An industrial unmanned aerial vehicle cluster intelligent scheduling management method based on multi-data analysis

Abstract

The application relates to the technical field of unmanned aerial vehicle cluster control and intelligent scheduling, and discloses an industrial unmanned aerial vehicle cluster intelligent scheduling management method based on multi-data analysis, which comprises the following steps: extracting a quantitative risk feature vector based on multi-source operation data; constructing and updating a four-dimensional dynamic risk field; constructing a rolling time domain optimization model to generate a cooperative scheduling scheme containing adaptive formation adjustment; decomposing and issuing scheduling instructions; and updating the dynamic risk field based on real-time feedback data and triggering compensatory adjustment. The application can realize real-time sensing and quantitative evaluation of multi-source and dynamic risks in a complex industrial environment, and dynamically balance flight safety and task efficiency through a rolling optimization and elastic compensation mechanism, thereby improving the robustness and intelligence of overall operation of the cluster.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) swarm control and intelligent scheduling technology, and in particular to an intelligent scheduling and management method for industrial UAV swarms based on multi-data analysis. Background Technology

[0002] With the widespread application of industrial drones in fields such as inspection, surveying, and logistics, multi-drone swarm collaborative operations have become an important means of improving efficiency.

[0003] Chinese invention patent publication number CN119962783A discloses a drone swarm management system and method based on intelligent algorithms. The method includes: acquiring operational data for each drone and attribute data for each flight mission across various dimensions; constructing constraints and a path optimization objective function based on the load distribution and time-dimensional attribute data in the operational data of each drone, combined with the drone flight path; constructing a path node graph based on a grid map; updating pheromones using Tent chaotic mapping and ant colony algorithm; and obtaining the optimal path planning scheme for the drone swarm based on the path optimization objective function. While this invention improves the collaborative efficiency of drone swarms through the joint optimization of task allocation and path planning, the combination of intelligent algorithms, considering obstacles, flight area limitations, and task priorities, enables real-time dynamic adjustment of path planning, ensuring efficient and safe flight of drones in complex environments.

[0004] However, factors such as electromagnetic interference, variable airflow, dense obstacles, and the performance degradation of UAVs themselves in complex industrial environments pose significant challenges to the safe and efficient scheduling of swarms. Existing scheduling methods mostly focus on static path planning or simple coordination rules, lacking real-time fusion analysis and quantitative assessment of multi-source, dynamic risk data, making it difficult to achieve a dynamic balance between risk avoidance and task efficiency at the global level. When sudden interference or deviations occur during task execution, existing methods often have a delayed response and limited adjustment capabilities, leading to task failure or increased risk.

[0005] Therefore, there is an urgent need for an intelligent cluster management method that can perceive risks in real time, dynamically optimize scheduling, and has elastic compensation capabilities. Summary of the Invention

[0006] To address these issues, this invention provides an intelligent scheduling and management method for industrial drone swarms based on multi-data analysis. This method overcomes the problems in existing technologies, such as the lack of real-time quantitative assessment and fusion of multi-source dynamic risks, the difficulty in dynamically balancing safety and efficiency in global scheduling, and the lag in response to sudden interference and deviations during execution, as well as insufficient compensation capabilities.

[0007] To achieve the above objectives, this invention provides an intelligent scheduling and management method for industrial drone swarms based on multi-data analysis, comprising:

[0008] Step S1: Obtain risk characteristic data for different interference types of each UAV based on multi-source operational data of the UAVs in the cluster. The multi-source operational data includes at least environmental perception data and UAV operational status data.

[0009] The types of interference include electromagnetic interference, airflow disturbance, geographical obstacles, and degradation of machine performance;

[0010] Step S2: Based on the preset interference type identification model, perform fusion analysis on the multi-source operation data and extract at least one quantitative risk feature vector of the interference type;

[0011] Step S3: Based on the local risk index of each UAV, an interpolation method is used to generate and update a four-dimensional dynamic risk field covering the entire mission area in real time to characterize the comprehensive risk intensity distribution in spatial location and time dimension. The local risk index is calculated and characterized based on the real-time location of each UAV and the corresponding quantified risk feature vector.

[0012] Step S4: Construct a rolling time-domain optimization model including dynamic risk field constraints to determine the collaborative scheduling scheme of the cluster in the next time domain, wherein...

[0013] The rolling time-domain optimization model is constructed by minimizing a comprehensive objective function that integrates the overall risk of the cluster and the efficiency of task completion.

[0014] The cooperative scheduling scheme is determined based on the optimal task allocation sequence, the optimal cooperative flight path, and the adaptive formation topology. The adaptive formation topology is dynamically adjusted according to the gradient of the dynamic risk field and the distribution of task points.

[0015] Step S5: Decompose the scheduling scheme into scheduling instructions for each UAV. Each UAV executes the task based on the instruction set and continuously transmits updated multi-source operation data back.

[0016] Step S6: Update the dynamic risk field based on the updated data transmitted back, and trigger compensatory adjustments for incomplete or interfered tasks according to the updated dynamic risk field and task execution status. The compensatory adjustments include reallocating tasks that are delayed or failed due to risk interference, and / or correcting the paths of affected drones.

[0017] Further, in step S2, based on a preset interference type identification model, the multi-source operation data is fused and analyzed to extract a quantitative risk feature vector, specifically including:

[0018] Using a trained neural network model, feature fusion and pattern recognition are performed on the environmental perception data and the UAV operating status data to output risk probability and intensity estimates corresponding to different types of interference, thus forming the quantified risk feature vector.

[0019] Furthermore, in step S3, an interpolation method is used to generate and update the four-dimensional dynamic risk field in real time, specifically as follows:

[0020] The Kriging spatial interpolation method is adopted, with the real-time position of each UAV as the interpolation point, its local risk index as the observation value, and combined with the geographic information system data of the mission area as the background field, to calculate and generate the predicted risk intensity value of any spatial location in the entire mission area in the current and short-term future time, so as to construct the four-dimensional dynamic risk field.

[0021] Furthermore, in step S4, a rolling time-domain optimization model is constructed based on the weighted summation structure of the minimized global risk value, task completion time, and cluster energy consumption value. The global risk value is represented by the result of integrating the risk intensity value of the predicted path point in the dynamic risk field.

[0022] Furthermore, the constraints of the rolling time-domain optimization model include:

[0023] The constraints include UAV dynamics constraints, endurance constraints, temporal logic constraints between tasks, collision avoidance constraints between UAVs, and communication connectivity constraints required by the adaptive formation topology.

[0024] Furthermore, in step S4, the dynamic adjustment of the adaptive formation topology specifically involves:

[0025] When the gradient value of the dynamic risk field exceeds the first preset threshold, the formation will be adjusted from a loose formation to a tight formation to enhance the cooperative anti-interference capability.

[0026] When the dispersion of task points exceeds the second preset threshold, the formation will be split from a centralized formation into multiple sub-cluster formations to execute tasks in parallel.

[0027] Further, in step S5, the scheduling scheme is decomposed and issued as scheduling instructions for each of the UAVs, specifically including:

[0028] The optimal task allocation sequence, optimal cooperative flight path, and formation instructions of the whole system are solved into independent waypoint sequences, speed instructions, and formation relative coordinates for each UAV, and then distributed through the cluster communication network.

[0029] Furthermore, in step S6, the conditions for triggering the compensatory adjustment include:

[0030] The actual position of any drone was detected to deviate from the planned path from the allowable tolerance.

[0031] Based on updated multi-source operational data, sudden high-risk areas that were not predicted in the dynamic risk field were identified;

[0032] Received a report of mission failure or drone malfunction.

[0033] Furthermore, the compensatory adjustment in step S6 specifically includes:

[0034] Starting from the current moment, freeze the tasks that have been successfully completed. With the remaining unfinished tasks, available drone resources, and the latest dynamic risk field as input, rerun the rolling time-domain optimization model to generate an adjusted scheduling scheme for the remaining task phase.

[0035] Optionally, the method further includes recording the entire process data to form a historical knowledge base after the completion of a task, which is used for offline training and tuning of the parameters of the interference type identification model and the rolling time domain optimization model.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] By comprehensively utilizing multi-source operational data, including environmental perception data and UAV status data, and performing feature fusion and pattern recognition based on a trained neural network model, it is possible to output risk probability and intensity estimates for different types of interference, forming a high-dimensional quantitative risk feature vector. This overcomes the shortcomings of existing methods, which suffer from single, lagging risk perception and lack of quantitative basis. By outputting a quantitative risk feature vector through a neural network, it provides a high-precision and computable input data foundation for subsequent risk field construction and optimized scheduling.

[0038] Furthermore, by using the local risk index of each UAV, a four-dimensional dynamic risk field covering the entire mission area is generated and updated in real time using Kriging spatial interpolation. This enables the extrapolation and prediction of the comprehensive risk intensity distribution at any location within the entire mission area, both now and in the future, from sparse UAV detection points using geostatistical methods. This allows the cluster to obtain a global and forward-looking risk perspective that surpasses the perception capabilities of a single UAV, providing a unified situation map for collaborative avoidance decision-making.

[0039] Furthermore, by constructing a rolling time-domain optimization model with multiple objectives of minimizing global risk, task completion time, and cluster energy consumption, and incorporating constraints such as UAV dynamics, endurance, collision avoidance, and communication connectivity, as well as the dynamic risk field itself, into the optimization framework, this model can solve for the optimal task allocation, cooperative path, and formation topology that satisfy all safety and physical constraints in real time. This achieves an automated and intelligent trade-off between cluster scheduling security, task execution efficiency, and energy economy in complex dynamic environments.

[0040] Furthermore, by forming a closed loop of perception-decision-execution-feedback, it is possible not only to calculate the optimized solution into precise individual instructions and issue them for execution, but also to continuously update the risk field based on real-time feedback data, and to trigger compensatory adjustments when conditions such as path deviation, sudden risks or task failures are detected. On this basis, by quickly re-running the optimization model, the remaining tasks are replanned, thereby effectively dealing with various uncertainties and interferences in the execution process, and ensuring that the task can still be promoted or completed efficiently and safely in the event of partial failure or sudden environmental changes.

[0041] Furthermore, by automatically triggering the formation to switch between tight formation and dispersed sub-cluster modes based on the gradient of the dynamic risk field and the distribution divergence of task points, the overall stability is enhanced when collaborative anti-interference is required, and execution efficiency is improved when parallel operations are needed. In addition, by building a historical task knowledge base and offline training and tuning of the parameters of the interference identification model and optimization model, the entire scheduling management system can learn from historical experience, continuously optimize its risk identification accuracy and scheduling strategy, and possess long-term performance evolution capabilities. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating the steps of an intelligent scheduling and management method for industrial drone swarms based on multi-data analysis, as described in an embodiment of the present invention.

[0043] Figure 2 This is a flowchart illustrating the steps involved in constructing a multi-task deep neural network model according to an embodiment of the present invention.

[0044] Figure 3 This is a logic block diagram illustrating the dynamic adjustment of the adaptive formation topology in this invention.

[0045] Figure 4 The flowchart illustrates the steps of decomposing the scheduling scheme and issuing scheduling instructions to each UAV in accordance with the present invention. Detailed Implementation

[0046] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0047] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0048] Please see Figure 1 The diagram shown is a flowchart illustrating the steps of an intelligent scheduling and management method for industrial drone clusters based on multi-data analysis, according to an embodiment of the present invention.

[0049] A method for intelligent scheduling and management of industrial drone swarms based on multi-data analysis includes:

[0050] Step S1: Obtain risk characteristic data of different interference types for each drone based on multi-source operational data of drones in the cluster. The multi-source operational data includes at least environmental perception data and drone operational status data.

[0051] Interference types include electromagnetic interference, airflow disturbance, geographical obstacles, and degradation of machine performance;

[0052] Step S2: Based on the preset interference type identification model, perform fusion analysis on multi-source operation data and extract quantitative risk feature vectors of at least one type of interference.

[0053] Step S3: Based on the local risk index of each UAV, an interpolation method is used to generate and update a four-dimensional dynamic risk field covering the entire mission area in real time to characterize the comprehensive risk intensity distribution in spatial location and time dimension. The local risk index is calculated and characterized based on the real-time location of each UAV and the corresponding quantitative risk feature vector.

[0054] Step S4: Construct a rolling time-domain optimization model including dynamic risk field constraints to determine the collaborative scheduling scheme of the cluster in the next time domain, wherein...

[0055] A rolling time-domain optimization model is constructed by minimizing a comprehensive objective function that integrates the overall risk of the cluster and the efficiency of task completion.

[0056] The collaborative scheduling scheme is determined based on the optimal task allocation sequence, the optimal collaborative flight path, and the adaptive formation topology. The adaptive formation topology is dynamically adjusted according to the gradient of the dynamic risk field and the distribution of task points.

[0057] Step S5: Decompose the scheduling scheme into scheduling instructions for each UAV. Each UAV executes the task based on the instruction set and continuously transmits updated multi-source operation data back.

[0058] Step S6: Update the dynamic risk field based on the updated data transmitted back, and trigger compensatory adjustments for incomplete or interfered tasks according to the updated dynamic risk field and task execution status. The compensatory adjustments include reallocating tasks that are delayed or failed due to risk interference, and / or correcting the paths of affected UAVs.

[0059] Specifically, in step S2, based on a preset interference type identification model, multi-source operation data is fused and analyzed to extract a quantitative risk feature vector, which includes:

[0060] By using a trained neural network model, feature fusion and pattern recognition are performed on environmental perception data and UAV operational status data to output risk probability and intensity estimates corresponding to different types of interference, thus forming a quantitative risk feature vector.

[0061] Please see Figure 2 The diagram shown is a flowchart illustrating the steps involved in constructing a multi-task deep neural network model according to an embodiment of the present invention.

[0062] In this embodiment of the invention, the interference type identification model is a multi-task deep neural network model trained offline. The core function of this model is to realize the fusion analysis of heterogeneous multi-source data and output a structured risk quantification assessment. The specific implementation method is as follows:

[0063] 1. Model Input and Preprocessing:

[0064] Environmental perception data includes, but is not limited to, point cloud data from airborne millimeter-wave radar used to detect geographical obstacles, image data from visual sensors used to identify terrain and visible obstacles, meteorological sensor data, including wind speed and direction, used to sense airflow disturbances, and radio frequency signal strength data used to assess the electromagnetic interference environment. All of the above data have undergone time alignment and normalization preprocessing.

[0065] Drone operating status data includes real-time data from the flight control system, such as battery voltage / current, which reflects power supply stability and remaining range; motor speed and temperature, which reflect the health of the power system; attitude angle and angular velocity, which reflect flight stability; and self-diagnostic error codes.

[0066] 2. Neural Network Model Architecture and Fusion Strategy:

[0067] The model employs a multi-branch feature fusion network based on an attention mechanism, specifically:

[0068] Feature extraction branch: For different types of data, such as images, point clouds, and time-series sensor data, set up dedicated sub-networks, such as Convolutional Neural Network (CNN), PointNet, and Long Short-Term Memory (LSTM), to perform preliminary feature extraction.

[0069] Feature Fusion and Interaction Layer: This layer concatenates the primary features extracted from each branch and introduces a cross-modal attention module. This module can dynamically calculate the correlation weights between features from different data sources. For example, it strengthens the correlation between attitude data and wind speed data features during periods of strong wind, thereby more accurately identifying airflow disturbances.

[0070] Multi-task output layer: The fused high-level features are input to multiple parallel sub-output layers. Each sub-output layer corresponds to a specific type of interference, such as electromagnetic interference or airflow disturbance, and its output consists of two parts:

[0071] Risk probability: This indicates the likelihood of this type of interference occurring in the current state. In this embodiment, the risk probability is between 0 and 1.

[0072] Intensity estimate: This indicates the expected level of impact or physical quantity of the disturbance if it occurs, such as the expected wind force or the number of times the electromagnetic field strength exceeds the standard.

[0073] 3. Composition of the quantitative risk feature vector:

[0074] The quantified risk feature vector is a structured data vector whose dimension is equal to twice the number of identified interference types (risk probability + intensity estimate). For example, in a system containing four types of interference—electromagnetic, airflow, obstacles, and performance degradation—this vector can be represented as:

[0075] Where P represents the risk probability (ranging from 0 to 1), and I represents the normalized intensity estimate (e.g., 0-1 or actual physical units). This vector serves as the unified risk input interface for subsequent risk field construction and optimization models.

[0076] 4. Model Training:

[0077] This neural network model is trained in a supervised manner using historical task data recordings of disturbance events. The loss function is a weighted sum of the losses for identifying each type of disturbance, ensuring that the model can accurately identify multiple risks simultaneously.

[0078] Specifically, in step S3, the four-dimensional dynamic risk field is generated and updated in real time using interpolation, as follows:

[0079] The Kriging spatial interpolation method is adopted, with the real-time position of each UAV as the interpolation point, its local risk index as the observation value, and combined with the geographic information system data of the mission area as the background field, to calculate and generate the predicted risk intensity value of any spatial location in the entire mission area in the current and short-term future time, so as to construct a four-dimensional dynamic risk field.

[0080] In this embodiment of the invention, the construction of the four-dimensional dynamic risk field is a spatial prediction and temporal extrapolation process based on geostatistics. Its core lies in using sparse, real-time measured UAV local risk indices to accurately estimate the continuous risk distribution of the entire mission area and predict its short-term evolution trend. The specific implementation method is as follows:

[0081] 1. Input data preparation:

[0082] Interpolation points and observations: Real-time GPS / BeiDou coordinates of each UAV This serves as a spatial interpolation point. The local risk index calculated at this location... This serves as the observed value at that point. Local risk index. The local risk index is calculated by weighting and aggregating the quantitative risk feature vector corresponding to the UAV obtained in step S2. The weights reflect the severity of the impact of different interference types on overall flight safety. Calculated using the following formula:

[0083]

[0084] Where N is the number of interference types, This is the intensity estimate of the k-th type of interference. The weighting coefficients are set according to the corresponding types of interference. The weighting coefficients can be set based on historical risk event data.

[0085] Background field data: Integrating geographic information system data of the task area, including but not limited to: Digital Elevation Model (DEM) to characterize terrain undulations, permanent obstacle layers such as high-voltage power line towers and buildings, and known distribution maps of fixed interference sources such as radar stations and communication base stations. This data serves as prior knowledge, used to correct and constrain the results of spatial interpolation, making them more consistent with physical reality.

[0086] 2. Kriging space interpolation calculation:

[0087] Using the ordinary kriging method as the core algorithm, the process includes the following steps:

[0088] Variation function modeling: analyzing the local risk index among all UAV observation points Spatial autocorrelation. A variogram model is calculated and fitted, which quantitatively describes how the risk index changes with increasing spatial distance; for example, the closer the distance, the higher the risk correlation. This is the key difference between Kriging and simple linear or inverse distance interpolation, as it fully utilizes the spatial structure information of the data.

[0089] Optimal Unbiased Estimation: For any point p(x,y,z) to be predicted within the task region, the Kriging method solves a system of equations based on the above variogram model, which is a set of known observation points. Assign a set of optimal weights This set of weights satisfies the conditions of unbiased estimation and minimum estimation variance. The predicted risk intensity value at point p. The result is obtained through weighted summation: Furthermore, this method can also provide the kriging variance of the estimate, serving as a measure of prediction uncertainty.

[0090] 3. Introduction of the time dimension:

[0091] To construct a four-dimensional risk field, time is treated as another dimension:

[0092] Short-term forecasting: Based on the dynamic risk field sequence of the current and previous few moments, time series forecasting methods, such as linear extrapolation or simple autoregressive models, are used to estimate the trend of risk intensity at fixed spatial grid points in the future short time domain, such as the next decision-making cycle.

[0093] Generating a four-dimensional field: The final four-dimensional dynamic risk field F(x,y,z,t) can be represented as a data cube, which contains the interpolation results of the risk intensity of each spatial point at the current time, as well as the predicted values ​​of the risk intensity of each spatial point at several future times.

[0094] 4. Dynamic update mechanism:

[0095] The risk field is not static, but updated at a frequency that matches the drone data transmission and scheduling decision cycle. Whenever new drone swarm status data is received, specifically new locations and local risk indices, a complete Kriging interpolation calculation is triggered, thereby refreshing the risk distribution map of the entire space with the latest risk sampling, realizing the dynamic nature of the risk field.

[0096] The Kriging spatial interpolation method specifically includes: constructing a variogram model based on the real-time location of the UAV, combining geographic information system data as the background field, calculating the predicted risk intensity value of any location within the task area through an unbiased optimal estimation method, and introducing a time series extrapolation method to construct a four-dimensional dynamic risk field.

[0097] Specifically, in step S4, a rolling time-domain optimization model is constructed based on the weighted summation structure of the minimized global risk value, task completion time, and cluster energy consumption value. The global risk value is represented by the result of integrating the risk intensity value of the predicted path point in the dynamic risk field.

[0098] In this embodiment of the invention, the rolling time-domain optimization model is a mixed integer optimization problem with multiple objectives and constraints. Its core is to plan the future time domain based on the latest dynamic risk field F(x,y,z,t) and cluster state within each decision cycle. The optimal coordinated scheduling scheme within the system. Its specific implementation is as follows:

[0099] 1. Model Inputs and Decision Variables

[0100] Model input:

[0101] The four-dimensional dynamic risk field F(x,y,z,t) is generated by step S3 and includes risk intensity predictions in both spatial and temporal dimensions.

[0102] Task set: Includes the location, priority, time window, and operational requirements of all tasks to be performed (such as inspection points and transportation targets).

[0103] Drone cluster status: current location, remaining battery power, maximum flight time, maximum payload, sensor status, etc. of each drone.

[0104] Formation cooperation constraints: such as minimum safe distance, maximum communication distance, minimum number of drones required for collaborative operations, etc.

[0105] Decision variables:

[0106] Task assignment variable: This is a binary variable that indicates which drone will perform which task at a specific time.

[0107] Path variables: These are continuous variables, representing a sequence of three-dimensional waypoints for each UAV u within the planning time domain. The path can be expressed as a function of time t. Its physical meaning is the predicted position coordinates of the UAV u at time t.

[0108] Formation topology variables: can be represented by an adjacency matrix A, where elements This indicates that drones u and v are directly connected in formation (such as a leader-follower relationship); otherwise, it is 0.

[0109] 2. Definition of Multi-Objective Weighted Summation Structure

[0110] The model's comprehensive objective function It consists of a weighted sum of three key indicators:

[0111] Global risk value The risk exposure of each UAV's predicted flight path in a dynamic risk field is calculated by integrating the risk. Specifically, for the predicted path of UAV u... Defined by the path variable, its risk intensity value at time t is The global risk value is defined as the sum of the risk indices of all drone paths:

[0112]

[0113] in, This represents the global risk value.

[0114] To sum over all drones u in the cluster;

[0115] From the current moment To the future time domain Time integral;

[0116] : Dynamic risk field function, the input is the predicted position of UAV u at time t Given time t, the output is the risk intensity value at that spatiotemporal point;

[0117] Time element;

[0118] This design ensures that the scheduling scheme can automatically avoid high-risk areas and time periods.

[0119] Task completion time Its purpose is to maximize task completion efficiency, and it is defined as the weighted sum of the completion times of all tasks (or critical tasks). The weights can be set based on task priority; the higher the priority, the greater the penalty for delayed completion.

[0120] Cluster energy consumption Its purpose is to extend the overall operating time of the cluster, which is defined as the sum of the predicted energy consumption of all drones. The energy consumption model is calculated based on flight distance, speed changes and load.

[0121] Finally, the overall objective function is:

[0122]

[0123] in, These are adjustable weighting coefficients that reflect the trade-offs between safety, efficiency, and battery life. In one exemplary implementation, a set of baseline weights can be determined using offline simulation, with the average completion rate of historical tasks and the level of risk exposure as optimization metrics, and a Bayesian optimization method employed.

[0124] In this embodiment, the following settings are provided: In practical applications, adjustments can be made based on this benchmark according to the safety level and efficiency requirements of specific tasks.

[0125] 3. Core constraints of the model

[0126] To ensure the feasibility and safety of the solution, the model includes the following main constraints:

[0127] Dynamics and maneuverability constraints: The path of each UAV must meet physical constraints such as maximum speed, maximum acceleration, and turning radius.

[0128] Endurance constraint: The planned flight range and operating time of each drone must not exceed the maximum endurance supported by its current battery power.

[0129] Time and logical constraints: There are dependencies between tasks (e.g., task B must start after task A is completed) and fixed time window requirements.

[0130] Collision avoidance constraint: At any given time, the three-dimensional distance between any two drones must be greater than a preset minimum safe distance. The minimum safe distance is the core parameter in the collision avoidance constraint, and its setting requires comprehensive consideration of factors such as the drone's physical dimensions, positioning error, control error, and external interference. The determination process is as follows:

[0131] Basic safety distance: determined based on the maximum dimensions of the drone (including rotor or wingspan). It is generally taken as 2 to 3 times the maximum diagonal length of the drone to ensure sufficient physical clearance.

[0132] For example, for a multi-rotor UAV with a wheelbase of 500mm, its It can be set as follows:

[0133] =3 × 0.5m = 1.5m

[0134] Error compensation: Considering GPS / RTK positioning error (e.g., ±0.05m), attitude control error, and instantaneous position drift caused by gusts (e.g., ±0.2m), a safety margin is added. :

[0135] =Positioning error + Control error + Wind disturbance estimate = 0.05 + 0.1 + 0.2 = 0.35m

[0136] Final safe distance:

[0137]

[0138] In practical applications, it can be rounded up to 2.0m.

[0139] Communication connectivity constraints: When performing formation missions that require coordination, the relevant UAVs must maintain a communicable distance to ensure the connectivity of the formation topology.

[0140] Dynamic risk field constraints: The path variables of the model must satisfy the objective function. It is computable, meaning the path point must lie within the domain of the dynamic risk field F. More importantly, by... As part of the optimization objective, the model directly incorporates the dynamic risk field as an environmental potential field into the optimization framework, driving the generated path to proactively avoid high-risk areas.

[0141] 4. Rolling Time Domain Optimization Framework and Solution

[0142] Rolling time-domain mechanism: The model is re-solved every fixed decision cycle; in this embodiment, it is set to be solved every 10 seconds. During each solution, the time domain is optimized. Scroll forward, but only implement the optimal decision for the current moment. Next waypoint. As new data is acquired, the dynamic risk field and cluster state are updated, and the model solves again in the next cycle based on the new information, thus achieving adaptive scheduling.

[0143] Solution Strategy: Since this model is a complex nonlinear mixed integer programming problem, this embodiment adopts a hierarchical-decomposition heuristic solution strategy for efficient approximate solution, specifically as follows:

[0144] High-level task allocation and formation decision-making: Using a distributed auction algorithm or a hybrid genetic algorithm based on a contract network protocol, task allocation and coarse-grained formation patterns are quickly determined.

[0145] Fine-grained path planning at the bottom layer: Under fixed task allocation and formation mode, for each UAV or each group of UAVs, a model predictive control framework combined with gradient descent method is used to optimize a smooth, low-risk flight path that meets all constraints under the guidance of a dynamic risk field.

[0146] Implementation of adaptive formation topology adjustment: During the optimization process, the model dynamically adjusts the topology based on the spatial gradient of F (edge ​​of high-risk zone) and the spatial distribution of task points. and The weights are set, and the formation topology variables are allowed to switch between several preset modes, which in this implementation are centralized diamond, dispersed star, and multiple sub-clusters, in order to achieve a balance between safely passing through high-risk areas and efficiently covering dispersed task points.

[0147] Specifically, the constraints of the rolling time-domain optimization model include:

[0148] The constraints include UAV dynamics constraints, endurance constraints, temporal logic constraints between tasks, collision avoidance constraints between UAVs, and communication connectivity constraints required by the adaptive formation topology.

[0149] Please see Figure 3 As shown, it is a logic block diagram of the dynamic adjustment of adaptive formation topology in this invention.

[0150] Specifically, in step S4, the dynamic adjustment of the adaptive formation topology is as follows:

[0151] When the gradient value of the dynamic risk field exceeds the first preset threshold, the formation will be adjusted from a loose formation to a tight formation to enhance the cooperative anti-interference capability.

[0152] When the dispersion of task points exceeds the second preset threshold, the formation will be split from a centralized formation into multiple sub-cluster formations to execute tasks in parallel.

[0153] In this embodiment of the invention, the specific implementation of the constraints for the rolling time-domain optimization model is as follows:

[0154] 1. Dynamics constraints of unmanned aerial vehicles

[0155] This constraint ensures that the planned path is physically feasible. In the model, this is typically achieved by limiting the continuity of the UAV's motion and the limits of its maneuverability, specifically...

[0156] Continuity constraint: Predicted location of the UAV ,speed It is a continuous function of time t.

[0157] Velocity and acceleration limits: At any time t, satisfying and ,in and These are the maximum speed and maximum acceleration of this drone model, respectively.

[0158] Turning radius limitation: For fixed-wing UAVs, the instantaneous radius of curvature of its path must be greater than the minimum turning radius R_min.

[0159] 2. Battery life constraints

[0160] This constraint ensures that drones do not lose contact or crash due to depleted power. It is achieved by limiting the total energy consumption of each drone to no more than its remaining available energy.

[0161] Predicted total energy consumption per drone Estimated by calculation models based on flight distance, hovering time, and payload power, the following must be met: ,in η represents the current remaining battery power, and η is the safety factor. In this embodiment, η = 0.8, which is used to reserve battery power for emergency return.

[0162] 3. Timing constraints between tasks

[0163] This constraint ensures that the order in which tasks are executed conforms to business logic. It is reflected by defining the temporal relationships between tasks:

[0164] Sequential relationship: Task B can only begin after Task A is completed, that is... .

[0165] Time window: A specific task j must be performed within a preset time window. Execute internally.

[0166] Exclusive resource usage: At any given time, at the same location or on the same equipment, such as a specific take-off and landing platform, only one task can occupy the space.

[0167] 4. Collision avoidance constraints between drones

[0168] This constraint is central to flight safety, ensuring that any two drones maintain a safe distance at all times.

[0169] For any two different drones u and v, at any time t within the planning time domain, the following must be satisfied:

[0170]

[0171] Among them, d_safety is the three-dimensional minimum safety interval set according to factors such as UAV size, speed, and positioning error.

[0172] 5. Communication connectivity constraints

[0173] This constraint guarantees the reliability of the cluster's internal communication network when performing formation tasks that require data sharing or collaborative control.

[0174] Point-to-point connectivity: For a UAV pair (u,v) using a specific formation topology (such as leader-follower), the following conditions must be met during the time period in which the relationship needs to be maintained:

[0175]

[0176] in, This represents the maximum reliable distance for inter-machine communication links.

[0177] Overall network connectivity: For formation modes that require global information synchronization, the communication topology of the entire UAV cluster must remain connected throughout the planning time domain, meaning that there must be at least one path consisting of several communication links between any two UAVs.

[0178] In this embodiment of the invention, the specific implementation method for the dynamic adjustment of the adaptive formation topology is as follows:

[0179] The dynamic adjustment of formation topology, as a high-level decision-making step in the optimization model, has the following triggering and execution logic:

[0180] Trigger condition calculation: Real-time calculation of the gradient field of the current dynamic risk field F(x,y,z,t) in space. The gradient magnitude The formula for calculating at any point (x, y, z) in space is:

[0181]

[0182] It is the modulus of the rate of change of risk intensity in three spatial directions, and is a dimensionless scalar value (rate of change of risk intensity).

[0183] First preset threshold Quantification determination method:

[0184] 1. Data Preparation: Collect full-process data from multiple historical typical tasks, and extract the gradient magnitudes calculated at all times and within the entire task region. The sample set.

[0185] 2. Statistical Analysis: Calculate the statistical distribution of the sample set. First preset threshold. The P-quantile is defined as the sample set, where P is a high percentile value set according to the safety redundancy requirement. It can be selected as P=85, 90 or 95. In this embodiment, P=85 is set.

[0186] Physical meaning: Choosing P=85 quantile means that in historical data, only 15% of spatiotemporal points have risk changes (gradients) exceeding this threshold. Setting this value as the trigger condition ensures that the system tightens formation only in the 15% of situations where the risk field changes most drastically and the potential danger is highest, thus achieving a balance between safety and formation flexibility.

[0187] Implementation Example: In one embodiment, based on data analysis of 10 inspection missions in a mining area, the gradient amplitude was determined. The 85th percentile is 0.5, therefore, it can be set to... The risk intensity value is a normalized dimensionless scalar with a range of [0,1].

[0188] Any point near the planned path of the drone When the real-time calculated value exceeds 0.5, the formation tightening adjustment is triggered.

[0189] Adjustment action: Trigger formation mode switch command to maintain the distance between aircraft at the normal collision avoidance safety distance. The formation was reduced to a distance between aircraft. A diamond or triangular formation, in which ;

[0190] The It can be preset according to the drone's performance and communication requirements, the purpose of which is:

[0191] Shorten the distance between machines and enhance collaborative sensing and anti-interference capabilities;

[0192] The overall spatial outline of the small team is reduced to facilitate joint passage through high-risk passages.

[0193] Model Implementation: In the optimization model, this adjustment is reflected in reducing the d_communication value in the communication connectivity constraint within the corresponding time period, and adjusting the value in the anti-collision constraint. Within permissible limits, the size is reduced, thereby driving the algorithm to generate tighter formation paths.

[0194] 2. Adjustment based on the dispersion of task points

[0195] Trigger condition calculation: Evaluate the geographic distribution of the remaining task points. Calculate the convex hull area of ​​these task point coordinates. and compare it with the total area of ​​the mission area. In comparison, the dispersion of task point distribution is obtained:

[0196]

[0197] when Exceeding the second preset threshold This indicates that the task points are too scattered;

[0198] In this embodiment, the following settings are provided: .

[0199] Adjustment Action: Triggers a formation split command, splitting a single, centralized large formation into two or more sub-clusters. Each sub-cluster contains at least... Deploy drones to ensure basic collaboration capabilities and assign them responsibility for a relatively concentrated mission area. The distance between sub-clusters should be greater than [missing information]. To avoid mutual interference, The value is usually greater than the distance between machines within the sub-cluster.

[0200] Model Implementation: In the optimized model, this adjustment is reflected as follows:

[0201] At the task allocation layer, task sets are clustered and divided according to geographical regions.

[0202] The drone resources are divided into several groups to form multiple independent optimization sub-problems (or the decision variables of the original problem are grouped and constrained).

[0203] Each sub-cluster maintains its own formation constraints, while the strict formation connectivity constraints between sub-clusters are lifted, leaving only the anti-collision constraints.

[0204] This adjustment can significantly improve the parallelism and overall efficiency of task completion.

[0205] Specifically, the first preset threshold Used to determine whether the dynamic risk field gradient is drastic. Its value can be determined by analyzing historical data: calculating the risk field gradient magnitude at all times in historical tasks. The statistical distribution will lead to cases where the task is high-risk or fails. The 85th percentile as The initial suggested value;

[0206] Second preset threshold To determine whether task points are too scattered, the ratio of the convex hull area formed by the task points to the total area of ​​the task region can be set. In this embodiment, splitting is triggered when the ratio is greater than 0.4.

[0207] Specifically, in step S6, the scheduling scheme is decomposed and issued as scheduling instructions for each UAV, including:

[0208] The optimal task allocation sequence, optimal cooperative flight path, and formation instructions are calculated into independent waypoint sequences, speed instructions, and formation relative coordinates for each UAV, and then distributed through the cluster communication network.

[0209] Please see Figure 4 The diagram shown is a flowchart illustrating the steps of decomposing the scheduling scheme and issuing scheduling instructions to each UAV in accordance with the present invention.

[0210] In this embodiment, the scheduling scheme is decomposed and issued as scheduling instructions for each UAV. The goal is to transform the abstract, global collaborative scheduling scheme generated in step S5 into a low-level, time-sequential, precise control instruction set that each UAV can independently understand and execute. The specific implementation method is as follows:

[0211] 1. Global scheduling scheme calculation and instruction generation:

[0212] According to the collaborative scheduling scheme, the central dispatch server generates an independent, timestamped instruction sequence for each UAV in the cluster. This sequence typically contains the following types of instructions:

[0213] Task instructions: Parse the portion of the globally optimal task allocation sequence belonging to UAV u into specific operational action instructions. For example: [Time t1: Arrive at inspection point A, activate sensor X, hover and collect data], [Time t2: Proceed to transport point B, perform grabbing operation].

[0214] Path instructions: The continuous trajectory P_u(t) belonging to UAV u in the globally optimal cooperative flight path is discretized and encoded into a series of four-dimensional waypoints with predetermined arrival time, speed, and heading angle. For example: [waypoint W1:(x1,y1,z1), arrival time t1, speed v1, heading head1]. This ensures precise spatiotemporal coordination among multiple UAVs.

[0215] Formation commands: Based on the adaptive formation topology, calculate and generate formation-maintaining relative coordinate commands for the UAV u. These may include:

[0216] Absolute position mode: When in close formation, directly specify its precise position in the formation reference frame.

[0217] Relative position mode: Specifies its relative distance, bearing, and elevation difference relative to the navigator or adjacent UAV in the formation, such as [Follow navigator U0, maintain relative position (Δx,Δy,Δz)].

[0218] Mode switching command: When the topology is adjusted, a formation mode switching command and new parameters are issued.

[0219] 2. Instruction encapsulation, verification, and distribution:

[0220] Instruction encapsulation: This process encapsulates the generated instructions. The data is encapsulated according to a predefined, simplified binary or JSON protocol format, and additional information such as frame header (command sequence number, target drone ID) and frame trailer (checksum) is added to form a transmission data packet.

[0221] Timing verification: Before distribution, the scheduling server will perform a final global verification of timing and resource conflicts for all drone instruction sets to ensure there are no contradictions.

[0222] Network distribution: Command packets are reliably distributed to each UAV via a dedicated time-division multiple access / frequency-division multiple access wireless communication network. For critical commands (such as emergency obstacle avoidance commands), a high-priority mechanism that may involve repeated transmission is employed to ensure delivery.

[0223] 3. Command reception and local execution for the drone:

[0224] Each drone's onboard controller receives and parses its own instruction packets.

[0225] The controller converts high-level waypoint and formation commands into real-time control signals for lower-level actuators, such as motor speed and rudder surface deflection, through a local closed-loop controller (such as a PID controller or model predictive controller).

[0226] The drone strictly follows the timestamps in the instructions to execute tasks and maintains millisecond-level time synchronization with the cluster through onboard clock synchronization mechanisms (such as the GPS-based PTP protocol), which is the foundation for achieving precise spatiotemporal coordination.

[0227] 4. Maintenance of the status feedback channel:

[0228] The command issuance link is also the uplink link for status feedback. While executing commands, each UAV compresses and packages its real-time status (position, speed, mission progress, sensor data) at a preset frequency and transmits it back to the central scheduler through the same communication network, providing data input for the dynamic risk field update and compensatory adjustment in step S6.

[0229] Specifically, in step S6, the conditions that trigger the compensatory adjustment include:

[0230] The actual position of any drone was detected to deviate from the planned path from the allowable tolerance.

[0231] Based on updated multi-source operational data, sudden high-risk areas that were not predicted in the dynamic risk field were identified;

[0232] Received a report of mission failure or drone malfunction.

[0233] In this embodiment, the decision period is 10 seconds, the optimization time domain length is 60 seconds, and the objective function weights are... By using offline simulation, benchmark values ​​can be obtained by using Bayesian optimization algorithms to fine-tune the data, with comprehensive evaluation indicators such as task completion rate and average risk exposure.

[0234] Specifically, the compensatory adjustment in step S6 is as follows:

[0235] Starting from the current moment, freeze the tasks that have been successfully completed, and take the remaining unfinished tasks, available drone resources, and the latest dynamic risk field as inputs to rerun the rolling time-domain optimization model and generate an adjusted scheduling scheme for the remaining task phases.

[0236] In this embodiment, compensatory adjustments are triggered for incomplete or disrupted tasks, ensuring the system maintains robustness and task completion rate in dynamic and uncertain environments—a core fault-tolerant mechanism. It is not simply replanning, but a closed-loop process based on real-time monitoring, intelligent judgment, and efficient optimization, specifically:

[0237] 1. Anomaly monitoring and adjustment triggering

[0238] The system continuously compares the task execution status with the expected status of the original scheduling plan. A compensatory adjustment process is triggered when any of the following quantitative conditions are met:

[0239] Schedule Deviation Condition: The completion time of any critical task is expected to be later than the originally planned time window by more than a threshold. .

[0240] Path deviation condition: The lateral or vertical deviation between the actual position of any UAV and the planned path continuously exceeds the preset tolerance range. .

[0241] Risk mutation condition: Based on the latest backhaul data, the dynamic risk field shows that a risk intensity value that was not predicted in the previous risk field appears in the area ahead of the drone's planned path, and this value exceeds the safety threshold. The security threshold This is used to define the high-risk threshold as perceived by the system, and can be set according to the risk tolerance required by the mission and the risk resistance level of the UAV. In this embodiment, it is set... (The risk intensity is a normalized value, ranging from [0,1]).

[0242] Path deviation condition: If the actual position of any UAV deviates from the planned path laterally or vertically and continuously exceeds the preset path deviation tolerance, the condition is deemed to be in effect. The settings need to comprehensively consider the accuracy of the UAV positioning system (such as GPS error) and the accuracy of flight control. In this embodiment, the settings are as follows: .

[0243] Schedule Deviation Condition: The estimated completion time of any critical task will be later than the original planned time window by more than the schedule delay threshold. The timeframe can be set according to the urgency of the task and the allowed flexible time. In this embodiment, the timeframe is set as follows: .

[0244] Hardware / Task Failure Conditions: Receive hardware fault codes reported by the drone (such as sensor failure, power unit alarm), or confirmation of mission failure (such as capture failure, invalid detection data, etc.).

[0245] 2. Core process of compensatory rescheduling

[0246] The adjustment process follows the principles of efficiency and minimal disturbance, and the specific steps are as follows:

[0247] State Snapshot and Problem Reconstruction: The system immediately freezes the current state and marks all successfully completed tasks as immutable. Taking the current moment as the new decision starting point, it reconstructs a scaled-down version of the rolling time-domain optimization problem by taking the remaining set of unfinished tasks, the currently available drones and their states (location, battery level), and the latest four-dimensional dynamic risk field as input.

[0248] Model reuse and fast solution: The rolling time-domain optimization model of step S5 is rerun, but a hot-start strategy is adopted. For example, the current feasible path of the affected UAV or the part of the previous optimization scheme that was not executed is used as the initial solution input to the solver, and the focus is on optimizing the affected area and the short-term future time domain, thereby significantly shortening the solution time and meeting the real-time requirements.

[0249] Generate a differentiated adjustment plan: The optimization result is output as an adjusted scheduling plan. This plan may include:

[0250] Task reassignment: Reassign tasks that failed due to a malfunctioning drone to other drones in good condition within the cluster.

[0251] Local path correction: To avoid sudden risk areas or cope with wind disturbances, new local detours or wind-resistant paths are generated for affected drones.

[0252] Flexible formation reorganization: When drones malfunction or sub-clusters need to be merged, the formation topology is recalculated and adjusted.

[0253] 3. Seamless instruction injection and smooth execution

[0254] Command Differentiation and Issuance: The system compares the new adjustment plan with the old plan currently being executed, generating and issuing only the incremental commands that have changed, rather than the entire set of commands. For example, it only sends the command packet to the relevant UAV, instructing it to change flight sequence starting from the next waypoint. This reduces communication load and the complexity of execution switching.

[0255] Execution state synchronization: After receiving the adjustment command, the UAV will smoothly switch to the trajectory defined by the new command at an appropriate time (such as when it reaches the next decision point), and confirm the switch is completed through the feedback channel to ensure that the state of the entire cluster is resynchronized with the expectations of the central scheduler.

[0256] Optionally, after a task is completed, the entire process data, scheduling decisions, and compensation adjustment logs of the task are recorded to form a historical task knowledge base, which is used to train and optimize the parameters of the interference type identification model and the rolling time domain optimization model offline.

[0257] In this embodiment of the invention, the method further includes data recording and model optimization steps, the specific implementation of which is as follows:

[0258] This step establishes the system's self-learning and continuous evolution capabilities, which is crucial for the intelligent scheduling and management method to evolve from initial design to continuous improvement. It collects real-world operational data and performs offline iterative optimization of the core algorithm model, thereby constantly improving the system's performance under specific operating environments and task types.

[0259] 1. Full-process data recording and knowledge base construction

[0260] During and after each task execution, the system automatically records and structures the following data to form a historical task knowledge base:

[0261] Raw environmental and status data: Original snapshots of multi-source operational data from all UAVs are stored synchronously by timestamp, serving as the factual basis for subsequent analysis.

[0262] Decision-making process data: Records the dynamic risk field generated in each decision-making cycle, the input parameters of the rolling time-domain optimization model, the output scheduling scheme and its performance indicators.

[0263] Execution and Compensation Log: Records in detail every issued instruction, the actual execution trajectory of the drone, all triggered compensatory adjustment events, and specific solutions.

[0264] Final performance evaluation: Record comprehensive performance indicators such as the final completion rate of the task, total time spent, total energy consumption, and average risk exposure level.

[0265] 2. Offline training process for model optimization

[0266] The system periodically initiates the offline training process, such as during task breaks or after accumulating a certain amount of data, as follows:

[0267] Data preprocessing and labeling: The historical knowledge base data is cleaned, aligned, and enhanced. In particular, based on the actual results of the task, such as success, failure, or high-risk events, the types of interference and risk consequences in the historical data can be labeled retrospectively to provide more accurate labels for supervised learning.

[0268] Optimization of the interference type identification model:

[0269] Training objective: To train a more accurate neural network model using abundant historical data.

[0270] Training method: Using historical environment / state data as input and the posterior-labeled interference types and intensities as training objectives, the backpropagation algorithm is used to update the model weights. Incremental learning or federated learning strategies can be introduced to enable the model to continuously absorb information from new scenarios and interference patterns without forgetting old knowledge.

[0271] Optimization effect: The model has a stronger ability to extract and identify features for newly emerging or historically unrecognized types of interference.

[0272] Parameter tuning of the rolling time-domain optimization model:

[0273] Optimization targets: primarily key parameters in the optimization model that are difficult to determine theoretically, such as the objective function. Weighting coefficients in Trigger threshold for adaptive formation adjustment And the boundary safety margins for various constraints.

[0274] Optimization methods: Employing simulation-based reinforcement learning or Bayesian optimization. Specifically, historical task scenarios are used as simulation environments, and the final comprehensive performance indicators (such as high completion rate and low total risk) are used as reward signals. The intelligent optimization algorithm repeatedly tries different parameter combinations in the simulation, thereby automatically finding a set of parameter configurations that optimizes long-term performance.

[0275] Optimization results: The scheduling strategy is made more aligned with the actual business preferences for the trade-offs of security, efficiency, and cost, and the timing of adaptive formation adjustments is more reasonable.

[0276] 3. Model Validation and Seamless Deployment

[0277] Validation: The optimized model is thoroughly tested on an independent historical validation set or in a high-fidelity simulation test environment to ensure that its performance is superior to the old version and that there are no major defects.

[0278] Deployment: Through methods such as shadow mode or A / B testing, the new model version is gradually and smoothly deployed to the online scheduling system to replace the original model and complete a full iterative upgrade.

[0279] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0280] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent scheduling and management of industrial drone swarms based on multi-data analysis, characterized in that, include: Risk characteristic data of different interference types of each UAV are obtained based on multi-source operation data of UAVs in the cluster. The multi-source operation data includes at least environmental perception data and UAV working status data. The interference types include electromagnetic interference, airflow disturbance, geographical obstacles and performance degradation of the UAV. Based on a preset interference type identification model, the multi-source operation data is fused and analyzed to extract at least one type of interference quantitative risk feature vector. Based on the local risk index of each UAV, the Kriging spatial interpolation method is adopted, with the real-time position of each UAV as the interpolation point, its local risk index as the observation value, and combined with the geographic information system data of the mission area as the background field, to calculate and generate the predicted risk intensity value of any spatial location in the entire mission area in the current and short-term future time, so as to construct a four-dimensional dynamic risk field to characterize the comprehensive risk intensity distribution in the spatial location and time dimensions. The local risk index is calculated and characterized based on the real-time position of each UAV and the corresponding quantitative risk feature vector. A rolling time-domain optimization model incorporating dynamic risk field constraints is constructed to determine the collaborative scheduling scheme for the cluster in the next time domain. The rolling time-domain optimization model is constructed by minimizing a comprehensive objective function that integrates the overall risk of the cluster and the efficiency of task completion; the cooperative scheduling scheme is determined based on the optimal task allocation sequence, the optimal cooperative flight path, and the adaptive formation topology, which is dynamically adjusted according to the gradient of the dynamic risk field and the distribution of task points. The scheduling scheme is decomposed and issued as scheduling instructions for each UAV. Each UAV executes tasks based on the instruction set and continuously transmits updated multi-source operation data back. The dynamic risk field is updated based on the updated data returned, and compensatory adjustments are triggered for incomplete or interfered tasks according to the updated dynamic risk field and task execution status. The compensatory adjustments include reallocating tasks that are delayed or failed due to risk interference, and / or correcting the paths of affected drones. The rolling time-domain optimization model is constructed based on a weighted summation structure of minimizing the global risk value, task completion time, and cluster energy consumption value. The global risk value is represented by the result of integrating the risk intensity value of the predicted path point in the dynamic risk field.

2. The intelligent scheduling and management method for industrial UAV swarms based on multi-data analysis according to claim 1, characterized in that, Based on a preset interference type identification model, the multi-source operation data is fused and analyzed to extract a quantitative risk feature vector, including: Using a trained neural network model, feature fusion and pattern recognition are performed on the environmental perception data and the UAV operating status data to output risk probability and intensity estimates corresponding to different types of interference, thus forming the quantified risk feature vector.

3. The intelligent scheduling and management method for industrial UAV swarms based on multi-data analysis according to claim 1, characterized in that, The constraints of the rolling time-domain optimization model include: The constraints include UAV dynamics constraints, endurance constraints, temporal logic constraints between tasks, collision avoidance constraints between UAVs, and communication connectivity constraints required by the adaptive formation topology.

4. The intelligent scheduling and management method for industrial UAV swarms based on multi-data analysis according to claim 1, characterized in that, The dynamic adjustment of the adaptive formation topology includes: When the gradient value of the dynamic risk field exceeds the first preset threshold, the formation will be adjusted from a loose formation to a tight formation to enhance the cooperative anti-interference capability. When the dispersion of task points exceeds the second preset threshold, the formation will be split from a centralized formation into multiple sub-cluster formations to execute tasks in parallel.

5. The intelligent scheduling and management method for industrial UAV swarms based on multi-data analysis according to claim 1, characterized in that, The scheduling scheme is decomposed and issued as scheduling instructions for each of the UAVs, including: The optimal task allocation sequence, optimal cooperative flight path, and formation instructions of the whole system are solved into independent waypoint sequences, speed instructions, and formation relative coordinates for each UAV, and then distributed through the cluster communication network.

6. The intelligent scheduling and management method for industrial UAV swarms based on multi-data analysis according to claim 1, characterized in that, The conditions for triggering the compensatory adjustment include: The actual position of any drone was detected to deviate from the planned path from the allowable tolerance. Based on updated multi-source operational data, sudden high-risk areas that were not predicted in the dynamic risk field were identified; Received a report of mission failure or drone malfunction.

7. The intelligent scheduling and management method for industrial UAV swarms based on multi-data analysis according to claim 1 or 6, characterized in that, The compensatory adjustment includes: Starting from the current moment, freeze the tasks that have been successfully completed. With the remaining unfinished tasks, available drone resources, and the latest dynamic risk field as input, rerun the rolling time-domain optimization model to generate an adjusted scheduling scheme for the remaining task phase.

8. The intelligent scheduling and management method for industrial UAV swarms based on multi-data analysis according to claim 1, characterized in that, The method further includes: After a task is completed, the entire process data, scheduling decisions, and compensation adjustment logs of the task are recorded to form a historical task knowledge base, which is used to train and optimize the parameters of the interference type identification model and the rolling time domain optimization model offline.

Images