Unmanned aerial vehicle dynamic precision decoupling method based on deep learning

Abstract

The invention discloses an unmanned aerial vehicle dynamic precision decoupling method based on deep learning, and the method comprises the following steps: 1, collecting and preprocessing the original flight data of an unmanned aerial vehicle, and obtaining a standardized flight data set; 2, inputting the standardized flight data set into a ConvLSTM network, extracting space and time features, and generating a space-time feature vector set; 3, inputting the spatio-temporal feature vector set into an improved GraphSAGE model for precision decoupling, and obtaining a decoupling feature vector; 4, calculating a flight precision error based on the decoupling feature vector, and obtaining real-time precision error feedback; 5, according to real-time precision error feedback, error processing is carried out through differential evolution, and an optimal precision control parameter is obtained; and 6, generating a flight control instruction set according to the optimal precision control parameter. According to the method, the ConvLSTM network and the improved GraphSAGE model are combined, so that accurate flight control of the unmanned aerial vehicle in a dynamic environment is realized.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) control technology, and in particular to a deep learning-based method for decoupling the dynamic precision of UAVs. Background Technology

[0002] With the rapid development of UAV technology, especially its widespread application in industry, agriculture, and transportation, the flight accuracy and stability of UAV control systems have become key technologies. Traditional UAV control methods typically rely on classical control algorithms such as PID control, Kalman filtering, and particle filtering. These methods primarily achieve target accuracy by adjusting the control input of the aircraft through feedback. However, with the increasing complexity of the flight environment, especially in dynamic environments and ever-changing tasks, traditional methods face significant challenges. A major drawback of traditional methods is their over-reliance on pre-set dynamic models for flight accuracy control. This makes it difficult for the system to make real-time and effective adjustments when faced with uncertainties, dynamically changing environments, or sensor noise. For example, factors such as wind speed changes, air pressure fluctuations, and terrain complexity can all affect the flight path. Traditional control algorithms often cannot cope with these external environmental disturbances, resulting in flight accuracy failing to meet expectations. To overcome the shortcomings of traditional technologies, deep learning methods have been increasingly applied to UAV flight control systems in recent years. Deep learning automatically learns features and makes predictions from large amounts of data, enabling it to better cope with changes in complex environments and possessing stronger adaptive capabilities compared to traditional algorithms. However, existing deep learning-based UAV control methods mostly focus on static path planning and attitude control, exhibiting weak real-time adjustment capabilities for flight accuracy and dynamic mission changes. Furthermore, many existing deep learning methods rely solely on simple neural networks for flight control, failing to adequately consider the interactions of multiple factors during flight and the decoupling of accuracy. Dynamic changes in flight status and external environmental disturbances often lead to the accumulation of flight errors, making it difficult for existing methods to provide accurate real-time adjustments in highly dynamic environments.

[0003] Therefore, how to provide a deep learning-based method for decoupling the dynamic accuracy of UAVs is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a deep learning-based dynamic precision decoupling method for unmanned aerial vehicles (UAVs). This invention combines deep learning and graph neural networks to propose a dynamic precision decoupling method for UAVs. Spatiotemporal features are extracted using a ConvLSTM network, and an improved GraphSAGE model performs precision decoupling of flight state and environmental data, calculating and optimizing flight control parameters in real time. A differential evolution algorithm is employed to handle precision errors, generating accurate flight control commands to ensure high-precision flight control of the UAV in dynamic environments. This overcomes the problem of traditional methods being unable to cope with complex environmental changes, improving the flight stability and mission execution efficiency of the UAV.

[0005] A deep learning-based method for decoupling the dynamic accuracy of unmanned aerial vehicles (UAVs) according to an embodiment of the present invention includes the following steps:

[0006] Step 1: Collect raw flight data from the UAV and preprocess the raw flight data to obtain a standardized flight dataset;

[0007] Step 2: Input the standardized flight dataset into the ConvLSTM network, extract spatial features through convolution operations, and extract time-dependent features to generate a spatiotemporal feature vector set;

[0008] Step 3: Input the spatiotemporal feature vector set into the improved GraphSAGE model, and decouple the precision through the dynamic topology relationship construction module, the neighbor information aggregation module, and the feature transformation module to obtain the decoupled feature vector;

[0009] Step 4: Based on the decoupled feature vector, calculate the flight accuracy error, which is the difference between the actual flight accuracy and the expected flight accuracy, and obtain real-time accuracy error feedback;

[0010] Step 5: Based on the real-time accuracy error feedback, differential evolution is used for error processing to obtain the optimal accuracy control parameters;

[0011] Step 6: Generate a flight control command set based on the optimal precision control parameters. The flight control command set includes speed control commands, attitude control commands, and position control commands.

[0012] Optionally, the raw flight data specifically includes:

[0013] The latitude, longitude, and altitude information of the drone are obtained through GPS positioning to obtain data for determining the drone's position in space;

[0014] The pitch, roll, and yaw angles of the UAV are acquired by the inertial measurement unit to obtain attitude data representing the UAV's spatial orientation.

[0015] The drone's speed data is obtained by acquiring its linear velocity, angular velocity, and linear and rotational velocities in various axes through motion sensors.

[0016] Environmental data of the UAV is acquired through meteorological sensors, temperature and humidity sensors, barometers and anemometers. This environmental data is used to monitor meteorological conditions and the impact of the external environment on the UAV's flight.

[0017] Optionally, the preprocessing steps specifically include:

[0018] Kalman filtering is applied to the raw flight data to remove environmental interference and sensor errors;

[0019] Perform time alignment and synchronization processing on different types of data in the raw flight data until they correspond on the same time axis;

[0020] Different types of data in the original flight data are normalized using Z-Score standardization to obtain a standardized flight dataset with a unified structure.

[0021] Optionally, step two specifically involves:

[0022] The standardized flight dataset is input into the convolutional layer of the ConvLSTM network. By using convolutional kernels of different sizes, spatial structure features are extracted to generate spatial feature maps.

[0023] The spatial feature map is input into the pooling layer, and the spatial feature map is downsampled through max pooling to reduce the size of the spatial feature map and generate a pooled spatial feature matrix.

[0024] The spatial feature matrix is ​​input into the LSTM layer. The LSTM layer extracts time-dependent features from the spatial feature matrix by capturing the dynamic changes between different moments in the flight process, and obtains a temporal feature vector. The temporal feature vector represents the time change information in the flight process.

[0025] The temporal feature vector is input into a fully connected layer, which combines the temporal feature vector with spatial features and performs feature fusion through linear transformation to obtain a spatiotemporal feature vector set, which represents the spatiotemporal information during flight.

[0026] Optionally, the improved GraphSAGE model is specifically as follows:

[0027] The spatiotemporal feature vector set is input into the dynamic topology construction module, and different spatiotemporal feature vectors in the spatiotemporal feature vector set are used as nodes to obtain a node set.

[0028] The Manhattan distance between the spatiotemporal feature vectors corresponding to every two different nodes in the node set is calculated. If the Manhattan distance is less than a preset distance threshold, an edge connection is established between the corresponding nodes. All node pairs are traversed to obtain the set of connection edges.

[0029] Based on the connection relationships in the node set and the edge set, a dynamic topology graph of UAV flight information is established;

[0030] Input the dynamic topology graph into the neighbor information aggregation module, select several neighbor nodes of each node in the dynamic topology graph to form a neighbor node set;

[0031] Aggregate the neighbor nodes in the neighbor node set with the current node to obtain the multi-source aggregated node feature vector;

[0032] The feature vectors of the multi-source aggregation nodes are input into the feature transformation module, and the Sigmoid nonlinear transformation function is used to map the feature vectors of the multi-source aggregation nodes to a new feature space.

[0033] In the new feature space, the feature vector of the current multi-source aggregation node is weighted and averaged with that of its neighboring nodes to obtain the decoupled feature vector after fusion.

[0034] Optionally, the step of aggregating the neighbor nodes in the neighbor node set with the current node to obtain a multi-source aggregated node feature vector specifically involves:

[0035] Calculate the cosine similarity between the spatiotemporal feature vector of the current node and the spatiotemporal feature vector of each neighboring node;

[0036] If the cosine similarity is less than the preset similarity threshold, the corresponding neighbor node is added to the set of neighbor nodes to be aggregated.

[0037] The spatiotemporal feature vector corresponding to the current node is weighted and added to the spatiotemporal feature vectors corresponding to the neighbor nodes in the set of neighbor nodes to be aggregated, to obtain the multi-source aggregated node feature vector. The weight of the weighting is the reciprocal of the corresponding cosine similarity.

[0038] Optionally, step four specifically involves:

[0039] Based on the decoupled feature vector, the flight accuracy error is calculated, whereby the flight accuracy error is the difference between the actual flight accuracy and the expected flight accuracy. The calculation steps include:

[0040] The expected flight accuracy is set according to the requirements of the UAV flight mission and the target accuracy. The expected flight accuracy includes position accuracy, attitude accuracy and velocity accuracy, and the expected flight accuracy vector is obtained. The expected flight accuracy vector represents the accuracy standard required by the mission.

[0041] Based on the current flight status data of the UAV in the decoupled feature vector, the actual flight accuracy is calculated through real-time feedback sensor data, and compensation is performed through sensor accuracy correction to obtain the actual flight accuracy vector, which represents the accuracy of the current flight status of the UAV.

[0042] The actual flight accuracy vector is compared with the expected flight accuracy vector, the difference between the two is calculated to obtain the flight accuracy error, and real-time accuracy error feedback is generated. The real-time accuracy error feedback is used to represent the accuracy adjustment requirements of flight control.

[0043] Optionally, the specific steps of the differential evolution include:

[0044] An initial population is generated based on the real-time accuracy error feedback. Each individual in the initial population represents a set of accuracy control parameters. Each individual contains multiple parameter dimensions, including the initial values ​​of position accuracy, attitude accuracy, and velocity accuracy control parameters.

[0045] The initial population is initialized by random generation to obtain the initial precision control parameter population;

[0046] Calculate the fitness value of each individual in the initial precision control parameter population;

[0047] The fitness value is obtained by minimizing the real-time accuracy error based on the accuracy control parameters corresponding to each individual according to a preset fitness function.

[0048] The fitness value of each individual is calculated, and a fitness value matrix is ​​obtained.

[0049] Individuals in the initial precision control parameter population are selected based on their fitness values. Individuals with fitness values ​​greater than a preset threshold are selected to generate a new candidate individual population.

[0050] Mutate each individual in the candidate population to generate new mutated individuals;

[0051] The mutation operation explores the solution space by adding random perturbations to the parent individuals, resulting in a population of mutated individuals;

[0052] A single-point crossover operation is performed between the mutant individual population and the candidate individual population to generate a new generation of individuals. Based on the fitness value of the new generation of individuals generated by the crossover, the optimal individual is selected as the new precision control parameter.

[0053] The control parameters corresponding to the new generation of individuals are the optimal solution under the current error minimization objective, and the optimal precision control parameters are output.

[0054] Optionally, step six specifically includes:

[0055] Based on the speed accuracy requirements in the optimal accuracy control parameters, the speed control amount required to adjust the flight accuracy error is determined, and a speed control command is generated. The speed control command is used to adjust the linear velocity and angular velocity of the UAV.

[0056] Based on the attitude accuracy requirements in the optimal accuracy control parameters, the attitude adjustment amount required to adjust the flight accuracy error is determined, and attitude control commands are generated. The attitude control commands are used to adjust the pitch angle, roll angle, and yaw angle of the UAV.

[0057] Based on the position accuracy requirements in the optimal accuracy control parameters, determine the spatial position adjustment amount required to adjust the flight accuracy error;

[0058] The speed control command, attitude control command, and position control command are integrated to obtain the flight control command set, which contains precise adjustment information for controlling the UAV at each stage of flight.

[0059] The beneficial effects of this invention are:

[0060] This invention provides a deep learning-based dynamic precision decoupling method for UAVs, overcoming several shortcomings of traditional UAV flight control methods and significantly improving flight accuracy and stability in complex dynamic environments. Traditional control methods, such as PID control and Kalman filtering, while providing effective control in certain static environments, often fail to cope with environmental changes (such as wind speed, air pressure changes, and complex terrain), leading to unstable flight paths and unsatisfactory accuracy. This invention, through the combination of deep learning and graph neural networks, can process and adapt to various dynamic factors during flight in real time, automatically adjusting the flight control strategy to effectively reduce flight accuracy errors. By employing a ConvLSTM network, this invention can extract spatiotemporal features from multimodal sensor data, not only extracting spatial information (such as obstacle and terrain changes) but also capturing temporal changes during flight, ensuring more precise control of the UAV along its flight path. Unlike traditional static model-based control methods, this invention's deep learning method dynamically adjusts the control strategy through real-time analysis of flight data, ensuring high accuracy for the UAV in different flight stages and environmental conditions. Meanwhile, the improved GraphSAGE model effectively performs correlation analysis on flight status and environmental data by constructing a dynamic topological relationship graph, further improving the accuracy of the decoupling process.

[0061] This invention ensures the optimization of precision control parameters by decoupling the precision error calculation of feature vectors and optimizing them using a differential evolution algorithm. This allows the flight control system to adjust in real time and obtain the optimal solution when faced with precision errors. Furthermore, the generated flight control command set includes precise adjustment commands for speed, attitude, and position, ensuring that the UAV executes its mission accurately during flight and flexibly adjusts its path when the environment changes, greatly improving the quality of mission completion. By combining deep learning, graph neural networks, and optimization algorithms, this invention overcomes the shortcomings of existing technologies, solves the instability problem of precision control in UAVs in dynamic and complex environments, and significantly improves the autonomous flight capability and adaptability of UAVs. It is particularly suitable for high-precision applications such as industrial inspection, agricultural plant protection, and environmental monitoring, providing an innovative solution for the further development of UAV flight control technology. Attached Figure Description

[0062] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0063] Figure 1 This is an overall flowchart of a deep learning-based UAV dynamic precision decoupling method proposed in this invention.

[0064] Figure 2 This is a schematic diagram of the improved GraphSAGE model processing steps for a deep learning-based UAV dynamic precision decoupling method proposed in this invention.

[0065] Figure 3 This is a differential evolution flowchart of a deep learning-based UAV dynamic accuracy decoupling method proposed in this invention. Detailed Implementation

[0066] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0067] refer to Figure 1-3 A deep learning-based method for decoupling the dynamic accuracy of unmanned aerial vehicles (UAVs) includes the following steps:

[0068] Step 1: Collect raw flight data from the UAV and preprocess the raw flight data to obtain a standardized flight dataset;

[0069] Step 2: Input the standardized flight dataset into the ConvLSTM network, extract spatial features through convolution operations, and extract time-dependent features to generate a spatiotemporal feature vector set;

[0070] Step 3: Input the spatiotemporal feature vector set into the improved GraphSAGE model, and decouple the precision through the dynamic topology relationship construction module, the neighbor information aggregation module, and the feature transformation module to obtain the decoupled feature vector;

[0071] Step 4: Based on the decoupled feature vector, calculate the flight accuracy error, which is the difference between the actual flight accuracy and the expected flight accuracy, and obtain real-time accuracy error feedback;

[0072] Step 5: Based on the real-time accuracy error feedback, differential evolution is used for error processing to obtain the optimal accuracy control parameters;

[0073] Step 6: Generate a flight control command set based on the optimal precision control parameters. The flight control command set includes speed control commands, attitude control commands, and position control commands.

[0074] In this embodiment, the raw flight data specifically includes:

[0075] By using GPS positioning to obtain the drone's latitude, longitude, and altitude information, data can be obtained to determine the drone's location in space.

[0076] The drone acquires latitude, longitude, and altitude information in real time via a high-precision GPS receiver. Latitude and longitude are typically accurate to six decimal places to ensure meter-level precision; altitude is generally accurate to 0.1 meters. This location data allows for efficient and accurate determination of the drone's position in three-dimensional space, ensuring real-time tracking and control during flight.

[0077] The pitch, roll, and yaw angles of the UAV are acquired by the inertial measurement unit to obtain attitude data representing the UAV's spatial orientation.

[0078] The inertial measurement unit (IMU) equipped on the UAV can acquire attitude information in real time during flight, including pitch, roll, and yaw angles. The accuracy of pitch, roll, and yaw angles is typically 0.01 degrees, ensuring that the UAV can maintain stable attitude control in complex flight environments. This attitude data provides the necessary directional basis for subsequent flight control and accuracy decoupling.

[0079] The drone's speed data is obtained by acquiring its linear velocity, angular velocity, and linear and rotational velocities along various axes using motion sensors.

[0080] The drone's built-in motion sensors (such as accelerometers and gyroscopes) acquire linear velocity and rotational speed in real time. The measurement accuracy of linear velocity is typically 0.1 m / s, and the measurement accuracy of angular velocity is 0.01 degrees / s. Using this data, the drone's speed changes along various axes can be accurately tracked, especially for dynamic adjustments to speed changes during flight.

[0081] Environmental data of the UAV is acquired through meteorological sensors, temperature and humidity sensors, barometers, and anemometers. This environmental data is used to monitor meteorological conditions and the impact of the external environment on the UAV's flight.

[0082] The weather sensors onboard the drone can monitor weather conditions in real time during flight, including temperature, humidity, air pressure, and wind speed. Specifically:

[0083] Temperature and humidity sensors typically have a measurement accuracy of 0.1°C and 1%RH, and are used to monitor changes in air temperature and humidity during flight.

[0084] The barometer is typically accurate to 0.1 hPa and is used to monitor changes in air pressure in real time and adjust the flight path according to altitude.

[0085] An anemometer can measure wind speed in real time with an accuracy of 0.1 m / s, and is used to sense changes in wind force and direction during flight.

[0086] With this environmental data, drones can better adapt to changes in the external environment during flight, avoid a decrease in flight accuracy due to weather changes, and provide more accurate feedback to the flight control system.

[0087] The raw flight data includes the UAV's spatial position, attitude information, speed information, and environmental data. This data is used for real-time monitoring and control of flight accuracy during real-time flight, thereby providing basic data support for subsequent flight control accuracy optimization and decoupling processes.

[0088] In this embodiment, the preprocessing step specifically includes:

[0089] Kalman filtering is applied to the raw flight data to remove environmental interference and sensor errors:

[0090] Raw flight data is typically affected by environmental disturbances (such as wind speed variations and temperature fluctuations) and sensor errors (such as noise and drift). To improve data quality and accuracy, the Kalman filter algorithm is used for noise reduction. The Kalman filter is a recursive algorithm based on a state-space model, which can effectively estimate the system state and reduce noise interference. The accuracy of the Kalman filter can typically control position errors to within 0.5 meters and velocity errors to within 0.1 meters per second, thus effectively removing errors from the flight data and making the data more reliable.

[0091] Different types of data in the raw flight data are time-aligned and synchronized until they correspond on the same timeline:

[0092] In the process of multi-sensor data fusion, the data acquisition frequencies and timestamps of different sensors may differ, thus requiring time alignment and synchronization. Specifically, for data from multiple sources such as GPS, IMU, and weather sensors, the timestamps must first be unified, and then interpolation or synchronization processing should be performed according to the sampling frequency of each sensor to ensure that the data from different sensors are aligned on the same timeline. This step ensures that the data at each moment is comprehensive and consistent, typically requiring data synchronization accuracy within milliseconds to achieve accurate time-series comparisons in subsequent analysis.

[0093] Different types of data in the raw flight data were normalized using Z-Score standardization to obtain a standardized flight dataset with a uniform structure.

[0094] Data collected from different sensors, such as position, attitude, velocity, and environmental data, may differ in dimensions and ranges, thus requiring data normalization. Z-Score normalization is a commonly used method that eliminates dimensional differences by subtracting the mean from the data and dividing by the standard deviation, thereby adjusting the mean to 0 and the standard deviation to 1. Specifically, position data, velocity data, and sensor data (such as temperature, air pressure, and wind speed) are all subjected to Z-Score normalization to ensure that all data fall within the same scale, facilitating subsequent data fusion and feature extraction. Standardized data has a uniform structure, providing higher accuracy and robustness for subsequent model training and flight accuracy decoupling.

[0095] Through the above preprocessing steps, the raw flight data has been effectively denoised, synchronized, and standardized, ensuring the reliability and consistency of the data and providing a precise data foundation for subsequent flight accuracy analysis, decoupling, and optimization.

[0096] In this embodiment, step two specifically includes:

[0097] The standardized flight dataset is input into the convolutional layer of the ConvLSTM network. By using convolutional kernels of different sizes, spatial structure features are extracted to generate spatial feature maps.

[0098] The spatial feature map is input into the pooling layer, and the spatial feature map is downsampled through max pooling to reduce the size of the spatial feature map and generate a pooled spatial feature matrix.

[0099] The spatial feature matrix is ​​input into the LSTM layer. The LSTM layer extracts time-dependent features from the spatial feature matrix by capturing the dynamic changes between different moments in the flight process, and obtains a temporal feature vector. The temporal feature vector represents the time change information in the flight process.

[0100] The temporal feature vector is input into a fully connected layer, which combines the temporal feature vector with spatial features and performs feature fusion through linear transformation to obtain a spatiotemporal feature vector set, which represents the spatiotemporal information during flight.

[0101] In this embodiment, the improved GraphSAGE model is specifically as follows:

[0102] The spatiotemporal feature vector set is input into the dynamic topology construction module, and different spatiotemporal feature vectors in the spatiotemporal feature vector set are used as nodes to obtain a node set.

[0103] The Manhattan distance between the spatiotemporal feature vectors corresponding to every two different nodes in the node set is calculated. If the Manhattan distance is less than a preset distance threshold, an edge connection is established between the corresponding nodes. All node pairs are traversed to obtain the set of connection edges.

[0104] Based on the connection relationships in the node set and the edge set, a dynamic topology graph of UAV flight information is established;

[0105] Input the dynamic topology graph into the neighbor information aggregation module, select several neighbor nodes of each node in the dynamic topology graph to form a neighbor node set;

[0106] Aggregate the neighbor nodes in the neighbor node set with the current node to obtain the multi-source aggregated node feature vector;

[0107] The feature vectors of the multi-source aggregation nodes are input into the feature transformation module, and the Sigmoid nonlinear transformation function is used to map the feature vectors of the multi-source aggregation nodes to a new feature space.

[0108] In the new feature space, the feature vector of the current multi-source aggregation node is weighted and averaged with that of its neighboring nodes to obtain the decoupled feature vector after fusion.

[0109] This step significantly improves the flight accuracy and control capabilities of UAVs in complex dynamic environments by introducing an improved GraphSAGE model. Unlike traditional flight control methods, this invention utilizes graph neural network technology from deep learning to dynamically construct the topological relationship between flight state and environmental variables, ensuring real-time decoupling and optimization of flight accuracy. By using spatiotemporal feature vectors as nodes and calculating the Manhattan distance between nodes, this invention can accurately capture the similarity and dependency between states during flight, effectively avoiding the limitations of static models in traditional methods that cannot adapt to dynamic environments. Furthermore, by establishing a dynamic topological relationship graph, the neighbor node set can be flexibly adjusted according to changes in flight state and environment, making the decoupling process more precise and dynamic. This approach not only improves the accuracy of decoupling feature vectors but also enhances the UAV system's adaptability to complex tasks and environmental changes. The feature transformation module optimizes the node feature space through nonlinear transformation, further improving the accuracy of decoupling, enabling the flight control system to precisely adjust various control parameters during flight, ensuring the stability and efficiency of mission execution. The generated fused and decoupled feature vectors provide accurate flight control commands, enabling UAVs to achieve high-precision flight missions in changing flight environments. This solves the problem of existing technologies being unable to adjust control strategies in real time and cope with environmental changes, and significantly improves the intelligence and autonomy of UAV control.

[0110] In this embodiment, the step of aggregating the neighbor nodes in the neighbor node set with the current node to obtain the multi-source aggregated node feature vector specifically involves:

[0111] Calculate the cosine similarity between the spatiotemporal feature vector of the current node and the spatiotemporal feature vector of each neighboring node;

[0112] If the cosine similarity is less than the preset similarity threshold, the corresponding neighbor node is added to the set of neighbor nodes to be aggregated.

[0113] The spatiotemporal feature vector corresponding to the current node is weighted and added to the spatiotemporal feature vectors corresponding to the neighbor nodes in the set of neighbor nodes to be aggregated, to obtain the multi-source aggregated node feature vector. The weight of the weighting is the reciprocal of the corresponding cosine similarity.

[0114] This step significantly improves the adaptability and accuracy of the flight control system in dynamic environments by introducing a multi-source aggregation node feature vector generation method based on cosine similarity. By calculating the cosine similarity between the current node and its neighbors, the similarity between nodes can be effectively evaluated, ensuring that only neighboring nodes highly correlated with the current node's state are selected during the aggregation process. This method improves the accuracy of feature aggregation through strict similarity threshold screening, making the decoupling process more precise and avoiding interference from irrelevant nodes on flight control accuracy. Based on this, an aggregation method using inverse similarity weighting can dynamically adjust the aggregation weights according to the similarity between neighboring nodes and the current node. Neighboring nodes with higher similarity are assigned higher weights, ensuring that their information has a greater impact on the current node, thereby optimizing flight control accuracy. Through this weighted aggregation strategy, this invention can accurately extract the correlation between flight state and environmental data, further improving the stability of the flight path and the reliability of mission execution. It can flexibly adjust the control strategy according to real-time changes in flight state and environment, exhibiting stronger adaptability and robustness compared to traditional static models, and is particularly suitable for high-precision flight control tasks in complex environments.

[0115] In this embodiment, step four specifically includes:

[0116] Based on the decoupled feature vector, the flight accuracy error is calculated, whereby the flight accuracy error is the difference between the actual flight accuracy and the expected flight accuracy. The calculation steps include:

[0117] Based on the requirements and target accuracy of the UAV flight mission, the expected flight accuracy is set, including position accuracy, attitude accuracy, and velocity accuracy, resulting in an expected flight accuracy vector. This expected flight accuracy vector represents the accuracy standard required by the mission.

[0118] Before a flight mission begins, the expected flight accuracy is set based on mission requirements and flight environment factors. Expected flight accuracy typically includes position accuracy (e.g., position error not exceeding 1 meter), attitude accuracy (e.g., pitch, roll, and yaw angle errors controlled within 0.1 degrees), and velocity accuracy (e.g., linear velocity and angular velocity errors controlled within 0.1 m / s and 0.01 degrees / s, respectively). These accuracy standards are set based on the UAV's mission objectives, environmental factors, and the required level of accuracy during mission execution, ultimately forming an expected flight accuracy vector that represents the accuracy requirements of the target mission.

[0119] Based on the current flight status data of the UAV in the decoupled feature vector, the actual flight accuracy is calculated through real-time feedback sensor data, and compensated by sensor accuracy correction to obtain the actual flight accuracy vector, which represents the accuracy of the UAV's current flight status.

[0120] Based on real-time flight status data obtained from UAV sensors (such as GPS, IMU, motion sensors, etc.), the actual flight accuracy is calculated by comparing the deviation between the actual flight status and the predetermined path. Actual flight accuracy refers to the difference between the accuracy of the UAV's current position, attitude, speed, etc., during flight and the target values. Furthermore, by calibrating the sensor accuracy to eliminate deviations caused by environmental changes or sensor errors, an actual flight accuracy vector is obtained. This vector reflects the accuracy of the UAV's current flight status and is used for subsequent error adjustment and optimization.

[0121] The actual flight accuracy vector is compared with the expected flight accuracy vector, the difference between the two is calculated to obtain the flight accuracy error, and real-time accuracy error feedback is generated. The real-time accuracy error feedback is used to represent the accuracy adjustment requirements of flight control.

[0122] The flight accuracy error is calculated by subtracting the actual flight accuracy vector from the expected flight accuracy vector. This error reflects the difference between the UAV's current flight accuracy and the target accuracy, and is typically expressed as Euclidean distance, Manhattan distance, or weighted distance. Calculating the flight accuracy error allows for the evaluation of the flight control system's performance and the identification of areas with insufficient accuracy. Subsequently, real-time accuracy error feedback is generated and transmitted to the flight control system to dynamically adjust the flight control strategy and accuracy decoupling parameters, thereby optimizing flight accuracy and ensuring that the UAV consistently meets mission accuracy requirements during flight.

[0123] Through the above steps, the flight accuracy error calculation and real-time feedback mechanism can accurately assess the flight accuracy of the UAV and provide feedback for subsequent control strategy adjustments and accuracy optimization, ensuring that the flight accuracy of the UAV is always kept within the range required by the mission.

[0124] In this embodiment, the specific steps of differential evolution include:

[0125] An initial population is generated based on the real-time accuracy error feedback. Each individual in the initial population represents a set of accuracy control parameters, and each individual contains multiple parameter dimensions, including initial values ​​for position accuracy, attitude accuracy, and velocity accuracy control parameters.

[0126] Based on real-time accuracy error feedback, an initial population is generated. Each individual represents a set of accuracy control parameters, including position accuracy, attitude accuracy, and velocity accuracy. The initial value of each accuracy control parameter is typically randomly generated based on prior knowledge of mission requirements and the flight environment. To ensure the diversity of control parameters, the accuracy control parameters of each individual in the initial population cover a wide range of values. For example, the initial value of position accuracy can be randomly generated between 1 meter and 10 meters, the initial value of attitude accuracy between 0.1 degrees and 5 degrees, and the initial value of velocity accuracy between 0.1 m / s and 2 m / s. This random generation method ensures the initial diversity of the population and enhances the ability of the differential evolution algorithm to explore the solution space.

[0127] The initial population is initialized through random generation to obtain the initial precision control parameter population:

[0128] The initial precision control parameter population is initialized using a random generation method. The control parameters of each individual in the population are randomly generated within a predetermined range, ensuring the randomness and global search capability of the algorithm. Specifically, each parameter of position precision, attitude precision, and velocity precision is generated through random methods such as uniform distribution or normal distribution, thus obtaining an initial population containing multi-dimensional precision control parameters.

[0129] Calculate the fitness value of each individual in the initial precision control parameter population:

[0130] For each initial individual, its fitness value is calculated. The fitness value represents the effect of the accuracy control parameters corresponding to that individual on minimizing flight accuracy error. A fitness function is typically used to evaluate the performance of each individual. The fitness function is defined by calculating the error between the actual flight accuracy and the expected flight accuracy; the smaller the error, the larger the fitness value. Specifically, the fitness value is used to determine the effectiveness of the individual in reducing accuracy error by calculating the magnitude of the flight accuracy error corresponding to each accuracy control parameter.

[0131] The fitness value is obtained by minimizing the real-time accuracy error for each individual based on the preset fitness function and the corresponding accuracy control parameters.

[0132] In differential evolution algorithms, the accuracy control parameters of each individual are evaluated using a pre-defined fitness function, which is typically a function that minimizes flight accuracy error. By comparing the fitness values ​​of different individuals, it can be determined which individuals' control parameters are best suited to the current flight accuracy error adjustment target.

[0133] By calculating the fitness value of each individual, a fitness value matrix is ​​obtained:

[0134] Calculate the fitness values ​​of all individuals and organize these fitness values ​​into a fitness value matrix. This matrix is ​​used for subsequent individual selection and mutation operations to select individuals with better fitness for crossover and mutation to generate new candidate individuals.

[0135] Individuals in the initial precision control parameter population are selected based on their fitness values. Individuals with fitness values ​​greater than a preset threshold are selected to generate a new candidate population.

[0136] In differential evolution, the selection operation is performed by comparing the fitness values ​​of individuals. Individuals with higher fitness values ​​are typically chosen as parents, as these individuals better meet the accuracy control objective. The selected individuals are used to generate a new candidate population, with individuals having fitness values ​​greater than a preset threshold being given priority to ensure the algorithm's convergence.

[0137] Mutate each individual in the candidate population to generate new mutated individuals:

[0138] Mutation is a process that generates new individuals by introducing random perturbations into parent individuals. Specifically, for each parent individual, its control parameters are randomly varied to explore different possible solutions in the solution space. This operation ensures population diversity and avoids getting trapped in local optima. The mutation magnitude is usually adjusted according to the error range of the flight mission; for example, in the mutation of position accuracy, attitude accuracy, and velocity accuracy, the perturbation magnitude is typically 20%-30% of the accuracy requirement.

[0139] The mutation operation explores the solution space by adding random perturbations to the parent individuals, resulting in a population of mutated individuals:

[0140] By adding perturbations to the control parameters of the parent individuals, a population of mutant individuals is obtained. The precision control parameters of each mutant individual are generated by perturbation based on the parent individuals. This process aims to find a better solution and explore regions in the solution space that are more likely to reduce flight precision errors.

[0141] A single-point crossover operation is performed between the mutated individual population and the candidate individual population to generate a new generation of individuals. Based on the fitness value of the new generation of individuals generated by the crossover, the optimal individual is selected as the new precision control parameter.

[0142] Single-point crossover involves crossing a mutant individual with a candidate individual to generate a new generation of individuals. The crossover operation produces new offspring by exchanging the genetic information (i.e., precision control parameters) of the parent individuals. The fitness value of the new generation is calculated, and the individual with the best fitness is selected as the final precision control parameter.

[0143] The control parameters corresponding to the new generation of individuals are the optimal solution under the current error minimization objective, and the optimal precision control parameters are output:

[0144] Based on the fitness values ​​of the new generation of individuals generated by crossover and mutation operations, the optimal individual is selected as the final accuracy control parameter. This accuracy control parameter is the optimal solution that minimizes the current flight accuracy error, effectively adjusting the UAV's flight accuracy and outputting the final optimal accuracy control parameter for subsequent flight control optimization.

[0145] Through multiple iterations and optimizations of the differential evolution algorithm, this implementation method can minimize flight accuracy errors by continuously adjusting the precision control parameters, thereby ensuring high-precision control of the UAV in complex flight missions.

[0146] In this embodiment, step six specifically includes:

[0147] Based on the speed accuracy requirements in the optimal accuracy control parameters, the speed control amount required to adjust the flight accuracy error is determined, and a speed control command is generated. This speed control command is used to adjust the linear velocity and angular velocity of the UAV.

[0148] Based on the speed accuracy requirements in the optimal precision control parameters, the speed control amount needed to adjust the flight accuracy error is first determined. By calculating the difference between the actual flight speed and the target flight speed, and considering the magnitude of the flight error, a precise speed control command is generated. This control command is used to adjust the UAV's linear velocity (i.e., velocity along each axis) and angular velocity (i.e., rotational speed around each axis). The accuracy of the speed control command can typically control the flight speed within a range of 0.1 m / s, ensuring accurate adjustment and stable flight of the UAV in dynamic environments.

[0149] Based on the attitude accuracy requirements in the optimal accuracy control parameters, the attitude adjustment amount required to adjust the flight accuracy error is determined, and attitude control commands are generated. These attitude control commands are used to adjust the pitch angle, roll angle, and yaw angle of the UAV.

[0150] Based on the attitude accuracy requirements in the optimal precision control parameters, the error between the current attitude (such as pitch, roll, and yaw) and the target attitude is calculated to obtain the attitude control commands needed for adjustment. These commands are used to adjust the UAV's pitch, roll, and yaw angles, thereby ensuring the UAV's stability and accuracy during flight. The accuracy of the attitude control commands can typically be adjusted within 0.1 degrees, ensuring that the UAV can accurately execute flight tasks according to the set attitude during flight.

[0151] Based on the position accuracy requirement in the optimal accuracy control parameters, determine the spatial position adjustment amount required to adjust the flight accuracy error:

[0152] Based on the position accuracy requirements in the optimal precision control parameters, the error between the current UAV position and the target position is calculated to obtain the spatial position adjustment amount. This step ensures that the UAV can accurately adjust its flight path and eliminate flight deviations caused by changes in the external environment (such as wind speed, air pressure, etc.). Position accuracy control typically requires an error of no more than 1 meter, thereby ensuring that the UAV can maintain a precise flight path during mission execution.

[0153] The speed control command, attitude control command, and position control command are integrated to obtain the flight control command set, which contains precise adjustment information for controlling the UAV at each stage of flight:

[0154] The generated speed control commands, attitude control commands, and position control commands are integrated to form a complete flight control command set. This command set not only includes precise adjustment information for the UAV at different flight stages but also allows for dynamic adjustments based on changes in flight accuracy errors. The flight control command set ensures that the UAV's speed, attitude, and position remain within the predetermined accuracy range throughout flight, guaranteeing successful mission execution and enabling necessary adjustments based on real-time flight conditions.

[0155] Through the above steps, the flight control command set can precisely adjust the flight path, speed, attitude and position of the UAV according to the optimal precision control parameters, thereby achieving high-precision flight control and ensuring the stability of the UAV in complex dynamic environments and the reliability of mission execution.

[0156] Example 1:

[0157] To verify the feasibility of this invention in practice, it was applied to an industrial inspection task scenario. This scenario involves the flight of a drone in a complex industrial environment. The invention's deep learning-based dynamic precision decoupling method aims to solve the problem of traditional drone control methods being unable to efficiently adjust precision in dynamic environments. This industrial inspection task requires the drone to perform automated inspections inside a large factory, demanding that the drone fly precisely through complex building structures, avoid collisions, and maintain stable flight precision. Especially under conditions of strong winds, localized climate changes, and large-scale machine interference, the flight precision requirements are extremely high, and traditional PID control methods struggle to cope with real-time changes in the environment and flight status.

[0158] In this scenario, the task requires a drone to fly from a starting point to various inspection points to complete equipment status checks. Traditional methods struggle to meet high-precision requirements, especially when adjustments to the inspection path, changes in flight status, and frequent environmental disturbances often lead to deviations from the predetermined path, affecting inspection quality and efficiency. Therefore, this invention provides a deep learning-based precision decoupling method that, by extracting flight status and environmental features in real time, can flexibly adjust the flight path and dynamically optimize precision control. First, sensors collect real-time data from the drone, including position, speed, attitude, and environmental data, at a frequency of 50 times per second. During flight, the GPS system updates the drone's position and altitude in real time, the IMU provides accurate attitude data, and meteorological sensors measure wind speed, air pressure, and temperature and humidity. After data preprocessing, Kalman filtering removes external noise and sensor errors to ensure data accuracy. Then, a standardized flight dataset is used as input to a ConvLSTM network for spatiotemporal feature extraction. ConvLSTM extracts spatial features through convolutional layers and captures temporal dependencies during flight through LSTM layers, generating spatiotemporal feature vectors.

[0159] These spatiotemporal feature vectors are fed into the improved GraphSAGE model, where they are decoupled through a dynamic topology construction module and a neighbor information aggregation module. By calculating the Manhattan distance between each flight state node and its neighboring nodes, neighboring nodes relevant to the current node are selected for feature aggregation. Through this process, we achieve decoupling of flight accuracy, eliminating errors caused by complex environmental factors (such as wind speed changes and temperature fluctuations). These decoupled feature vectors provide accurate references for subsequent accuracy error calculations. Based on this, flight accuracy errors are calculated in real time and optimized using a differential evolution algorithm to obtain optimal accuracy control parameters. After optimization, the generated flight control command set includes adjustments for speed, attitude, and position. Throughout the flight, the UAV adjusts its flight state according to these commands, ensuring that the error between the flight path and the predetermined path is controlled within 1 meter, the attitude error within 0.1 degrees, and the speed error within 0.1 meters per second.

[0160] To verify the effectiveness of this invention, a series of comparative experiments were conducted. In the experiments, the method of this invention was compared with that of the UAV using traditional PID control, Kalman filtering, and particle filtering methods. During flight, the UAV's flight path, speed, attitude, and other data were recorded in real time, and flight accuracy errors were calculated. The table below shows the experimental results for different methods:

[0161] Table 1 Performance Comparison of Flight Control Methods

[0162] method Average position error (meters) Mean attitude error (degrees) Average speed error (m / s) Flight time (minutes) Adaptability to dynamic environments PID control method 3.5 1.2 0.25 25 Poor Kalman filtering method 2.0 1.0 0.15 24 A certain degree of adaptability Particle filtering method 1.5 0.8 0.12 23 good Method of the present invention 0.9 0.1 0.08 22 excellent

[0163] As shown in Table 1, the PID control method has significant limitations in handling dynamic environmental changes, especially in high-precision tasks, where position and attitude errors are large and flight times are relatively long. Its adaptability to dynamic environments is poor, and it cannot effectively cope with external disturbances during flight. The Kalman filter method, a classic estimation method, can reduce noise interference and improve flight accuracy, but its flight accuracy remains unsatisfactory in dynamic environments (such as strong winds and air pressure changes), particularly in attitude and speed control, and its flight time is also long. The particle filter method is relatively effective in handling nonlinear systems and can cope well with dynamic environments, but it still has certain errors in high-precision control and requires a long computation time to estimate the flight path. Compared with other traditional methods, this invention, through the combination of deep learning models and graph neural networks, significantly improves flight and speed accuracy, and can adjust the control strategy in real time according to the dynamic environment during flight. It improves flight accuracy while shortening flight time, has a better ability to adapt to dynamic environments, and can provide high-precision flight control in complex environments.

[0164] The method of this invention far surpasses traditional methods in flight accuracy and control capabilities. Particularly in position, attitude, and velocity accuracy, errors are significantly reduced, and flight time is shortened compared to traditional methods. The flight time using this method is 22 minutes, 3 minutes faster than the PID control method. This is mainly due to the fact that this invention optimizes the flight control strategy in real time through deep learning and graph neural networks, reducing accuracy errors and control adjustment delays during flight. These experimental results fully verify the efficiency and accuracy of this invention in complex environments, solving the technical problem of existing control methods' inability to accurately adjust control strategies in dynamic environments. Through the deep learning accuracy decoupling method of this invention, UAVs can maintain high-precision control continuously in ever-changing flight environments, significantly improving the execution efficiency and reliability of flight missions.

[0165] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based method for decoupling the dynamic accuracy of unmanned aerial vehicles (UAVs), characterized in that, Includes the following steps: Step 1: Collect raw flight data from the UAV and preprocess the raw flight data to obtain a standardized flight dataset; Step 2: Input the standardized flight dataset into the ConvLSTM network, extract spatial features through convolution operations, and extract time-dependent features to generate a spatiotemporal feature vector set; Step 3: Input the spatiotemporal feature vector set into the improved GraphSAGE model, and decouple the precision through the dynamic topology relationship construction module, the neighbor information aggregation module, and the feature transformation module to obtain the decoupled feature vector; Step 4: Based on the decoupled feature vector, calculate the flight accuracy error, which is the difference between the actual flight accuracy and the expected flight accuracy, and obtain real-time accuracy error feedback; Step 5: Based on the real-time accuracy error feedback, differential evolution is used for error processing to obtain the optimal accuracy control parameters; Step 6: Generate a flight control command set based on the optimal precision control parameters. The flight control command set includes speed control commands, attitude control commands, and position control commands.

2. The method for decoupling the dynamic accuracy of unmanned aerial vehicles based on deep learning according to claim 1, characterized in that, The raw flight data specifically includes: The latitude, longitude, and altitude information of the drone are obtained through GPS positioning to obtain data for determining the drone's position in space; The pitch, roll, and yaw angles of the UAV are acquired by the inertial measurement unit to obtain attitude data representing the UAV's spatial orientation. The drone's speed data is obtained by acquiring its linear velocity, angular velocity, and linear and rotational velocities in various axes through motion sensors. Environmental data of the UAV is acquired through meteorological sensors, temperature and humidity sensors, barometers and anemometers. This environmental data is used to monitor meteorological conditions and the impact of the external environment on the UAV's flight.

3. The method for decoupling the dynamic accuracy of unmanned aerial vehicles based on deep learning according to claim 1, characterized in that, The preprocessing steps are specifically as follows: Kalman filtering is applied to the raw flight data to remove environmental interference and sensor errors; Perform time alignment and synchronization processing on different types of data in the raw flight data until they correspond on the same time axis; Different types of data in the original flight data are normalized using Z-Score standardization to obtain a standardized flight dataset with a unified structure.

4. The deep learning-based UAV dynamic accuracy decoupling method according to claim 1, characterized in that, Step two specifically involves: The standardized flight dataset is input into the convolutional layer of the ConvLSTM network. By using convolutional kernels of different sizes, spatial structure features are extracted to generate spatial feature maps. The spatial feature map is input into the pooling layer, and the spatial feature map is downsampled through max pooling to reduce the size of the spatial feature map and generate a pooled spatial feature matrix. The spatial feature matrix is ​​input into the LSTM layer. The LSTM layer extracts time-dependent features from the spatial feature matrix by capturing the dynamic changes between different moments in the flight process, and obtains a temporal feature vector. The temporal feature vector represents the time change information in the flight process. The temporal feature vector is input into a fully connected layer, which combines the temporal feature vector with spatial features and performs feature fusion through linear transformation to obtain a spatiotemporal feature vector set, which represents the spatiotemporal information during flight.

5. The deep learning-based UAV dynamic accuracy decoupling method according to claim 1, characterized in that, The improved GraphSAGE model is specifically as follows: The spatiotemporal feature vector set is input into the dynamic topology construction module, and different spatiotemporal feature vectors in the spatiotemporal feature vector set are used as nodes to obtain a node set. The Manhattan distance between the spatiotemporal feature vectors corresponding to every two different nodes in the node set is calculated. If the Manhattan distance is less than a preset distance threshold, an edge connection is established between the corresponding nodes. All node pairs are traversed to obtain the set of connection edges. Based on the connection relationships in the node set and the edge set, a dynamic topology graph of UAV flight information is established; Input the dynamic topology graph into the neighbor information aggregation module, select several neighbor nodes of each node in the dynamic topology graph to form a neighbor node set; Aggregate the neighbor nodes in the neighbor node set with the current node to obtain the multi-source aggregated node feature vector; The feature vectors of the multi-source aggregation nodes are input into the feature transformation module, and the Sigmoid nonlinear transformation function is used to map the feature vectors of the multi-source aggregation nodes to a new feature space. In the new feature space, the feature vector of the current multi-source aggregation node is weighted and averaged with that of its neighboring nodes to obtain the decoupled feature vector after fusion.

6. The deep learning-based UAV dynamic accuracy decoupling method according to claim 5, characterized in that, The process of aggregating the neighbor nodes in the neighbor node set with the current node to obtain the multi-source aggregated node feature vector is as follows: Calculate the cosine similarity between the spatiotemporal feature vector of the current node and the spatiotemporal feature vector of each neighboring node; If the cosine similarity is less than the preset similarity threshold, the corresponding neighbor node is added to the set of neighbor nodes to be aggregated. The spatiotemporal feature vector corresponding to the current node is weighted and added to the spatiotemporal feature vectors corresponding to the neighbor nodes in the set of neighbor nodes to be aggregated, to obtain the multi-source aggregated node feature vector. The weight of the weighting is the reciprocal of the corresponding cosine similarity.

7. The method for decoupling the dynamic accuracy of unmanned aerial vehicles based on deep learning according to claim 1, characterized in that, Step four specifically involves: Based on the decoupled feature vector, the flight accuracy error is calculated, whereby the flight accuracy error is the difference between the actual flight accuracy and the expected flight accuracy. The calculation steps include: The expected flight accuracy is set according to the requirements of the UAV flight mission and the target accuracy. The expected flight accuracy includes position accuracy, attitude accuracy and velocity accuracy, and the expected flight accuracy vector is obtained. The expected flight accuracy vector represents the accuracy standard required by the mission. Based on the current flight status data of the UAV in the decoupled feature vector, the actual flight accuracy is calculated through real-time feedback sensor data, and compensation is performed through sensor accuracy correction to obtain the actual flight accuracy vector, which represents the accuracy of the current flight status of the UAV. The actual flight accuracy vector is compared with the expected flight accuracy vector, the difference between the two is calculated to obtain the flight accuracy error, and real-time accuracy error feedback is generated. The real-time accuracy error feedback is used to represent the accuracy adjustment requirements of flight control.

8. The method for decoupling the dynamic accuracy of unmanned aerial vehicles based on deep learning according to claim 1, characterized in that, The specific steps of differential evolution include: An initial population is generated based on the real-time accuracy error feedback. Each individual in the initial population represents a set of accuracy control parameters. Each individual contains multiple parameter dimensions, including the initial values ​​of position accuracy, attitude accuracy, and velocity accuracy control parameters. The initial population is initialized by random generation to obtain the initial precision control parameter population; Calculate the fitness value of each individual in the initial precision control parameter population; The fitness value is obtained by minimizing the real-time accuracy error based on the accuracy control parameters corresponding to each individual according to a preset fitness function. The fitness value of each individual is calculated, and a fitness value matrix is ​​obtained. Individuals in the initial precision control parameter population are selected based on their fitness values. Individuals with fitness values ​​greater than a preset threshold are selected to generate a new candidate individual population. Mutate each individual in the candidate population to generate new mutated individuals; The mutation operation explores the solution space by adding random perturbations to the parent individuals, resulting in a population of mutated individuals; A single-point crossover operation is performed between the mutant individual population and the candidate individual population to generate a new generation of individuals. Based on the fitness value of the new generation of individuals generated by the crossover, the optimal individual is selected as the new precision control parameter. The control parameters corresponding to the new generation of individuals are the optimal solution under the current error minimization objective, and the optimal precision control parameters are output.

9. The deep learning-based UAV dynamic accuracy decoupling method according to claim 1, characterized in that, Step six specifically involves: Based on the speed accuracy requirements in the optimal accuracy control parameters, the speed control amount required to adjust the flight accuracy error is determined, and a speed control command is generated. The speed control command is used to adjust the linear velocity and angular velocity of the UAV. Based on the attitude accuracy requirements in the optimal accuracy control parameters, the attitude adjustment amount required to adjust the flight accuracy error is determined, and attitude control commands are generated. The attitude control commands are used to adjust the pitch angle, roll angle, and yaw angle of the UAV. Based on the position accuracy requirements in the optimal accuracy control parameters, determine the spatial position adjustment amount required to adjust the flight accuracy error; The speed control command, attitude control command, and position control command are integrated to obtain the flight control command set, which contains precise adjustment information for controlling the UAV at each stage of flight.

Images