Unmanned aerial vehicle flight task allocation method and system based on artificial intelligence

Abstract

The invention discloses an unmanned aerial vehicle flight task allocation method and system based on artificial intelligence, and the method comprises the steps: carrying out the coding of an unmanned aerial vehicle state of feature matrix data through an RNN recurrent neural network, and constructing a comprehensive state feature vector; a distribution model is established based on a GAN generative adversarial network, a generator inputs the comprehensive state feature vector into the generator to output a task distribution sequence, and a discriminator evaluates the rationality of the task distribution sequence by adopting a convolutional neural network (CNN) structure to obtain an initial task distribution sequence; and executing tasks according to the initial task allocation sequence, setting each unmanned aerial vehicle as an agent based on MARL multi-agent reinforcement learning, and performing dynamic adjustment according to a real-time condition to obtain a target task allocation sequence. The task completion rate and the resource utilization efficiency of the unmanned aerial vehicle are remarkably improved, and human intervention requirements are reduced.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of unmanned aerial vehicle cooperative control, in particular to an unmanned aerial vehicle flight task allocation method and system based on artificial intelligence. BACKGROUND

[0002] With the rapid development of unmanned aerial vehicle technology, unmanned aerial vehicle clusters have become an important application in various fields. In a complex dynamic battlefield environment, an unmanned aerial vehicle cluster needs to complete complex tasks such as large-area cooperative reconnaissance and surveillance, cooperative multi-target saturation attack, etc. In the prior art, unmanned aerial vehicle task allocation and flight path planning are usually processed separately, resulting in limited cooperative effect. Existing unmanned aerial vehicle task allocation methods are mainly based on graph theory and heuristic algorithms, while flight path planning mainly uses genetic algorithms and simulated annealing algorithms. Although these methods have achieved certain results in a single link, they lack a cooperative optimization mechanism for task allocation and flight path planning. In particular, in a dynamically changing battlefield environment, existing methods are difficult to respond to environmental changes in real time, resulting in low task allocation efficiency and low utilization rate of unmanned aerial vehicle resources. SUMMARY

[0003] The application aims to solve the above problems and designs an unmanned aerial vehicle flight task allocation method and system based on artificial intelligence.

[0004] To achieve the above-mentioned purpose, the technical solution of the application is as follows: further, in the unmanned aerial vehicle flight task allocation method based on artificial intelligence, the unmanned aerial vehicle flight task allocation method comprises the following steps: Collecting unmanned aerial vehicle state information, task information and environmental information, establishing a state space representation according to the collected information, and obtaining feature matrix data; Encoding the unmanned aerial vehicle state of the feature matrix data using an RNN recurrent neural network to construct a comprehensive state feature vector; Establishing an allocation model based on a GAN generative adversarial network, a generator based on an LSTM-Seq2Seq structure, inputting the comprehensive state feature vector into the generator to output a task allocation sequence, a discriminator adopting a convolutional neural network (CNN) structure to evaluate the rationality of the task allocation sequence, and obtaining an initial task allocation sequence; According to the initial task allocation sequence, performing a task, setting each unmanned aerial vehicle as an agent based on MARL multi-agent reinforcement learning, and using a GDM-MADDPG algorithm to dynamically adjust according to real-time conditions to obtain a target task allocation sequence.

[0005] Further, in the unmanned aerial vehicle flight task allocation method based on artificial intelligence, the collecting of unmanned aerial vehicle state information, task information and environmental information, the establishment of a state space representation according to the collected information, and the obtaining of feature matrix data comprise: Collecting unmanned aerial vehicle state information, task information and environment information, the unmanned aerial vehicle state information includes position, power, load and communication capability, the task information includes position, priority, completion time requirement; the environment information includes obstacle distribution and weather condition; The collected data is sequentially subjected to data cleaning, normalization and time sequence alignment, the dimensions and value ranges of the state space are defined with unmanned aerial vehicles, tasks and environment as cores, assuming that there are N unmanned aerial vehicles and M tasks in the system, the state space dimensions are: N unmanned aerial vehicle state dimensions + M task state dimensions + 1 environment global dimensions, and the feature matrix data is obtained.

[0006] Further, in the unmanned aerial vehicle flight task allocation method based on artificial intelligence, the unmanned aerial vehicle state of the feature matrix data is encoded by using the RNN recurrent neural network to construct a comprehensive state feature vector, including: An LSTM long short-term memory network is used as the RNN core structure, 1 input layer, 2 LSTM hidden layers are set, the number of hidden units in each layer is set to 64, and the activation function uses tanh; The time sequence correlation of the encoded unmanned aerial vehicle state is taken as the target, the loss function uses MSE mean square error, the optimizer uses Adam, the training iteration number is set to 100 rounds, 80% data is used for training and 20% data is used for verification in each round, and the verification set loss is stable until convergence; The trained LSTM network is used as an encoder, the last time step LSTM hidden layer output is extracted after inputting a three-dimensional time sequence tensor, and a comprehensive state feature vector is constructed.

[0007] Further, in the unmanned aerial vehicle flight task allocation method based on artificial intelligence, the allocation model is established based on the GAN generative adversarial network, the generator is based on the LSTM-Seq2Seq structure, the comprehensive state feature vector is input into the generator to output the task allocation sequence, the discriminator uses a convolutional CNN neural network structure to evaluate the rationality of the task allocation sequence, and an initial task allocation sequence is obtained, including: The generator uses an LSTM-Seq2Seq encoder and decoder structure, and the comprehensive state feature vector is input into the generator; The encoder converts the comprehensive state feature vector into a fixed-dimension context vector, captures the adaptation features of the unmanned aerial vehicle and the task, and the decoder uses 1 LSTM hidden layer and 1 fully connected output layer to output the allocation sequence in an element-by-element generation manner.

[0008] Furthermore, in the aforementioned AI-based UAV flight mission allocation method, the allocation model is established based on a GAN (Generative Adversarial Network). The generator is based on an LSTM-Seq2Seq structure. The comprehensive state feature vector is input into the generator to output a mission allocation sequence. The discriminator uses a convolutional CNN (Convolutional Neural Network) structure to evaluate the rationality of the mission allocation sequence, thus obtaining an initial mission allocation sequence, including: The discriminator uses a CNN to extract local rationality features of the allocation sequence. The input is the allocation sequence output by the generator. The input layer converts the allocation sequence into a one-dimensional matrix, and then expands it into a three-dimensional feature map to obtain the initial task allocation sequence.

[0009] Furthermore, in the aforementioned AI-based UAV flight mission allocation method, the step of executing tasks according to the initial mission allocation sequence, where each UAV is designated as an agent based on MARL multi-agent reinforcement learning, and the GDM-MADDPG algorithm is used to dynamically adjust the sequence based on real-time conditions to obtain the target mission allocation sequence, includes: Each of the N drones in the system is designated as an independent intelligent agent, and the core objective of each intelligent agent is defined. The state space is defined as the real-time state of each agent, including the current remaining battery power, the progress of the executed tasks, the distance between the current position and the task target position, real-time environmental changes, and the task execution status of other agents; The action space is defined as the actions of each agent, including continuing to perform the current task, requesting to adjust the task, and rejecting the current task. The action space has a dimension of 3 and uses discrete action encoding.

[0010] Furthermore, in the aforementioned AI-based UAV flight mission allocation method, the step of executing tasks according to the initial mission allocation sequence, where each UAV is designated as an agent based on MARL multi-agent reinforcement learning, and the GDM-MADDPG algorithm is used to dynamically adjust the sequence based on real-time conditions to obtain the target mission allocation sequence, includes: Each agent dynamically adjusts itself based on its real-time state using the GDM-MADDPG algorithm, and autonomously makes decisions and outputs actions through the Actor network to perform distributed drone missions.

[0011] Furthermore, in the AI-based drone flight mission allocation system, the drone flight mission allocation system includes the following modules: The feature data acquisition module is used to collect UAV status information, mission information, and environmental information, and to establish a state space representation based on the collected information to obtain feature matrix data. The state vector encoding module is used to encode the UAV state of the feature matrix data using an RNN recurrent neural network to construct a comprehensive state feature vector. The initial task allocation module is used to establish an allocation model based on the GAN generative adversarial network. The generator is based on the LSTM-Seq2Seq structure. The comprehensive state feature vector is input into the generator to output the task allocation sequence. The discriminator uses a convolutional CNN structure to evaluate the rationality of the task allocation sequence and obtain the initial task allocation sequence. The target task allocation module is used to execute tasks according to the initial task allocation sequence. Based on MARL multi-agent reinforcement learning, each UAV is set as an agent, and the GDM-MADDPG algorithm is used to dynamically adjust according to the real-time situation to obtain the target task allocation sequence.

[0012] Furthermore, in the AI-based UAV flight mission allocation system, the target mission allocation module includes the following sub-modules: The target submodule is used to set up N drones in the system as N independent intelligent agents and define the core target of each intelligent agent; The state submodule is used to define the state space as the real-time state of each agent, including the current remaining battery power, the progress of the executed tasks, the distance between the current position and the task target position, real-time environmental changes, and the task execution status of other agents. The action submodule is used to define the actions of each agent in the action space, including continuing to perform the current task, requesting to adjust the task, and rejecting the current task. The action space has a dimension of 3 and uses discrete action encoding.

[0013] Furthermore, in the AI-based UAV flight mission allocation system, the target mission allocation module includes the following sub-modules: The output submodule is used by each agent to dynamically adjust according to its real-time state using the GDM-MADDPG algorithm, and to autonomously make decisions and output actions through the Actor network to perform distributed tasks for the drone.

[0014] Its beneficial effects lie in the fact that by integrating RNN, GAN, and multi-agent reinforcement learning (MARL), it achieves deep encoding and sequential decision-making of multi-dimensional information, effectively improving the accuracy and adaptability of task allocation in complex environments. The LSTM-Seq2Seq generator in the generative adversarial network can output high-quality initial allocation schemes, which, combined with a CNN discriminator for rationality evaluation, ensure the feasibility and efficiency of the schemes. The introduction of the GDM-MADDPG algorithm for real-time dynamic adjustment enables the UAV to autonomously coordinate and replan based on environmental changes and aircraft status during mission execution, significantly enhancing the system's anti-interference capability and robustness. This structure not only supports large-scale UAV cooperative control but also significantly improves UAV mission completion rate and resource utilization efficiency, reducing the need for human intervention. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 This is a schematic diagram of the first embodiment of the artificial intelligence-based drone flight mission allocation method in this invention. Figure 2 This is a schematic diagram of the second embodiment of the artificial intelligence-based drone flight mission allocation method in this invention. Figure 3 This is a schematic diagram of the first embodiment of the artificial intelligence-based drone flight mission allocation system in this invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms "one," "an," and "this" used herein may also include the plural forms. It should be further understood that the terminology used in this specification includes the presence of features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0019] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 As shown, an artificial intelligence-based drone flight mission allocation method includes the following steps: Step 101: Collect UAV status information, mission information, and environmental information; establish a state space representation based on the collected information to obtain feature matrix data. Specifically, in this embodiment, UAV status information, task information, and environmental information are collected. UAV status information includes location, battery level, payload, and communication capabilities; task information includes location, priority, and completion time requirements; and environmental information includes obstacle distribution and weather conditions. The collected data is sequentially cleaned, normalized, and time-series aligned. The dimensions and value ranges of the state space are defined with UAVs, tasks, and environment as the core. Assuming there are N UAVs and M tasks in the system, the state space dimension is: N × UAV state dimension + M × task state dimension + 1 × global environment dimension, resulting in feature matrix data.

[0020] Information collection is executed in layers. UAV status information collection: Static and dynamic parameters of a single UAV are collected in real time via onboard sensors (GPS, battery sensor, load sensor) and a ground control terminal. Static parameters include the UAV model, which determines maximum load, maximum endurance, and mission suitability type (reconnaissance / transportation / rescue). Dynamic parameters include current remaining battery power (accurate to 1%), real-time flight speed (m / s), current location coordinates (latitude, longitude, and altitude), completed mission progress, percentage of reconnaissance area completed, equipment health status (motor temperature), and communication module signal strength. The collection frequency is set to once every 30 seconds to avoid data redundancy. Mission information collection: Core mission attributes are entered through a mission release terminal or obtained from a satellite remote sensing system and categorized by mission priority and requirements. Core attributes include mission type (emergency reconnaissance, material delivery, area inspection), target coordinates (accurate to within 10 meters), mission time limit (completed within 2 hours), mission resource requirements (transportation missions require a 5kg load capacity), and mission priority (levels 1-5, with level 1 being the highest). For multi-task scenarios, the dependencies between tasks must be marked; reconnaissance of area A must be completed before delivery to area B can be executed. Environmental information collection: Combining meteorological satellite data, ground radar, and real-time UAV perception data, dynamic environmental parameters of the mission area must be collected. This includes real-time weather (wind speed > 8 m / s must be marked as high risk, rainfall > 5 mm / h must be recorded), obstacle distribution, building height, no-fly zone coordinates, terrain slope, communication environment, 5G / satellite signal coverage strength in the mission area (in areas with weak coverage, signal interruption probability must be marked), airspace conflict risk, and the real-time location of other aircraft to avoid airspace overlap.

[0021] The state space representation is constructed around the three dimensions of UAV, task, and environment, defining the dimensions and value ranges of the state space. Assuming there are N UAVs and M tasks in the system, the state space dimension is N×, consisting of the UAV state dimension + M×, the task state dimension + 1×, and the global environment dimension. For example, if the UAV state dimension is set to 8 (model, remaining battery power, speed, location, etc.), the task state dimension is set to 6 (type, coordinates, time limit, etc.), and the global environment dimension is set to 5 (weather, communication, obstacles, etc.), then the total state space dimension is 8N+6M+5. The value range of each dimension is constrained by actual physical meaning, with remaining battery power ∈ [0,100] and wind speed ∈ [0,20].

[0022] Feature matrix data generation and data preprocessing: The collected raw data is cleaned to remove sensor outliers, such as a sudden drop in battery power from 50% to 0%. Normalization is performed, mapping all parameters to the [0,1] interval, such as remaining battery power 50% → 0.5, wind speed 8m / s → 0.4. Time-series alignment is ensured to guarantee a one-to-one correspondence between UAV, mission, and environmental data at the same timestamp. Matrix structure definition: A two-dimensional feature matrix of sample × feature is constructed. The sample dimension consists of UAV-mission pairings, totaling N × M samples. The feature dimension represents the comprehensive attributes of each pairing, including UAV static / dynamic parameters, mission requirement parameters, and corresponding environmental parameters, totaling 8 + 6 + 5 = 19 features. The final feature matrix has a size of (N × M) × 19.

[0023] Step 102: Encode the UAV state of the feature matrix data using an RNN recurrent neural network to construct a comprehensive state feature vector; Specifically, in this embodiment, an LSTM (Long Short-Term Memory) network is used as the core structure of the RNN, with one input layer and two LSTM hidden layers. The number of hidden units in each layer is set to 64, and the activation function is tanh. The goal is to encode the temporal correlation of the UAV state. The loss function is MSE (Mean Squared Error), the optimizer is Adam, and the number of training iterations is set to 100 rounds. In each round, 80% of the data is used for training and 20% of the data is used for validation until the validation set loss is stable and converged. The trained LSTM network is used as an encoder. After inputting a three-dimensional temporal tensor, the output of the LSTM hidden layer at the last time step is extracted to construct a comprehensive state feature vector.

[0024] In the network architecture design phase, considering the significant temporal dependencies of drone states, such as remaining battery consumption and real-time position movement, and the tendency of traditional RNNs to suffer from gradient vanishing problems, an LSTM (Long Short-Term Memory) network was chosen as the core structure of the RNN. Its gating mechanism—input gate, forget gate, and output gate—effectively captures state correlation features over long time periods. Specifically, a single input layer is set to receive a temporally processed three-dimensional feature tensor, with dimensions of sample number × time step × feature number. This corresponds to drone-task paired samples, continuous acquisition time windows, and multi-dimensional state features, ensuring... The original time-series data can be accurately fed into the network. At the same time, two LSTM hidden layers are constructed, with 64 hidden units in each layer. The two-layer structure can improve the network's ability to fit complex time-series patterns and avoid the problem of insufficient capture of deep correlations by a single-layer network. The 64 hidden units can balance the computational complexity of the network while ensuring the accuracy of feature extraction and prevent overfitting. The activation function is tanh, and its output range is [-1,1]. It can not only adapt to the input data that is in the [0,1] interval after normalization, but also enhance the network's ability to express time-series features through nonlinear transformation and alleviate the gradient saturation phenomenon.

[0025] During model training, the core objective is to encode the temporal correlation of UAV states. This involves the LSTM network learning the state changes of the UAV within a continuous time window, such as the gradual decrease of remaining battery power over task execution time and the dynamic shift of position along the flight trajectory. This provides temporally meaningful features for subsequent task allocation. The loss function chosen is the mean squared error (MSE), which calculates the mean squared difference between the network's predicted temporal features and the actual UAV state features. This accurately measures the deviation of continuous value predictions, aligning with the regression task requirements for encoding the temporal correlation of UAV states. The optimizer used is Adam, which combines the advantages of momentum gradient descent and adaptive learning rate, dynamically adjusting the learning rate based on the gradient magnitude of different parameters. This approach avoids the tedious manual parameter tuning and accelerates training convergence, making it particularly suitable for optimizing a large number of parameters in LSTM networks. The training iteration count is set to 100 epochs, ensuring the network fully learns data patterns while avoiding overfitting due to excessive iteration. Data is divided into 80% for training and 20% for validation. The training set is used for updating network parameters, while the validation set is used to monitor the model's generalization ability, preventing the model from only adapting to training data and failing to handle new UAV state data. Training terminates when the validation set loss stably converges. Specifically, the criterion is that the fluctuation range of the validation set loss is less than 1% for 10 consecutive epochs. This indicates that the model has fully learned temporal correlation features, and further training will not significantly improve performance, at which point training can be stopped.

[0026] In the comprehensive state feature vector construction stage, the trained and validated LSTM network is used as a dedicated encoder. Its core function is to transform the high-dimensional, dynamic three-dimensional temporal tensor into a low-dimensional, concise static feature vector. The input data is the three-dimensional temporal tensor constructed above, which consists of the number of samples × time steps × number of features. Each sample corresponds to a set of continuous time window data for UAV-task pairing. The LSTM encoder gradually accumulates the state information of each time step when processing temporal data through the gating mechanism of the hidden layer. The output of the LSTM hidden layer at the last time step has integrated the temporal features of the UAV state of all time steps within the time window, such as the trend of power change and position movement within the entire window, which can comprehensively and compactly represent the comprehensive state of the UAV at this stage. Finally, the output of the hidden layer at the last time step of each UAV-task pairing sample is used as the comprehensive state feature vector. The vector dimension is consistent with the number of LSTM hidden units, i.e., 64 dimensions, providing high-quality feature input for the subsequent generation of task assignment sequences by the GAN generative adversarial network.

[0027] Step 103: Establish an assignment model based on GAN generative adversarial network. The generator is based on LSTM-Seq2Seq structure. The comprehensive state feature vector is input into the generator to output the task assignment sequence. The discriminator uses convolutional CNN structure to evaluate the rationality of the task assignment sequence and obtain the initial task assignment sequence. Specifically, in this embodiment, the generator adopts an LSTM-Seq2Seq encoder and decoder structure, inputting the comprehensive state feature vector into the generator. The encoder converts the comprehensive state feature vector into a fixed-dimensional context vector, capturing the adaptation features of the UAV and the task. The decoder uses one LSTM hidden layer and one fully connected output layer to output the allocation sequence in an element-wise generation manner. The discriminator uses a CNN to extract local rationality features of the allocation sequence. The input is the allocation sequence output by the generator. The input layer converts the allocation sequence into a one-dimensional matrix, and then expands it into a three-dimensional feature map to obtain the initial task allocation sequence.

[0028] To achieve both constraint adaptability and global rationality in UAV task allocation, a dual-module collaborative allocation model needs to be constructed based on a Generative Adversarial Network (GAN). Its core logic is to use dynamic adversarial training between the generator and the discriminator to ensure that the generated task allocation sequence satisfies the matching conditions of UAV capabilities, task requirements, and environmental constraints, while also possessing global optimality. The generator employs an LSTM-Seq2Seq encoder-decoder architecture, which is naturally suited for sequence generation tasks. First, the 64-dimensional comprehensive state feature vector constructed earlier, encompassing the temporal correlation features of UAV-task pairing and environmental adaptation features, is input into the encoder. The encoder further refines key features through the gating mechanism of the LSTM unit, outputting a fixed-length context vector of dimension 64. This vector condenses the core information of whether a UAV is suitable for a certain task. Subsequently, the decoder uses this context vector as the initial input and generates a sequence through the time-step iteration of the LSTM unit. Each time step corresponds to the allocation decision of one task, and the output is a discrete value of 0 or 1. 1 indicates that the current task is assigned to the corresponding UAV, and 0 indicates that it is not assigned. Finally, a task allocation sequence with the same length as the number of tasks is generated, and the built-in task uniqueness constraint layer ensures that each task is assigned to only one UAV to avoid conflicts. The discriminator, working in conjunction with the generator, employs a CNN (Convolutional Neural Network) structure. Its core function is to evaluate the rationality of the generated sequences. The discriminator's input includes two types of data: one is the task assignment sequence to be evaluated output by the generator, which needs to be converted into a 1×M one-dimensional feature map, where M is the number of tasks; the other is a manually labeled rational assignment sequence, which serves as a real sample and must meet all constraints such as drone load, battery power, and environmental adaptability. Through two convolutional layers, using 1×3 and 1×5 convolutional kernels respectively, the local correlation features of adjacent task assignments and the global correlation features across tasks are extracted from the sequence. A max pooling layer is used to compress the feature dimension while retaining key information. The system also undergoes stepwise processing with two fully connected layers, ultimately outputting a rationality score between 0 and 1, where 1 indicates completely rational and 0 indicates completely unreasonable. During the adversarial training of GANs, the generator aims to maximize the probability of the discriminator misclassifying the generated sequence, that is, to make the generated sequence as close as possible to the real and reasonable sequence. The discriminator aims to minimize the discrimination error between the generated sequence and the real sequence. The two are trained alternately and iteratively. In each round, the generator is fixed and the discriminator is optimized, and then the discriminator is fixed and the generator is optimized, until the difference between the scores of the discriminator for the two types of sequences is less than 0.1 and the constraint satisfaction rate of the generated sequence, such as the proportion of drones whose battery power can support task allocation reaches more than 95%, is the initial task allocation sequence output by the generator that takes into account both constraints and reasonableness.

[0029] Step 104: Execute tasks according to the initial task allocation sequence. Based on MARL multi-agent reinforcement learning, each UAV is set as an agent. The GDM-MADDPG algorithm is used to dynamically adjust according to the real-time situation to obtain the target task allocation sequence.

[0030] Specifically, in this embodiment, the N drones in the system are set as N independent intelligent agents, and the core objective of each intelligent agent is defined; The state space is defined as the real-time state of each agent, including the remaining battery power, progress of executed tasks, distance between the current location and the task target location, real-time environmental changes, and the task execution status of other agents. The action space is defined as the actions of each agent, including continuing the current task, requesting task adjustment, and rejecting the current task. The action space has three dimensions and uses discrete action encoding. Each agent dynamically adjusts itself based on its real-time state using the GDM-MADDPG algorithm, and autonomously decides and outputs actions through the Actor network to perform distributed task execution for the UAV.

[0031] MARL Multi-Agent System Modeling: Agent Definition: N drones within the system are defined as N independent agents. The core objective of each agent is to complete its assigned tasks and maximize global gains while satisfying its own constraints. State Space Definition: The real-time state of each agent includes: remaining battery power, progress of executed tasks, distance between current position and task target position, real-time environmental changes, and task execution status of other agents, totaling 7 dimensions, all normalized to [0,1]. Action Space Definition: Each agent's actions include 3 categories: continue executing the current task, request task adjustment (e.g., lower task priority, reject the current task), and insufficient battery power. The action space has 3 dimensions and uses discrete action encoding: 0 = continue, 1 = request adjustment, 2 = reject.

[0032] Reward function design: A hybrid reward mechanism of global benefit + local constraint is adopted: Positive reward: 10 points for completing the task, 5 extra points for completing the task ahead of time, and 8 points for assisting other agents in completing emergency tasks; Negative penalty: 15 points are deducted for task timeout, 20 points are deducted for running out of power and failing to complete the task, and 10 points are deducted for airspace conflict and overlapping with the position of other drones; Reward normalization: The reward for each step is mapped to the interval [-1,1] to avoid the reward value being too large and causing training instability.

[0033] The GDM-MADDPG algorithm execution steps are as follows: GDM-MADDPG, a multi-agent deep deterministic policy gradient algorithm based on gradient descent, adopts a centralized training and distributed execution mode. The specific steps are as follows: Network initialization: For each agent, an Actor network is built to output action probabilities, and a Critic network is built to evaluate action value. Simultaneously, a global Critic network is built to evaluate the joint action value of all agents. The network initialization parameters adopt a Xavier normal distribution, with a learning rate of 0.001 and a discount factor of γ of 0.9, emphasizing long-term returns. Experience replay pool construction: Each agent stores its real-time state-action-reward-next state experience tuple into a shared experience replay pool with a pool size of 1×10n. 5 To avoid the influence of experience-related factors on training.

[0034] Centralized training: After accumulating 1000 experiences, 32 experiences are randomly sampled from the replay pool for batch training: Critic network training: The Q-value of the joint action is calculated through the global Critic network to minimize the difference between the Q-value and the target Q-value. The Critic network parameters are updated based on the error calculated in the next state. Actor network training: The action value output by the Critic network is maximized through policy gradient ascent, while adding an action regularization term to avoid frequent action adjustments by the agent. Target network update: Every 10 training epochs, the Actor / Critic network parameters are copied to the target network to mitigate training fluctuations. Distributed execution: After training, each agent autonomously decides its output action based on its real-time state through the Actor network, without relying on global information, achieving distributed task execution.

[0035] Dynamic adjustment of trigger and target sequence generation, and adjustment of trigger conditions: Dynamic adjustment is triggered when the following scenarios occur: ① The drone's remaining battery power is <20% and the mission has not been completed; ② Sudden environmental changes, such as wind speed >12m / s or the appearance of a temporary no-fly zone; ③ Change in mission priority, such as the addition of a Level 1 emergency mission; ④ Drone equipment failure.

[0036] Adjustment Logic: After adjustment is triggered, the system re-collects the current UAV-task-environment information and evaluates the current value of the initial sequence through the global Critic network of the GDM-MADDPG algorithm. If the value is <0.3, it indicates that the sequence is no longer reasonable. In this case, the Actor network outputs the adjusted actions, such as assigning the tasks of the faulty UAV to idle UAVs or postponing low-priority tasks. Target Task Allocation Sequence: The adjustment is repeated until the global Critic network evaluates the value >0.8, indicating that the sequence meets the real-time constraints and maximizes the benefit. The output sequence at this time is the target task allocation sequence and is distributed to each UAV for execution.

[0037] After the drone mission execution process is started based on the initial task allocation sequence, due to the large number of dynamic variables in the actual flight scenario, such as some drones suddenly consuming power faster, strong winds or no-fly zones appearing temporarily in the mission area, and the addition of new high-priority emergency tasks, the static initial sequence is difficult to continuously adapt to real-time requirements. Therefore, it is necessary to introduce the MARL multi-agent reinforcement learning framework to build a dynamic adjustment mechanism.

[0038] Its beneficial effects lie in the fact that by integrating RNN, GAN, and multi-agent reinforcement learning (MARL), it achieves deep encoding and sequential decision-making of multi-dimensional information, effectively improving the accuracy and adaptability of task allocation in complex environments. The LSTM-Seq2Seq generator in the generative adversarial network can output high-quality initial allocation schemes, which, combined with a CNN discriminator for rationality evaluation, ensure the feasibility and efficiency of the schemes. The introduction of the GDM-MADDPG algorithm for real-time dynamic adjustment enables the UAV to autonomously coordinate and replan based on environmental changes and aircraft status during mission execution, significantly enhancing the system's anti-interference capability and robustness. This structure not only supports large-scale UAV cooperative control but also significantly improves UAV mission completion rate and resource utilization efficiency, reducing the need for human intervention.

[0039] Please see Figure 2 In the AI-based UAV flight mission allocation method, the UAV state is encoded using an RNN recurrent neural network to construct a comprehensive state feature vector, which includes the following steps: Step 201: Use LSTM (Long Short-Term Memory) network as the core structure of RNN, set up 1 input layer and 2 LSTM hidden layers, set the number of hidden units in each layer to 64, and use tanh as the activation function; Step 202: With the goal of encoding the temporal correlation of UAV state, the loss function is MSE (mean squared error), the optimizer is Adam, the number of training iterations is set to 100 rounds, and each round uses 80% of the data for training and 20% of the data for validation, until the validation set loss is stable and converged. Step 203: Using the trained LSTM network as an encoder, input the three-dimensional temporal tensor and extract the output of the LSTM hidden layer at the last time step to construct a comprehensive state feature vector.

[0040] The above describes embodiments of the artificial intelligence-based UAV flight mission allocation method of the present invention. Please refer to [link / reference]. Figure 3 In an AI-based drone flight mission allocation system, the drone flight mission allocation system includes the following modules: The feature data acquisition module is used to collect UAV status information, mission information, and environmental information, and to establish a state space representation based on the collected information to obtain feature matrix data. The state vector encoding module is used to encode the UAV state of the feature matrix data using an RNN recurrent neural network to construct a comprehensive state feature vector. The initial task allocation module is used to establish an allocation model based on the GAN generative adversarial network. The generator is based on the LSTM-Seq2Seq structure. The comprehensive state feature vector is input into the generator to output the task allocation sequence. The discriminator uses a convolutional CNN structure to evaluate the rationality of the task allocation sequence and obtain the initial task allocation sequence. The target task allocation module is used to execute tasks according to the initial task allocation sequence. Based on MARL multi-agent reinforcement learning, each UAV is set as an agent, and the GDM-MADDPG algorithm is used to dynamically adjust according to the real-time situation to obtain the target task allocation sequence.

[0041] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-based method for allocating unmanned aerial vehicle (UAV) flight missions, characterized in that: The method for allocating UAV flight missions includes the following steps: Collect UAV status information, mission information, and environmental information, establish a state space representation based on the collected information, and obtain feature matrix data; The UAV state is encoded using a recurrent neural network (RNN) to construct a comprehensive state feature vector from the feature matrix data. A task allocation model is established based on a generative adversarial network (GAN). The generator is based on an LSTM-Seq2Seq structure. The comprehensive state feature vector is input into the generator to output a task allocation sequence. The discriminator uses a convolutional neural network (CNN) structure to evaluate the rationality of the task allocation sequence and obtain the initial task allocation sequence. The task is executed according to the initial task allocation sequence. Each UAV is set as an agent based on MARL multi-agent reinforcement learning, and the GDM-MADDPG algorithm is used to dynamically adjust according to the real-time situation to obtain the target task allocation sequence.

2. The AI-based UAV flight mission allocation method as described in claim 1, characterized in that, The process involves collecting UAV status information, mission information, and environmental information, establishing a state-space representation based on the collected information, and obtaining feature matrix data, including: Collect drone status information, mission information, and environmental information. Drone status information includes location, battery level, payload, and communication capabilities. Mission information includes location, priority, and completion time requirements. Environmental information includes obstacle distribution and weather conditions. The collected data is cleaned, normalized, and time-series aligned sequentially. The dimensions and value range of the state space are defined with UAVs, tasks, and environment as the core. Assuming there are N UAVs and M tasks in the system, the dimensions of the state space are: N × UAV state dimension + M × task state dimension + 1 × global environment dimension, thus obtaining the feature matrix data.

3. The AI-based UAV flight mission allocation method as described in claim 1, characterized in that, The process of encoding the UAV state of the feature matrix data using an RNN (Recurrent Neural Network) to construct a comprehensive state feature vector includes: The LSTM (Long Short-Term Memory) network is used as the core structure of the RNN, with one input layer and two LSTM hidden layers. The number of hidden units in each layer is set to 64, and the activation function is tanh. The goal is to encode the temporal correlation of UAV states. The loss function is MSE (mean squared error), the optimizer is Adam, and the number of training iterations is set to 100 rounds. In each round, 80% of the data is used for training and 20% of the data is used for validation until the loss on the validation set is stable and converges. The trained LSTM network is used as an encoder. After inputting a three-dimensional temporal tensor, the output of the hidden layer of the LSTM at the last time step is extracted to construct a comprehensive state feature vector.

4. The AI-based UAV flight mission allocation method as described in claim 1, characterized in that, The assignment model is established based on a GAN (Generative Adversarial Network). The generator is based on an LSTM-Seq2Seq structure. The comprehensive state feature vector is input into the generator to output a task assignment sequence. The discriminator uses a convolutional CNN (Convolutional Neural Network) structure to evaluate the rationality of the task assignment sequence, thus obtaining an initial task assignment sequence, including: The generator uses an LSTM-Seq2Seq encoder and decoder structure, and the comprehensive state feature vector is input into the generator; The encoder transforms the comprehensive state feature vector into a fixed-dimensional context vector to capture the adaptation features of the drone and the mission. The decoder uses a 1-layer LSTM hidden layer and a 1-layer fully connected output layer to output the assignment sequence in an element-wise generation manner.

5. The AI-based UAV flight mission allocation method as described in claim 1, characterized in that, The assignment model is established based on a GAN (Generative Adversarial Network). The generator is based on an LSTM-Seq2Seq structure. The comprehensive state feature vector is input into the generator to output a task assignment sequence. The discriminator uses a convolutional CNN (Convolutional Neural Network) structure to evaluate the rationality of the task assignment sequence, thus obtaining an initial task assignment sequence, including: The discriminator uses a CNN to extract local rationality features of the allocation sequence. The input is the allocation sequence output by the generator. The input layer converts the allocation sequence into a one-dimensional matrix, and then expands it into a three-dimensional feature map to obtain the initial task allocation sequence.

6. The AI-based UAV flight mission allocation method as described in claim 1, characterized in that, The step of executing tasks according to the initial task allocation sequence involves setting each UAV as an agent based on MARL multi-agent reinforcement learning, and dynamically adjusting the sequence according to real-time conditions using the GDM-MADDPG algorithm to obtain the target task allocation sequence, including: Each of the N drones in the system is designated as an independent intelligent agent, and the core objective of each intelligent agent is defined. The state space is defined as the real-time state of each agent, including the current remaining battery power, the progress of the executed tasks, the distance between the current position and the task target position, real-time environmental changes, and the task execution status of other agents; The action space is defined as the actions of each agent, including continuing to perform the current task, requesting to adjust the task, and rejecting the current task. The action space has a dimension of 3 and uses discrete action encoding.

7. The AI-based UAV flight mission allocation method as described in claim 1, characterized in that, The step of executing tasks according to the initial task allocation sequence involves setting each UAV as an agent based on MARL multi-agent reinforcement learning, and dynamically adjusting the sequence according to real-time conditions using the GDM-MADDPG algorithm to obtain the target task allocation sequence, including: Each agent dynamically adjusts itself based on its real-time state using the GDM-MADDPG algorithm, and autonomously makes decisions and outputs actions through the Actor network to perform distributed drone missions.

8. An artificial intelligence-based unmanned aerial vehicle (UAV) flight mission allocation system, characterized in that, The UAV flight mission allocation system includes the following modules: The feature data acquisition module is used to collect UAV status information, mission information, and environmental information, and to establish a state space representation based on the collected information to obtain feature matrix data. The state vector encoding module is used to encode the UAV state of the feature matrix data using an RNN recurrent neural network to construct a comprehensive state feature vector. The initial task allocation module is used to establish an allocation model based on the GAN generative adversarial network. The generator is based on the LSTM-Seq2Seq structure. The comprehensive state feature vector is input into the generator to output the task allocation sequence. The discriminator uses a convolutional CNN structure to evaluate the rationality of the task allocation sequence and obtain the initial task allocation sequence. The target task allocation module is used to execute tasks according to the initial task allocation sequence. Based on MARL multi-agent reinforcement learning, each UAV is set as an agent, and the GDM-MADDPG algorithm is used to dynamically adjust according to the real-time situation to obtain the target task allocation sequence.

9. The AI-based UAV flight mission allocation system as described in claim 8, characterized in that, The target task allocation module includes the following sub-modules: The target submodule is used to set up N drones in the system as N independent intelligent agents and define the core target of each intelligent agent; The state submodule is used to define the state space as the real-time state of each agent, including the current remaining battery power, the progress of the executed tasks, the distance between the current position and the task target position, real-time environmental changes, and the task execution status of other agents. The action submodule is used to define the actions of each agent in the action space, including continuing to perform the current task, requesting to adjust the task, and rejecting the current task. The action space has a dimension of 3 and uses discrete action encoding.

10. The AI-based UAV flight mission allocation system as described in claim 8, characterized in that, The target task allocation module includes the following sub-modules: The output submodule is used by each agent to dynamically adjust according to its real-time state using the GDM-MADDPG algorithm, and to autonomously make decisions and output actions through the Actor network to perform distributed tasks for the drone.

Images