Shared unmanned aerial vehicle task scheduling method based on big data analysis

Abstract

The invention discloses a shared unmanned aerial vehicle task scheduling method based on big data analysis, and the method comprises the following steps: collecting multi-source data of a shared unmanned aerial vehicle platform, and carrying out the preprocessing; performing time sequence feature extraction and multi-dimensional feature fusion processing; inputting an improved Crossform model, sequentially carrying out feature coding, time sequence modeling and multi-task prediction processing, and then carrying out feature splicing; task request information is extracted, and feature fusion, correlation analysis and matching score calculation are carried out in combination with a task energy consumption comprehensive result; establishing a multi-objective optimization function to perform state and action modeling and reinforcement learning iterative optimization processing; a task instruction is generated and issued, feedback is collected and executed, an optimal scheduling strategy is updated in real time, and the unmanned aerial vehicle is driven to execute adaptive scheduling. According to the invention, big data analysis and reinforcement learning technologies are fused, intelligent cooperative scheduling of tasks and energy consumption of the shared unmanned aerial vehicle is realized, and the method has the advantages of high efficiency, energy saving and adaptive optimization.

Description

Technical Field

[0001] This invention relates to the field of intelligent scheduling of unmanned aerial vehicle (UAV) missions, and in particular to a shared UAV mission scheduling method based on big data analysis. Background Technology

[0002] With the widespread application of shared drone platforms in logistics delivery, inspection and monitoring, and emergency rescue, the intelligence and efficiency of task scheduling have become key factors affecting the overall system performance. Existing drone task scheduling methods typically rely on preset rules or optimization models based on static weights, making it difficult to adapt to dynamic changes in multi-user, multi-task, and complex environments. Especially in scenarios where task requirements are frequently updated and energy consumption is greatly affected by environmental factors, traditional scheduling algorithms often ignore the coupling relationship between task execution and energy consumption, leading to low task allocation efficiency, insufficient energy utilization, and uneven distribution of drone resources.

[0003] Current research has attempted to improve UAV mission scheduling using machine learning or deep learning methods, but most models only predict from a single mission dimension and lack the ability to fuse multi-source data and collaboratively model energy consumption. Furthermore, existing reinforcement learning scheduling strategies are mostly based on simplified environmental parameters and cannot achieve joint optimization of mission and resource states in real-world multi-mission environments. Therefore, there is an urgent need for a shared UAV mission scheduling method based on big data analytics to achieve joint modeling of mission load and energy consumption characteristics, mission resource matching, and adaptive updates of dynamic optimal scheduling strategies.

[0004] Therefore, how to provide a shared drone mission scheduling method based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a shared drone task scheduling method based on big data analysis. This invention integrates big data analysis and reinforcement learning techniques to achieve intelligent collaborative scheduling of shared drone tasks and energy consumption, and has the advantages of high efficiency, energy saving and adaptive optimization.

[0006] A shared drone task scheduling method based on big data analysis according to an embodiment of the present invention includes the following steps:

[0007] Collect multi-source data from shared drone platforms, preprocess it, and generate standardized multi-source datasets;

[0008] The standardized multi-source dataset is subjected to temporal feature extraction and multi-dimensional feature fusion to generate task load feature vector and energy consumption feature vector.

[0009] Input the task load feature vector and energy consumption feature vector into the improved Crossformer model, and perform feature encoding, time series modeling and multi-task prediction processing in sequence to generate task demand prediction results and energy consumption assessment results respectively. Then, perform feature concatenation to generate a comprehensive task energy consumption result.

[0010] Task request information is extracted from a standardized multi-source dataset, and feature fusion, correlation analysis and matching score calculation are performed in combination with the comprehensive results of task energy consumption to generate a task resource mapping matrix.

[0011] Based on the task resource mapping matrix and standardized multi-source dataset, a multi-objective optimization function is constructed to perform state and action modeling and reinforcement learning iterative optimization to generate the optimal scheduling strategy.

[0012] Task instructions are generated based on the optimal scheduling strategy and sent to each UAV. During the task execution process, feedback data is continuously collected, the optimal scheduling strategy is updated in real time, and the shared UAVs are driven to perform adaptive dynamic scheduling.

[0013] Optionally, the multi-source data includes task request information, UAV operating status parameters, environmental monitoring data, and historical task records, and the preprocessing includes cleaning, format unification, and time alignment.

[0014] Optionally, the generation of the task load feature vector and energy consumption feature vector specifically includes:

[0015] Multi-scale convolutional kernel groups are used to perform local feature extraction operations on the time dimension of time data in standardized multi-source datasets. The task status data and energy consumption status data in adjacent time segments are processed by sliding window convolution to obtain a preliminary convolutional feature sequence.

[0016] The initial convolutional feature sequence is processed by time-dependent modeling. At each time step, the updated temporal state vector is obtained by weighting the hidden state of the previous time step with the current input state.

[0017] The updated temporal state vector is updated, the current input weighted sum and the nonlinear transformation result of the previous hidden state are calculated, and the temporal feature output sequence is obtained after fusion operation by activation function;

[0018] The time-series feature output sequence is normalized and concatenated, and feature components related to task execution load and energy consumption change are extracted respectively.

[0019] The task-related feature components and energy consumption-related feature components are linearly mapped and dimensionally reduced respectively to generate task load feature vectors and energy consumption feature vectors.

[0020] Optionally, the generation of the overall task energy consumption result specifically includes:

[0021] The task load feature vector and energy consumption feature vector are respectively input into the dual-channel feature input layer of the improved Crossformer model for time encoding and dimensional embedding processing, resulting in task time embedding sequences and energy consumption time embedding sequences, respectively. Compared with the original Crossformer model, the improved Crossformer model introduces a dual-channel feature input layer, performs time encoding and feature embedding processing on the task load feature vector and energy consumption feature vector respectively, adds a shared encoding layer on the basis of the original time encoding structure, performs multi-scale convolutional feature extraction and hierarchical time encoding on the dual-channel input, replaces the cross-dimensional attention structure with a cross-task association attention layer, establishes an interaction weight matrix between task features and energy consumption features and performs cross-task attention calculation, introduces a dynamic interaction bias term to adjust the association strength in the attention calculation, and sets a multi-task branch output layer at the output end to perform linear transformation and weighted convergence operations respectively. The dual-channel feature input layer includes a task channel input unit, an energy consumption channel input unit, a time encoding unit and a feature embedding unit.

[0022] The task time embedding sequence and the energy consumption time embedding sequence are input into the shared coding layer, and multi-scale temporal convolution processing and hierarchical temporal coding operations are performed to obtain the temporal joint feature sequence. The shared coding layer includes multi-scale convolutional units, temporal coding units and feature fusion units.

[0023] The temporal joint feature sequence is input into the cross-task association attention layer for feature mapping processing to obtain the interaction weight matrix. A dynamic interaction bias term is then introduced for attention adjustment to obtain the cross-task association feature sequence. The cross-task association attention layer includes a query mapping unit, a key mapping unit, an interaction weight calculation unit, and a cross-task feature fusion unit. The dynamic interaction bias term is a dynamic adjustment factor generated by linear mapping and normalization based on the time change rate of task feature components and energy consumption feature components.

[0024] The cross-task associated feature sequences are input into the multi-task branch output layer to perform linear transformation and weighted convergence operations respectively, so as to obtain the task demand prediction results and energy consumption assessment results. The features are then spliced ​​to generate the comprehensive task energy consumption result. The multi-task branch output layer includes a task prediction branch and an energy consumption assessment branch.

[0025] Optionally, the generation of the task resource mapping matrix specifically includes:

[0026] Task request information is extracted from a standardized multi-source dataset. The task request records of multiple users are sorted by timestamp and standardized in format to obtain a unified task request sequence.

[0027] For each task request in the unified task request sequence, extract the task demand vector and energy consumption demand vector, and perform element-wise weighted summation and amplitude normalization according to preset weight coefficients to obtain the fused task feature vector.

[0028] The task feature vector and the comprehensive result of task energy consumption are matched according to the task number. The predicted features are extracted and the UAV operating status parameters in the standardized multi-source dataset are used to establish the task feature matrix and the resource feature matrix. The weight of the task feature matrix is ​​adjusted according to the predicted features. The two matrices are then feature-mapped and their similarity is calculated.

[0029] The task feature matrix and resource feature matrix are sorted and paired according to similarity values. The corresponding task numbers and UAV numbers are associated and mapped according to the matching degree. The matching score of each task and resource pair is calculated and recorded to generate a task-resource mapping matrix.

[0030] Optionally, the generation of the optimal scheduling strategy specifically includes:

[0031] Based on the task resource mapping matrix, key feature parameters are extracted from task allocation data and path planning data, and each parameter is normalized to obtain a standardized scheduling feature set.

[0032] A linear weighted operation is performed on each parameter in the standardized scheduling feature set according to a preset weight coefficient, and the weighted results are summed to obtain a multi-objective optimization function.

[0033] Reinforcement learning modeling is performed on the multi-objective optimization function. A state vector space is jointly constructed based on the task state characteristics and UAV state parameters. A set of task allocation actions and path planning actions is defined and an action space is established. The policy network structure is initialized. The state vector space is used as the input layer and the action space is used as the output layer. The network weights and bias parameters are randomly initialized to obtain the initialized policy network.

[0034] In each iteration, task allocation actions and path planning actions are selected based on the state vector space. The selected actions are input into the multi-objective optimization function to calculate the corresponding reward value. The weights and bias parameters of the initial policy network are updated based on the reward value feedback. The above process is repeated until the policy network converges and the optimal scheduling policy is generated.

[0035] Optionally, the task instructions include a task number, target location coordinates, task type, task priority, estimated execution time, path planning parameters, energy consumption control parameters, and feedback upload frequency.

[0036] The beneficial effects of this invention are:

[0037] This invention constructs a shared UAV mission scheduling method based on big data analysis, achieving unified modeling of multi-source data fusion, mission demand prediction, energy consumption assessment, and dynamic scheduling optimization. This significantly improves the resource utilization efficiency and execution reliability of shared UAV swarms in complex mission environments. Compared with traditional scheduling methods based on static rules or single-dimensional feature optimization, this invention can simultaneously consider mission load characteristics and energy consumption status characteristics. It utilizes an improved Crossformer model to jointly predict mission demands and energy consumption changes, thereby achieving collaborative modeling of mission execution requirements and energy consumption trends. This provides more accurate input data support for subsequent mission allocation and path planning, significantly reducing energy consumption errors and uneven allocation problems in mission scheduling.

[0038] In terms of model structure, this invention enhances the model's ability to represent complex temporal relationships by introducing a dual-channel feature input layer and a shared encoding layer to perform parallel processing and multi-scale fusion of task load feature vectors and energy consumption feature vectors. The introduction of a cross-task association attention layer and a dynamic interaction bias term enables the model to dynamically capture the temporal dependencies and correlations between task and energy consumption features, ensuring the adaptability of feature mapping and high-precision prediction performance. Simultaneously, the multi-task branch output layer enables the synchronous generation of task demand prediction results and energy consumption assessment results, providing more comprehensive feature information for subsequent task resource matching and significantly improving the accuracy of task prediction and the real-time performance of energy consumption modeling.

[0039] In the task allocation and scheduling optimization phase, this invention constructs a task feature matrix and a resource feature matrix based on the comprehensive results of task energy consumption. It then combines UAV operating status parameters with feature mapping and similarity calculation to generate a task resource mapping matrix, achieving optimal matching between tasks and UAV resources. Building upon this, a multi-objective optimization function is constructed using reinforcement learning. Through joint modeling and dynamic iterative optimization of multi-dimensional constraint parameters such as task priority, flight distance, execution time, and energy cost, an optimal scheduling strategy that can be updated in real time according to changes in task status and environment is generated. This strategy can adaptively adjust based on feedback data during task execution, achieving dynamic allocation and path optimization of UAVs. This improves the overall task completion rate, reduces system energy consumption, decreases UAV idle rate, and significantly enhances the scheduling intelligence and autonomous decision-making capabilities of the shared UAV system in complex multi-task environments. Attached Figure Description

[0040] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0041] Figure 1 This is a flowchart of a shared UAV task scheduling method based on big data analysis proposed in this invention;

[0042] Figure 2 This is a schematic diagram of the improved Crossformer model structure of a shared UAV task scheduling method based on big data analysis proposed in this invention. Detailed Implementation

[0043] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0044] refer to Figure 1-2 A shared drone task scheduling method based on big data analysis includes the following steps:

[0045] Collect multi-source data from shared drone platforms, preprocess it, and generate standardized multi-source datasets;

[0046] The standardized multi-source dataset is subjected to temporal feature extraction and multi-dimensional feature fusion to generate task load feature vector and energy consumption feature vector.

[0047] Input the task load feature vector and energy consumption feature vector into the improved Crossformer model, and perform feature encoding, time series modeling and multi-task prediction processing in sequence to generate task demand prediction results and energy consumption assessment results respectively. Then, perform feature concatenation to generate a comprehensive task energy consumption result.

[0048] Task request information is extracted from a standardized multi-source dataset, and feature fusion, correlation analysis and matching score calculation are performed in combination with the comprehensive results of task energy consumption to generate a task resource mapping matrix.

[0049] Based on the task resource mapping matrix and standardized multi-source dataset, a multi-objective optimization function is constructed to perform state and action modeling and reinforcement learning iterative optimization to generate the optimal scheduling strategy.

[0050] Task instructions are generated based on the optimal scheduling strategy and sent to each UAV. During the task execution process, feedback data is continuously collected, the optimal scheduling strategy is updated in real time, and the shared UAVs are driven to perform adaptive dynamic scheduling.

[0051] In this embodiment, the multi-source data includes task request information, UAV operating status parameters, environmental monitoring data, and historical task records. The preprocessing includes cleaning, format unification, and time alignment.

[0052] In this embodiment, the generation of the task load feature vector and energy consumption feature vector specifically includes:

[0053] Multi-scale convolutional kernels are used to extract local features in the time dimension of time-series data in standardized multi-source datasets. Sliding window convolution is applied to the task status data and energy consumption status data in adjacent time segments to obtain a preliminary convolutional feature sequence. The task status data includes task execution parameters, flight operation parameters, and task load parameters. The task execution parameters include task type, execution duration, priority, and task geographical location. The flight operation parameters include flight speed, altitude, attitude angle, and trajectory coordinate sequence. The task load parameters include payload mass, task power, and load operation status. The energy consumption status data includes energy consumption parameters, electrical sampling parameters, and environmental impact parameters. The energy consumption parameters include power consumption rate, power output, and unit energy consumption value. The electrical sampling parameters include voltage, current, and inverter frequency. The environmental impact parameters include wind speed, temperature, and air pressure.

[0054] The initial convolutional feature sequence is processed by time-dependent modeling. At each time step, the updated temporal state vector is obtained by weighting the hidden state of the previous time step with the current input state.

[0055] The updated temporal state vector is updated, the current input weighted sum and the nonlinear transformation result of the previous hidden state are calculated, and the temporal feature output sequence is obtained after fusion operation by the activation function. The activation function is used to control the temporal propagation ratio of feature information.

[0056] The time-series feature output sequence is normalized and concatenated, and feature components related to task execution load and energy consumption change are extracted respectively.

[0057] The task-related feature components and energy consumption-related feature components are linearly mapped and dimensionally reduced respectively to generate task load feature vectors and energy consumption feature vectors.

[0058] In this embodiment, the generation of the overall task energy consumption result specifically includes:

[0059] The task load feature vector and energy consumption feature vector are respectively input into the dual-channel feature input layer of the improved Crossformer model for time encoding and dimensional embedding processing, resulting in task time embedding sequences and energy consumption time embedding sequences, respectively. Compared with the original Crossformer model, the improved Crossformer model introduces a dual-channel feature input layer, performs time encoding and feature embedding processing on the task load feature vector and energy consumption feature vector respectively, adds a shared encoding layer on the basis of the original time encoding structure, performs multi-scale convolutional feature extraction and hierarchical time encoding on the dual-channel input, replaces the cross-dimensional attention structure with a cross-task association attention layer, establishes an interaction weight matrix between task features and energy consumption features and performs cross-task attention calculation, introduces a dynamic interaction bias term to adjust the association strength in the attention calculation, and sets a multi-task branch output layer at the output end to perform linear transformation and weighted convergence operations respectively. The dual-channel feature input layer includes a task channel input unit, an energy consumption channel input unit, a time encoding unit and a feature embedding unit.

[0060] The specific steps for obtaining the task time embedding sequence and the energy consumption time embedding sequence are as follows: input the task load feature vector and the energy consumption feature vector into the dual-channel feature input layer of the improved Crossformer model, respectively; receive the input vectors in the task channel input unit and the energy consumption channel input unit, respectively; perform one-dimensional convolution to increase the dimensionality and perform amplitude standardization processing; input the enhanced feature sequence into the time encoding unit; perform sine and cosine function encoding on each time position according to the time step index; add the encoding result to the corresponding feature vector element by element to obtain the time encoding sequence; input the time encoding sequence into the feature embedding unit; perform linear transformation and feature mapping processing respectively; project the mapping result uniformly into the high-dimensional feature space; and output the corresponding time embedding sequence in channel order.

[0061] The task time embedding sequence and the energy consumption time embedding sequence are input into the shared coding layer, and multi-scale temporal convolution processing and hierarchical temporal coding operations are performed to obtain the temporal joint feature sequence. The shared coding layer includes multi-scale convolutional units, temporal coding units and feature fusion units.

[0062] The process of obtaining the temporal joint feature sequence specifically includes: performing multi-scale one-dimensional convolution on the task time embedding sequence and the energy consumption time embedding sequence using different convolution kernel sizes through a multi-scale convolution unit and concatenating them to obtain a multi-scale convolution feature sequence; inputting the multi-scale convolution feature sequence into a time coding unit; performing hierarchical time coding calculation according to the time step order; weighted combination of features of adjacent time segments; inputting the time-coded task channel sequence and energy consumption channel sequence into a feature fusion unit; and performing weighted fusion operation according to the channel correlation weight to obtain the temporal joint feature sequence.

[0063] The temporal joint feature sequence is input into the cross-task association attention layer for feature mapping processing to obtain the interaction weight matrix. A dynamic interaction bias term is then introduced for attention adjustment to obtain the cross-task association feature sequence. The cross-task association attention layer includes a query mapping unit, a key mapping unit, an interaction weight calculation unit, and a cross-task feature fusion unit. The dynamic interaction bias term is a dynamic adjustment factor generated by linear mapping and normalization based on the time change rate of task feature components and energy consumption feature components.

[0064] The process of obtaining the cross-task associated feature sequence specifically includes: performing feature mapping on the task feature components and energy consumption feature components in the temporal joint feature sequence through query mapping units and query mapping units respectively, to obtain a task feature query matrix and an energy consumption feature key matrix; performing a weighted product operation on the task feature query matrix and the energy consumption feature key matrix and performing normalization processing to obtain an interaction weight matrix; introducing a dynamic interaction bias term during the attention calculation process to dynamically adjust the correlation between the task feature components and the energy consumption feature components in the interaction weight matrix to obtain the cross-task associated feature sequence; the dynamic interaction bias term is generated by linear mapping and normalization of the time change rate of the task feature components and the energy consumption feature components.

[0065] The cross-task associated feature sequence is input into the multi-task branch output layer to perform linear transformation and weighted convergence operations respectively, so as to obtain the task demand prediction result and energy consumption assessment result. The feature is then spliced ​​to generate the comprehensive task energy consumption result. The multi-task branch output layer includes a task prediction branch and an energy consumption assessment branch.

[0066] The specific steps to obtain the task requirement prediction result include: weighting and channel aggregation of the feature vectors at each time step in the task prediction branch, and then obtaining the result through a fully connected transformation.

[0067] The task demand prediction results include the predicted number of tasks, the probability distribution of task types, the estimated time requirement of tasks, and the task priority score.

[0068] The energy consumption assessment result is obtained by performing channel-weighted aggregation operation on the input sequence in the energy consumption assessment branch, calculating the energy consumption contribution vector according to the weight distribution of the feature dimension, and performing a linear transformation operation on the aggregated feature result.

[0069] The energy consumption assessment results include total energy consumption estimate, unit task energy consumption, energy efficiency index, and energy consumption anomaly risk score.

[0070] In this embodiment, the generation of the task resource mapping matrix specifically includes:

[0071] Task request information is extracted from a standardized multi-source dataset. The task request records of multiple users are sorted by timestamp and standardized in format to obtain a unified task request sequence.

[0072] For each task request in the unified task request sequence, extract the task demand vector and energy consumption demand vector, and perform element-wise weighted summation and amplitude normalization according to preset weight coefficients to obtain the fused task feature vector.

[0073] The task feature vector and the comprehensive result of task energy consumption are matched according to the task number. The predicted features are extracted and the UAV operating status parameters in the standardized multi-source dataset are used to establish the task feature matrix and the resource feature matrix. The weight of the task feature matrix is ​​adjusted according to the predicted features. The two matrices are feature mapped and the similarity is calculated. The UAV operating status parameters include flight distance, payload capacity and remaining energy information. The similarity is obtained by normalizing the Euclidean distance between the two matrices.

[0074] The establishment of the task feature matrix and resource feature matrix specifically includes: extracting predicted features and UAV operating status parameters from standardized multi-source datasets, matching them according to the task number and UAV number, normalizing the predicted values ​​of task requirements and energy consumption assessment values ​​to obtain the task feature matrix, and standardizing and weighting the flight distance, payload capacity and remaining energy information to obtain the resource feature matrix.

[0075] The task feature matrix and resource feature matrix are sorted and paired according to similarity values. The corresponding task numbers and UAV numbers are associated and mapped according to the matching degree. The matching score value of each task and resource pair is calculated and recorded to generate a task-resource mapping matrix. The matching score value is obtained by calculating and normalizing each task sample and UAV resource sample according to Euclidean distance based on the feature similarity results between the task feature matrix and the resource feature matrix. The similarity score value is obtained by weighted summation of the similarity scores according to task priority, energy consumption weight, and flight distance weight. The task priority includes time urgency, task importance, energy consumption cost, and service level. The time urgency weight and task importance weight are derived from the task request information, the energy consumption cost weight is derived from the energy consumption assessment results, and the service level weight is derived from the platform system settings.

[0076] In this embodiment, the generation of the optimal scheduling strategy specifically includes:

[0077] Based on the task resource mapping matrix, key feature parameters are extracted from task allocation data and path planning data, and each parameter is normalized to obtain a standardized scheduling feature set. The task allocation data is generated based on the task number, resource number, and matching score in the task resource mapping matrix, including task demand prediction results, energy consumption assessment results, UAV execution capability parameters, and task resource correspondence. The path planning data is generated based on the task geographical location information in the task request information and the current position in the UAV operating status parameters, including flight distance, execution time, energy consumption cost, and environmental impact parameters. The feature parameters include task priority, flight distance, execution time, and energy consumption cost.

[0078] A linear weighted operation is performed on each parameter in the standardized scheduling feature set according to a preset weight coefficient, and the weighted results are summed to obtain a multi-objective optimization function. The objective function value is used to measure the comprehensive performance of the scheduling scheme.

[0079] Reinforcement learning modeling is performed on the multi-objective optimization function. A state vector space is jointly constructed based on the task state characteristics and UAV state parameters. A set of task allocation actions and path planning actions is defined and an action space is established. The policy network structure is initialized. The state vector space is used as the input layer and the action space is used as the output layer. The network weights and bias parameters are randomly initialized to obtain the initialized policy network.

[0080] The construction of the state vector space specifically includes: normalizing the predicted values ​​of task requirements and energy consumption assessment values ​​in the task state features, arranging them according to the task number to obtain the task feature sequence, extracting the flight distance, payload capacity and remaining energy information from the UAV operating state parameters from the standardized multi-source dataset, performing standardization and feature encoding to obtain the UAV feature sequence, and concatenating the task feature sequence and the UAV feature sequence according to their corresponding numbers and performing feature fusion to obtain a state vector space that includes joint features of the task and the UAV.

[0081] The establishment of the action space specifically includes: obtaining a set of task allocation actions based on the correspondence between task numbers and UAV numbers in the task resource mapping matrix; extracting flight distance, track nodes, and geographical location information based on UAV operating status parameters in the standardized multi-source dataset; generating a corresponding set of path planning actions for each task allocation action; merging the task allocation action set and the path planning action set according to their corresponding numbers; performing action coding and index numbering processing; and establishing an action space that includes joint actions of task allocation and path planning.

[0082] In each iteration, task allocation actions and path planning actions are selected based on the state vector space. The selected actions are input into a multi-objective optimization function to calculate the corresponding reward value. The weights and bias parameters of the initial policy network are updated based on the reward value feedback. The above process is repeated until the policy network converges and the optimal scheduling policy is generated. The convergence condition is that the update magnitude of network weights and the rate of change of reward value are both lower than a preset threshold in multiple consecutive iterations. The task allocation action includes task number, UAV number, matching score value and task allocation probability weight. The path planning action includes starting position coordinates, target position coordinates, track node sequence, flight distance and direction angle parameters, energy consumption prediction value and obstacle avoidance flag parameters.

[0083] In this embodiment, the task instruction includes task number, target location coordinates, task type, task priority, estimated execution time, path planning parameters, energy consumption control parameters, and feedback upload frequency.

[0084] Example 1:

[0085] To verify the feasibility of this invention in practice, it was applied to a drone sharing management platform in a large city. This city has a wide airspace and complex task types. The platform daily dispatches hundreds of drones to perform various tasks, including urban road inspection, emergency relief material delivery, environmental monitoring, and traffic flow observation. Due to the diverse sources of task requests, varying time requirements, and significant differences in weather changes, wind speed fluctuations, and drone energy consumption levels, traditional dispatching methods based on fixed rules or static parameters often result in uneven drone allocation, task delays, and energy waste. For example, during peak hours, some drones repeatedly perform similar tasks, leading to increased energy consumption, while tasks in other areas cannot be responded to in a timely manner, significantly reducing overall dispatching efficiency.

[0086] In this application scenario, the shared UAV task scheduling method based on big data analytics of this invention achieves a deep integration of task allocation and energy consumption optimization through centralized management and intelligent analysis of multi-source data from the platform. First, the platform continuously collects multi-source data such as UAV operating status, task request information, meteorological parameters, and historical flight records. A preprocessing module performs data cleaning, format unification, and time alignment operations to form a standardized multi-source dataset. Subsequently, the system utilizes an improved Crossformer model to perform dual-channel feature modeling on task load feature vectors and energy consumption feature vectors, performing temporal feature encoding, cross-task correlation modeling, and multi-task prediction, thereby achieving joint analysis of task requirements and energy consumption trends. By introducing a shared encoding layer and a cross-task correlation attention layer, this model can adaptively capture the temporal dependencies between task and energy consumption features, effectively improving the model's prediction accuracy and generalization ability under complex operating conditions, and providing a high-confidence decision-making basis for task resource allocation.

[0087] In actual operation, the system performs matching calculations on the UAV resource pool based on the predicted task requirements and energy consumption assessment results. The scheduling engine extracts task request information and UAV operating status parameters, constructs a task feature matrix and a resource feature matrix, and generates a task-resource mapping matrix through feature mapping and similarity calculation to achieve optimal matching between task numbers and UAV numbers. This process comprehensively considers multiple dimensions such as flight distance, payload capacity, remaining energy, energy consumption cost, and task priority, avoiding the deviations caused by single energy consumption or distance parameters in previous scheduling strategies, and enabling a more balanced allocation of UAV resources in both spatial and temporal dimensions.

[0088] During the scheduling optimization phase, the system constructs a multi-objective optimization function based on the task resource mapping matrix. Combining constraints such as task priority, execution time, flight distance, and energy cost, it utilizes reinforcement learning algorithms for dynamic optimization and self-iteration. The policy network continuously adjusts parameter weights through multiple rounds of learning, gradually generating optimal scheduling strategies adapted to different task densities, geographical distributions, and energy consumption conditions. This strategy can be corrected online based on real-time feedback data. When a UAV is unable to continue its mission due to sudden weather conditions or insufficient energy, the system can automatically recalculate the task allocation scheme, achieving adaptive dynamic scheduling of the UAV.

[0089] In a multi-day urban environmental monitoring mission, the scheduling system implemented in this invention demonstrated high stability and coordination capabilities. The system can automatically adjust the drone deployment order based on the time urgency of different tasks, dynamically monitor the task execution status, and promptly update flight data and energy consumption model parameters upon task completion, achieving data-driven continuous optimization. In this way, the system effectively alleviates the delay problems caused by concentrated tasks or uneven resource allocation in traditional scheduling strategies, significantly improving the collaborative operation efficiency and energy efficiency of drone swarms.

[0090] Overall, this invention successfully solves the problems of inaccurate energy consumption prediction, delayed scheduling response, and low resource matching efficiency in traditional shared drone scheduling methods in practical applications. By deeply integrating multi-source data with intelligent optimization algorithms, it achieves a closed-loop process from task demand prediction, energy consumption assessment, task matching to dynamic scheduling.

[0091] Table 1. Performance Comparison of the Invention and Traditional Shared UAV Task Scheduling Methods

[0092] Indicator Categories Traditional methods Method of the present invention Task assignment accuracy (%) 84.6 94.2 Average energy efficiency improvement rate (%) 78.3 91.5 Scheduling response time (s) 4.7 2.6 Resource utilization rate (%) 76.9 89.8 Task delay rate (%) 12.4 4.3 System stability index 0.81 0.93 Adaptive scheduling success rate (%) 73.2 91.1

[0093] As can be clearly seen from Table 1, the method of the present invention is superior to the traditional method in many indicators.

[0094] In terms of task allocation accuracy, the method of this invention achieves 94.2%, which is nearly 10 percentage points higher than the 84.6% of the traditional method. The main reason for this is that the present invention uses an improved Crossformer model to jointly model task load characteristics and energy consumption characteristics, and realizes dynamic matching of task requirements and energy consumption assessment through a cross-task association attention mechanism, thereby making the task allocation process more accurate and avoiding the problems of duplicate allocation and misallocation.

[0095] In terms of average energy efficiency improvement, the method of this invention achieves 91.5%, significantly higher than the 78.3% of the traditional method. This is because the system introduces a multi-objective optimization function and reinforcement learning mechanism during the scheduling phase, which can comprehensively consider factors such as flight distance, task priority, and energy cost to achieve dynamic optimization that minimizes energy consumption and balances resource allocation. Through the self-iterative training of the reinforcement learning algorithm, the system gradually learns to adjust the UAV's energy consumption strategy under different task load conditions, thereby significantly improving the overall energy efficiency level.

[0096] The scheduling response time was reduced from 4.7 seconds using traditional methods to 2.6 seconds, demonstrating a significant improvement in response speed. This invention significantly shortens the computation time for task matching and path planning by constructing a standardized multi-source dataset and a task resource mapping matrix. Simultaneously, the model employs time-series organization and feature embedding mechanisms in the task request processing stage, enabling the scheduling engine to process multiple task requests in parallel and ensuring rapid response capabilities in high-frequency dynamic task environments.

[0097] In terms of resource utilization, the method of this invention achieves 89.8%, an improvement of nearly 13 percentage points compared to the 76.9% of the traditional method. This result demonstrates that the task resource matching mechanism of this invention effectively avoids the problems of drone idleness and task duplication. Through a comprehensive evaluation of drone operating parameters (such as flight distance, payload capacity, and remaining energy), the system achieves optimal task-resource matching, enabling each drone to operate in the most suitable task scenario and improving the overall scheduling system's operational efficiency.

[0098] Regarding task delay rates, traditional methods result in a 12.4% delay rate, while this invention achieves only 4.3%, significantly reducing the risk of task delays. This is attributed to the introduction of a task priority factor during the scheduling process, assigning higher weight to urgent tasks, and employing a real-time update mechanism through reinforcement learning strategies. This ensures that the system can automatically adjust task order and path planning in the event of sudden tasks or changes in equipment status, guaranteeing the priority completion of critical tasks.

[0099] In terms of system stability index, this invention achieves 0.93, a significant improvement compared to the traditional method's 0.81. This enhanced system stability is primarily attributed to the multi-scale convolution and hierarchical temporal coding design employed in the temporal modeling layer, enabling the model to maintain the stability and consistency of prediction results when facing complex environmental changes. Simultaneously, the continuous acquisition of task feedback data and the dynamic parameter update mechanism ensure that the system maintains high robustness and reliability during long-term operation.

[0100] Finally, in terms of adaptive scheduling success rate, the method of this invention achieves 91.1%, while the traditional method only achieves 73.2%. This demonstrates that the reinforcement learning scheduling module of this invention can continuously correct the optimal strategy based on feedback information in a real-time changing environment, achieving truly adaptive dynamic scheduling. Through continuous state monitoring and strategy updates, the system ensures that the UAV can flexibly adjust its flight path and task execution order under different environmental and mission conditions, thereby significantly improving the intelligence level of scheduling.

[0101] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A shared drone task scheduling method based on big data analysis, characterized in that, Includes the following steps: Collect multi-source data from shared drone platforms, preprocess it, and generate standardized multi-source datasets; The standardized multi-source dataset is subjected to temporal feature extraction and multi-dimensional feature fusion to generate task load feature vector and energy consumption feature vector. Input the task load feature vector and energy consumption feature vector into the improved Crossformer model, and perform feature encoding, time series modeling and multi-task prediction processing in sequence to generate task demand prediction results and energy consumption assessment results respectively. Then, perform feature concatenation to generate a comprehensive task energy consumption result. Task request information is extracted from a standardized multi-source dataset, and feature fusion, correlation analysis and matching score calculation are performed in combination with the comprehensive results of task energy consumption to generate a task resource mapping matrix. Based on the task resource mapping matrix and standardized multi-source dataset, a multi-objective optimization function is constructed to perform state and action modeling and reinforcement learning iterative optimization to generate the optimal scheduling strategy. Task instructions are generated based on the optimal scheduling strategy and sent to each UAV. During the task execution process, feedback data is continuously collected, the optimal scheduling strategy is updated in real time, and the shared UAVs are driven to perform adaptive dynamic scheduling.

2. The shared drone task scheduling method based on big data analysis according to claim 1, characterized in that, The multi-source data includes mission request information, UAV operating status parameters, environmental monitoring data, and historical mission records. The preprocessing includes cleaning, format unification, and time alignment.

3. The shared drone task scheduling method based on big data analysis according to claim 1, characterized in that, The generation of the task load feature vector and energy consumption feature vector specifically includes: Multi-scale convolutional kernel groups are used to perform local feature extraction operations on the time dimension of time data in standardized multi-source datasets. The task status data and energy consumption status data in adjacent time segments are processed by sliding window convolution to obtain a preliminary convolutional feature sequence. The initial convolutional feature sequence is processed by time-dependent modeling. At each time step, the updated temporal state vector is obtained by weighting the hidden state of the previous time step with the current input state. The updated temporal state vector is updated, the current input weighted sum and the nonlinear transformation result of the previous hidden state are calculated, and the temporal feature output sequence is obtained after fusion operation by activation function; The time-series feature output sequence is normalized and concatenated, and feature components related to task execution load and energy consumption change are extracted respectively. The task-related feature components and energy consumption-related feature components are linearly mapped and dimensionally reduced respectively to generate task load feature vectors and energy consumption feature vectors.

4. The shared drone task scheduling method based on big data analysis according to claim 1, characterized in that, The generation of the overall energy consumption result for the task specifically includes: The task load feature vector and energy consumption feature vector are respectively input into the dual-channel feature input layer of the improved Crossformer model for time encoding and dimensional embedding processing, resulting in task time embedding sequences and energy consumption time embedding sequences, respectively. Compared with the original Crossformer model, the improved Crossformer model introduces a dual-channel feature input layer, performs time encoding and feature embedding processing on the task load feature vector and energy consumption feature vector respectively, adds a shared encoding layer on the basis of the original time encoding structure, performs multi-scale convolutional feature extraction and hierarchical time encoding on the dual-channel input, replaces the cross-dimensional attention structure with a cross-task association attention layer, establishes an interaction weight matrix between task features and energy consumption features and performs cross-task attention calculation, introduces a dynamic interaction bias term to adjust the association strength in the attention calculation, and sets a multi-task branch output layer at the output end to perform linear transformation and weighted convergence operations respectively. The dual-channel feature input layer includes a task channel input unit, an energy consumption channel input unit, a time encoding unit and a feature embedding unit. The task time embedding sequence and the energy consumption time embedding sequence are input into the shared coding layer, and multi-scale temporal convolution processing and hierarchical temporal coding operations are performed to obtain the temporal joint feature sequence. The shared coding layer includes multi-scale convolutional units, temporal coding units and feature fusion units. The temporal joint feature sequence is input into the cross-task association attention layer for feature mapping processing to obtain the interaction weight matrix. A dynamic interaction bias term is then introduced for attention adjustment to obtain the cross-task association feature sequence. The cross-task association attention layer includes a query mapping unit, a key mapping unit, an interaction weight calculation unit, and a cross-task feature fusion unit. The dynamic interaction bias term is a dynamic adjustment factor generated by linear mapping and normalization based on the time change rate of task feature components and energy consumption feature components. The cross-task associated feature sequences are input into the multi-task branch output layer to perform linear transformation and weighted convergence operations respectively, so as to obtain the task demand prediction results and energy consumption assessment results. The features are then spliced ​​to generate the comprehensive task energy consumption result. The multi-task branch output layer includes a task prediction branch and an energy consumption assessment branch.

5. The shared drone task scheduling method based on big data analysis according to claim 1, characterized in that, The generation of the task resource mapping matrix specifically includes: Task request information is extracted from a standardized multi-source dataset. The task request records of multiple users are sorted by timestamp and standardized in format to obtain a unified task request sequence. For each task request in the unified task request sequence, extract the task demand vector and energy consumption demand vector, and perform element-wise weighted summation and amplitude normalization according to preset weight coefficients to obtain the fused task feature vector. The task feature vector and the comprehensive result of task energy consumption are matched according to the task number. The predicted features are extracted and the UAV operating status parameters in the standardized multi-source dataset are used to establish the task feature matrix and the resource feature matrix. The weight of the task feature matrix is ​​adjusted according to the predicted features. The two matrices are then feature-mapped and their similarity is calculated. The task feature matrix and resource feature matrix are sorted and paired according to similarity values. The corresponding task numbers and UAV numbers are associated and mapped according to the matching degree. The matching score of each task and resource pair is calculated and recorded to generate a task-resource mapping matrix.

6. The shared drone task scheduling method based on big data analysis according to claim 1, characterized in that, The generation of the optimal scheduling strategy specifically includes: Based on the task resource mapping matrix, key feature parameters are extracted from task allocation data and path planning data, and each parameter is normalized to obtain a standardized scheduling feature set. A linear weighted operation is performed on each parameter in the standardized scheduling feature set according to a preset weight coefficient, and the weighted results are summed to obtain a multi-objective optimization function. Reinforcement learning modeling is performed on the multi-objective optimization function. A state vector space is jointly constructed based on the task state characteristics and UAV state parameters. A set of task allocation actions and path planning actions is defined and an action space is established. The policy network structure is initialized. The state vector space is used as the input layer and the action space is used as the output layer. The network weights and bias parameters are randomly initialized to obtain the initialized policy network. In each iteration, task allocation actions and path planning actions are selected based on the state vector space. The selected actions are input into the multi-objective optimization function to calculate the corresponding reward value. The weights and bias parameters of the initial policy network are updated based on the reward value feedback. The above process is repeated until the policy network converges and the optimal scheduling policy is generated.

7. The shared drone task scheduling method based on big data analysis according to claim 1, characterized in that, The task instructions include task number, target location coordinates, task type, task priority, estimated execution time, path planning parameters, energy consumption control parameters, and feedback upload frequency.

Images