as a drone dynamic behavior matrix in the Python environment. The matrix dimension is M×F, where M is the number of drones and F is the 5 flight dynamic dimensions. The dynamic behavior matrix is ​​directly embedded in the multi-layer scenic area environment map generated in step S154, and the dynamic behavior field is added to the drone node object of each layer of the map. In this way, the generated drone cluster dynamic behavior data exists in the form of GraphML extended fields, ensuring the coupling of drone flight real-time dynamic data and the multi-layer structure of scenic area environmental risks.

[0059] Step S156: Based on the multi-layer scenic area environment map and the dynamic behavior data of the drone cluster, joint visualization and map layer mapping are performed to obtain an integrated situation map of the scenic area environment, and a drone cluster situation map is constructed based on the community dynamic characteristics in the scenic area environment community risk matrix and the dynamic behavior data of the drone cluster.

[0060] In this embodiment, based on the multi-layer scenic area environment map and the dynamic behavior data of the drone cluster, the WebGL visualization framework CesiumJS is used for three-dimensional visualization rendering. In the specific operation, each layer in the multi-layer scenic area environment map is marked with a different color, each community node is represented as a sphere or a cylinder, and the animated trajectory of the drone flight dynamic data is attached to the map visualization. In the drone cluster situation map, the trajectory of each drone node is presented in the form of a dynamic path, with a heading angle and a bar indicator of the remaining power, and the flight altitude is used as a three-dimensional attribute in the three-dimensional scene. Finally, the scenic area environment integrated situation map is output in the Cesium3DTiles format, and the drone cluster situation map is output in the GeoJSON extended format, which supports the scenic area management end to switch the situation layer, community risk level and drone cluster behavior dynamics in real time, so as to achieve intuitive visualization and multi-dimensional data mapping effects.

[0061] Optionally, step S153 is specifically as follows:

[0062] Extract node features in each community based on multimodal community structure data to obtain a set of community node feature vectors;

[0063] In this example, for multimodal community structure data, Python's pandas and networkx are first used to traverse and extract the node set of each community. Node features include: flight speed (m / s), heading angle (°), and remaining battery (%) of drone nodes; PM2.5 concentration (μg / m 3 ), temperature and humidity values ​​(%). To ensure data consistency, all node attributes are aligned by timestamp, and the sampling frequency is unified to once per second to ensure the integrity of temporal dynamic features. All eigenvalues ​​of each node are spliced ​​by column to form a node feature vector, which is uniformly saved in float32 format. The final generated community node feature vector set is output in the form of a two-dimensional matrix, where each row represents the feature vector of a single node, and each column represents a single feature dimension (such as speed, concentration, etc.). The matrix dimension is the number of nodes × the number of feature dimensions (for example: 20 × 5) for subsequent processing.

[0064] The normalization factor threshold is set to 0.85, and the edge weights of the community structure are normalized using the node spatial distribution characteristics and time dynamic fluctuation characteristics of the community node feature vector to obtain the normalized community edge weight matrix;

[0065] In this embodiment, in the generated node feature vector set, the edge weights of the community structure are first normalized using spatial distribution features (such as relative distance between nodes) and time dynamic fluctuation features (such as speed change rate). The spatial distribution feature uses the longitude and latitude distance of the node, and the time dynamic fluctuation feature uses the standard deviation of the node feature change. The normalization factor threshold is set to 0.85. When the edge weight normalization coefficient is lower than the threshold, it is set as the threshold to avoid over-scaling. All processing results are stored in a two-dimensional matrix. The rows and columns of the matrix are the total number of nodes in the community, and the matrix elements represent the normalized values ​​of the edge weights between the nodes in the community (0 to 1). The matrix is ​​exported to Numpy format (.npy file) to facilitate direct loading and calling in subsequent multidimensional data fusion.

[0066] The fusion dimension is set to 64, and the community node feature vector set and the community edge weight normalized matrix are fused to obtain the community node interaction feature data;

[0067] In this embodiment, the node feature vector set and the edge weight normalization matrix are combined to perform multidimensional data fusion processing. During the operation, the fusion dimension is set to 64, and the feature vector of each node is combined with the normalized edge weights of its adjacent nodes by splicing. Then, the principal component projection is processed to uniformly compress the dimensions to 64 dimensions to ensure computability and processing efficiency when the data volume is large. The interaction feature data of each community node is saved in CSV format, and each row contains the node number, the 64-dimensional fusion feature vector, and the community number label of the node for subsequent node risk analysis.

[0068] Utilize the interactive characteristic data of community nodes to perform multi-index weighted risk calculation of node risk, thereby calculating the comprehensive risk value of each point, and weighting the comprehensive risk value to generate the community node risk weight matrix;

[0069] In this embodiment, based on the generated community node interaction feature data, the node's flight speed fluctuation value, PM2.5 concentration, and remaining power, and other multiple indicators are used to set indicator weights (flight speed 0.3, PM2.5 concentration 0.4, remaining power 0.3) for multi-indicator weighted processing. First, each indicator of each node is normalized). Finally, the comprehensive risk value is smoothed using a weighted average method to avoid excessive risk output due to abnormal values. The final comprehensive risk value of the node is saved in CSV format, and the file contains the node number and the corresponding risk value. Based on the node comprehensive risk value, for each node within the community, the node risk value is represented in matrix form according to its number and the community label to which it belongs. The rows and columns of the matrix are the number of community nodes. Each element in the matrix represents the risk weight of node i to node j, with a value range of 0 to 1. The weight value is obtained by weighted multiplication of the node risk value and the normalized value of the edge weight. The self-risk value of the node is retained on the diagonal position of the matrix. The matrix data is stored in Numpy (.npy) format for easy and fast loading. Each matrix file is named after the community number, such as "Community A_Risk Weight Matrix.npy", to ensure the convenience of file management and the efficiency of subsequent analysis.

[0070] Based on the community node risk weight matrix, the risk weight values ​​in each community are aggregated to obtain the scenic area environment community risk matrix.

[0071] In this embodiment, based on the community node risk weight matrix, the risk weight values ​​in each community are aggregated in the order of node numbers. The specific operation is to sum the risk weight vector of each node and divide it by the number of nodes to obtain the average risk value of the community as a whole, and then aggregate the average risk values ​​of all communities to form a scenic area environment community risk matrix, and the matrix dimension is the number of communities × the number of communities. The elements in the matrix represent the risk coupling intensity of community i to community j (range 0 to 1), and are output in CSV format. The first line is the community number index, which is convenient for subsequent association analysis and multi-layer graph construction. The final community risk matrix serves as the risk input for subsequent multi-layer graph modeling and dynamic embedding of drones.

[0072] Optionally, the dynamic evolution of the drone-scene in step S2 is specifically as follows:

[0073] Perform real-time twin scene mapping of the scenic area environment integrated situation map and the drone cluster situation map, extract environmental elements, and obtain a twin scene environmental element set;

[0074] In this embodiment, based on the integrated situation map of the scenic area environment (including multimodal perception information: temperature and humidity, PM2.5, noise, tourist density, etc.) and the drone cluster situation map (including drone position, speed, heading angle, remaining power, etc.), the real-time data synchronization function of the IoT edge node is used to align the two maps in the time dimension in seconds. After alignment, the multimodal environmental features of each node in the environmental map (such as PM2.5: 35.5μg / m 3 , temperature: 28.3℃, tourist density: 15 people / m 2 ) is extracted from the flight characteristics of each node in the drone map (such as speed: 5.5m / s, heading angle: 270°, remaining battery: 60%) to generate a twin scene environment feature set. The twin scene environment feature set is stored in a CSV file with columns including node number, node type, timestamp, and multimodal environment characteristics to facilitate subsequent correlation analysis.

[0075] The UAV dynamic flight behavior data is collected through the Internet of Things, and the feature set of the UAV dynamic flight behavior data is aligned with the twin scene environment element set to obtain the UAV dynamic behavior feature set;

[0076] In this embodiment, the UAV dynamic flight behavior data, including real-time speed, heading angle, attitude angle, remaining power and flight status mark, is obtained through the edge node of the Internet of Things. Each piece of data contains a unique UAV ID, a timestamp and a behavior data field. The UAV dynamic flight behavior data is aligned with the generated twin scene environment element set, and the timestamp alignment granularity is set to 1 second to ensure data synchronization. The alignment process adopts a multimodal feature standardization method to ensure that the data range of dimensions such as speed, attitude angle, PM2.5 is consistent (0~1). The final output UAV dynamic behavior feature set is represented in the form of a two-dimensional matrix, where each row represents the dynamic behavior characteristics of a single UAV at a certain timestamp, and each column corresponds to a different feature dimension. The matrix dimension is the number of nodes × the number of features, for example 50×6.

[0077] Combining the twin scene environmental element set with the UAV dynamic behavior feature set to model the scene temporal evolution relationship and obtain the UAV-scene dynamic relationship map;

[0078] In this embodiment, the NetworkX library in Python is used to construct a dynamic relationship graph between drones and scenes, combining the twin scene environmental element set and the drone dynamic behavior feature set. First, the environmental element nodes (PM2.5, temperature and humidity, etc.) and drone nodes (dynamic behavior features) are respectively used as node sets in the graph, and the node attributes include node type, timestamp, feature vector and other information. Then, the environmental nodes and the corresponding drone nodes are connected in sequence in the graph according to the timestamp order. The edge weights are calculated based on the time dynamic change rate (speed, concentration, heading angle). The edge weight range is 0 to 1, and edges below 0.1 will be filtered to improve the sparsity of the graph. The final output drone-scene dynamic relationship graph is saved in the form of a GraphML file. Each node contains a timestamp and a multidimensional feature vector, and the edge contains a weight value and time dynamic information.

[0079] Based on the UAV-scene dynamic relationship graph, the trigger conditions of scene dynamic events are extracted, and a multi-dimensional dynamic event triggering mechanism is constructed in combination with the UAV dynamic behavior feature set;

[0080] In this embodiment, based on the UAV-scene dynamic relationship map, key features of the environment dynamic node and the UAV dynamic node are extracted (such as the PM2.5 concentration threshold ≥ 75 μg / m 3 , flight speed ≥ 6m / s, remaining power ≤ 20%) as dynamic event trigger conditions. The parameters of the multi-dimensional dynamic event trigger mechanism are set as follows: trigger condition detection window 10 seconds, threshold sliding step 1 second, dynamic event label (such as "pollution warning", "low battery avoidance", "high-speed flight warning"). The flight speed, attitude angle, and remaining power fields in the UAV dynamic behavior feature set are subjected to multi-dimensional condition detection with the above trigger conditions to generate an event trigger identifier for each node at each timestamp. The output results are saved in a CSV file, including the node number, timestamp, and event label fields.

[0081] The multi-dimensional dynamic event triggering mechanism is used to simulate the event triggering of the UAV-scene dynamic relationship graph to obtain the UAV-scene dynamic interaction triggering data;

[0082] In this example, based on the generated multi-dimensional dynamic event triggering mechanism, event triggering conditions are applied to the generated drone-scene dynamic relationship graph to perform event-triggered simulation. The simulation traverses the graph nodes and edges, labels each node's event label according to a time series, and simulates dynamic interactions based on the changes in edge weights when the event is triggered. The simulation time step is set to 1 second, and the total simulation duration is 300 seconds. The simulation results are saved as a CSV file containing the node number, timestamp, event label, simulated interaction edge number, and edge weight change value, facilitating subsequent drone-environment situational analysis.

[0083] The UAV-scene dynamic interaction trigger data is used in combination with the UAV-scene dynamic relationship map to deduce the real-time interaction situation between the UAV and the environment and obtain a twin interaction situation map.

[0084] In this embodiment, based on the generated drone-scene dynamic interaction trigger data, combined with the drone-scene dynamic relationship graph, the real-time interaction situation of the drone-environment is deduced. The deduction adopts a graph convolution structure to perform feature propagation of the dynamic behavior data of the drone node and the multimodal elements of the environment node. During the deduction process, the feature vector of each node is represented in matrix form (number of nodes × feature dimension, such as 50×10), and weighted feature aggregation is performed using edge weights. The deduction outputs a twin interaction situation graph, which is saved as a GraphML file. The node information includes node number, node type, comprehensive risk value, timestamp and situation level label (such as safe, warning, dangerous). This situation graph will be used for dynamic scheduling and task allocation of drone clusters.

[0085] Optionally, the event-driven dynamic scheduling in step S2 is specifically as follows:

[0086] Based on the twin interaction situation graph, dynamic behavior nodes and event trigger conditions are extracted. The execution node-event condition graph is convolutionally aggregated based on the real-time task execution status in the UAV dynamic flight behavior data to obtain the task event fusion feature matrix.

[0087] In this embodiment, dynamic behavior nodes and event trigger condition information are extracted from the twin interaction situation graph, including node ID, timestamp, node dynamic behavior vector (such as speed, heading angle, remaining power) and event trigger identifier (such as PM2.5 is too high, low power alarm). Combined with the real-time task execution status (including task ID, task stage, remaining execution time, execution priority identifier) ​​in the dynamic flight behavior data of the drone, a graph convolution aggregation process of execution node-event condition is constructed. First, the feature vector of the dynamic behavior node (dimension is 10, including PM2.5, temperature, speed, remaining power, etc.) is spliced ​​with the task vector of the real-time task execution status (dimension is 5, including task ID, task stage, remaining time, etc.) to form the node input feature and form a node input matrix (number of nodes × 15-dimensional features). Then, weighted feature transfer is performed according to the edge relationship of the twin interaction situation graph, the initial value of the edge weight is set to 0.6, and three convolution aggregations are performed. The output task event fusion feature matrix dimension is the number of nodes × 32-dimensional features (fusion dimension 32), which is used for subsequent task priority evaluation.

[0088] The UAV mission priority is evaluated using the mission event fusion feature matrix to obtain the UAV mission priority dataset;

[0089] In this example, the obtained task event fusion feature matrix (number of nodes × 32-dimensional features) is used to extract key features such as the task phase, remaining execution time, and event trigger label of each node. The task urgency parameter threshold is set to 0.75, and the urgency of the task execution status is calculated. Through matrix row normalization, the remaining time is mapped to (0, 1.0), and the task urgency and event risk are weighted and fused to obtain the drone task priority score for each node. The priority scores are sorted to generate a drone task priority dataset in the form of a CSV file containing the node ID, timestamp, task ID, and priority score for subsequent event-driven time window division.

[0090] Based on the UAV mission priority dataset, combined with the event sequence and node dynamic behavior characteristics in the twin interaction situation diagram, event-driven time window division is performed to obtain the event-driven scheduling window set;

[0091] In this embodiment, based on the UAV task priority data set, the event time series data (node ​​ID, timestamp, event label) in the twin interaction situation diagram is synchronized with the node dynamic behavior characteristics (speed, attitude angle, remaining power), and the time window division step is set to 10 seconds and the sliding step is set to 2 seconds. For each time window, the task priority mean and event risk mean in the window are extracted to construct the initial label of the event-driven scheduling window (such as "high priority high risk" and "low priority low risk"). Generate an event-driven scheduling window set in the format of a JSON file, with fields including window ID, start time, end time, priority mean and event risk mean, to facilitate subsequent UAV task allocation processing.

[0092] Combine the event-driven scheduling window set and the UAV task priority data set to perform UAV task-event window matching allocation and obtain a dynamic task allocation table;

[0093] In this embodiment, based on the generated event-driven scheduling window set, the task ID in the UAV task priority data set is mapped to the window ID. Using a priority-based allocation method, tasks with a priority greater than 0.85 are assigned to the high-priority window, tasks with a priority of 0.5 to 0.85 are assigned to the medium-priority window, and tasks with a priority lower than 0.5 are assigned to the low-priority window. Tasks in each window are sorted according to the task ID and marked with the allocation status (successfully allocated or waiting for allocation). Finally, a dynamic task allocation table is generated and saved in a CSV file, which contains the task ID, window ID, allocation status, node ID and priority label for UAV task scheduling simulation.

[0094] Perform task scheduling simulation on the dynamic task allocation table, and use the twin interaction situation diagram to evaluate the scheduling feasibility of the scheduling simulation results to obtain scheduling simulation verification data;

[0095] In this embodiment, based on the dynamic task allocation table, a scheduling simulation method based on multi-agent simulation technology is used to simulate task execution. Multi-agent simulation technology regards each drone as an independent simulation agent by building a multi-agent collaborative simulation environment. Each agent has its own task status (executing, waiting to be executed, completed), dynamic behavior characteristics (speed, remaining power, attitude angle) and interaction characteristics (edge ​​relationship with nodes in the twin interaction situation diagram). In the simulation environment, a multi-agent simulation framework (such as FLAME or Mesa framework) is used to build a simulation scene, and the dynamic behavior characteristics of the nodes in the twin interaction situation diagram are imported into the simulation system. The simulation interaction network is constructed through the edge weight relationship between the nodes. The simulation time step is set to 1 second, and the total simulation time is set to 300 seconds. In each simulation cycle: (1) the drone agent performs task scheduling according to the task allocation result of the dynamic task allocation table, (2) the status of the drone agent is updated in real time according to information such as task priority and remaining power, and (3) the simulation agents interact with data and coordinate tasks through the edge relationship of the twin interaction situation diagram. After each simulation time step, the UAV's task execution status, remaining battery power, task completion rate, average task execution time, and number of event risk triggers are collected to generate scheduling simulation output data. This data is then compared with the event node data in the twin interaction situation diagram, using the event risk trigger rate and task completion rate as feasibility assessment indicators. Scheduling is considered feasible if the task completion rate is greater than 95% and the average event risk trigger rate is less than 0.05; otherwise, the next step, iterative optimization of the dynamic task allocation table, is initiated. The simulation results are output as a CSV file containing fields such as node ID, task ID, simulation time, task execution status (0 / 1), remaining battery power, average event risk trigger rate, and a scheduling feasibility flag (true / false). This is used for scheduling feasibility assessment and subsequent task scheduling optimization.

[0096] The dynamic task allocation table is iteratively optimized using scheduling simulation verification data until the scheduling performance is greater than a preset performance threshold, thereby obtaining a dynamic task scheduling table.

[0097] In this embodiment, based on scheduling simulation verification data, tasks with a completion rate below 95% are reallocated. A performance threshold is set at 95%, and the task allocation table with a scheduling completion rate below the threshold is iteratively optimized. Each round of optimization adjusts the task allocation order and re-divides the event-driven scheduling window. The task allocation order is adjusted by up to 10% per iteration to prevent large-scale changes from affecting system stability. After the optimized allocation table is simulated and verified, if the scheduling performance is greater than 95%, the final dynamic task scheduling table is output as a CSV file containing the task ID, final allocation window ID, task priority, node ID, and optimization round label for invocation by the drone execution terminal.

[0098] Of particular importance is the assessment of drone mission priorities:

[0099] Perform multi-dimensional deconstruction on the task-event fusion feature matrix, extract task behavior characteristics and event dynamic trigger indicators, and generate a task-event multi-dimensional deconstruction dataset;

[0100] In this embodiment, the output task event fusion feature matrix is ​​subjected to multidimensional data processing, the matrix is ​​expanded dimension by dimension, and the task behavior characteristics (including task type, task execution time, task completion status, task risk value) and event dynamic trigger indicators (including event type, trigger probability, trigger frequency, trigger timestamp) contained therein are hierarchically deconstructed. During deconstruction, the task behavior characteristics and event dynamic trigger indicators are extracted separately using the matrix column index method and rearranged in the form of row vectors to form a multidimensional deconstructed data set. An example of the data set structure is as follows: a task behavior feature matrix (4 columns × N rows) and an event dynamic trigger indicator matrix (4 columns × N rows), each row corresponding to a task event sample; through row vector splicing, a task-event multidimensional deconstructed data set is generated, and the output format is a CSV file. The fields include task ID, task type, task execution time, task risk value, event type, trigger probability, trigger frequency, and trigger timestamp.

[0101] Perform time series modeling on the task behavior characteristics in the task-event multidimensional deconstruction dataset to generate task dynamic behavior vectors;

[0102] In this embodiment, the task behavior feature matrix in the output task-event multidimensional deconstruction dataset is grouped moment by moment. The data is divided into time series subsets with 5-second time steps according to the timestamp field, and time segmentation is completed using a sliding window. Then, within each time step, the task behavior feature vectors are stacked and reorganized to construct a task dynamic behavior vector set. This vector set adopts a four-dimensional structure: task type, task execution time, task completion status, and task risk value. Each vector represents the average task dynamic behavior status within the current time step. The output data is stored in JSON format, with fields including the time step ID, task dynamic behavior vector, and the corresponding drone task number.

[0103] Perform spatiotemporal modeling based on the event dynamic trigger indicators in the task dynamic behavior vector and output the event dynamic trigger vector;

[0104] In this embodiment, based on the generated task dynamic behavior vector set, time step information and event dynamic trigger indicators are used for spatiotemporal modeling. First, the time step ID of each task dynamic behavior vector is mapped one-to-one with the trigger timestamp field of the event dynamic trigger indicator, and matched according to the spatial distribution label (drone ID or task area ID). Then, through matrix reorganization, the event dynamic trigger vector is generated. Each vector contains the event type, trigger probability, trigger frequency and the corresponding spatial node ID. The event dynamic trigger vector is output in matrix form with a dimension of (4 columns × M rows), where M is the number of event samples.

[0105] Perform multi-dimensional feature concatenation on the task dynamic behavior vector and the event dynamic trigger vector, and set the task behavior weight to 0.6 and the event dynamic trigger weight to 0.4 to calculate the task-event comprehensive impact score, and output the task-event comprehensive impact score set;

[0106] In this embodiment, the task dynamic behavior vector and the event dynamic trigger vector are concatenated vector by vector according to the drone ID. The concatenation method uses column concatenation to construct a task-event joint feature vector, with each vector dimension being 8 (4 dimensions for task dynamic behavior + 4 dimensions for event dynamic trigger). The task behavior feature weight is then set to 0.6, and the event dynamic trigger feature weight is set to 0.4. Each joint feature vector is element-wise multiplied and weighted summed to calculate the task-event comprehensive impact score. The output is a task-event comprehensive impact score set in JSON format, with fields including task ID, drone ID, task-event comprehensive impact score, and timestep ID.

[0107] The task-event comprehensive impact score set is classified according to the corresponding UAV task number, and the UAV task priority score is calculated based on the mission event characteristics of each UAV. The task priority score is normalized to generate a UAV task priority dataset.

[0108] In this example, the task-event comprehensive impact score set is grouped and aggregated by drone ID. The task event score for each drone is classified and statistically analyzed by task ID, and the average comprehensive impact score for each task is calculated. Then, based on the drone task event characteristics (task type, task priority, and event risk level), the average task score is multiplied by a characteristic coefficient (e.g., a task type weight of 0.3, a task priority weight of 0.5, and an event risk level weight of 0.2) to obtain the drone task priority score. Finally, the drone task priority scores are normalized, with the normalization threshold set to 1.0. This generates a drone task priority dataset and outputs it as a CSV file with fields including drone ID, task ID, and drone task priority score (a decimal between 0 and 1).

[0109] Optionally, the agent's trajectory strategy in step S3 is specifically as follows:

[0110] Analyze tourists' photography needs based on tourist photography order data, and encode the dynamic task scheduling table and tourists' photography needs into tags to generate a tourist task tag set;

[0111] In this embodiment, personalized tourist photography requirements are extracted from scenic spot photography order data, including fields such as attraction location, shooting angle, desired time, shooting scene type, and visitor duration. Using string parsing and field matching, this requirement data is converted into a structured matrix. The matrix structure is set as a six-dimensional matrix: tourist ID, attraction ID, desired shooting time, shooting angle, scene type, and duration, with each row corresponding to a tourist order. Next, the task ID, task type, task execution duration, task priority, and scheduling time fields in the dynamic task scheduling table are labeled and encoded with the tourist ID in the tourist demand matrix, using the tourist ID as the primary key for label alignment. The labeled encoding format uses a one-hot format, with the label vector dimension being (tourist ID × task ID × task type dimension). Finally, a tourist task label set is generated and output as a JSON file. The fields include the tourist ID, task ID, labeled encoding vector, and label weight parameter (default is 1.0, adjustable). The tourist task label set here refers to the task set generated by real-time ordering and is different from the drone's existing or currently executing task set, and is chronologically ordered.

[0112] According to the tourist task label set and the dynamic task scheduling table, task features are fused to obtain a multi-objective task feature matrix;

[0113] In this embodiment, the generated tourist task tag set is fused with the dynamic task scheduling table for task features. In the fusion step, the tourist task tag set and the dynamic task scheduling table are first connected by row index based on the tourist ID, and the task features such as task priority, task type, task duration, and scheduling time in the scheduling table are spliced ​​row by row. The tourist task label vector (such as one-hot encoding) and the task feature vector (including task priority, task type, etc.) are spliced ​​into a multi-target task feature matrix by matrix splicing. The matrix dimension is (tourist ID × task ID × feature dimension), where the feature dimension includes the task label dimension (such as 10 dimensions) and the task feature dimension (such as 5 dimensions), totaling 15 dimensions. The output is in CSV format, including the tourist ID, task ID, and each column field of the multi-target task feature matrix.

[0114] Matching drone tasks with tourist orders is performed based on the multi-objective task feature matrix to obtain the initial task allocation strategy set;

[0115] In this embodiment, the multi-objective task feature matrix is ​​grouped by tourist ID, using the tourist ID as the primary key. The task ID and the task priority field in the task feature matrix are combined to match drone tasks to tourist orders. By comparing the tourist ID, task priority, and task label dimensions, the optimal task combination for each tourist is selected, and an initial task allocation strategy set is output. The initial task allocation strategy set structure is set as follows: tourist ID, drone ID, task ID, task priority, pre-allocated time, and task label code. The output format is a JSON file, and a filtering threshold (e.g., 0.7) is set for the task priority, retaining only high-priority tasks to ensure task execution efficiency.

[0116] Based on the initial task allocation strategy set, the real-time scenic environment data and the UAV dynamic flight behavior data are used to plan the UAV trajectory path to obtain the initial trajectory planning data set;

[0117] In this embodiment, the generated initial task allocation strategy set is associated with real-time scenic area environmental data (such as scenic area weather, visitor density, and traffic conditions) and UAV dynamic flight behavior data (including flight speed, current location, remaining battery power, and flight heading). First, using the UAV ID as the primary key, the initial task allocation strategy and the UAV flight behavior data are concatenated to generate an initial UAV flight behavior set. Subsequently, trajectory path modeling is performed using the initial UAV flight behavior set in conjunction with the scenic area environmental data. A modified 3D GridMap is used for scene modeling. The map resolution is set to 1 meter, and the grid data format is (X, Y, Z, environmental factors). Environmental factors include terrain elevation, obstacle weight, visitor density index, and meteorological risk factors. To enhance the dynamic adaptability of trajectory generation, a dynamic A-replanning technique based on A-search is selected, integrating UAV battery constraints and trajectory safety indicators. Each initial trajectory is represented as a node sequence, with node attributes including X, Y, Z coordinates, flight time, flight speed, remaining battery power, and risk score. During the planning process, the mission's starting point is used as the root node. A dynamic heuristic function (including remaining flight distance, power consumption rate, and risk indicators) is used to generate track nodes layer by layer. The resulting track node matrix has a dimension of (number of nodes × 7-dimensional features). To ensure flight safety, the obstacle crossing threshold is set to 0.2 (a higher value allows for detours). A node-by-node risk score is accumulated for each track, ultimately outputting the initial track planning dataset in a JSON file format. The fields include the drone ID, mission ID, track node matrix, and dynamic risk score.

[0118] Perform multi-objective coupling optimization based on the initial trajectory planning data set and the initial task allocation strategy set to obtain the task trajectory optimization data set;

[0119] In this embodiment, the generated initial trajectory planning data set and the initial task allocation strategy set are jointly modeled, and the track weight parameters and task priorities in the initial trajectory node matrix are used to perform multi-objective coupling optimization. The initial trajectory planning data set and the initial task allocation strategy set are subjected to multi-objective coupling optimization, and the core purpose is to take into account task priority, trajectory safety and energy consumption constraints at the same time. First, the risk score, node height change and track length in the initial trajectory node matrix are calculated as a single-objective indicator vector (track risk vector, energy consumption vector, path length vector), and then the task priority vector is introduced to construct a multi-objective optimization vector set. In order to achieve collaborative optimization of multiple objectives, a multi-objective optimization technology based on non-dominated sorting (NSGA-II) is adopted for model construction. The model consists of a Pareto frontier extraction subset and a crowding distance sorting module to ensure the uniformity of multiple solutions. The model is structured as a four-layer architecture: the input layer includes task priority, trajectory risk, energy consumption, and path length (four dimensions). The first layer weights task priorities (with a weight threshold of 0.5), the second layer normalizes risk (with a normalization factor threshold of 0.85), and the third layer normalizes energy consumption (with a normalization threshold of 0.9). The output layer calculates a comprehensive optimization weight vector (with a dimension of 1), groups and aggregates tasks by task ID and drone ID, and outputs an optimized coupling weight matrix (with dimensions of task ID × drone ID × optimization weight). The output is saved in JSON format, including the task ID, drone ID, optimized trajectory node matrix, optimized weight matrix, and risk indicator, for subsequent dynamic task scheduling and simulation.

[0120] The initial task allocation strategy set and the task trajectory optimization dataset are jointly modeled, and the Internet of Things is used to perform real-time visualization and dynamic scheduling simulation on the joint modeling results to generate a multi-agent strategy set.

[0121] In this embodiment, the initial task allocation strategy set and the task trajectory optimization data set are jointly modeled. During modeling, a multidimensional data structure is used to splice the task ID, drone ID, task priority, trajectory node matrix and coupling weight matrix row by row to form a joint modeling data set with the dimension of (task ID × drone ID × feature dimension). The feature dimension includes 20-dimensional features such as task label, task priority, trajectory node, optimization weight, etc. Subsequently, the IoT edge computing node is used to push the joint modeling data to the visualization system in real time, and a three-dimensional dynamic scheduling simulation is constructed in combination with the Unity3D visualization engine. The scene simulation is used to display the drone task execution status, trajectory planning and tourist distribution information in real time. At the same time, the scheduling simulation process is analyzed for performance and a multi-agent strategy set is generated. The output format is a visualization scene file and a JSON data set, which contains task ID, drone ID, task priority, scheduling performance indicators and visualization node data.

[0122] Optionally, step S4 is specifically:

[0123] Step S41: Deconstruct the multi-agent strategy set to extract the spatiotemporal interaction relationship and strategy features between drones to obtain a multi-agent feature matrix;

[0124] In this example, the multi-agent strategy set is deconstructed based on each drone's mission characteristics, dynamic behavior data, and historical execution records to extract the spatiotemporal interaction relationships and strategy features between drones. First, the real-time mission data of each drone is parsed from the strategy set, extracting timestamps, spatial location information (latitude, longitude, altitude), flight speed, and heading angle. A node vector of the form [UAV_i = [t, x, y, z, v, θ, P]] is generated, where t represents the timestamp, x, y, z represent the three-dimensional coordinates, v represents the speed, θ represents the heading angle, and P represents the mission priority. Then, combining the interactions between drones (such as communication links, collision avoidance distances, and dynamic formation relationships), an interaction matrix W_ij is constructed between drone pairs, recording the real-time distance, signal strength, and task collaboration weight between drone pairs. Finally, all drone node vectors are concatenated with the interaction matrix to form a multi-agent feature matrix M with dimensions [N x (7 + N)], where N is the number of drones. Each row in the matrix contains the node features of a single drone and the interaction features with other drones.

[0125] Step S42: performing global trajectory feasibility analysis and constraint modeling based on the multi-agent feature matrix and scenic area environment data to obtain a global trajectory constraint set;

[0126] In this embodiment, a comprehensive analysis is performed using the multi-agent feature matrix M and the digital twin environmental data of the scenic area. The environmental data includes a three-dimensional terrain model, obstacle point cloud data (such as buildings, trees, and no-fly zone grids), and meteorological data (wind speed, wind direction, temperature, and humidity). First, the real-time positions of the multi-agents are spatially mapped with the terrain model to generate a visual obstacle map. The flight safety area is then calculated using a collision detection module (based on Voxel-Grid discretization). Second, the dynamic flight envelope of each drone is extracted (a minimum safety radius R_min = 5m is set based on the drone size and safety distance requirements), and a dynamic flight restriction area is generated by combining it with the environmental meteorological data. Finally, a graph convolutional structure is used to spatially correlate the obstacle grid and the drone dynamic envelope, outputting a global trajectory constraint set A, which includes information such as the three-dimensional obstacle grid, meteorological area restrictions, flight altitude restrictions (H_min = 20m, H_max = 120m), minimum turning radius and speed constraints (V_min = 3m / s, V_max = 20m / s) for subsequent trajectory planning.

[0127] Step S43: performing global trajectory planning based on the global trajectory constraint set to generate an initial global trajectory matrix;

[0128] In this embodiment, a trajectory generation method based on a hierarchical random tree (L-RRT) combined with an improved ant colony search is used for global trajectory planning. First, based on the global trajectory constraint set A, the scenic area is spatially divided into a navigation layer (for global path guidance) and a local obstacle avoidance layer (for real-time dynamic obstacle avoidance between drones). The navigation layer samples nodes using a L-RRT to prioritize feasible path segments and prioritizes nodes using a local weighted heuristic factor f_h (set to 0.75). The local obstacle avoidance layer embeds an ant colony search sub-process within the initial path segments of the navigation layer. It refines and dynamically fine-tunes the path based on pheromone intensity and path length penalty weights to ensure path safety and smoothness. Finally, the path segments of the navigation and obstacle avoidance layers are combined to generate the initial global trajectory matrix P, which is in the form of [P = {p_ij}] (i is the drone number, j is the track point number). Each p_ij contains three-dimensional coordinates (x, y, z), flight speed v, and heading angle θ.

[0129] Step S44: performing multi-objective coupled optimization based on the initial global trajectory matrix and combining the multi-agent strategy set with the scenic area environment data to obtain global trajectory optimization data;

[0130] In this embodiment, a multi-objective coupled optimization is performed using an initial global trajectory matrix P, a multi-agent strategy set, and scenic area environmental data. First, a multi-objective function (task completion time, energy consumption, risk exposure rate, and flight distance) is set based on the drone's mission priority, energy consumption requirements, and scenic area passenger flow hotspot density. The target weight for mission completion time is 0.4, the target weight for energy consumption is 0.3, the target weight for risk exposure rate is 0.2, and the target weight for flight distance is 0.1. Next, based on a multi-objective evolutionary optimization framework, the drone's mission priority data is used as the initial population seed. Combined with a distributed multi-objective ant colony collaboration mechanism, dynamic pheromone regulation (pheromone volatility is set to 0.65) and a multidimensional constraint update module are used to achieve iterative multi-objective optimization of the global path. During the optimization process, the node speeds, heading angles, and obstacle avoidance radius of the trajectory matrix P are updated in real time, generating an optimized global trajectory matrix P_opt. This matrix has the same format as the initial matrix but incorporates improved multi-objective performance metrics for use in subsequent scheduling steps.

[0131] Step S45: Dynamically distribute and schedule the global trajectory optimization data, generate an individualized execution instruction set for the drone, and send the individualized execution instruction set to each drone terminal through the edge node deployed by the Internet of Things.

[0132] In this embodiment, based on the global track optimization matrix P_opt, the track nodes of each drone are analyzed node by node, and the individualized execution instruction set of the drone is generated by combining the task execution progress and real-time environmental dynamics. Each instruction set contains: node serial number, flight coordinates (x, y, z), target speed v, heading angle θ, task trigger flag and dynamic obstacle avoidance weight. To ensure real-time and security, the MQTT protocol message queue service is deployed through the IoT edge node, and the individualized execution instruction set of the drone is packaged according to priority and encrypted for transmission. The IoT edge node cluster sets the heartbeat detection cycle to 2s to ensure that the drone can receive task updates and dynamic adjustment instructions in real time during the task execution process, realizing closed-loop control and dynamic scheduling of task execution.

[0133] It is particularly important that step S45 is specifically as follows:

[0134] Step S451: Perform multi-dimensional deconstruction on the global track optimization data to extract the individualized flight track of the UAV, dynamic posture change characteristics, track key point sequence, and dynamic interaction weight associated with the scenic area environment, and generate individualized dynamic track data of the UAV;

[0135] In this embodiment, when performing multi-dimensional deconstruction on the global trajectory optimization data, the global trajectory optimization data containing the flight trajectory of each drone is first read through the data structured interface. Each trajectory data is stored in the four-dimensional array A = [x, y, z, t] in the form of a three-dimensional coordinate point sequence (x, y, z) and a timestamp. In addition, the trajectory optimization data also contains the drone attitude angle information (pitch angle θ, roll angle and yaw angle ψ), these attitude information are integrated into the three-dimensional matrix In order to extract dynamic posture changes. For the track key point sequence, points where the curvature change rate is greater than a threshold of 0.05 are selected as key points and sorted in track timestamp order to form a track key point matrix C = [x, y, z, k], where k represents the key point number. Scenic area environmental data uses a multi-layer geographic information database to extract the dynamic interaction weights of scenic areas that intersect with the track path in space or are less than 20 meters away. These are marked in the track dataset as environmental dynamic interaction features in the form of W = [w1, w2, ..., wn]. Finally, the integrated matrix D = [A, B, C, W] is used as the individualized dynamic track data of the drone for subsequent analysis.

[0136] Step S452: performing hierarchical multi-objective weighted sorting based on the individualized dynamic track data of the UAV and the UAV mission priority data set to obtain a UAV track priority list;

[0137] In this embodiment, when the obtained UAV individualized dynamic track data is subjected to hierarchical multi-objective weighted sorting with the UAV task priority dataset, the task priority dataset is first normalized and all task priority scores are mapped to the range of 0 to 1. In the flight track matrix A in the UAV individualized dynamic track data, the task priority associated with each track is correlated with the flight stability in the dynamic attitude change matrix B (given by θ, A comprehensive analysis is performed using the standard deviation of ψ (calculated from the flight stability score and the task priority score). The flight stability weight is set to 0.3, and the task priority weight is set to 0.7. Each track is weighted and ranked using the formula P = 0.7 * T + 0.3 * (1 - σ), where T is the task priority score and σ is the standard deviation of flight attitude variation. The resulting ranking result is stored as a matrix P[i] = priority score, track ID, forming a priority list for the UAV track, which is used for subsequent task scheduling.

[0138] Step S453: Analyze the spatial conflict relationship of the UAV tracks in combination with the UAV track priority list and the global track optimization data, and perform dynamic scheduling allocation on the spatial conflict relationship to obtain the UAV allocation track matrix;

[0139] In this embodiment, after combining the drone track priority list with the global track optimization data, the spatial conflict relationship between the tracks is analyzed using a three-dimensional spatial conflict analysis module. Specifically, for any two tracks A[i] and A[j], the three-dimensional distance calculation formula D = sqrt((xi-xj)^2+(yi-yj)^2+(zi-zj)^2) is used, and the spatial conflict threshold is set to 10 meters. When the distance between the two tracks in the same time period is less than the threshold, it is determined to be a conflict. By establishing a conflict relationship matrix M, where M[i][j] = 1 indicates that there is a conflict between track i and track j, otherwise it is 0. Then, a hierarchical scheduling mechanism based on conflict relationships is adopted to allocate time slices according to the priority order in the track priority list, and offset adjustments are made to low-priority tracks in the conflict relationship matrix (each offset does not exceed 10% of the track interval) to avoid conflicts, and the output drone track allocation matrix N = [track ID, time slice, adjustment amplitude].

[0140] Step S454: Based on each UAV assigned track in the UAV assigned track matrix and the corresponding track key points and dynamic posture change characteristics, the UAV execution instructions are assigned by integrating the flight safety zone information in the scenic area environment data to generate an individualized UAV execution instruction set;

[0141] In this embodiment, based on the obtained drone track distribution matrix, each track and the corresponding track key point sequence are extracted together with the dynamic attitude change characteristics, and the flight safety area information in the scenic area environment data is combined to generate the distribution instructions. Specifically, for each key point in each track, the scenic area flight safety area matrix S = [area ID, spatial coordinate range] is queried, and it is determined whether the key point is located inside the safety area. If it is not in the safety area, it is marked as a risk point, and a safety adjustment instruction is added to the flight instruction (for example, limiting the flight speed to 60% of the maximum speed). Each drone execution instruction contains a track ID, a list of key points, an attitude adjustment value, and a safety mark, and finally outputs the drone's individualized execution instruction set E = [track ID, key point sequence, attitude parameters, safety instructions].

[0142] Step S455: The drone individualized execution instruction set is distributed and distributed through the edge nodes deployed by the Internet of Things to execute the drone control task.

[0143] In this embodiment, when the individualized execution instruction set of the drone is distributed and distributed through the edge nodes deployed by the Internet of Things, the MQTT message middleware is first configured on the edge node, the topic is set to / uav / control / command, and QoS = 2 is used to ensure reliable transmission of instructions. Each instruction set is encapsulated in a JSON structure, including track ID, key point sequence, attitude instructions and safety control instructions. The edge node is pre-configured with a local fault-tolerant processing mechanism. After each instruction is successfully sent, an ACK confirmation is returned. Otherwise, it will automatically retry three times, each with an interval of 200ms. The drone terminal uses PX4 firmware and ROS interface to receive and parse instructions, write them into the local control buffer, and execute them in real time in the flight control system. Through the real-time monitoring module of the edge node, the drone status and execution feedback information are synchronously transmitted back to support subsequent scheduling simulation and performance evaluation.

[0144] Optionally, step S5 is specifically:

[0145] Step S51: Using the Internet of Things to call the high-resolution camera of each drone to collect real-time image data, and preliminarily classify and label the image data to generate a real-time drone image set;

[0146] In this embodiment, multiple drones are deployed in the scenic flight area. Each drone is equipped with a high-resolution camera (model: Sony IMX700 sensor, resolution: 8192×4320 pixels, frame rate: 30fps). The IoT edge computing node calls the MQTT protocol to subscribe to the topic / uav / camera / stream to obtain real-time image data streams. During flight, the drone uses GPS and IMU data to dynamically sample camera attitude parameters (including pitch, roll, and yaw). These attitude parameters and image frames are then labeled at the frame level with a tag structure of [frame ID, shooting time, location coordinates, attitude parameters]. The image data is initially classified by the edge node into categories such as "panoramic", "close-up", and "aerial photography". Each type of image data is labeled, for example, "panoramic shooting" is labeled T1, "close-up shooting" is labeled T2, and so on. Finally, a real-time drone-captured image set is generated. The data format is a multi-layer index structure: image ID, label, location coordinates, attitude parameters, and image data file path, which facilitates subsequent order label mapping and analysis.

[0147] Step S52: combining the tourist order data with the real-time drone image set to perform label alignment mapping to obtain tourist order mapping data;

[0148] In this embodiment, in the generated real-time drone-photographed image set, each image data contains a frame ID, a label, a position coordinate, and attitude parameters, and the tourist order data is stored in JSON format, including the order ID, shooting time, the desired shooting angle (pitch, yaw), the desired position coordinate, and the order label. Through the data alignment module deployed on the edge node of the Internet of Things, multi-dimensional label mapping is performed based on the shooting timestamp and the position coordinates, and each frame of data in the image set is aligned with the tourist order according to the matching criteria of a time difference of less than 2 seconds and a spatial distance of less than 5 meters. The data after label alignment is stored in the form of a mapping matrix M = [order ID, image ID, matching score, label] and bound to the original order data, and the tourist order mapping data is output for subsequent shooting angle error analysis.

[0149] Step S53: Calculating the angle difference between the tourist's desired shooting angle and the actual shooting angle based on the tourist order mapping data to obtain shooting order error data;

[0150] In this embodiment, based on the output tourist order mapping data, for each tourist order, the angle difference is calculated by parsing the expected shooting angle (pitch angle, yaw angle) contained in the order and the actual shooting angle in the drone image matching the order (derived from IMU data). The angle difference is expressed in matrix form as E = [order ID, pitch angle difference, yaw angle difference], where the pitch angle difference calculation formula is: Δθ = |θ expected - θ actual|, and the yaw angle difference calculation formula is: Δψ = |ψ expected - ψ actual|. To ensure the rationality of the shooting error, a threshold is set. When Δθ or Δψ is greater than 15 degrees, it is marked as "needs adjustment", otherwise it is marked as "qualified". Each error data is associated with the order ID and image ID, and finally the shooting order error dataset is output, providing basic data for the construction of the subsequent angle correction model.

[0151] Step S54: constructing an adaptive angle correction model based on the shooting order error data and combined with the real-time acquired UAV dynamic parameters;

[0152] In this embodiment, based on the shooting order error data, the real-time dynamic parameters of the drone are collected, including flight speed, acceleration, attitude change rate, and real-time adjustment capability of the camera (such as gimbal response rate). When constructing an adaptive angle correction model, based on the multi-input dynamic parameter matrix P = [speed, acceleration, attitude change rate, gimbal response rate], P and the error matrix E are subjected to multi-dimensional data analysis to construct a multi-target dynamic angle correction function. This function superimposes the influence coefficients of each dynamic parameter in the form of linear weighting (for example, speed weight 0.3, acceleration weight 0.2, attitude change rate weight 0.25, gimbal response rate weight 0.25), and dynamically generates a correction coefficient matrix K = [order ID, corrected pitch angle coefficient, corrected yaw angle coefficient] according to the threshold. The model structure is represented by a multi-layer data flow graph (DataFlowGraph), and each node corresponds to a dimensional input and weighted output of a dynamic parameter. The output model supports angle correction based on the real-time input of dynamic parameters.

[0153] Step S55: Utilize the adaptive angle correction model to perform optimized shooting parameter inference on the UAV dynamic parameters and tourist order mapping data, obtain the UAV optimized shooting parameter set, and distribute the UAV optimized shooting parameter set to each UAV terminal.

[0154] In this embodiment, based on the obtained adaptive angle correction model, the correction coefficient matrix K is applied to each image record in the tourist order mapping data to dynamically generate an optimized shooting parameter set O = [order ID, image ID, corrected pitch angle, corrected yaw angle, flight speed correction]. The optimized shooting parameter set is distributed and distributed via the MQTT protocol using the topic / uav / control / photo_adjust. Each instruction is packaged in a JSON structure, including the track ID, image ID, corrected camera attitude, and gimbal response speed. The drone terminal receives the shooting instruction set through the ROS node and calls the local shooting control interface to dynamically adjust the camera angle to ensure that the shooting angle is consistent with the tourist's expectations. The edge node sets QoS to 2 to ensure reliable data delivery, monitors the execution status in real time, and implements closed-loop control through feedback messages.

[0155] Optionally, this specification also provides an Internet of Things-based UAV collaborative control system, which is used to execute the Internet of Things-based UAV collaborative control method described above. The Internet of Things-based UAV collaborative control system includes:

[0156] The multimodal integrated perception module is used to collect scenic area sensor data through the Internet of Things, and perform multimodal integrated perception based on the scenic area sensor data to obtain an integrated situation map of the scenic area environment and a situation map of the drone cluster;

[0157] The task allocation module is used to dynamically evolve the drone-scene based on the integrated situation map of the scenic area environment and the drone cluster situation map to obtain a twin interaction situation map; the twin interaction situation map is used to perform event-driven dynamic scheduling to obtain a dynamic task allocation table;

[0158] The trajectory strategy optimization module is used to obtain tourist photo-taking order data through the Internet of Things, and iteratively optimize the agent trajectory strategy of the dynamic task allocation table based on the tourist photo-taking order data to obtain a multi-agent strategy set;

[0159] The global flight path optimization module is used to collaboratively optimize the global flight path of the drone cluster based on the multi-agent strategy set, obtain the drone's individual execution instruction set, and use the edge nodes deployed by the Internet of Things to distribute the drone's individual execution instruction set to each drone terminal;

[0160] The shooting angle correction module is used to collect real-time drone shooting image sets, and evaluate the order shooting angle offset of the real-time drone shooting image sets based on tourist order data to obtain shooting order error data; use the shooting order error data to perform angle offset adaptive correction to obtain the drone optimized shooting parameter set, and then send the drone optimized shooting parameter set to each drone terminal.

[0161] The present invention is therefore intended to be illustrative and non-restrictive in all respects, with the scope of the invention being defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the application documents are intended to be embraced therein.

[0162] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. A UAV collaborative control method based on the Internet of Things, characterized in that: The following steps are involved: Step S1: collecting scenic area sensor data through the Internet of Things, and performing multimodal integrated perception based on the scenic area sensor data to obtain a scenic area environment integrated situation map and a drone cluster situation map; Step S2: Based on the integrated situation map of the scenic area environment and the situation map of the drone cluster, the drone-scene dynamic evolution is performed to obtain a twin interaction situation map; the twin interaction situation map is used to perform event-driven dynamic scheduling to obtain a dynamic task allocation table; Step S3: Obtain tourist photo-taking order data through the Internet of Things, and iteratively optimize the agent trajectory strategy of the dynamic task allocation table based on the tourist photo-taking order data to obtain a multi-agent strategy set; Step S4: Based on the multi-agent strategy set, the global flight path collaborative optimization of the drone cluster is performed to obtain the drone individual execution instruction set, and the drone individual execution instruction set is distributed to each drone terminal using the edge node deployed by the Internet of Things; Step S5: Collect a real-time drone image set, and perform order shooting angle offset evaluation on the real-time drone image set based on the tourist order data to obtain shooting order error data; use the shooting order error data to perform angle offset adaptive correction to obtain the drone optimized shooting parameter set, and send the drone optimized shooting parameter set to each drone terminal.

2. The method for cooperative control of unmanned aerial vehicles based on the Internet of Things according to claim 1, characterized in that: Step S1 is specifically as follows: Step S11: collecting scenic area sensor data through the IoT edge node set deployed in the scenic area, and converting the scenic area sensor data into a format to obtain a multi-source heterogeneous original data set; Step S12: performing multimodal data preprocessing and feature standardization on the multi-source heterogeneous original data sets to obtain a standardized multimodal feature data set; Step S13: performing feature alignment and time series synchronization on the standardized multimodal feature dataset to obtain multimodal time series fusion data; Step S14: performing multimodal feature fusion on the multimodal time series fusion data, and performing spatial correlation analysis based on the feature fusion results to obtain a multimodal spatial correlation matrix; Step S15: Construct a scenic area situation map based on the multimodal spatial association matrix classification, thereby obtaining a scenic area environment integrated situation map and a drone cluster situation map.

3. The method for cooperative control of unmanned aerial vehicles based on the Internet of Things according to claim 2, characterized in that: Step S15 is specifically as follows: Step S151: performing spatial node partitioning on the multimodal spatial association matrix, using the node partitioning results as independent graph subgraphs, and calculating the edge weights between spatial nodes to obtain a multimodal regional subgraph; Step S152: Perform community segmentation and hierarchical analysis based on the multimodal regional sub-graph to obtain multimodal community structure data; Step S153: Calculate the community comprehensive risk using the node and edge weights in each community structure in the multimodal community structure data to obtain the scenic area environment community risk matrix; Step S154: combining the risk values ​​in the scenic area environment community risk matrix and the community division results, performing multi-layer graph modeling on each community to obtain a multi-layer scenic area environment graph; Step S155: extracting the real-time dynamic behavior features of the drones based on the multimodal time series fusion data, embedding the real-time dynamic behavior features of the drones into the multi-layer scenic area environment map, modeling the dynamic interaction behavior between drones, and obtaining the dynamic behavior data of the drone cluster; Step S156: Based on the multi-layer scenic area environment map and the dynamic behavior data of the drone cluster, joint visualization and map layer mapping are performed to obtain an integrated situation map of the scenic area environment, and a drone cluster situation map is constructed based on the community dynamic characteristics in the scenic area environment community risk matrix and the dynamic behavior data of the drone cluster.

4. The method for cooperative control of UAVs based on the Internet of Things according to claim 3, characterized in that: Step S153 is specifically as follows: Extract node features in each community based on multimodal community structure data to obtain a set of community node feature vectors; The normalization factor threshold is set to 0.85, and the edge weights of the community structure are normalized using the node spatial distribution characteristics and time dynamic fluctuation characteristics of the community node feature vector to obtain the normalized community edge weight matrix; The fusion dimension is set to 64, and the community node feature vector set and the community edge weight normalized matrix are fused to obtain the community node interaction feature data; Utilize the interactive characteristic data of community nodes to perform multi-index weighted risk calculation of node risk, thereby calculating the comprehensive risk value of each point, and weighting the comprehensive risk value to generate the community node risk weight matrix; Based on the community node risk weight matrix, the risk weight values ​​in each community are aggregated to obtain the scenic area environment community risk matrix.

5. The method for cooperative control of UAVs based on the Internet of Things according to claim 1, characterized in that: The dynamic evolution of the drone-scene in step S2 is specifically as follows: Perform real-time twin scene mapping of the scenic area environment integrated situation map and the drone cluster situation map, extract environmental elements, and obtain a twin scene environmental element set; The UAV dynamic flight behavior data is collected through the Internet of Things, and the feature set of the UAV dynamic flight behavior data is aligned with the twin scene environment element set to obtain the UAV dynamic behavior feature set; Combining the twin scene environmental element set with the UAV dynamic behavior feature set to model the scene temporal evolution relationship and obtain the UAV-scene dynamic relationship map; Based on the UAV-scene dynamic relationship graph, the trigger conditions of scene dynamic events are extracted, and a multi-dimensional dynamic event triggering mechanism is constructed in combination with the UAV dynamic behavior feature set; The multi-dimensional dynamic event triggering mechanism is used to simulate the event triggering of the UAV-scene dynamic relationship graph to obtain the UAV-scene dynamic interaction triggering data; The UAV-scene dynamic interaction trigger data is used in combination with the UAV-scene dynamic relationship map to deduce the real-time interaction situation between the UAV and the environment and obtain a twin interaction situation map.

6. The method for cooperative control of unmanned aerial vehicles based on the Internet of Things according to claim 1, characterized in that: The event-driven dynamic scheduling in step S2 is specifically as follows: Based on the twin interaction situation graph, dynamic behavior nodes and event trigger conditions are extracted. The execution node-event condition graph is convolutionally aggregated based on the real-time task execution status in the UAV dynamic flight behavior data to obtain the task event fusion feature matrix. The UAV mission priority is evaluated using the mission event fusion feature matrix to obtain the UAV mission priority dataset; Based on the UAV mission priority dataset, combined with the event sequence and node dynamic behavior characteristics in the twin interaction situation diagram, event-driven time window division is performed to obtain the event-driven scheduling window set; Combine the event-driven scheduling window set and the UAV task priority data set to perform UAV task-event window matching allocation and obtain a dynamic task allocation table; Perform task scheduling simulation on the dynamic task allocation table, and use the twin interaction situation diagram to evaluate the scheduling feasibility of the scheduling simulation results to obtain scheduling simulation verification data; The dynamic task allocation table is iteratively optimized using scheduling simulation verification data until the scheduling performance is greater than a preset performance threshold, thereby obtaining a dynamic task scheduling table.

7. The method for cooperative control of UAVs based on the Internet of Things according to claim 1, characterized in that: The specific trajectory strategy of the agent in step S3 is: Analyze tourists' photography needs based on tourist photography order data, and encode the dynamic task scheduling table and tourists' photography needs into tags to generate a tourist task tag set; According to the tourist task label set and the dynamic task scheduling table, task features are fused to obtain a multi-objective task feature matrix; Matching drone tasks with tourist orders is performed based on the multi-objective task feature matrix to obtain the initial task allocation strategy set; Based on the initial task allocation strategy set, the real-time scenic environment data and the UAV dynamic flight behavior data are used to plan the UAV trajectory path to obtain the initial trajectory planning data set; Perform multi-objective coupling optimization based on the initial trajectory planning data set and the initial task allocation strategy set to obtain the task trajectory optimization data set; The initial task allocation strategy set and the task trajectory optimization dataset are jointly modeled, and the Internet of Things is used to perform real-time visualization and dynamic scheduling simulation on the joint modeling results to generate a multi-agent strategy set.

8. The method for cooperative control of unmanned aerial vehicles based on the Internet of Things according to claim 1, characterized in that: Step S4 is specifically as follows: Step S41: Deconstruct the multi-agent strategy set to extract the spatiotemporal interaction relationship and strategy features between drones to obtain a multi-agent feature matrix; Step S42: performing global trajectory feasibility analysis and constraint modeling based on the multi-agent feature matrix and scenic area environment data to obtain a global trajectory constraint set; Step S43: performing global trajectory planning based on the global trajectory constraint set to generate an initial global trajectory matrix; Step S44: performing multi-objective coupled optimization based on the initial global trajectory matrix and combining the multi-agent strategy set with the scenic area environment data to obtain global trajectory optimization data; Step S45: Dynamically distribute and schedule the global trajectory optimization data, generate an individualized execution instruction set for the drone, and send the individualized execution instruction set to each drone terminal through the edge node deployed by the Internet of Things.

9. The method for cooperative control of unmanned aerial vehicles based on the Internet of Things according to claim 1, characterized in that: Step S5 is specifically as follows: Step S51: Using the Internet of Things to call the high-resolution camera of each drone to collect real-time image data, and preliminarily classify and label the image data to generate a real-time drone image set; Step S52: combining the tourist order data with the real-time drone image set to perform label alignment mapping to obtain tourist order mapping data; Step S53: Calculating the angle difference between the tourist's desired shooting angle and the actual shooting angle based on the tourist order mapping data to obtain shooting order error data; Step S54: constructing an adaptive angle correction model based on the shooting order error data and combined with the real-time acquired UAV dynamic parameters; Step S55: Utilize the adaptive angle correction model to perform optimized shooting parameter inference on the UAV dynamic parameters and tourist order mapping data, obtain the UAV optimized shooting parameter set, and distribute the UAV optimized shooting parameter set to each UAV terminal.

10. A UAV collaborative control system based on the Internet of Things, characterized in that: The method for controlling a UAV based on the Internet of Things according to claim 1 is used to execute the method, wherein the UAV coordinated control system based on the Internet of Things comprises: The multimodal integrated perception module is used to collect scenic area sensor data through the Internet of Things, and perform multimodal integrated perception based on the scenic area sensor data to obtain an integrated situation map of the scenic area environment and a situation map of the drone cluster; The task allocation module is used to dynamically evolve the drone-scene based on the integrated situation map of the scenic area environment and the drone cluster situation map to obtain a twin interaction situation map; the twin interaction situation map is used to perform event-driven dynamic scheduling to obtain a dynamic task allocation table; The trajectory strategy optimization module is used to obtain tourist photo-taking order data through the Internet of Things, and iteratively optimize the agent trajectory strategy of the dynamic task allocation table based on the tourist photo-taking order data to obtain a multi-agent strategy set; The global flight path optimization module is used to collaboratively optimize the global flight path of the drone cluster based on the multi-agent strategy set, obtain the drone's individual execution instruction set, and use the edge nodes deployed by the Internet of Things to distribute the drone's individual execution instruction set to each drone terminal; The shooting angle correction module is used to collect real-time drone shooting image sets, and evaluate the order shooting angle offset of the real-time drone shooting image sets based on tourist order data to obtain shooting order error data; use the shooting order error data to perform angle offset adaptive correction to obtain the drone optimized shooting parameter set, and then send the drone optimized shooting parameter set to each drone terminal.

Images