An optimal path planning method and system for unmanned aerial vehicle swarm

Abstract

The present invention relates to the technical field of unmanned aerial vehicle flight planning, and discloses an optimal path planning method and system for unmanned aerial vehicle clusters. The method first constructs an unmanned aerial vehicle flight state prediction model and a path planning decision model to respectively predict the future state of the unmanned aerial vehicle and decide on the optimal path. The innovative use of a dual machine learning algorithm to train the two models not only improves the accuracy of model prediction and decision-making, but also effectively corrects and optimizes the path planning results by constructing a residual fitting model, thereby enhancing the robustness of the planning. The present invention comprehensively considers the dynamic flight state, environmental uncertainty and mission objective diversity of the unmanned aerial vehicle, and can generate in real time the optimal path planning instructions for the unmanned aerial vehicle cluster that meet the actual flight conditions. The method improves the task execution efficiency and safety of the unmanned aerial vehicle cluster, is suitable for military reconnaissance, disaster relief, logistics distribution and other fields, and has broad application prospects and important practical value.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) flight planning, and in particular to an optimal path planning method and system for an unmanned aerial vehicle (UAV) cluster. Background Art

[0002] With the rapid development of drone technology, drone swarms have shown great application potential in many fields such as military reconnaissance, disaster relief, and logistics distribution. However, how to efficiently and accurately plan the optimal path of drone swarms to ensure efficient execution of tasks and safe flight of drones is a major challenge currently faced.

[0003] Current UAV path planning methods often rely on a single algorithm or model, such as rule-based methods, heuristic search algorithms, or classic optimization algorithms, etc. These methods may perform well in simple or static environments, but in complex and changing actual scenarios, they often fail to fully consider the dynamic flight state of the UAV, the uncertainty of the environment, and the diversity of mission objectives, and thus it is often difficult to obtain the optimal path planning results.

[0004] In addition, drone swarms are affected by many factors during flight, such as wind direction, wind speed, obstacles, other aircraft, etc. These factors can cause the actual flight status of drones to deviate from expectations. Traditional path planning methods often lack effective processing mechanisms for these deviations, resulting in the planned path not being able to achieve the expected results in actual execution. Summary of the invention

[0005] The purpose of the present invention is to provide a method and system for optimal path planning of a drone cluster to solve the problems raised in the above-mentioned background technology.

[0006] To achieve the above object, the present invention provides the following technical solution: a method for optimal path planning of a drone cluster, the method comprising:

[0007] Step A: Construct a UAV flight state prediction model and a UAV path planning decision model, wherein the UAV flight state prediction model is used to predict the future flight state of each UAV in the UAV cluster, and the UAV path planning decision model is used to determine the optimal path of the UAV cluster based on current and historical flight state data, environmental data, and mission target data;

[0008] Step B: Use a dual machine learning algorithm to train the UAV flight state prediction model and the UAV path planning decision model respectively, and use the residuals of the two models to construct a residual fitting model, which is used to correct and optimize the results of the UAV path planning;

[0009] Step C: Use the trained model, combined with real-time flight status data, environmental data, and mission target data, to generate the optimal path planning instructions for the drone cluster.

[0010] Preferably, a long short-term memory network (LSTM) algorithm is used to construct a UAV flight status prediction model, and the flight status of the UAV is predicted through time series analysis.

[0011] Preferably, the training step of the UAV flight state prediction model includes:

[0012] Data preprocessing: Collect and clean UAV flight data, including position, speed, acceleration, attitude, and environmental information, and encode and standardize the collected data;

[0013] Construct LSTM network structure: Design LSTM model, including input layer, LSTM layer, fully connected layer and output layer; the input layer receives the preprocessed UAV flight data, the LSTM layer is used to capture the long-term dependencies in the time series data, the fully connected layer is used to map the output of the LSTM layer to the predicted target state, and the output layer outputs the predicted result of the UAV's future flight state;

[0014] Model training: The preprocessed UAV flight data is input into the LSTM model, the model parameters are optimized by the back propagation algorithm and the gradient descent algorithm, and the cross-validation method is used to select the best network structure and hyperparameters; wherein, the formula for optimizing the model parameters by the gradient descent algorithm is:

[0015]

[0016] in, Represents the model parameters, including all weights and biases that the model needs to learn; Represents the learning rate, which is used to control the step size of parameter update; Represents the loss function About parameters The gradient of , which is a vector pointing in the direction of fastest growth of the loss function; Indicates an assignment operation, that is, updating parameters The value of

[0017] Model evaluation: Evaluate the predictive performance of the model on the test set and adjust model parameters or feature selection based on the evaluation results.

[0018] Preferably, a deep reinforcement learning (DRL) algorithm is used to construct a UAV path planning decision model, and the optimal path planning instructions for the UAV cluster are output through the policy gradient approximation method.

[0019] Preferably, the training steps of the UAV path planning decision model include:

[0020] Dataset preparation: Integrate historical flight status data, environmental data, mission target data, and corresponding path planning results to form a complete data set, including state characteristics, action space, and reward function;

[0021] Construct the DRL model structure: Design the DRL model, including the policy network and the value network. The policy network directly outputs the action probability distribution, and the value network evaluates the state or action state value.

[0022] Model training: Input the data set into the DRL model, generate experience samples by interacting with the environment, use the policy gradient method to update the model parameters, and use experience replay and target network algorithm to stabilize the training process;

[0023] Model optimization: Evaluate the decision-making performance of the model on the validation set, and adjust the network structure, learning rate, and exploration strategy based on the evaluation results.

[0024] Preferably, the step of obtaining training data for the residual fitting model includes:

[0025] Calculate the residual between the actual output and the expected output of the UAV flight state prediction model and the UAV path planning decision model on the training set, specifically:

[0026] For the UAV flight state prediction model, the difference between the predicted state and the actual state at each time step is calculated to obtain the state prediction residual;

[0027] For the UAV path planning decision model, the deviation or cost difference between the planned path and the actual optimal path is calculated to obtain the decision residual;

[0028] The calculated state prediction residuals and decision residuals are used as new features and combined with the original flight state data, environmental data and mission target data to form a residual feature set; the residual feature set is used as training data for the residual fitting model.

[0029] Preferably, an extreme gradient boosting XGBoost algorithm is selected to train the residual feature set to construct a residual fitting model, wherein the structure of the residual fitting model includes an input layer, a plurality of base learners and an output layer:

[0030] The input layer receives the residual feature set as input;

[0031] Multiple base learners process the input data in turn, and each base learner makes up for the residual left by the previous round of models;

[0032] The output layer summarizes the outputs of all base learners to obtain the corrected optimal path planning result of the drone cluster.

[0033] Preferably, in the training step of the UAV flight state prediction model, the method of encoding the collected data includes: performing one-hot encoding conversion on discrete features.

[0034] Preferably, in the training step of the UAV flight state prediction model, the method of standardizing the collected data includes: using the Z-score standardization method to standardize the continuous features, specifically: subtracting the mean of each numerical data and dividing it by its standard deviation, so that the processed data conforms to the standard normal distribution, and the formula for standardization is: ,in, is the original data, is the mean of the original data, is the standard deviation of the original data, is the standardized data.

[0035] Preferably, an optimal path planning system for a drone swarm comprises:

[0036] A model building module, used to build a UAV flight state prediction model and a UAV path planning decision model, wherein the UAV flight state prediction model is configured to predict the future flight state of each UAV in the UAV cluster, and the UAV path planning decision model is configured to determine the optimal path of the UAV cluster based on current and historical flight state data, environmental data, and mission target data;

[0037] A model training and optimization module, connected to the model building module, configured to use a dual machine learning algorithm to train the UAV flight state prediction model and the UAV path planning decision model respectively, and calculate the residuals between the actual outputs and the expected outputs of the two models, and further use these residuals to build a residual fitting model to correct and optimize the results of the UAV path planning;

[0038] The real-time path planning instruction generation module is connected to the model training and optimization module, and is configured to use the trained UAV flight state prediction model, UAV path planning decision model and residual fitting model, combined with the real-time acquired flight state data, environmental data and mission target data, to generate the optimal path planning instructions for the UAV cluster.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] By constructing a UAV flight state prediction model and a UAV path planning decision model, the present invention can comprehensively consider the dynamic flight state of the UAV, the uncertainty of the environment, and the diversity of mission objectives, thereby generating a path planning result that is more in line with actual flight conditions. At the same time, the model is trained using a dual machine learning algorithm, and a residual fitting model is constructed to correct and optimize the path planning results, further improving the accuracy and robustness of path planning.

[0041] Current path planning methods often have difficulty dealing with complex and changing environmental factors, such as wind direction, wind speed, obstacles, etc. However, the present invention obtains environmental data in real time and integrates it into the path planning decision model as input information, so that the drone cluster can better adapt to complex environments, ensure flight safety and complete tasks efficiently.

[0042] The present invention reduces unnecessary movement and waiting time of drone clusters during flight through precise path planning, thereby improving the efficiency of mission execution. This is particularly important for time-sensitive missions such as military reconnaissance and disaster relief, which can respond quickly and effectively. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A step diagram of an optimal path planning method for a drone cluster according to the present invention;

[0044] Figure 2 Training flow chart for the UAV flight status prediction model;

[0045] Figure 3 Training flowchart for fitting a model to residuals. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] See also Figure 1-3 The present invention provides an optimal path planning method for a drone cluster, the method comprising:

[0048] Step A: Build the model

[0049] UAV flight status prediction model: This model is based on the UAV's historical flight data, dynamic model and environmental factors (such as wind direction, wind speed, etc.), and is trained using machine learning algorithms (such as neural networks, support vector machines, etc.) to predict the flight status of each UAV in the UAV cluster in the future, including but not limited to position, speed, attitude, etc.

[0050] UAV path planning decision model: This model comprehensively considers current and historical flight status data, environmental data (such as obstacle distribution, weather conditions, etc.) and mission target data (such as target point location, arrival time requirements, etc.), and uses optimization algorithms (such as genetic algorithms, particle swarm optimization, etc.) or machine learning algorithms (such as deep reinforcement learning, etc.) for training to decide the optimal path for the UAV cluster.

[0051] Step B: Train the model and build a residual fitting model

[0052] Use dual machine learning algorithm training: First, use appropriate machine learning algorithms to train the UAV flight status prediction model and the UAV path planning decision model respectively to obtain preliminary prediction and decision-making capabilities.

[0053] Calculate residuals: During the training process, the difference between the actual output and the expected output of the two models is recorded, that is, the residual. For the flight state prediction model, the residual may represent the deviation between the predicted state and the actual state; for the path planning decision model, the residual may represent the cost difference between the planned path and the ideal path.

[0054] Build a residual fitting model: The calculated residual is used as a new feature, combined with the original input data (such as flight status data, environmental data, mission target data, etc.), and trained using another machine learning algorithm (such as extreme gradient boosting XGBoost, etc.) to obtain a residual fitting model. This model is used to correct and optimize the results of drone path planning and improve the accuracy and robustness of path planning.

[0055] Step C: Generate optimal path planning instructions

[0056] Real-time data acquisition: During the execution of a mission by a drone cluster, the flight status data, environmental data, and mission target data of the drones are acquired in real time.

[0057] Model prediction and decision-making: Input real-time data into the trained UAV flight status prediction model and UAV path planning decision model to perform flight status prediction and path planning decisions.

[0058] Residual correction: Input the prediction and decision results into the residual fitting model to correct and optimize the path planning results.

[0059] Generate instructions: Based on the corrected path planning results, generate the optimal path planning instructions for the drone cluster and send them to the drone cluster for execution.

[0060] Through the above steps, the present invention can comprehensively consider the dynamic flight state of the UAV, the uncertainty of the environment and the diversity of mission objectives, and realize the optimization and correction of the path planning results of the UAV cluster.

[0061] The present invention will be further described below in conjunction with Examples 1 to 3:

[0062] Embodiment 1:

[0063] This embodiment provides a method for building a UAV flight status prediction model using a long short-term memory network (LSTM) algorithm. The method accurately predicts the flight status of the UAV through time series analysis. The following is a detailed description of the specific implementation of the method:

[0064] Step 1: Data preprocessing

[0065] Data Collection: Collect drone flight data from the drone flight control system, including but not limited to position (longitude, latitude, altitude), velocity (three-dimensional velocity vector), acceleration (three-dimensional acceleration vector), attitude (roll angle, pitch angle, yaw angle), and environmental information (such as wind direction, wind speed, temperature, humidity, etc.).

[0066] Data cleaning: De-noise, remove duplicates and process missing values ​​on the collected data to ensure the accuracy and completeness of the data.

[0067] Data encoding: Perform one-hot encoding on discrete features to convert them into numerical data so that the model can process them.

[0068] Data standardization: The Z-score standardization method is used to standardize continuous features. Specifically, each numerical data is subtracted from its mean and divided by its standard deviation so that the processed data conforms to the standard normal distribution. The formula for standardization is: ,in, is the original data, is the mean of the original data, is the standard deviation of the original data, is the standardized data.

[0069] Step 2: Build LSTM network structure

[0070] Design LSTM model: Build an LSTM model consisting of input layer, LSTM layer, fully connected layer, and output layer.

[0071] Input layer: Receives preprocessed UAV flight data, arranges the data in time series, and sets the appropriate input time step.

[0072] LSTM layer: contains one or more layers of LSTM units, which are used to capture long-term dependencies in time series data. LSTM units control the flow of information through forget gates, input gates, and output gates, effectively solving the gradient vanishing and gradient exploding problems in traditional recurrent neural networks (RNNs).

[0073] Fully connected layer: maps the output of the LSTM layer to the predicted target state, such as the future position, speed, acceleration, etc. of the drone.

[0074] Output layer: According to the mission requirements, the prediction results of the future flight status of the UAV are output in the form of regression output or classification output.

[0075] Set model parameters: According to the data scale and task complexity, set the LSTM model's hyperparameters such as the number of hidden units, number of layers, learning rate, batch size, etc.

[0076] Step 3: Model training

[0077] Data partitioning: The preprocessed UAV flight data is divided into training set, validation set and test set for model training, validation and evaluation.

[0078] Model training: Input the training set data into the LSTM model and optimize the model parameters through the back propagation algorithm and the gradient descent algorithm. The formula for optimizing the model parameters by the gradient descent algorithm is:

[0079]

[0080] in, Represents the model parameters, including all weights and biases that the model needs to learn; Represents the learning rate, which is used to control the step size of parameter update; Represents the loss function About parameters The gradient of , which is a vector pointing in the direction of fastest growth of the loss function; Indicates an assignment operation, that is, updating parameters The value of .

[0081] Use the cross-validation method to evaluate the performance of the model on the validation set, and adjust the hyperparameters such as learning rate, number of hidden units, number of layers, etc. based on the evaluation results to obtain the best network structure and prediction performance. Save the trained LSTM model as a record in a file or database for subsequent prediction and evaluation.

[0082] Step 4: Model Evaluation

[0083] Input the test set data into the trained LSTM model to predict the flight status. Calculate the error between the prediction result and the true value, such as mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), etc., to evaluate the prediction performance of the model. According to the evaluation results, analyze the prediction performance of the model and the existing problems, such as overfitting, underfitting, gradient disappearance, etc., and adjust the model parameters or feature selection according to the analysis results to further improve the prediction performance of the model.

[0084] Embodiment 2:

[0085] This embodiment provides a method for building a drone path planning decision model using a deep reinforcement learning (DRL) algorithm. The method outputs the optimal path planning instructions for the drone cluster through a policy gradient approximation method. The following is a detailed description of the specific implementation of the method:

[0086] Step 1: Dataset preparation

[0087] Integrate historical flight status data, environmental data, mission target data and corresponding path planning results from UAV flight records, environmental monitoring systems and mission planning systems.

[0088] Flight status data: including the drone’s position, speed, acceleration, attitude, etc.

[0089] Environmental data: including obstacle locations, weather conditions (such as wind speed and direction), terrain information, etc.

[0090] Mission target data: including the location of the target point, arrival time requirement, mission priority, etc.

[0091] Path planning results: the optimal path output by the historical path planning algorithm or the path planning example given by the expert.

[0092] Extract state features from the integrated data for input to the DRL model. State features can include a combination of the current state of the drone, the state of the environment, and the state of the mission goal. Define the possible action space of the drone, such as turning, accelerating, decelerating, climbing, descending, etc., based on the control capability of the drone and the needs of path planning. Design a reward function based on factors such as mission objectives, safety, and efficiency to evaluate the quality of the drone after performing an action. The reward function can include rewards for reaching the target point, rewards for avoiding obstacles, rewards for reducing flight time, etc.

[0093] Step 2: Build the DRL model structure

[0094] Construct a policy network (Actor network), whose input is state features and output is action probability distribution. The policy network is implemented using a neural network (such as a convolutional neural network CNN or a fully connected neural network FCN), and learns the mapping relationship from state to action through training. Construct a value network (Critic network), whose input is state features or action-state pairs, and output is state values ​​or action-state values. The value network is used to evaluate the quality of the current state or action state, and provide guidance for the update of the policy network.

[0095] Randomly initialize the parameters of the policy network and value network, including weights and biases.

[0096] Step 3: Model training

[0097] The DRL model interacts with the environment, selects actions based on the current state, observes the new state and reward after executing the action, and generates experience samples (state-action-reward-new state).

[0098] The generated experience samples are stored in the experience replay buffer for subsequent model training. Experience replay can break the temporal correlation between samples and improve the training efficiency of the model. Sample a batch of experience samples from the experience replay buffer, use the policy gradient method (such as the Actor-Critic method) to calculate the gradient, and update the parameters of the policy network. The gradient calculation depends on the value network's evaluation of the state value or action state value.

[0099] Target network stable training: In order to stabilize the training process, the target network is introduced, whose structure is the same as the value network, but the parameter update lags behind the value network. When calculating the gradient, the output of the target network is used as the stable target value.

[0100] The steps of environment interaction, experience replay, policy gradient update, and target network update are repeated until the preset number of training rounds is reached or the stopping condition (such as reward convergence) is met.

[0101] Step 4: Model optimization

[0102] Evaluate the decision-making performance of the model on the validation set and calculate indicators such as success rate, average reward, path length, etc. According to the evaluation results, adjust hyperparameters such as network structure (such as increasing the number of layers, changing the number of neurons), learning rate (such as using learning rate decay), and exploration strategy (such as the ε value in the ε-greedy strategy). Retrain the DRL model based on the adjusted parameters and evaluate the performance again on the validation set until satisfactory decision-making performance is obtained.

[0103] Save the trained DRL model as a record in a file or database for subsequent path planning decisions.

[0104] Embodiment 3:

[0105] This embodiment provides a method for obtaining training data of a residual fitting model and constructing the model, aiming to fit the residuals of the UAV flight state prediction model and the UAV path planning decision model through the extreme gradient boosting (XGBoost) algorithm, thereby further improving the accuracy of path planning. The following is a detailed description of the specific implementation of the method:

[0106] Step 1: Obtain residual data

[0107] For the UAV flight state prediction model, the difference between its predicted state and the actual state is calculated on the training set at each time step. Specifically, for each time step t, the Euclidean distance or Manhattan distance between the predicted state vectors such as position, velocity, acceleration and the corresponding actual state vector is calculated as the state prediction residual.

[0108] For the UAV path planning decision model, the difference between its planned path and the actual optimal path is compared on the training set. This can be achieved by calculating the deviation distance between the planned path and the actual optimal path, the difference in path length, or the cost difference calculated according to a specific cost function, thereby obtaining the decision residual.

[0109] Step 2: Construct residual feature set

[0110] The calculated state prediction residuals and decision residuals are used as new features, which reflect the inaccuracy of model prediction and decision. These residual features are combined with the original flight state data (such as position, speed, acceleration, etc.), environmental data (such as obstacle location, weather conditions, etc.) and mission target data (such as target point location, arrival time requirements, etc.) to form a residual feature set. The residual feature set contains richer information, which helps the residual fitting model learn the residual distribution more accurately.

[0111] Step 3: Construct residual fitting model

[0112] The extreme gradient boosting (XGBoost) algorithm was selected as the construction method of the residual fitting model. XGBoost is an efficient ensemble learning method that can significantly improve the prediction performance of the model by constructing multiple base learners (such as decision trees) and fitting the residuals in sequence.

[0113] The model structure includes:

[0114] Input layer: receives the residual feature set as input data.

[0115] Multiple base learners: The XGBoost algorithm processes the input data sequentially by constructing multiple base learners (such as decision trees). Each base learner attempts to make up for the residuals left by the previous round of models, thereby gradually approaching the true residual distribution.

[0116] Output layer: Summarize the outputs of all base learners to obtain the modified optimal path planning result of the drone cluster. This result is based on the original model prediction and decision, and is corrected by residual fitting, so it has higher accuracy.

[0117] The residual feature set is used as training data to train the XGBoost model. During the training process, the parameters and structure of the base learner are continuously adjusted by optimizing the loss function (such as square loss) so that the model can better fit the residual data. The model is verified and tuned using methods such as cross-validation to ensure the generalization ability and stability of the model.

[0118] Step 4: Model application

[0119] The trained residual fitting model is applied to the actual UAV path planning task. After obtaining the prediction and decision results of the original model (UAV flight state prediction model and UAV path planning decision model), the residual fitting model is used to correct the results to obtain a more accurate optimal path planning result.

[0120] The present invention also includes an optimal path planning system for a drone cluster, the system comprising:

[0121] A model building module, used to build a UAV flight state prediction model and a UAV path planning decision model, wherein the UAV flight state prediction model is configured to predict the future flight state of each UAV in the UAV cluster, and the UAV path planning decision model is configured to determine the optimal path of the UAV cluster based on current and historical flight state data, environmental data, and mission target data;

[0122] A model training and optimization module, connected to the model building module, configured to use a dual machine learning algorithm to train the UAV flight state prediction model and the UAV path planning decision model respectively, and calculate the residuals between the actual outputs and the expected outputs of the two models, and further use these residuals to build a residual fitting model to correct and optimize the results of the UAV path planning;

[0123] The real-time path planning instruction generation module is connected to the model training and optimization module, and is configured to use the trained UAV flight state prediction model, UAV path planning decision model and residual fitting model, combined with the real-time acquired flight state data, environmental data and mission target data, to generate the optimal path planning instructions for the UAV cluster.

[0124] The implementation of the system refers to the above-mentioned embodiment and will not be described in detail in the specification.

[0125] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0126] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An optimal path planning method for a drone cluster, characterized in that: The method comprises: Step A: Construct a UAV flight state prediction model and a UAV path planning decision model, wherein the UAV flight state prediction model is used to predict the future flight state of each UAV in the UAV cluster, and the UAV path planning decision model is used to determine the optimal path of the UAV cluster based on current and historical flight state data, environmental data, and mission target data; Step B: Use a dual machine learning algorithm to train the UAV flight state prediction model and the UAV path planning decision model respectively, and use the residuals of the two models to construct a residual fitting model, which is used to correct and optimize the results of the UAV path planning; Step C: Use the trained model, combined with real-time flight status data, environmental data, and mission target data, to generate the optimal path planning instructions for the drone cluster.

2. The optimal path planning method for a drone cluster according to claim 1 is characterized in that: The long short-term memory network (LSTM) algorithm is used to build a UAV flight status prediction model, and the flight status of the UAV is predicted through time series analysis.

3. The optimal path planning method for a drone cluster according to claim 2 is characterized in that: The training steps of the UAV flight state prediction model include: Data preprocessing: Collect and clean UAV flight data, including position, speed, acceleration, attitude, and environmental information, and encode and standardize the collected data; Construct LSTM network structure: Design LSTM model, including input layer, LSTM layer, fully connected layer and output layer; the input layer receives the preprocessed UAV flight data, the LSTM layer is used to capture the long-term dependencies in the time series data, the fully connected layer is used to map the output of the LSTM layer to the predicted target state, and the output layer outputs the predicted result of the UAV's future flight state; Model training: The preprocessed UAV flight data is input into the LSTM model, the model parameters are optimized by the back propagation algorithm and the gradient descent algorithm, and the cross-validation method is used to select the best network structure and hyperparameters; wherein, the formula for optimizing the model parameters by the gradient descent algorithm is: in, Represents the model parameters, including all weights and biases that the model needs to learn; Represents the learning rate, which is used to control the step size of parameter update; Represents the loss function About parameters The gradient of , which is a vector pointing in the direction of fastest growth of the loss function; Indicates an assignment operation, that is, updating parameters The value of Model evaluation: Evaluate the predictive performance of the model on the test set and adjust model parameters or feature selection based on the evaluation results.

4. The optimal path planning method for a drone cluster according to claim 1 is characterized in that: The deep reinforcement learning (DRL) algorithm is used to build a UAV path planning decision model, and the optimal path planning instructions for the UAV cluster are output through the policy gradient approximation method.

5. The optimal path planning method for a drone cluster according to claim 4 is characterized in that: The training steps of the UAV path planning decision model include: Dataset preparation: Integrate historical flight status data, environmental data, mission target data, and corresponding path planning results to form a complete data set, including state characteristics, action space, and reward function; Construct the DRL model structure: Design the DRL model, including the policy network and the value network. The policy network directly outputs the action probability distribution, and the value network evaluates the state or action state value. Model training: Input the data set into the DRL model, generate experience samples by interacting with the environment, use the policy gradient method to update the model parameters, and use experience replay and target network algorithm to stabilize the training process; Model optimization: Evaluate the decision-making performance of the model on the validation set, and adjust the network structure, learning rate, and exploration strategy based on the evaluation results.

6. The optimal path planning method for a drone cluster according to claim 5 is characterized in that: The steps to obtain training data for the residual fitting model include: Calculate the residual between the actual output and the expected output of the UAV flight state prediction model and the UAV path planning decision model on the training set, specifically: For the UAV flight state prediction model, the difference between the predicted state and the actual state at each time step is calculated to obtain the state prediction residual; For the UAV path planning decision model, the deviation or cost difference between the planned path and the actual optimal path is calculated to obtain the decision residual; The calculated state prediction residuals and decision residuals are used as new features and combined with the original flight state data, environmental data and mission target data to form a residual feature set; the residual feature set is used as training data for the residual fitting model.

7. The optimal path planning method for a drone cluster according to claim 6 is characterized in that: The extreme gradient boosting XGBoost algorithm is selected to train the residual feature set to construct a residual fitting model. The structure of the residual fitting model includes an input layer, multiple base learners and an output layer: The input layer receives the residual feature set as input; Multiple base learners process the input data in turn, and each base learner makes up for the residual left by the previous round of models; The output layer summarizes the outputs of all base learners to obtain the corrected optimal path planning result of the drone cluster.

8. The optimal path planning method for a drone cluster according to claim 3 is characterized in that: In the training step of the UAV flight state prediction model, the method of encoding the collected data includes: performing one-hot encoding conversion on discrete features.

9. The optimal path planning method for a drone cluster according to claim 3 is characterized in that: In the training step of the UAV flight state prediction model, the method of standardizing the collected data includes: using the Z-score standardization method to standardize the continuous features, specifically: subtracting the mean of each numerical data and dividing it by its standard deviation, so that the processed data conforms to the standard normal distribution, and the formula for standardization is: ,in, is the original data, is the mean of the original data, is the standard deviation of the original data, is the standardized data.

10. An optimal path planning system for drone swarms, characterized in that: The system comprises: A model building module, used to build a UAV flight state prediction model and a UAV path planning decision model, wherein the UAV flight state prediction model is configured to predict the future flight state of each UAV in the UAV cluster, and the UAV path planning decision model is configured to determine the optimal path of the UAV cluster based on current and historical flight state data, environmental data, and mission target data; A model training and optimization module, connected to the model building module, configured to use a dual machine learning algorithm to train the UAV flight state prediction model and the UAV path planning decision model respectively, and calculate the residuals between the actual outputs and the expected outputs of the two models, and further use these residuals to build a residual fitting model to correct and optimize the results of the UAV path planning; The real-time path planning instruction generation module is connected to the model training and optimization module, and is configured to use the trained UAV flight state prediction model, UAV path planning decision model and residual fitting model, combined with the real-time acquired flight state data, environmental data and mission target data, to generate the optimal path planning instructions for the UAV cluster.

Images