Cross-scene aircraft path planning method and system based on transfer learning

Abstract

The invention discloses a cross-scene aircraft path planning method and system based on transfer learning. The six-rotor variable-pitch aircraft control method comprises the steps of S110, collecting flight data of an aircraft in different scenes and performing preprocessing; s120, performing model training by taking the urban environment as a source scene according to the collected flight data; s130, carrying out transfer learning on the basis of the source scene model, and constructing a target scene model; s140, performing fine tuning training on the target scene model; s150, performing path planning and execution by using the target scene model; through transfer learning and cross-scene intelligent path planning, the autonomous obstacle avoidance capability, flight stability and task execution efficiency of the aircraft in a complex environment are improved.

Description

Technical Field

[0001] The present invention relates to transfer learning and path planning, and particularly to a cross-scenario aircraft path planning method and system based on transfer learning. Background Art

[0002] With the wide application of unmanned aerial vehicles (UAVs) in fields such as inspection, rescue, and mapping, their autonomous flight and path planning capabilities in different scenarios become particularly important. However, traditional path planning methods usually rely on data training in a specific environment and are difficult to be directly applied to new scenarios, resulting in poor adaptability of the model in different environments, high training costs, and low efficiency. As a machine learning technology, transfer learning can utilize the knowledge obtained from training in the source scenario to quickly adapt to the target scenario, reduce the need for training data, and improve the generalization ability of the model. The cross-scenario aircraft path planning method and system based on transfer learning of the present invention realize efficient path planning of the aircraft in different environments by extracting key features in the source scenario and migrating them to the target scenario, thereby improving the autonomous flight ability and task execution efficiency. Summary of the Invention

[0003] The present invention provides a cross-scenario aircraft path planning method based on transfer learning, including:

[0004] S110. Collect flight data of the aircraft in different scenarios and perform preprocessing;

[0005] S120. Use the urban environment as the source scenario for model training according to the collected flight data;

[0006] S130. Perform transfer learning on the basis of the source scenario model to construct a target scenario model;

[0007] S140. Perform fine-tuning training on the target scenario model;

[0008] S150. Use the target scenario model for path planning and execution.

[0009] In the cross-scenario aircraft path planning method based on transfer learning as described above, for multiple different aircraft flight scenarios, such as urban environment, mountain environment, and forest environment, flight data of the aircraft in these scenarios is collected using sensor devices such as lidar and cameras. These data include environmental information around the aircraft: obstacle position, height, shape, state information of the aircraft itself: position, speed, attitude, and existing planned path information.

[0010] In the cross-scenario aircraft path planning method based on transfer learning as described above, using the urban environment as the source scenario for model training according to the collected flight data specifically includes the following sub-steps:

[0011] Build a training environment;

[0012] Perform model training;

[0013] Store the state, action, reward, and new state of each time into the experience replay pool;

[0014] Perform training iterations.

[0015] A cross-scenario aircraft path planning method based on transfer learning as described above, in which on the basis of the model trained in the source scenario, its key features are extracted to improve the adaptability of the model in the new environment. The key features include the core information of aircraft path planning, such as obstacle avoidance strategies, flight trajectory patterns, response mechanisms to different terrains, etc. The features reflect the effective strategies and experiences learned by the aircraft in the source scenario and can provide good initialization information for the new scenario.

[0016] A cross-scenario aircraft path planning method based on transfer learning as described above, in which the target scenario model is used for path planning and execution, specifically including the following sub-steps:

[0017] Obtain real-time state;

[0018] Make path planning decisions using the target scenario model;

[0019] Perform path execution and adjustment according to the path planning decision.

[0020] This application also provides a cross-scenario aircraft path planning system based on transfer learning, including:

[0021] Data collection module: Collect the flight data of the aircraft in different scenarios and perform preprocessing;

[0022] Source scenario model training module: Use the collected flight data to train the model with the urban environment as the source scenario;

[0023] Target scenario model training module: Perform transfer learning based on the source scenario model to build the target scenario model;

[0024] Fine-tuning module: Perform fine-tuning training on the target scenario model;

[0025] Execution module: Use the target scenario model for path planning and execution.

[0026] A cross-scenario aircraft path planning system based on transfer learning as described above, in which for various different aircraft flight scenarios, such as urban environment, mountain environment, and forest environment, sensor devices, such as lidar and cameras, are used to collect flight data of the aircraft in these scenarios. This data includes environmental information around the aircraft: obstacle position, height, shape, the aircraft's own state information: position, speed, attitude, and existing planned path information.

[0027] A cross-scenario aircraft path planning system based on transfer learning as described above, in which according to the collected flight data, the urban environment is used as the source scenario for model training, which specifically includes the following sub-steps:

[0028] Construct a training environment;

[0029] Perform model training;

[0030] Store each state, action, reward, and new state into the experience replay pool;

[0031] Perform training iterations.

[0032] A cross-scenario aircraft path planning system based on transfer learning as described above, in which on the basis of the model trained in the source scenario, its key features are extracted to improve the adaptability of the model in the new environment. The key features include the core information of aircraft path planning, such as obstacle avoidance strategies, flight trajectory patterns, response mechanisms to different terrains, etc. The features reflect the effective strategies and experiences learned by the aircraft in the source scenario and can provide good initialization information for the new scenario.

[0033] A cross-scenario aircraft path planning system based on transfer learning as described above, in which the target scenario model is used for path planning and execution, which specifically includes the following sub-steps:

[0034] Obtain the real-time state;

[0035] Use the target scenario model to make path planning decisions;

[0036] According to the path planning decision, perform path execution and adjustment.

[0037] The beneficial effects achieved by the present invention are as follows: Through transfer learning and cross-scenario intelligent path planning, the present invention improves the autonomous obstacle avoidance ability, flight stability, and task execution efficiency of the aircraft in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments recorded in the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings.

[0039] Figure 1 It is a flowchart of a cross-scenario aircraft path planning method based on transfer learning provided in Embodiment 1 of this application;

[0040] Figure 2 It is a schematic diagram of a cross-scenario aircraft path planning system based on transfer learning provided in Embodiment 2 of this application; Detailed implementation manners

[0041] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0042] Embodiment 1

[0043] As Figure 1 shown, Embodiment 1 of this application provides a cross-scenario aircraft path planning method based on transfer learning, including:

[0044] Step S110: Collect the flight data of the aircraft in different scenarios and perform preprocessing;

[0045] Collecting the flight data of the aircraft in different scenarios and performing preprocessing specifically includes the following sub-steps:

[0046] Step S111: Collect the flight data of the aircraft in different scenarios;

[0047] For a variety of different aircraft flight scenarios, such as urban environments, mountainous environments, forest environments, etc., use sensor devices, such as lidar, cameras, etc., to collect the flight data of the aircraft in these scenarios. These data include the environmental information around the aircraft: obstacle position, height, shape, the state information of the aircraft itself: position, speed, attitude, and the existing planned path information.

[0048] In the urban environment, by setting data collection points in multiple different urban areas, let the aircraft fly in different time periods and different weather conditions to obtain environmental data such as urban street layouts and building distributions, as well as corresponding flight path data.

[0049] In a mountainous environment, select mountainous areas with different topographical features, such as steep mountain ranges, gentle slope mountainous areas, etc., for data collection, and record information such as the undulation of the mountain terrain, the location of valleys, etc. and the flight path of the aircraft.

[0050] In a forest environment, select different types of forest areas, such as dense primary forests, sparse plantations, etc., for data collection, and record information such as the canopy height of the forest, vegetation density, open space distribution, etc. and the flight path of the aircraft.

[0051] Step S112: Perform data preprocessing on the collected data;

[0052] Clean the collected raw data to remove outliers and noise data. For the distance data collected by lidar, if there are outliers that deviate significantly from the surrounding data, mark and delete them.

[0053] Normalize the data to unify different types of data (such as distance data, speed data, etc.) to the same numerical range. All distance data is normalized to the interval [0, 1], and the speed data is normalized to the interval [-1, 1] for subsequent model processing.

[0054] Step S120: Based on the collected flight data, use the urban environment as the source scenario for model training;

[0055] Based on the collected flight data, use the urban environment as the source scenario for model training, which specifically includes the following sub-steps:

[0056] Step S121: Construct a training environment;

[0057] In the source scenario (urban environment), construct a training environment based on the preprocessed data. Use the environmental information around the aircraft and its own state information as the input state of the model, and the actions that the aircraft can perform: fly forward, fly left, fly right, ascend, descend, as the action space of the model.

[0058] Step S122: Perform model training;

[0059] Initialize the model parameters, including the weights of the neural network, etc. In the training environment, let the aircraft start from the initial position and select an action to fly according to the current state. After performing the action, the aircraft reaches a new state and obtains a corresponding reward value. The setting of the reward value is determined according to factors such as whether the aircraft successfully avoids obstacles and whether it approaches the target position. Give a certain positive reward for successfully avoiding an obstacle, a large negative reward for colliding with an obstacle, and an increasingly larger positive reward for approaching the target position. Calculate the total reward value. Specifically, use the formula: Calculate the reward value, where r tRepresents the reward value at the current moment. A represents the number of times of successfully avoiding obstacles, with one added for each obstacle avoided. D represents the degree of proximity of the aircraft to the target position, and d prev Represents the previous distance, d current Represents the current distance. C represents the number of times the aircraft collides with an obstacle, with one added for each collision. ω 1 , ω 2 , ω 3 Represents the weight used to adjust the influence degree of each factor.

[0060] Step S123: Store each state, action, reward, and new state into the experience replay pool;

[0061] The experience replay pool represents a buffer for storing the states, actions, rewards, and new states experienced by the aircraft during training. Its core function is to randomly sample past experiences for learning during training to improve training stability and data utilization. Storing each state, action, reward, and new state into the experience replay pool, specifically, using the formula: ERB = (s t , a t , r t , s t+1 ) represents the storage process, where ERB represents the experience pool, s t Represents the state, a t Represents the action, r t Represents the reward, s t+1 Represents the new state. When the data in the experience replay pool reaches a certain quantity, a batch of data is randomly sampled from the pool for updating the model. When the experience replay pool reaches the set maximum capacity, new experiences will replace the oldest experiences to keep the data fresh.

[0062] Step S124: Perform training iterations;

[0063] Repeat the above process. After a large number of training iterations, enable the model to learn a better path planning strategy in the source scenario. Specifically, use the formula:

[0064] To evaluate the quality of the path planning strategy, where S represents the score of the path planning strategy, and the larger the value, the better the strategy. T represents the total number of flight time steps, r t Represents the immediate reward value at the t-th time step, L represents the actual flight path length of the aircraft, L max Represents the set maximum allowable flight path length, C represents the number of collisions that occur to the aircraft throughout the path, C max Represents the set maximum allowable number of collisions, γ 1 , γ 2 , γ 3It represents the weight coefficient. By calculating the weighted sum of the reward mean term, path efficiency term, and safety term during the flight of the aircraft, the quality of the path planning strategy is evaluated, so as to select the optimal flight path for the aircraft.

[0065] Step S130: Perform transfer learning based on the source scene model to construct a target scene model;

[0066] Based on the model trained in the source scene, extract its key features to improve the adaptability of the model in the new environment. The key features include the core information of the aircraft path planning, such as obstacle avoidance strategies, flight trajectory patterns, response mechanisms to different terrains, etc. The features reflect the effective strategies and experiences learned by the aircraft in the source scene and can provide good initialization information for the new scene.

[0067] In the target scene (such as a mountainous environment), transfer these extracted features to the new path planning model so that they serve as the initial feature representation. The model of the target scene can utilize the knowledge learned in the source scene at the initial stage of training without having to explore from scratch, thus avoiding the repeated learning of inefficient strategies.

[0068] Step S140: Fine-tune and train the target scene model;

[0069] In the target scene, build a fine-tuning training environment based on the transferred features. Similarly, use the environmental information and the aircraft's own state in the target scene as inputs, keep the action space consistent with the source scene, and initialize other parameters of the target scene model. In the fine-tuning training environment, let the aircraft perform path planning according to the transferred features and initial parameters, and give corresponding rewards according to the actual situation of the target scene. For example, in a mountainous environment, set the reward mechanism considering factors such as mountain peak height and valley orientation.

[0070] Repeat the experience replay and model update steps in the training of the source scene model, but set the learning rate relatively small at this time to avoid excessive adjustment of the transferred features. Through fine-tuning training, make the target scene model adapt to the characteristics of the target scene and optimize the path planning strategy.

[0071] Step S150: Use the target scene model for path planning and execution;

[0072] Using the target scene model for path planning and execution specifically includes the following sub-steps:

[0073] Step S151: Obtain the real-time state;

[0074] During actual flight, the aircraft relies on a variety of sensors to continuously perceive the environmental information and its own state. These sensors include lidar, cameras, ultrasonic sensors, radars, etc., which are used to obtain the distribution of obstacles and terrain features, and GPS, IMU, barometers, wind speed sensors, etc., which are used to monitor the aircraft's own position, attitude, and motion state. The collected data is processed through fusion algorithms to improve the perception accuracy and robustness.

[0075] Step S152: Use the target scenario model for path planning decision-making;

[0076] Input the real-time obtained state information into the target scenario model after transfer learning and fine-tuning training. According to the learned path planning strategy, the model quickly calculates the actions that the aircraft should execute currently, in order to plan the optimal or approximately optimal path to avoid obstacles and reach the target position.

[0077] Step S153: According to the path planning decision, perform path execution and adjustment;

[0078] The aircraft executes flight operations according to the action instructions output by the model. During flight, it continuously monitors environmental changes and its own state. If it is found that the actual situation deviates from the model prediction, for example, new obstacles appear, it promptly feeds back the new state information to the model, and the model re-plans the path to ensure that the aircraft can reach the target position safely and efficiently.

[0079] Embodiment 2

[0080] As Figure 2 shown, Embodiment 2 of the present application provides a cross-scenario aircraft path planning system based on transfer learning, including:

[0081] Data collection module 21: Collect the flight data of the aircraft in different scenarios and perform preprocessing;

[0082] Collect the flight data of the aircraft in different scenarios and perform preprocessing, which specifically includes the following sub-steps:

[0083] Collection module: Collect the flight data of the aircraft in different scenarios;

[0084] For a variety of different aircraft flight scenarios, such as urban environment, mountain environment, forest environment, etc., use sensor devices, such as lidar, cameras, etc., to collect the flight data of the aircraft in these scenarios. These data include the environmental information around the aircraft: the position, height, and shape of obstacles, the state information of the aircraft itself: position, speed, attitude, and the existing planned path information.

[0085] In the urban environment, by setting data collection points in multiple different urban areas, the aircraft is made to fly in different time periods and different weather conditions to obtain environmental data such as the urban street layout and building distribution, as well as the corresponding flight path data.

[0086] In the mountainous environment, mountains with different terrain features are selected, such as steep mountain areas and gentle slope mountain areas, for data collection, and information such as the terrain undulation and valley positions in the mountains, as well as the flight path of the aircraft, are recorded.

[0087] In the forest environment, different types of forest areas are selected, such as dense primary forests and sparse plantations, for data collection, and information such as the tree crown height, vegetation density, and open space distribution in the forest, as well as the flight path of the aircraft, are recorded.

[0088] Preprocessing module: Perform data preprocessing on the collected data;

[0089] Clean the collected raw data to remove outliers and noise data. For the distance data collected by lidar, if there are outliers that deviate significantly from the surrounding data, they are marked and deleted.

[0090] Normalize the data to unify different types of data (such as distance data, speed data, etc.) to the same numerical range. All distance data is normalized to the interval [0, 1], and the speed data is normalized to the interval [-1, 1] for subsequent model processing.

[0091] Source scene model training module 22: Use the urban environment as the source scene to train the model based on the collected flight data;

[0092] Use the urban environment as the source scene to train the model based on the collected flight data, which specifically includes the following sub-steps:

[0093] Training environment construction module: Construct the training environment;

[0094] In the source scene (urban environment), construct the training environment based on the preprocessed data. Use the environmental information around the aircraft and its own state information as the input state of the model, and the actions that the aircraft can perform: fly forward, fly left, fly right, ascend, descend, as the action space of the model.

[0095] Model training module: Conduct model training;

[0096] Initialize the model parameters, including the weights of the neural network, etc. In the training environment, let the aircraft start from the initial position and select an action to fly according to the current state. After performing the action, the aircraft reaches a new state and obtains a corresponding reward value. The setting of the reward value is determined by factors such as whether the aircraft successfully avoids obstacles and whether it approaches the target position. A certain positive reward is given for successfully avoiding an obstacle, a large negative reward is given for colliding with an obstacle, and an increasingly larger positive reward is given for approaching the target position. Calculate the total reward value. Specifically, use the formula: Calculate the reward value, where r t represents the reward value at the current moment, A represents the number of times of successfully avoiding obstacles, incremented by one for each obstacle avoided, D represents the degree of proximity of the aircraft to the target position, d prev represents the previous distance, d current represents the current distance, C represents the number of times the aircraft collides with an obstacle, incremented by one for each collision with an obstacle, ω 1 , ω 2 , ω 3 represents the weight, used to adjust the influence degree of each factor.

[0097] Storage module: Store each state, action, reward, and new state into the experience replay pool;

[0098] The experience replay pool represents a buffer for storing the states, actions, rewards, and new states experienced by the aircraft during training. Its core role is to randomly sample past experiences for learning during training to improve training stability and data utilization rate. Store each state, action, reward, and new state into the experience replay pool. Specifically, use the formula: ERB = (s t , a t , r t , s t+1 ) represents the storage process, where ERB represents the experience pool, s t represents the state, a t represents the action, r t represents the reward, s t+1 represents the new state. When the data in the experience replay pool reaches a certain quantity, randomly sample a batch of data from the pool for updating the model. When the experience replay pool reaches the set maximum capacity, new experiences will replace the oldest experiences to keep the data fresh.

[0099] Iteration module: Perform training iterations;

[0100] Repeat the above process. After a large number of training iterations, enable the model to learn a better path planning strategy in the source scenario. Specifically, use the formula:

[0101] To evaluate the quality of the path planning strategy, where S represents the score of the path planning strategy, and the larger the value, the better the strategy; T represents the total number of flight time steps, and r t represents the immediate reward value at the t-th time step, L represents the actual flight path length of the aircraft, and L max represents the set maximum allowable flight path length, C represents the number of collisions that occur to the aircraft during the entire path, and C max represents the set maximum allowable number of collisions, and γ 1 , γ 2 , γ 3 represents the weight coefficient. By calculating the weighted sum of the reward mean term, path efficiency term, and safety term during the flight of the aircraft, the quality of the path planning strategy is evaluated, so as to select the optimal flight path for the aircraft.

[0102] Target scene model training module 23: Perform transfer learning based on the source scene model to construct the target scene model;

[0103] Based on the model trained in the source scene, extract its key features to improve the adaptability of the model in the new environment. The key features include the core information of the aircraft path planning, such as obstacle avoidance strategies, flight trajectory patterns, response mechanisms to different terrains, etc. The features reflect the effective strategies and experiences learned by the aircraft in the source scene and can provide good initialization information for the new scene.

[0104] In the target scene (such as a mountainous environment), transfer these extracted features to the new path planning model and use them as the initial feature representation. The model of the target scene can utilize the knowledge learned in the source scene at the initial stage of training without having to start exploring from scratch, thus avoiding the repeated learning of inefficient strategies.

[0105] Fine-tuning module 24: Perform fine-tuning training on the target scene model;

[0106] In the target scene, construct a fine-tuning training environment based on the transferred features. Similarly, use the environmental information and its own state of the aircraft in the target scene as input, and keep the action space consistent with the source scene. Initialize other parameters of the target scene model. In the fine-tuning training environment, let the aircraft perform path planning according to the transferred features and initial parameters, and give corresponding rewards according to the actual situation of the target scene. For example, in a mountainous environment, set the reward mechanism considering factors such as mountain peak height and valley direction.

[0107] Repeat the experience replay and model update steps in the source scene model training, but at this time, set the learning rate relatively small to avoid excessive adjustment of the transferred features. Through fine-tuning training, make the target scene model adapt to the characteristics of the target scene and optimize the path planning strategy.

[0108] Execution Module 25: Perform path planning and execution using the target scenario model;

[0109] Perform path planning and execution using the target scenario model, which specifically includes the following sub-steps:

[0110] Status Acquisition Module: Real-time status acquisition;

[0111] During actual flight, the aircraft relies on a variety of sensors to continuously sense the environmental information and its own status in real time. These sensors include lidar, cameras, ultrasonic sensors, radars, etc. to obtain the distribution of obstacles and terrain features, and GPS, IMU, barometers, wind speed sensors, etc. to monitor its own position, attitude, and motion state. The collected data is processed through a fusion algorithm to improve the perception accuracy and robustness.

[0112] Decision-making Module: Make path planning decisions using the target scenario model;

[0113] Input the real-time acquired status information into the target scenario model that has undergone transfer learning and fine-tuning training. The model quickly calculates the actions that the aircraft should execute currently according to the learned path planning strategy to plan the optimal or approximately optimal path to avoid obstacles and reach the target position.

[0114] Execution and Adjustment Module: Perform path execution and adjustment according to the path planning decision;

[0115] The aircraft executes flight operations according to the action instructions output by the model. During flight, it continuously monitors environmental changes and its own status. If it is found that the actual situation deviates from the model prediction, for example, new obstacles appear, the new status information is promptly fed back to the model, and the model re-plans the path to ensure that the aircraft can reach the target position safely and efficiently.

[0116] The specific implementation manners described above further elaborate on the purpose, technical solutions, and beneficial effects of the present invention. It should be understood that the above are only the specific implementation manners of the present invention and are not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made on the basis of the technical solutions of the present invention shall be included within the protection scope of the present invention.

Claims

1. A cross-scenario aircraft path planning method based on transfer learning, characterized in that: include: S110, collecting flight data of the aircraft in different scenarios and performing preprocessing; S120, using the urban environment as a source scene for model training based on the collected flight data; S130, performing transfer learning based on the source scene model to construct a target scene model; S140, fine-tuning the target scene model; S150: Use the target scene model to perform path planning and execution.

2. A cross-scenario aircraft path planning method based on transfer learning as claimed in claim 1, characterized in that: For a variety of different aircraft flight scenarios, such as urban environments, mountainous environments, and forest environments, sensor equipment such as lidar and cameras are used to collect flight data of aircraft in these scenarios. This data includes environmental information around the aircraft: obstacle location, height, shape, and the aircraft's own status information: location, speed, attitude, and existing planned path information.

3. A cross-scenario aircraft path planning method based on transfer learning as claimed in claim 1, characterized in that: Based on the collected flight data, the urban environment is used as the source scene for model training, which includes the following sub-steps: Build a training environment; Conduct model training; Store each state, action, reward and new state into the experience replay pool; Perform training iterations.

4. The cross-scenario aircraft path planning method based on transfer learning as claimed in claim 1, characterized in that: Based on the model trained in the source scene, extract its key features. To improve the adaptability of the model in new environments, key features include core information of the aircraft path planning, such as obstacle avoidance strategy, flight trajectory mode, response mechanism to different terrains, etc. Features reflect the effective strategies and experience learned by the aircraft in the source scene, and can provide good initialization information for the new scene.

5. The cross-scenario aircraft path planning method based on transfer learning as claimed in claim 1, characterized in that: Using the target scene model to plan and execute the path includes the following sub-steps: Real-time status acquisition; Use the target scene model to make path planning decisions; Execute and adjust the path based on the path planning decision.

6. A cross-scenario aircraft path planning system based on transfer learning, characterized in that: include: Data collection module: collects flight data of aircraft in different scenarios and performs pre-processing; Source scene model training module: Based on the collected flight data, the urban environment is used as the source scene for model training; Target scene model training module: performs transfer learning based on the source scene model to build the target scene model; Fine-tuning module: fine-tuning the target scene model; Execution module: Uses the target scene model for path planning and execution.

7. A cross-scenario aircraft path planning system based on transfer learning as claimed in claim 6, characterized in that: For a variety of different aircraft flight scenarios, such as urban environments, mountainous environments, and forest environments, sensor equipment such as lidar and cameras are used to collect flight data of aircraft in these scenarios. This data includes environmental information around the aircraft: obstacle location, height, shape, and the aircraft's own status information: location, speed, attitude, and existing planned path information.

8. The cross-scenario aircraft path planning system based on transfer learning as claimed in claim 6, characterized in that: Based on the collected flight data, the urban environment is used as the source scene for model training, which includes the following sub-steps: Build a training environment; Conduct model training; Store each state, action, reward and new state into the experience replay pool; Perform training iterations.

9. The cross-scenario aircraft path planning system based on transfer learning as claimed in claim 6, characterized in that: Based on the model trained in the source scene, extract its key features. To improve the adaptability of the model in new environments, key features include core information of the aircraft path planning, such as obstacle avoidance strategy, flight trajectory mode, response mechanism to different terrains, etc. Features reflect the effective strategies and experience learned by the aircraft in the source scene, and can provide good initialization information for the new scene.

10. A cross-scenario aircraft path planning system based on transfer learning as claimed in claim 6, characterized in that: Using the target scene model to plan and execute the path includes the following sub-steps: Real-time status acquisition; Use the target scene model to make path planning decisions; Execute and adjust the path based on the path planning decision.

Images