A low-altitude aircraft flight control system and method based on AI algorithm

Abstract

The present invention discloses a low-altitude aircraft flight control system and method based on an AI algorithm, which relates to the field of AI algorithm technology, and includes: a data processing module, which collects environmental data of the aircraft flight airspace and performs denoising on the collected data; an airspace unit capacity dynamic adjustment module, which dynamically adjusts the airspace unit capacity based on the aircraft flight state, wind speed and communication accuracy, and predicts the unit capacity demand in the future time period through an LSTM model. The method of the present invention generates a path with the optimal total cost in a complex environment based on a heuristic search strategy through an AI algorithm, and further cooperates with the population diversity of a genetic algorithm, and combines the cumulative cost and the predicted cost through a heuristic search, so that the generated path has global optimality, and can flexibly adjust the path generation strategy in scenes with sparse or dense obstacles to adapt to the complexity of low-altitude flight.

Description

Technical Field

[0001] The present invention relates to the field of AI algorithm technology, and in particular to a low-altitude aircraft flight control system and method based on AI algorithm. Background Art

[0002] With the rapid development of unmanned aerial vehicle technology, the application scenarios of low-altitude aircraft are constantly expanding, covering multiple fields such as logistics transportation, disaster relief, patrol monitoring, etc. The complexity and dynamics of the low-altitude flight environment and the limited resource allocation have put forward higher requirements for the flight control of aircraft. Traditional flight control methods are mostly based on rules or static algorithms, which have obvious limitations in path planning, task scheduling and environmental adaptability.

[0003] When dealing with real-time multi-dimensional environmental variables and large-scale aircraft collaborative tasks, they often exhibit the shortcomings of low computational efficiency and poor adaptability. In terms of path planning and dynamic adjustment, existing systems mostly rely on static environmental models and are unable to perceive and adjust environmental changes during flight (such as wind speed, obstacle position changes, etc.) in real time, resulting in insufficient flexibility and safety in path planning. Secondly, in terms of airspace capacity management, existing methods usually lack predictive capabilities for dynamic environments and are unable to efficiently allocate flight resources, which can easily lead to airspace congestion or waste of resources. In addition, traditional path planning results usually lack smoothness, which reduces the safety requirements for low-altitude flight. Summary of the invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a low-altitude aircraft flight control system based on AI algorithm to solve the shortcomings of low computational efficiency and poor adaptability when dealing with real-time multi-dimensional environmental variables and large-scale aircraft collaborative tasks. In terms of path planning and dynamic adjustment, most existing systems rely on static environmental models and are unable to perceive and adjust environmental changes during flight in real time, resulting in insufficient flexibility and safety in path planning; secondly, in terms of airspace capacity management, existing methods usually lack predictive capabilities for dynamic environments and are unable to efficiently allocate flight resources, which can easily lead to airspace congestion or waste of resources. In addition, traditional path planning results usually lack smoothness, which reduces the safety requirements for low-altitude flight.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a low-altitude aircraft flight control system based on an AI algorithm, comprising:

[0008] The data processing module collects environmental data of the aircraft's flight airspace and performs denoising on the collected data;

[0009] The airspace unit capacity dynamic adjustment module dynamically adjusts the airspace unit capacity based on the aircraft flight status, wind speed and communication accuracy, and predicts the unit capacity demand in the future time period through the LSTM model;

[0010] The path planning module generates the total generation value path based on the AI ​​algorithm, calculates the path fitness value in combination with the genetic algorithm, and optimizes the path selection;

[0011] The path optimization module uses deep reinforcement learning to calculate the environmental impact factors of path nodes, optimize path segments, and perform path smoothing interpolation based on Bezier curves;

[0012] The data security transmission module encrypts and transmits data through the communication module for secure data storage.

[0013] As a preferred solution of the low-altitude aircraft flight control system based on AI algorithm described in the present invention, wherein: the environmental data of the aircraft flight airspace is collected and the collected data is denoised, including:

[0014] Obtain environmental data from the aircraft’s lidar, vision sensors, multispectral cameras, anemometers, millimeter-wave radars, and GPS, including weather conditions, obstacle locations, and terrain information;

[0015] The extended Kalman filter (EKF) is used to align the sensor data in time and space to remove redundant data and noise.

[0016] As a preferred solution of the low-altitude aircraft flight control system based on AI algorithm described in the present invention, the airspace unit capacity is dynamically adjusted based on the aircraft flight status, wind speed and communication accuracy, and the unit capacity demand in the future time period is predicted by the LSTM model, including:

[0017] The low-altitude area is divided into a three-dimensional grid of N×M×L, each grid cell is a cubic cell, and each cubic cell defines a capacity C, which is determined by the ratio of the cell volume to the average occupied space of a single aircraft's flight time step;

[0018] The unit volume is determined by the product of the aircraft speed and the aircraft's safety reaction time;

[0019] Dynamically adjust the cube unit capacity based on environmental parameters, expressed as:

[0020] ;

[0021] in Represents the unit capacity, and Respectively represent the current weather wind speed and standard weather wind speed, and They represent the average values ​​of Beidou navigation communication accuracy and standard communication accuracy, Indicates the historical unit capacity value;

[0022] Based on the airspace flight altitude standards of aircraft, the airspace is vertically layered to clarify the low-altitude flight range;

[0023] Build an LSTM learning model based on the long short-term memory network, including input layer, hidden layer and output layer;

[0024] The input layer inputs the aircraft quantity change data of each grid cell collected from the time dimension and the historical capacity occupancy data of adjacent grid cells in the space dimension respectively;

[0025] The hidden layer processes and learns the dependencies in the sequence data, and the output layer outputs the predicted capacity requirements of each grid unit in the next time step;

[0026] Use the training data to train the LSTM learning model, select the cross entropy loss function to calculate the computational loss between the prediction of the LSTM learning model and the actual label, use the Adam optimizer for gradient descent optimization, update the parameters of the LSTM learning model, and stop iterating and outputting the model parameters when the model loss no longer decreases significantly during continuous iterations;

[0027] Based on the real-time collected time dimension data and space dimension data, the LSTM learning model is input to predict the capacity requirements of each grid unit in the next time step and adjust the unit capacity;

[0028] If the predicted capacity demand exceeds the maximum limit of the current unit capacity, path optimization is triggered.

[0029] As a preferred solution of the low-altitude aircraft flight control system based on AI algorithm described in the present invention, wherein: the total generation value path based on AI algorithm is generated, the path fitness value is calculated in combination with genetic algorithm, and the path selection is optimized, including:

[0030] The improved Euclidean distance is used to calculate the prediction cost from the current node n to the target node, which is expressed as:

[0031] ;

[0032] in Represents the predicted cost from the current node n to the target node, , and are the three-dimensional coordinates of the current aircraft position, , and are the three-dimensional coordinates of the target position, and Based on the current time and target time of the flight mission, Indicates the maximum operational flight time of the aircraft;

[0033] At the same time, the cumulative cost from the starting point to the current node n is calculated, expressed as:

[0034] ;

[0035] in represents the cumulative cost from the starting point to the current node n, Indicates the node before the current node The cumulative cost of Representation Node The Euclidean distance value with node n;

[0036] The cost function is determined by combining the predicted cost and the accumulated cost, which is expressed as:

[0037] ;

[0038] in Represents the total cost of node n;

[0039] Based on the needs of path optimization, AI algorithm is used for heuristic path generation. The nodes to be expanded in the grid unit are used as the development list. Based on the adjacent expansion nodes of the three-dimensional coordinates of the current starting point of the aircraft, the node with the smallest total cost value is selected to add to the closed list, and repeated selection is performed until the selected node is the target node, forming a cost evaluation path;

[0040] Based on the path obstacles as key nodes, nodes with the same total cost value at the key nodes are formed into different cost evaluation paths;

[0041] The initial population is determined based on the cost evaluation path generated by the AI ​​algorithm, and the genetic algorithm is used for path optimization. The fitness function is defined to evaluate the fitness value of the path, which is expressed as:

[0042] ;

[0043] in represents the fitness value of the evaluation path p, represents the total number of path nodes in the population, represents the Euclidean distance value from the i-th node n to the i+1-th node of path P, represents the distance from the i-th node n of path P to the nearest obstacle;

[0044] At the same time, a crossover operation is performed through the genetic algorithm, two paths are randomly selected to exchange some path segments, a new path is generated, and a mutation operation is performed to randomly change the node positions in the path to offset the path, and repeated iterations are performed. When the fitness value of the best individual in the population no longer changes significantly in continuous iterations, the iteration is stopped and the path with the highest fitness value is selected as the optimized path.

[0045] As a preferred solution of the low-altitude aircraft flight control system based on AI algorithm described in the present invention, the method of using deep reinforcement learning to calculate the environmental impact factors of path nodes, optimizing path segments, and performing path smoothing interpolation based on Bezier curves includes:

[0046] A deep reinforcement learning algorithm is used for path optimization. The current path node position, surrounding obstacle distribution, and target position are used as model state data. The flight direction of the aircraft from the current node to the next node is used as action data. The reward function is determined based on the path length and obstacle avoidance, which is expressed as:

[0047] ;

[0048] Where R represents the reward value, Indicates the path increment caused by the current action, Indicates the distance to the nearest obstacle after the action;

[0049] The starting point of the optimization path obtained by the genetic algorithm is used as the initial state of reinforcement learning, and the Q learning algorithm is used for training. The expected value of the cumulative reward after the state s takes action a is calculated for training;

[0050] Use the training set for training, and use the ϵ-greedy strategy to select actions for training. Randomly select actions with probability ϵ to explore new paths. Select the action with the maximum current Q function value with probability 1−ϵ, and ϵ gradually decreases with the training process until it reaches the historical threshold and stops updating. Execute actions based on the action selection and update the state to calculate the immediate reward and update the Q value until the change in the Q function value is no longer obvious, then stop repeating the training;

[0051] In the actual path adjustment, the trained Q function is used to select the optimal action according to the current state s, and the path nodes are gradually expanded according to the selected optimal action, and the path is updated until the target position is reached, which is used as the reinforcement learning path;

[0052] Based on the Q-learning updated path, the key node sequence is discretized to generate the key node sequence, and the real-time environmental impact factor of the path node is calculated, which is expressed as:

[0053] ;

[0054] in represents the environmental impact factor, represents the distance from the i-th key node to the nearest obstacle, represents the aircraft speed at the i-th key node, represents the wind force on the i-th key node, represents the mass of the aircraft, Indicates the maximum acceleration of the aircraft;

[0055] Will The largest key node is used as the priority adjustment node, and the path nodes before and after the node are selected to form a path segment. If there is an obstacle in the path segment, it is determined that the path segment needs to be optimized. If the direction change angle of the path segment is less than the minimum turning angle of the aircraft, it is determined that the path segment needs to be optimized.

[0056] The path segment that needs to be optimized is modeled using a cubic Bezier curve, which is expressed as:

[0057] ;

[0058] in Represents the coordinates of the path point corresponding to the interpolation parameter t on the Bezier curve, represents the starting point of the path segment, represents the end point of the path segment, and represents the curvature and smoothness control points of the path segment, t represents the interpolation parameter, and the interpolation range is determined based on historical data;

[0059] Define the smooth adjustment vector of the control point as the ratio of the obstacle direction vector to the obstacle vector modulus, and based on the smooth adjustment vector and of and as The coordinate values ​​of the smoothing vector and of and as The coordinate value of

[0060] The interpolation points are evenly distributed through the interpolation range of t, and the coordinates of the path points are calculated using the Bezier curve formula. If the angle between the coordinates of the interpolation points is greater than or equal to the minimum steering angle of the aircraft, the optimized interpolation point sequence is replaced into the reinforcement learning path to generate an updated global path.

[0061] As a preferred solution of the low-altitude aircraft flight control system based on AI algorithm described in the present invention, the encrypted transmission of data through the communication module includes:

[0062] Deploy edge computing modules on aircraft and configure point-to-point communication protocols between aircraft based on 5G communication modules;

[0063] Use task decomposition algorithm to decompose global tasks into small computing units and distribute them to each aircraft;

[0064] MQTT is used to transmit mission status, position, and environmental perception data between multiple aircraft, and AES256 encryption is used to encrypt communication data for transmission.

[0065] As a preferred solution of the low-altitude aircraft flight control system based on AI algorithm described in the present invention, the data security storage includes:

[0066] Recording flight mission execution data in a distributed ledger and storing it in the aircraft’s local storage;

[0067] The aircraft flight data is transmitted to the operating terminal based on the 5G communication module for data visualization and backup storage at the same time. The operating terminal uses AES256 to encrypt the backup storage data and transmit it wirelessly to the cloud storage.

[0068] In a second aspect, the present invention provides a method for a low-altitude aircraft flight control system based on an AI algorithm, comprising:

[0069] Collect environmental data of the aircraft's flight airspace and perform denoising processing;

[0070] Dynamically adjust airspace unit capacity based on communication accuracy and wind speed, and use LSTM model to predict unit capacity demand;

[0071] The AI ​​algorithm is used to calculate the total generation value to form a cost evaluation path, and the genetic algorithm is used to calculate the fitness value for path optimization;

[0072] Perform deep reinforcement learning on the optimized path to calculate the environmental impact factors of the path nodes, use the Bezier curve algorithm to interpolate the path, and generate an optimized global path;

[0073] Data is transmitted through encrypted communication modules for secure data storage.

[0074] The beneficial effects of the present invention are as follows: an AI algorithm is used to generate a path with the optimal total cost in a complex environment based on a heuristic search strategy, and the population diversity of the genetic algorithm is further coordinated. The generated path has global optimality through heuristic search combined with the cumulative cost and the predicted cost. The path generation strategy can be flexibly adjusted in scenes with sparse or dense obstacles to adapt to the complexity of low-altitude flight. Through reinforcement learning, environmental changes are dynamically perceived to generate an optimized path that adapts to a dynamic environment. Dynamic optimization of the path can be achieved in scenes with dense obstacles or multiple tasks. The path segments generated by the Bezier curve avoid sharp turns, improve the stability and controllability of the aircraft's track, and the smooth path reduces the frequent adjustments of the aircraft. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0076] Figure 1 This is a schematic diagram of the structure of the low-altitude aircraft flight control system based on the AI ​​algorithm in Example 1.

[0077] Figure 2 This is a flow chart of the low-altitude aircraft flight control method based on the AI ​​algorithm in Example 1. DETAILED DESCRIPTION

[0078] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0079] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0080] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0081] Example 1, reference Figure 1 and Figure 2 , which is the first embodiment of the present invention, and which provides a low-altitude aircraft flight control system based on an AI algorithm, comprising the following steps:

[0082] S1, collect environmental data of the aircraft's flight airspace and perform denoising on the collected data;

[0083] Preferably, collecting environmental data of the flight airspace of the aircraft and performing denoising on the collected data include:

[0084] Obtain environmental data from the aircraft’s lidar, vision sensors, multispectral cameras, anemometers, millimeter-wave radars, and GPS, including meteorological conditions (wind speed, visibility), obstacle locations, and terrain information;

[0085] The extended Kalman filter (EKF) is used to align the sensor data in time and space to remove redundant data and noise.

[0086] Through multi-sensor perception, different dimensions of the environment (spatial position, weather, obstacles, etc.) are covered to provide comprehensive environmental information. Sensor data with different timestamps are aligned through EKF to avoid erroneous perception due to time difference. Through spatial calibration, the observation results of each sensor are mapped to a unified coordinate system to improve the accuracy of environmental perception. EKF filters the noise data in the dynamic model to extract more accurate environmental information, improve the effectiveness of perception, remove redundant information, reduce the burden of data processing and storage, and improve the utilization efficiency of system resources.

[0087] S2, dynamically adjusts the airspace unit capacity based on the aircraft flight status, wind speed and communication accuracy, and predicts the unit capacity demand in the future time period through the LSTM model;

[0088] Preferably, based on the flight status of the aircraft, wind speed and communication accuracy, the airspace unit capacity is dynamically adjusted, and the unit capacity demand in the future time period is predicted by the LSTM model, including:

[0089] The low-altitude area is divided into a three-dimensional grid of N×M×L, each grid cell is a cubic cell, and each cubic cell defines a capacity C, which is determined by the ratio of the cell volume to the average occupied space of a single aircraft's flight time step;

[0090] The unit volume is determined by the product of the aircraft speed and the aircraft's safety reaction time;

[0091] Dynamically adjust the cube unit capacity based on environmental parameters, expressed as:

[0092] ;

[0093] in Represents the unit capacity, and Respectively represent the current weather wind speed and standard weather wind speed, and They represent the average values ​​of Beidou navigation communication accuracy and standard communication accuracy, Indicates the historical unit capacity value;

[0094] The ratio of the actual number of aircraft in each grid cell to the cell capacity is used as the monitoring ratio, and the sum of the average value and the standard deviation of the monitoring ratio in the area is used as the monitoring threshold. If the monitoring ratio is greater than or equal to the monitoring threshold, dynamic capacity adjustment will be triggered, where the aircraft position information is obtained through ADS-B signals and the broadcast position information of autonomous aircraft;

[0095] Based on the airspace flight altitude standards of aircraft, the airspace is vertically layered to clarify the low-altitude flight range;

[0096] Build an LSTM learning model based on the long short-term memory network, including input layer, hidden layer and output layer;

[0097] The input layer inputs the aircraft quantity change data of each grid cell collected from the time dimension and the historical capacity occupancy data of adjacent grid cells in the space dimension respectively;

[0098] The hidden layer processes and learns the dependencies in the sequence data, and the output layer outputs the predicted capacity requirements of each grid unit in the next time step;

[0099] Use the training data to train the LSTM learning model, select the cross entropy loss function to calculate the computational loss between the prediction of the LSTM learning model and the actual label, use the Adam optimizer for gradient descent optimization, update the parameters of the LSTM learning model, and stop iterating and outputting the model parameters when the model loss no longer decreases significantly during continuous iterations;

[0100] Based on the real-time collected time dimension data and space dimension data, the LSTM learning model is input to predict the capacity requirements of each grid unit in the next time step and adjust the unit capacity;

[0101] If the predicted capacity demand exceeds the maximum limit of the current unit capacity, path optimization is triggered.

[0102] By dividing the airspace into cubic units, accurately monitoring the capacity and aircraft distribution of each unit, improving the utilization rate of low airspace, and dynamically adjusting the unit capacity according to environmental parameters (such as wind speed and communication accuracy), excessive congestion can be avoided and safety can be improved. By vertically stratifying based on flight altitude standards, the flight range of low airspace can be clarified, and the long short-term memory network (LSTM) can be used to predict future unit capacity requirements, reduce capacity overload and idleness, improve resource utilization, provide highly reliable capacity prediction in highly dynamic environments (such as weather changes and mission adjustments), and support dynamic scheduling.

[0103] S3, generates the total generation value path based on the AI ​​algorithm, calculates the path fitness value in combination with the genetic algorithm, and optimizes the path selection;

[0104] Preferably, the total generation value path is generated based on the AI ​​algorithm, and the path fitness value is calculated in combination with the genetic algorithm to optimize the path selection, including:

[0105] The improved Euclidean distance is used to calculate the prediction cost from the current node n to the target node, which is expressed as:

[0106] ;

[0107] in Represents the predicted cost from the current node n to the target node, , and are the three-dimensional coordinates of the current aircraft position, , and are the three-dimensional coordinates of the target position, and Based on the current time and target time of the flight mission, Indicates the maximum operational flight time of the aircraft;

[0108] At the same time, the cumulative cost from the starting point to the current node n is calculated, expressed as:

[0109] ;

[0110] in represents the cumulative cost from the starting point to the current node n, Indicates the node before the current node The cumulative cost of Representation Node The Euclidean distance value with node n;

[0111] The cost function is determined by combining the predicted cost and the accumulated cost, which is expressed as:

[0112] ;

[0113] in Represents the total cost of node n;

[0114] Based on the needs of path optimization, AI algorithm is used. A-star algorithm can be used for heuristic path generation. The nodes to be expanded in the grid unit are used as the development list. Based on the adjacent expansion nodes of the three-dimensional coordinates of the current starting point of the aircraft, the nodes with the smallest total cost value are selected to add to the closed list, and repeated selection is performed until the selected node is the target node, forming a cost evaluation path;

[0115] Based on the path obstacles as key nodes, nodes with the same total cost value at the key nodes are formed into different cost evaluation paths;

[0116] The initial population is determined based on the cost evaluation path generated by the AI ​​algorithm, and the genetic algorithm is used for path optimization. The fitness function is defined to evaluate the fitness value of the path, which is expressed as:

[0117] ;

[0118] in represents the fitness value of the evaluation path p, represents the total number of path nodes in the population, represents the Euclidean distance value from the i-th node n to the i+1-th node of path P, represents the distance from the i-th node n of path P to the nearest obstacle;

[0119] By considering the Euclidean distance value of different path nodes, the total length of the path is comprehensively considered, and the distance cost of the path is directly measured. The shorter the path, the lower the fitness value and the better the path. At the same time, the sum of the reciprocals of the distances from each node in the path to the nearest obstacle is considered as an inverse indicator of safety. The closer the node is to the obstacle, the higher the fitness value and the path is penalized. This increases the distance between the node and the obstacle during optimization and improves the path safety. This is suitable for scenarios with sparse and dense obstacles. Compared with the traditional fitness function, it can optimize the dynamic switching of targets in scenarios with low-density obstacles and dense obstacles to ensure that both path length and safety are taken into consideration.

[0120] At the same time, a crossover operation is performed through the genetic algorithm, two paths are randomly selected to exchange some path segments, a new path is generated, and a mutation operation is performed to randomly change the node positions in the path to offset the path, and repeated iterations are performed. When the fitness value of the best individual in the population no longer changes significantly in continuous iterations, the iteration is stopped and the path with the highest fitness value is selected as the optimized path.

[0121] Through the AI ​​algorithm based on the heuristic search strategy, the path with the optimal total cost is generated in a complex environment. In addition, with the population diversity of the genetic algorithm, the fitness function is used to simultaneously optimize the path length, safety (obstacle avoidance) and other goals, making up for the deficiency that AI can only be optimized based on a single heuristic function. The path provided by AI already has a high initial feasibility in a static obstacle environment. The genetic algorithm can meet the demand by generating a better path through global optimization. Through heuristic search combined with cumulative cost and predicted cost, the generated path has global optimality, can flexibly adjust the path generation strategy in scenes with sparse or dense obstacles, adapt to the complexity of low-altitude flight, and can adjust the path planning according to the real-time updated cost evaluation, support path optimization in the middle of the mission, and optimize the initial path generated by AI in combination with the genetic algorithm. Genetic algorithm generates a new path through crossover and mutation operations, iteratively optimizes the fitness value, measures the quality of the path through the fitness value, and quickly converges to the optimal path in limited iterations.

[0122] S4, uses deep reinforcement learning to calculate the environmental impact factors of path nodes, optimizes path segments, and performs path smoothing interpolation based on Bezier curves;

[0123] Preferably, deep reinforcement learning is used to calculate the environmental impact factors of the path nodes, optimize the path segments, and perform path smoothing interpolation based on Bezier curves, including:

[0124] A deep reinforcement learning algorithm is used for path optimization. The current path node position, surrounding obstacle distribution, and target position are used as model state data. The flight direction of the aircraft from the current node to the next node is used as action data. The reward function is determined based on the path length and obstacle avoidance, which is expressed as:

[0125] ;

[0126] Where R represents the reward value, Indicates the path increment caused by the current action, Indicates the distance to the nearest obstacle after the action;

[0127] The starting point of the optimization path obtained by the genetic algorithm is used as the initial state of reinforcement learning, and the Q learning algorithm is used for training. The expected value of the cumulative reward after the state s takes action a is calculated for training, which is expressed as:

[0128] ;

[0129] in and Represent the Q value of state s and action a before and after the update, respectively. represents the learning rate, Represents the reward value, represents the discount factor, determined based on historical data, Indicates the state after executing action a Q value;

[0130] Use the training set for training, and use the ϵ-greedy strategy to select actions for training. Randomly select actions with probability ϵ to explore new paths. Select the action with the maximum current Q function value with probability 1−ϵ, and ϵ gradually decreases with the training process until it reaches the historical threshold and stops updating. Execute actions based on the action selection and update the state to calculate the immediate reward and update the Q value until the change in the Q function value is no longer obvious, then stop repeating the training;

[0131] In the actual path adjustment, the trained Q function is used to select the optimal action according to the current state s, and the path nodes are gradually expanded according to the selected optimal action, and the path is updated until the target position is reached, which is used as the reinforcement learning path;

[0132] Based on the Q-learning updated path, the key node sequence is discretized to generate the real-time environmental impact factor of the path node, which is expressed as:

[0133] ;

[0134] in represents the environmental impact factor, represents the distance from the i-th key node to the nearest obstacle, represents the aircraft speed at the i-th key node, represents the wind force on the i-th key node, represents the mass of the aircraft, Indicates the maximum acceleration of the aircraft;

[0135] Will The largest key node is used as the priority adjustment node, and the path nodes before and after the node are selected to form a path segment. If there is an obstacle in the path segment, it is determined that the path segment needs to be optimized. If the direction change angle of the path segment is less than the minimum turning angle of the aircraft, it is determined that the path segment needs to be optimized.

[0136] The path segment that needs to be optimized is modeled using a cubic Bezier curve, which is expressed as:

[0137] ;

[0138] in Represents the coordinates of the path point corresponding to the interpolation parameter t on the Bezier curve, represents the starting point of the path segment, represents the end point of the path segment, and represents the curvature and smoothness control points of the path segment, t represents the interpolation parameter, and the interpolation range is determined based on historical data;

[0139] Define the smooth adjustment vector of the control point as the ratio of the obstacle direction vector to the obstacle vector modulus, and based on the smooth adjustment vector and of and as The coordinate values ​​of the smoothing vector and of and as The coordinate value of

[0140] The interpolation points are evenly distributed through the interpolation range of t, and the coordinates of the path points are calculated using the Bezier curve formula. If the angle between the coordinates of the interpolation points is greater than or equal to the minimum steering angle of the aircraft, the optimized interpolation point sequence is replaced into the reinforcement learning path to generate an updated global path.

[0141] Through reinforcement learning, it dynamically perceives environmental changes (such as obstacles and wind speed changes) and generates optimized paths that adapt to dynamic environments. It can realize dynamic optimization of paths in obstacle-dense or multi-task scenarios and improve the task completion rate. Through the ϵ-greedy strategy, it increases the diversity of path planning and reduces the possibility of falling into local optimality. The cumulative reward function comprehensively considers the global characteristics of the path and generates a more efficient global path. The path segments generated by the Bezier curve avoid sharp turns and improve the stability and controllability of the aircraft track. The smooth path reduces the frequent adjustment of the aircraft, reduces energy consumption, and extends the mission endurance. In areas with dense obstacles, the path generated by the Bezier curve has higher passability. The globality of the path is ensured by reinforcement learning. The genetic algorithm and the Bezier curve are combined to optimize the local path details, so that the optimized path has high safety, low risk and high efficiency, which significantly improves the flight mission completion rate.

[0142] S5, encrypts flight data through the communication module to achieve secure storage and transmission of data;

[0143] Preferably, encrypting and transmitting data through the communication module includes:

[0144] Deploy edge computing modules on aircraft and configure point-to-point communication protocols between aircraft based on 5G communication modules;

[0145] Use task decomposition algorithm to decompose global tasks into small computing units and distribute them to each aircraft;

[0146] MQTT is used to transmit mission status, position, and environmental perception data between multiple aircraft, and AES256 encryption is used to encrypt communication data for transmission.

[0147] The edge computing module processes mission data locally on the aircraft, reducing dependence on the cloud and communication latency. Local computing enables rapid path adjustment and mission decision-making, which is particularly suitable for complex dynamic environments. Some computing tasks are processed by the edge module, reducing requests to the central server and improving the overall stability of the system. 5G communication provides high-bandwidth and low-latency point-to-point data transmission to ensure fast and reliable information synchronization between multiple aircraft. The lightweight feature of MQTT reduces the communication protocol's demand for bandwidth and computing resources, making it suitable for real-time data synchronization between multiple aircraft. AES256 encryption is used to encrypt and transmit communication data to protect the security of mission status, location, and environmental perception data.

[0148] Furthermore, data security storage is performed, including:

[0149] Recording flight mission execution data in a distributed ledger and storing it in the aircraft’s local storage;

[0150] The aircraft flight data is transmitted to the operating terminal based on the 5G communication module for data visualization and backup storage at the same time. The operating terminal uses AES256 to encrypt the backup storage data and transmit it wirelessly to the cloud storage.

[0151] The authenticity and tamper-proofness of flight mission data records are ensured through distributed ledger features (such as blockchain), and single point failures are avoided through a multi-node storage mechanism to ensure the long-term availability and security of mission data. Local storage supports the aircraft to independently perform missions in an environment without network coverage. Backup data can be used for subsequent tasks such as mission analysis, path optimization, and environmental modeling. Data visualization presents the mission status in an intuitive way, helping operators quickly understand mission progress and anomalies.

[0152] This embodiment also provides a method for a low-altitude aircraft flight control system based on an AI algorithm, including:

[0153] Collect environmental data of the aircraft's flight airspace and perform denoising processing;

[0154] Dynamically adjust airspace unit capacity based on communication accuracy and wind speed, and use LSTM model to predict unit capacity demand;

[0155] The AI ​​algorithm is used to calculate the total generation value to form a cost evaluation path, and the genetic algorithm is used to calculate the fitness value for path optimization;

[0156] Perform deep reinforcement learning on the optimized path to calculate the environmental impact factors of the path nodes, use the Bezier curve algorithm to interpolate the path, and generate an optimized global path;

[0157] Data is transmitted through encrypted communication modules for secure data storage.

[0158] Example 2, referring to Table 1, is the second example of the present invention. In order to further verify the technical solution of the present invention, experimental simulation data of a method for a low-altitude aircraft flight control system based on an AI algorithm are provided.

[0159] First, the unit capacity is calculated based on the collected wind speed data and communication accuracy, and the data on the change in the number of aircraft in the grid unit and the spatial dimension data of the historical capacity occupancy data of the adjacent grid units are collected respectively;

[0160] Table 1 The data table of the change of the number of aircraft in the grid unit is shown in the following table:

[0161] Time step Grid cell A1 Grid cell A2 Grid unit A3 Grid cell B1 Grid cell B2 Grid cell B3 t-4 2.5 3.2 1.8 4.1 3.6 2.9 t-3 3.1 3.4 2.0 4.3 3.9 3.1 t-2 3.6 3.8 2.4 4.7 4.1 3.5 t-1 4.0 4.2 2.8 5.0 4.5 3.8 t 4.5 4.6 3.1 5.3 4.8 4.0

[0162] Table 2 The historical capacity occupancy data of adjacent grid units is shown in the following table:

[0163] Time step Nearby A1 Capacity Nearby A2 Capacity Nearby A3 capacity Adjacent B1 Capacity Adjacent B2 capacity Adjacent B3 capacity t-4 70.5 65.3 72.4 80.3 76.5 71.2 t-3 72.5 67.2 73.8 82.4 77.9 72.8 t-2 73.5 69.0 75.0 84.0 79.4 74.5 t-1 75.2 70.8 76.4 85.5 80.6 76.2 t 76.8 72.3 78.0 87.0 82.1 77.8

[0164] Based on the tabular data input, the LSTM model outputs the capacity requirement for the next time step;

[0165] Table 3 The data table of capacity demand for the next time step is shown in the following table:

[0166] Grid Number A1 A2 A3 B1 B2 B3 Predicted value 82.3 85.7 78.9 90.2 88.1 83.6

[0167] After completing the calculation of the capacity demand for the next time step, start optimizing the aircraft's path and obtain relevant data of the current aircraft;

[0168] Table 4 The current aircraft data table is shown in the following table:

[0169] Current node coordinates Target node coordinates Current time Target time Maximum flight time （10.5,20.3,5.2） （50,60,15） 10 50 100

[0170] Determine the total cost from the current node to the target node respectively, expressed as:

[0171] ;

[0172] ;

[0173] ;

[0174] Clarify the coordinates of the adjacent extended nodes of the current node to calculate the total cost value;

[0175] Table 5 The total cost value data table of adjacent extended node coordinates is shown in the following table:

[0176] Adjacent extension node coordinates (12.5, 21.0, 5.5) (11.0, 22.3, 5.1) (13.0, 19.8, 6.0) (10.0, 20.0, 6.2) Total value 124.56 124.71 124.39 124.62

[0177] Based on the needs of path optimization, the AI ​​algorithm is used for heuristic path generation. The nodes to be expanded in the grid unit are used as the development list. Based on the adjacent expansion nodes of the three-dimensional coordinates of the current starting point of the aircraft, the node with the smallest total cost value is selected to add to the closed list, and repeated selection is performed until the selected node is the target node, forming a cost evaluation path. The path coordinates are expressed as:

[0178] [(10.5, 20.3, 5.2),(12.5, 21.0, 5.5),(30.0, 35.0, 10.0),(50.0, 60.0,15.0)];

[0179] [(10.5, 20.3, 5.2),(11.0, 22.3, 5.1),(25.0, 30.0, 12.0),(50.0, 60.0,15.0)];

[0180] And determine the obstacle distance sequence, expressed as:

[0181] [5.5, 6.0, 4.8, 7.2, 8.0];

[0182] Genetic algorithm is used to calculate the fitness value of the evaluation path for path optimization. At the same time, the genetic algorithm is used for crossover operation, and two paths are randomly selected to exchange part of the path segments to generate a new path, which is expressed as:

[0183] [(10.5, 20.3, 5.2),(11.0, 22.3, 5.1),(25.0, 30.0, 12.0),(50.0, 60.0,15.0)];

[0184] And perform mutation operation to randomly change the node position in the path to offset the path, which is expressed as:

[0185] [(10.5, 20.3, 5.2),(11.0, 22.3, 5.1),(26.0, 30.5, 12.2),(50.0, 60.0,15.0)];

[0186] Repeat the iterations. When the fitness value of the best individual in the population no longer changes significantly in the continuous iterations, stop the iteration and select the path with the highest fitness value as the optimization path, which is expressed as:

[0187] [(10.5,20.3,5.2) starting point, (11.0,22.3,5.1) first segment node, (26.0,30.5,12.2) mutation node, (50.0,60.0,15.0) target node];

[0188] A deep reinforcement learning algorithm is used for path optimization, and the reward function is determined based on the path length and obstacle avoidance, which is expressed as:

[0189] ;

[0190] The Q learning algorithm is used for training. The expected value of the cumulative reward after the state s takes action a is calculated for training. The Q value is calculated based on the current state data coordinates (26.0, 30.5, 12.2), and the Q function is used to determine the optimal action from the current state to the next state;

[0191] Table 6 The optimal Q value table for the next state is shown in the following table:

[0192] Available action nodes Corresponding Q value (30.0, 35.0, 15.0) 9.4 (28.0, 32.0, 14.0) 8.9 (27.0, 31.5, 13.0) 9.1

[0193] Select the optimal action and update the current state coordinates to (30.0, 35.0, 15.0);

[0194] Repeatedly select the optimal action update path and determine the reinforcement learning path as:

[0195] [(26.0,30.5,12.2),(30.0,35.0,15.0),(35.0,40.0,18.0),(50.0,60.0,15.0)];

[0196] The environmental impact factor of the reinforcement learning path is calculated based on the collected environmental data and expressed as:

[0197] ;

[0198] ;

[0199] ;

[0200] ;

[0201] Select node 3 with the largest influencing factor, and determine the starting point of the path segment as node 2, the adjustment node as node 3, and the end point as node 4;

[0202] The interpolation is calculated using the Bezier curve algorithm, expressed as:

[0203] ;

[0204] ;

[0205] ;

[0206] ;

[0207] After verifying the angle between the coordinates, the global path is determined, which is expressed as: [ (26.0, 30.5, 12.2), (30.0, 35.0, 15.0), (35.5, 40.0, 16.2), (40.0, 44.5, 16.5), (45.5, 50.0,15.8), (50.0, 60.0, 15.0) ], and the final global path is used to optimize the smoothness of each path and improve the feasibility of the aircraft's trajectory, especially in a dynamic environment, which meets the optimization expectations of path smoothness and aircraft operability.

[0208] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A low-altitude aircraft flight control system based on AI algorithm, characterized in that: include: The data processing module collects environmental data of the aircraft's flight airspace and performs denoising on the collected data; The airspace unit capacity dynamic adjustment module dynamically adjusts the airspace unit capacity based on the aircraft flight status, wind speed and communication accuracy, and predicts the unit capacity demand in the future time period through the LSTM model; The path planning module generates the total generation value path based on the AI ​​algorithm, calculates the path fitness value in combination with the genetic algorithm, and optimizes the path selection; The path optimization module uses deep reinforcement learning to calculate the environmental impact factors of path nodes, optimize path segments, and perform path smoothing interpolation based on Bezier curves; Data security transmission module, which transmits data through encrypted communication module and performs data security storage; The collecting of environmental data of the flight airspace of the aircraft and denoising of the collected data include: Obtain environmental data from the aircraft’s lidar, vision sensors, multispectral cameras, anemometers, millimeter-wave radars, and GPS, including weather conditions, obstacle locations, and terrain information; Use the extended Kalman filter (EKF) to align the sensor data in time and space to remove redundant data and noise; The method dynamically adjusts the airspace unit capacity based on the aircraft flight status, wind speed and communication accuracy, and predicts the unit capacity demand in the future time period through the LSTM model, including: The low-altitude area is divided into a three-dimensional grid of N×M×L, each grid cell is a cubic cell, and each cubic cell defines a capacity C, which is determined by the ratio of the cell volume to the average occupied space of a single aircraft's flight time step; The unit volume is determined by the product of the aircraft speed and the aircraft's safety reaction time; Dynamically adjust the cube unit capacity based on environmental parameters, expressed as: ； in Represents the unit capacity, and Respectively represent the current weather wind speed and standard weather wind speed, and They represent the average values ​​of Beidou navigation communication accuracy and standard communication accuracy, Indicates the historical unit capacity value; Based on the airspace flight altitude standards of aircraft, the airspace is vertically layered to clarify the low-altitude flight range; Build an LSTM learning model based on the long short-term memory network, including input layer, hidden layer and output layer; The input layer inputs the aircraft quantity change data of each grid cell collected from the time dimension and the historical capacity occupancy data of adjacent grid cells in the space dimension respectively; The hidden layer processes and learns the dependencies in the sequence data, and the output layer outputs the predicted capacity requirements of each grid unit in the next time step; Use the training data to train the LSTM learning model, select the cross entropy loss function to calculate the computational loss between the prediction of the LSTM learning model and the actual label, use the Adam optimizer for gradient descent optimization, update the parameters of the LSTM learning model, and stop iterating and outputting the model parameters when the model loss no longer decreases significantly during continuous iterations; Based on the real-time collected time dimension data and space dimension data, the LSTM learning model is input to predict the capacity requirements of each grid unit in the next time step and adjust the unit capacity; If the predicted capacity demand exceeds the current maximum unit capacity limit, path optimization is triggered; The method of generating the total generation value path based on the AI ​​algorithm, calculating the path fitness value in combination with the genetic algorithm, and optimizing the path selection includes: The improved Euclidean distance is used to calculate the prediction cost from the current node n to the target node, which is expressed as: ； in Represents the predicted cost from the current node n to the target node, , and are the three-dimensional coordinates of the current aircraft position, , and are the three-dimensional coordinates of the target position, and Based on the current time and target time of the flight mission, Indicates the maximum operational flight time of the aircraft; At the same time, the cumulative cost from the starting point to the current node n is calculated, expressed as: ； in represents the cumulative cost from the starting point to the current node n, Indicates the node before the current node The cumulative cost of Representation Node The Euclidean distance value with node n; The cost function is determined by combining the predicted cost and the accumulated cost, which is expressed as: ； in Represents the total cost of node n; Based on the needs of path optimization, AI algorithm is used for heuristic path generation. The nodes to be expanded in the grid unit are used as the development list. Based on the adjacent expansion nodes of the three-dimensional coordinates of the current starting point of the aircraft, the node with the smallest total cost value is selected to add to the closed list, and repeated selection is performed until the selected node is the target node, forming a cost evaluation path; Based on the path obstacles as key nodes, nodes with the same total cost value at the key nodes are formed into different cost evaluation paths; The initial population is determined based on the cost evaluation path generated by the AI ​​algorithm, and the genetic algorithm is used for path optimization. The fitness function is defined to evaluate the fitness value of the path, which is expressed as: ； in represents the fitness value of the evaluation path p, represents the total number of path nodes in the population, represents the Euclidean distance value from the i-th node n to the i+1-th node of path P, represents the distance from the i-th node n of path P to the nearest obstacle; At the same time, the genetic algorithm is used to perform crossover operations, randomly select two paths to exchange some path segments, generate a new path, and perform mutation operations to randomly change the node positions in the path to offset the path, and repeat iterations. When the fitness value of the best individual in the population no longer changes significantly in consecutive iterations, the iteration is stopped, and the path with the highest fitness value is selected as the optimized path; The method of using deep reinforcement learning to calculate environmental impact factors of path nodes, optimizing path segments, and performing path smoothing interpolation based on Bezier curves includes: A deep reinforcement learning algorithm is used for path optimization. The current path node position, surrounding obstacle distribution, and target position are used as model state data. The flight direction of the aircraft from the current node to the next node is used as action data. The reward function is determined based on the path length and obstacle avoidance, which is expressed as: ； Where R represents the reward value, Indicates the path increment caused by the current action, Indicates the distance to the nearest obstacle after the action; The starting point of the optimization path obtained by the genetic algorithm is used as the initial state of reinforcement learning, and the Q learning algorithm is used for training. The expected value of the cumulative reward after the state s takes action a is calculated for training; Use the training set for training, and use the ϵ-greedy strategy to select actions for training. Randomly select actions with probability ϵ to explore new paths. Select the action with the maximum current Q function value with probability 1−ϵ, and ϵ gradually decreases with the training process until it reaches the historical threshold and stops updating. Execute actions based on the action selection and update the state to calculate the immediate reward and update the Q value until the change in the Q function value is no longer obvious, then stop repeating the training; In the actual path adjustment, the trained Q function is used to select the optimal action according to the current state s, and the path nodes are gradually expanded according to the selected optimal action, and the path is updated until the target position is reached, which is used as the reinforcement learning path; Based on the Q-learning updated path, the key node sequence is discretized to generate the key node sequence, and the real-time environmental impact factor of the path node is calculated, which is expressed as: ； in represents the environmental impact factor, represents the distance from the i-th key node to the nearest obstacle, represents the aircraft speed at the i-th key node, represents the wind force on the i-th key node, represents the mass of the aircraft, Indicates the maximum acceleration of the aircraft; Will The largest key node is used as the priority adjustment node, and the path nodes before and after the node are selected to form a path segment. If there is an obstacle in the path segment, it is determined that the path segment needs to be optimized. If the direction change angle of the path segment is less than the minimum turning angle of the aircraft, it is determined that the path segment needs to be optimized. The path segment that needs to be optimized is modeled using a cubic Bezier curve, which is expressed as: ； in Represents the coordinates of the path point corresponding to the interpolation parameter t on the Bezier curve, represents the starting point of the path segment, represents the end point of the path segment, and represents the curvature and smoothness control points of the path segment, t represents the interpolation parameter, and the interpolation range is determined based on historical data; Define the smooth adjustment vector of the control point as the ratio of the obstacle direction vector to the obstacle vector modulus, and based on the smooth adjustment vector and of and as The coordinate values ​​of the smoothing vector and of and as The coordinate value of The interpolation points are evenly distributed through the interpolation range of t, and the coordinates of the path points are calculated using the Bezier curve formula. If the angle between the coordinates of the interpolation points is greater than or equal to the minimum steering angle of the aircraft, the optimized interpolation point sequence is replaced into the reinforcement learning path to generate an updated global path.

2. A low-altitude aircraft flight control system based on AI algorithm as claimed in claim 1, characterized in that: The method of encrypting and transmitting data through the communication module includes: Deploy edge computing modules on aircraft and configure point-to-point communication protocols between aircraft based on 5G communication modules; Use task decomposition algorithm to decompose global tasks into small computing units and distribute them to each aircraft; MQTT is used to transmit mission status, position, and environmental perception data between multiple aircraft, and AES256 encryption is used to encrypt communication data for transmission.

3. A low-altitude aircraft flight control system based on AI algorithm as claimed in claim 2, characterized in that: The secure data storage includes: Recording flight mission execution data in a distributed ledger and storing it in the aircraft’s local storage; The aircraft flight data is transmitted to the operating terminal based on the 5G communication module for data visualization and backup storage at the same time. The operating terminal uses AES256 encryption based on the backup storage data and transmits it wirelessly to the cloud storage.

4. A low-altitude aircraft flight control method based on an AI algorithm, based on the method of the low-altitude aircraft flight control system based on an AI algorithm according to any one of claims 1 to 3, characterized in that: include: Collect environmental data of the aircraft's flight airspace and perform denoising processing; Dynamically adjust airspace unit capacity based on communication accuracy and wind speed, and use LSTM model to predict unit capacity demand; The AI ​​algorithm is used to calculate the total generation value to form a cost evaluation path, and the genetic algorithm is used to calculate the fitness value for path optimization; Perform deep reinforcement learning on the optimized path to calculate the environmental impact factors of the path nodes, use the Bezier curve algorithm to interpolate the path, and generate an optimized global path; Data is transmitted through encrypted communication modules for secure data storage.

Images