Unmanned aerial vehicle lossless operation method based on AI and low-interference close-range interaction trajectory planning

Abstract

The invention discloses an unmanned aerial vehicle non-destructive operation method based on AI and low-interference close-range interaction trajectory planning, and aims to solve the problem that in close-range interaction operation of an unmanned aerial vehicle, aerodynamic interference is caused by rotor wing downwash airflow and ground effect, mapping is inaccurate due to a reflective area or a weak texture area, and the surface of a target is prone to being damaged. According to the method, environment measurement data is collected and state estimation is carried out, an implicit scene model is constructed or updated to generate surface geometric features and a geometric uncertainty graph, local geometric input is extracted and input into a geometric condition neural operator aerodynamic prediction model to obtain an aerodynamic interference prediction value, uncertainty and disturbance sensitivity are predicted, and the aerodynamic interference prediction accuracy is improved. A disturbance control barrier function constraint containing geometric uncertainty and prediction uncertainty is constructed, a nonlinear model is introduced to predict and control a rolling solution control sequence, and the technical effect of quantitative constraint and close-range lossless interaction operation on aerodynamic disturbance intensity is achieved.

Description

Technical Field

[0001] This invention relates to the field of intelligent control of unmanned aerial vehicles (UAVs), and in particular to a non-destructive operation method for UAVs based on AI and low-interference close-range interactive trajectory planning. Background Technology

[0002] Unmanned aerial vehicles (UAVs), with their advantages of high maneuverability and low deployment cost, have been widely used in close-range operation scenarios on target surfaces, such as industrial equipment inspection and mapping, close-range observation of cultural relics and precision components, spraying and cleaning close to structural surfaces, exploration in confined spaces, and interactive operations requiring stable attitude or movement along surface boundaries near the target. To improve the autonomy of such tasks, existing technologies typically employ airborne sensors to acquire environmental information and achieve localization and mapping through methods such as visual inertial odometry and laser mapping, thereby performing local path planning and trajectory tracking control. At the control level, model predictive control has attracted attention because it can explicitly handle dynamic constraints and actuator constraints. Meanwhile, with the development of learning methods, wind field regression, disturbance estimation, and compensation based on neural networks have also been used to improve the stability and tracking accuracy of UAVs in complex airflow environments. In recent years, schemes utilizing implicit scene representations for 3D reconstruction have also emerged, expressing surface geometry through symbolic distance fields or volume density fields, providing more refined environmental models for close-range navigation and obstacle avoidance.

[0003] Existing technologies still have shortcomings in applications for "near-range lossless interactive operations," mainly in the following aspects:

[0004] 1. Lack of constrained quantitative modeling of near-wall aerodynamic disturbances: Many schemes compensate for near-distance aerodynamic effects such as rotor downwash and ground effect as external disturbances, or simply impose a general penalty in the cost function. It is difficult to form clear disturbance indicators and hard constraints for the target surface, thus making it difficult to guarantee the non-destructive requirements.

[0005] 2. Insufficient coupling between aerodynamic prediction and local geometric boundary: Existing learning-based wind field predictions often use UAV status or a small number of environmental parameters as inputs, and often fail to use near-wall geometric boundary conditions as key inputs. This results in insensitivity to the differences in backflow and ground effects caused by geometry such as corners, grooves, slits and curvature abrupt changes, limiting prediction accuracy and generalization ability.

[0006] 3. Uncertainty in perception mapping is not included in safety constraints: Reflective areas and weak texture areas in close-range operations can significantly reduce reconstruction accuracy. Existing planning controls usually do not explicitly quantify mapping uncertainty and introduce it into safety margins or constraint tightening mechanisms. This can easily lead to the system still running close to the target when the geometric error is large, increasing the risk of damage to the target surface.

[0007] Therefore, a non-destructive operation method for drones that can overcome the shortcomings of the existing technology is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0008] One objective of this invention is to propose a non-destructive operation method for unmanned aerial vehicles (UAVs) based on artificial intelligence and low-interference close-range interactive trajectory planning. Addressing the problems of existing technologies in close-range interactive operations, such as the difficulty in quantifying aerodynamic interference caused by rotor downwash and ground effects, insufficient sensitivity of aerodynamic prediction to near-wall geometric boundaries, and the difficulty in incorporating mapping errors into safety margins in reflective or weakly textured areas, thus easily damaging the target surface, the following technical solution is proposed: Obtaining operational task parameters and low-interference thresholds; collecting environmental measurement data and estimating UAV state variables; constructing or updating an implicit scene model and generating surface geometric features and geometric uncertainty maps; extracting local geometric input data corresponding to the interactive operation area; inputting the local geometric input data and UAV state variables into a geometric conditional neural operator aerodynamic prediction model to obtain aerodynamic interference prediction values, prediction uncertainty terms, and disturbance sensitivity; constructing a disturbance control barrier function constraint including geometric uncertainty terms and prediction uncertainty terms; and introducing a nonlinear model predictive control to solve the control sequence in a rolling manner to output control commands. This invention has the technical effect of tightening the confidence margin of the aerodynamic disturbance intensity of the target surface and implementing verifiable hard constraint control under uncertain environments, thereby achieving low-interference and non-destructive operation in close-range interactive work.

[0009] This invention provides a non-destructive operation method for drones based on AI and low-interference short-range interactive trajectory planning, including:

[0010] S1. Obtain task parameters, including interactive task area parameters and low-interference threshold; S2. Based on task parameters, control the UAV sensors to collect environmental measurement data and estimate UAV state variables; S3. Based on environmental measurement data and UAV state variables, construct or update the implicit scene model, generate surface geometric features within the interactive task area, and simultaneously generate a geometric uncertainty map to determine the geometric uncertainty term; S4. Based on the implicit scene model, surface geometric features, and UAV state variables, extract local geometric input data corresponding to the interactive task area; S5. Input the local geometric input data and UAV state variables into a geometrically condition-based neural operator aerodynamic prediction model to obtain aerodynamic interference prediction values ​​and prediction uncertainties. S6. Based on the low disturbance threshold, aerodynamic disturbance prediction value, prediction uncertainty term, and geometric uncertainty term, calculate the disturbance barrier function value. Construct the control barrier function constraint with the disturbance barrier function value not less than zero and introduce nonlinear model predictive control. Linearize the control barrier function constraint with the basic disturbance disturbance sensitivity and solve it in the prediction time domain to obtain the control sequence. Output the first control quantity of the control sequence as the control command. S7. Based on the control command, control the UAV to perform close-range interactive operations in the interactive operation area. Repeat steps S2 to S6 with a preset control cycle to continuously update the implicit scene model, local geometric input data, and the solution results of nonlinear model predictive control.

[0011] Optionally, S1 includes:

[0012] Obtain the target surface position parameters of the target to be operated on, wherein the target surface position parameters are used to characterize the position and orientation of the target surface in a preset coordinate system;

[0013] Obtain the interactive operation area parameters, which are used to characterize the boundary of the permitted flight area and the boundary of the prohibited area surrounding the target surface;

[0014] The low interference threshold is obtained, which is a threshold characterizing the upper limit of the aerodynamic interference intensity that the target surface can withstand.

[0015] The target surface position parameters, the interactive operation area parameters, and the low interference threshold are then output as the operation task parameters.

[0016] Optionally, S2 includes:

[0017] Based on the task parameters, the UAV's onboard sensors are controlled to collect environmental measurement data within the interactive work area. The environmental measurement data includes inertial measurement data and measurement data used to characterize the environmental geometry.

[0018] The environmental measurement data are synchronized in time and transformed in coordinates.

[0019] Based on the synchronized environmental measurement data, the UAV state variables are obtained through multi-sensor fusion state estimation, and the UAV state variables include position and attitude.

[0020] Optionally, S3 includes:

[0021] Based on the environmental measurement data and the UAV state variables, the measurement data used to characterize the environmental geometry is converted into three-dimensional spatial observation data in a preset coordinate system through coordinate transformation; the three-dimensional spatial observation data is fused into an implicit scene model to realize the construction or update of the implicit scene model; the interactive operation area is sampled in the updated implicit scene model, and for each sampling point, the implicit geometric quantities output by the implicit scene model are queried to obtain the implicit geometric quantities, including symbolic distance values ​​and / or volume density values;

[0022] The surface geometric features are generated by calculating the gradient and / or performing numerical difference on the implicit geometric quantities to determine the surface normal vector, calculating the second derivative of the implicit geometric quantities and / or performing numerical approximation based on the rate of change of the normal vector to determine the curvature.

[0023] A geometric uncertainty map is generated based on at least one of the observation count, reprojection error, measurement residual, photometric consistency, and view coverage of the model unit in the implicit scene model. The geometric uncertainty term is then obtained by querying the location corresponding to the UAV state variable based on the geometric uncertainty map.

[0024] Optionally, S4 includes:

[0025] A local extraction region is determined based on the UAV state variables. The local extraction region is located within the interactive operation area and centered on the position corresponding to the UAV state variables, and has a preset spatial range. Multiple sampling points are selected within the local extraction region according to preset sampling rules. For each sampling point, the implicit geometric value of the sampling point is obtained by querying based on the implicit scene model. The implicit geometric value includes a symbolic distance value and / or a volume density value. The surface normal vector and curvature corresponding to the sampling point are obtained based on the surface geometric features. The spatial positions, implicit geometric values, surface normal vectors, and curvatures of the multiple sampling points are combined according to a preset data structure to form the local geometric input data.

[0026] Optionally, S5 includes:

[0027] The local geometric input data is normalized and converted into geometric condition input for the neural operator aerodynamic prediction model. Simultaneously, the UAV state variables are normalized and converted into state input for the neural operator aerodynamic prediction model. The geometric condition input and the state input are then input into the geometric condition-based neural operator aerodynamic prediction model to obtain aerodynamic interference prediction values ​​characterizing the aerodynamic interference intensity caused by the rotor downwash and ground effect on the target surface. The aerodynamic interference intensity is at least one of the following indicators at the target surface: wind speed, dynamic pressure / wind pressure, shear stress, pulsation intensity, vorticity, force, and torque.

[0028] Furthermore, the aerodynamic interference intensity is at least one of the maximum value, peak value, root mean square value, average value, and integral value of the index on the target surface; based on at least one of the variance output of the geometrically based neural operator aerodynamic prediction model for the same input, the dispersion of the integrated inference output, and the dispersion of the Monte Carlo random inactivation inference output, the prediction uncertainty term is determined; based on the differentiable calculation of the UAV state variables by the geometrically based neural operator aerodynamic prediction model, the partial derivative of the aerodynamic interference prediction value with respect to the UAV state variables is obtained, and the disturbance sensitivity is obtained.

[0029] Optionally, S6 includes:

[0030] Based on the low disturbance threshold, the predicted aerodynamic disturbance value, the prediction uncertainty term, and the geometric uncertainty term, a disturbance barrier function value is calculated. The disturbance barrier function value is equal to the low disturbance threshold minus the sum of the predicted aerodynamic disturbance value, the geometric uncertainty term, and the prediction uncertainty term.

[0031] The disturbance barrier function value is not less than zero to construct a control barrier function constraint, and the control barrier function constraint is linearized to the first order based on the disturbance sensitivity to obtain a linearized constraint.

[0032] The linearization constraints, along with the UAV dynamics constraints and actuator constraints, are input into the nonlinear model predictive control. The control sequence is then solved in the prediction time domain, so that the cost function includes a trajectory tracking error term and an aerodynamic disturbance cost term.

[0033] The first control quantity of the control sequence is output as the control instruction.

[0034] Optionally, the S7 includes:

[0035] Based on the control commands, the drone is controlled to fly within the interactive work area and perform the close-range interactive operation;

[0036] During the execution of the close-range interaction operation, steps S2 to S6 are repeated with a preset control cycle to reacquire the environmental measurement data and update the UAV state variables, thereby updating the implicit scene model, the surface geometric features and the geometric uncertainty term, updating the local geometric input data, recalculating the aerodynamic disturbance prediction value, the prediction uncertainty term and the disturbance sensitivity, and resolving the nonlinear model predictive control to output the updated control command.

[0037] Optionally, the control barrier function constraint includes at least one of the following: a control barrier function constraint based on continuous-time form:

[0038] ;

[0039] in, For extended class K functions;

[0040] and / or control barrier function constraints based on discrete-time form:

[0041] ;

[0042] in, These are preset parameters. and These represent the system states for adjacent control cycles.

[0043] Optionally, within the preset control period, when the geometric uncertainty term and / or the prediction uncertainty term are greater than the preset uncertainty threshold, the UAV is controlled to execute an active perception maneuvering strategy to increase the spectral coverage of reflective areas and / or weak texture areas and reduce the uncertainty in the geometric uncertainty map.

[0044] The beneficial effects of this invention are:

[0045] 1. Achieve quantifiable hard constraint control of aerodynamic disturbances for non-destructive targets: By constructing disturbance control barrier function constraints using the aerodynamic disturbance prediction values ​​output by neural operators and combined with low disturbance thresholds, the disturbance intensity of rotor downwash airflow and ground effect on the target surface is transformed into calculable and verifiable constraints, which are then solved in a rolling nonlinear model predictive control, thereby reducing the risk of damage to the target surface during close-range interactive operations.

[0046] 2. Improve the accuracy and adaptability of disturbance prediction in complex near-wall geometric scenarios: By extracting local geometric conditions such as symbolic distance or volume density, surface normal and curvature generated by the implicit scene model as input to the neural operator aerodynamic prediction model, the prediction model becomes more sensitive to the differences in backflow and ground effects caused by geometric boundaries such as corners, grooves, slits and curvature abrupt changes, thereby improving the accuracy and generalization ability of aerodynamic disturbance prediction in complex near-wall environments.

[0047] 3. Incorporating mapping and prediction uncertainties into the safety margin to enhance robustness: By generating a geometric uncertainty map and determining the geometric uncertainty term, and simultaneously calculating the prediction uncertainty term, the above uncertainties are introduced into the disturbance barrier function to achieve threshold tightening. Even under unreliable perception conditions such as reflective areas or weak texture areas, conservative and safe trajectory planning and control can still be maintained, thereby improving the stability and safety of the system in uncertain environments. Attached Figure Description

[0048] 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:

[0049] Figure 1 The flowchart shows a method for non-destructive operation of a drone based on AI and low-interference close-range interactive trajectory planning proposed in this invention.

[0050] Figure 2 This is a flowchart illustrating step S6 of the present invention. Detailed Implementation

[0051] 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.

[0052] refer to Figure 1 A non-destructive operation method for drones based on AI and low-interference short-range interactive trajectory planning includes:

[0053] S1. Obtain task parameters, including interactive task area parameters and low-interference threshold; S2. Based on task parameters, control the UAV sensors to collect environmental measurement data and estimate UAV state variables; S3. Based on environmental measurement data and UAV state variables, construct or update the implicit scene model, generate surface geometric features within the interactive task area, and simultaneously generate a geometric uncertainty map to determine the geometric uncertainty term; S4. Based on the implicit scene model, surface geometric features, and UAV state variables, extract local geometric input data corresponding to the interactive task area; S5. Input the local geometric input data and UAV state variables into a geometrically condition-based neural operator aerodynamic prediction model to obtain aerodynamic interference prediction values ​​and prediction uncertainties. S6. Based on the low disturbance threshold, aerodynamic disturbance prediction value, prediction uncertainty term, and geometric uncertainty term, calculate the disturbance barrier function value. Construct the control barrier function constraint with the disturbance barrier function value not less than zero and introduce nonlinear model predictive control. Linearize the control barrier function constraint with the basic disturbance disturbance sensitivity and solve it in the prediction time domain to obtain the control sequence. Output the first control quantity of the control sequence as the control command. S7. Based on the control command, control the UAV to perform close-range interactive operations in the interactive operation area. Repeat steps S2 to S6 with a preset control cycle to continuously update the implicit scene model, local geometric input data, and the solution results of nonlinear model predictive control.

[0054] In this specific embodiment, S1 includes:

[0055] The ground mission configuration terminal completes and generates operational mission parameters for use by both the flight control and onboard computing units, with the world coordinate system selected as the preset coordinate system. Its origin is located at the drone's takeoff point. The axis points due east. The axis points due north. The axis is vertically upward;

[0056] The target surface position parameters are obtained by reading the pre-calibrated target surface pose, which is stored in the task configuration file, through the task configuration terminal. The pose is in the target surface coordinate system. Compared to Rigid body transformation representation, where the target surface coordinate system The origin is defined as the geometric center point of the target surface. The axis is defined as the direction of the outward normal to the target surface. The axis is defined as the tangential direction of a preset reference edge on the target surface. The axis is determined by the right-hand rule;

[0057] The target surface position parameters include a position vector. With attitude quaternions ,in They are respectively The origin is Middle The coordinate components of the direction and the unit is They are respectively Compared to The real part and three imaginary parts of the unit quaternion satisfy the unit length constraint and are used to uniquely characterize the orientation of the target surface.

[0058] The interactive operation area parameters are used to simultaneously specify the boundaries of the permitted flight area and the prohibited area. The mission configuration terminal reads these parameters from the mission configuration file and expresses them in a solidified form using a half-space inequality of a convex polyhedron. The mathematical expression of this expression is as follows:

[0059] ;

[0060] in For permitted flight areas, This is a restricted area. In order to be in The following represents the coordinates of a point in space and Units are For three-dimensional real space, and These are the surface normal constraint matrix and constraint bias vector corresponding to the permitted flight area, respectively. To determine the number of permitted flight zone boundary surfaces, and These are the surface normal constraint matrix and constraint bias vector corresponding to the restricted area, respectively. The symbol represents the number of boundary surfaces of the restricted area. To express element-wise inequalities to limit The half-space constraints of each boundary surface must be satisfied;

[0061] The The task configuration terminal configures the target surface based on the 3D model. The known poses in the data are constructed and solidified offline. The construction process is to discretize the boundary surface of the allowed flight area and the restricted area into several planar polygonal patches, take the outward normal unit vector for each patch and write the outward normal into the corresponding row of the matrix, and write the threshold of the inner product between any point on the patch and the outward normal into the corresponding element of the vector, so as to ensure that the intersection of the half space forms a closed convex polyhedron boundary.

[0062] In this embodiment, the low interference threshold uses the dynamic pressure at the target surface as an index of aerodynamic interference intensity and is expressed as a scalar. The terminal reads the task configuration file to represent the task configuration. And set as Where Pa is Pascal's upper limit for characterizing pressure intensity per unit area, so that subsequent steps can directly use it when constructing the disturbance control barrier function. Hard constraints are applied to the predicted aerodynamic disturbance values;

[0063] The task configuration terminal will set the target surface position parameters and Interactive work area parameters and corresponding and low interference threshold The parameters are encapsulated as job task parameters and written into the task cache area of ​​the airborne computing unit for unified use by the sensor acquisition strategy and coordinate transformation in step S2, as well as the constraint construction and rolling optimization in subsequent steps.

[0064] In this specific embodiment, S2 includes:

[0065] On the airborne computing unit with a fixed control cycle Execution and output: The airborne sensors include an inertial measurement unit (IMU) and a lidar for characterizing the geometry of the environment. The IMU's mounting coordinate system is defined as... Furthermore, it overlaps with the UAV system, and the lidar installation coordinate system is defined as follows: ;

[0066] The inertial measurement unit outputs angular velocity and specific force measurement data at 200 Hz, denoted as follows: and And the timestamp is The lidar outputs point cloud measurement data at 10 Hz, which is recorded as... And the timestamp is ,in In order to be in The following is a three-dimensional angular velocity measurement vector expressed in units of . In order to be in The following is a three-dimensional force measurement vector expressed in units of . Let be a set of points consisting of several three-dimensional points, and each point is in The following expressions are in units of m;

[0067] Time synchronization uses the monotonically increasing clock of the onboard computing unit as a unified time reference and records the hardware arrival time at the sensor driver layer to form a system timestamp. Fixed delay compensation obtained from offline calibration is applied to both inertial measurement data and lidar point cloud data. and To correct its timestamps so that the inertial measurement sequence used for fusion within each control cycle corresponds to the same system time as the point cloud measurement;

[0068] The coordinate transformation uses an extrinsic rigid body transformation obtained from offline calibration to align the lidar coordinate system with the inertial coordinate system. The extrinsic parameters are... arrive The rigid body transformation is represented and solidified as a translation vector. With unit quaternion ,in The unit is m and they represent respectively The origin is lower edge directional components, It is an attitude quaternion component and satisfies the unit length constraint;

[0069] Multi-sensor fusion state estimation employs an error-state extended Kalman filter and uses a world coordinate system. As a navigation coordinate system, the nominal state of the filter is defined as follows:

[0070] ;

[0071] in For a moment The nominal status of the drone For drones The position vector below and Units are For drones The velocity vector below and Units are , for Compared to Unit quaternions are used to characterize the attitude of drones. The zero bias vector of the inertial measurement unit accelerometer is given in units of . The zero bias vector of the inertial measurement unit gyroscope is given in units of 1. ;

[0072] The filter's state propagation performs debiasing and discrete integration on the inertial measurement sequence within each control cycle. The debiased angular velocity is... And the ratio after removing the bias is and in the gravity vector Update under effect and At the same time and Propagation follows a random walk model, and the standard deviation of the process noise is taken as accelerometer white noise. Gyroscope white noise Accelerometer zero-bias random walk Zero bias random walk of gyroscope Based on this, the noise covariance matrix of the discrete process is constructed;

[0073] The filter measurement update is performed when each frame of point cloud arrives; the onboard computing unit updates the current point cloud. Through external parameters Transform to And based on the current nominal state Transform the initial value of the point cloud to Subsequently, point-to-plane matching is performed in the sliding local subgraph to construct residuals and update them. and The sliding local submap is composed of the point clouds from the most recent 5 frames. The point cloud is accumulated and voxel downsampled before entering the sub-image. The voxel side length is set to 0.05 m to fix the point cloud density and stably match the residual statistical characteristics.

[0074] The airborne computing unit at the end of each control cycle... The system reads and outputs UAV state parameters, which include... The lower position with posture And write it to the shared cache for use in step S3.

[0075] In this specific embodiment, S3 includes:

[0076] By the onboard computing unit in each control cycle Internal execution constructs or updates the implicit scene model and outputs surface geometry features and geometric uncertainty terms. The implicit scene model uses a truncated symbolic distance field based on a voxel hash table and is denoted as... ,in World coordinate system The three-dimensional spatial position below and The unit is For position The symbolic distance value at the location, in meters, and satisfying... Indicates being in the space outside the obstacle and This indicates the space inside the obstacle, with the voxel side length set to... And the cutoff distance is set to To limit the impact of a single frame observation on the range field update;

[0077] The airborne computing unit reads the lidar point cloud after synchronization in step S2. With UAV state variables And complete the construction of three-dimensional spatial observation data, including For a moment The point cloud set and each point in the lidar coordinate system The following expression, For drones The position vector below and the unit is The attitude unit quaternion for UAVs is used to characterize the inertial coordinate system. Compared to The rotation;

[0078] Three-dimensional spatial observation data through Each point in the data first utilizes extrinsic parameters. from Transform to Reuse Transform to Obtain and record as ,in for arrive The translation vector is in units of m. for Compared to The unit quaternion;

[0079] The construction or updating of the implicit scene model is accomplished using a point-by-point ray fusion method, with the airborne computing unit utilizing lidar in... The instantaneous position below is used as the starting point of the ray and is used for At each measurement point in the ray, a voxel traversal is performed, at the center of each traversed voxel. The truncation sign distance from the voxel to the measurement surface is calculated and updated using a weighted recursive method. Meanwhile, the number of model unit observations is maintained in each voxel. Root mean square of measurement residuals To support the generation of geometric uncertainty diagrams, where voxel center The cumulative number of times the ray has been updated, and this number is a non-negative integer. The root mean square of the matching residual from the points participating in the fusion within the voxel to the local plane is given in m units, and the local plane is obtained by least-squares fitting of the fused points in the neighborhood of the voxel.

[0080] The interactive operation area is the permitted flight area given in step S1. With restricted areas Determine and record as The airborne computing unit in Inner space step size Perform rule-based sampling to obtain a set of sampling points and then perform rule-based sampling on each sampling point. Query The query uses the centers of 8 adjacent voxels in the voxel hash table. Trilinear interpolation is performed to obtain continuous field values;

[0081] Surface geometric features satisfy Generated at the sampling point and To determine whether a sampling point is in the surface neighborhood, the airborne computing unit uses a differential step size. right exist The gradient is obtained by central difference of the direction and normalized to obtain the surface normal vector. At the same time, the curvature is obtained by numerically approximating the spatial rate of change of the normal vector field in the same surface neighborhood. The normal vector and curvature are written into the cache as the surface geometric features.

[0082] Geometric uncertainty diagram with scalar field Form and They share the same voxel hash index and are written to or flushed after each fusion update, where For position The geometric uncertainty at a given point is equivalent to a tightening factor in Pa, and is used for threshold tightening of the subsequent disturbance control barrier function. Determined by the number of observations and the measurement residuals, and satisfying the following:

[0083] ;

[0084] in To map the geometric error length dimension to a proportionality coefficient for the equivalent tightening amount of aerodynamic disturbance. A reference geometric error scale consistent with the voxel side length. For position The cumulative number of observations for the given voxel, plus +1 to avoid division by zero. To measure the dimensionless weighting coefficients of the residual term, For position The root mean square of the measurement residuals of the voxel is given in m.

[0085] The geometric uncertainty term is determined by the current position of the drone. Geometric uncertainty diagram The result obtained by performing trilinear interpolation is denoted as It is output along with the surface geometry features.

[0086] In this specific embodiment, S4 includes:

[0087] Step S4 is performed by the onboard computing unit in each control cycle. Internal execution generates local geometric input data corresponding to the interactive operation area, and the onboard computing unit reads the UAV state variables. and And read the implicit scene model and surface geometric feature field, the World coordinate system The position vector of the UAV is given in m. The unit quaternion is used to characterize the attitude of the UAV. World coordinate system Next position The distance value of the truncation symbol at the location, in meters;

[0088] Airborne computing unit with Centered on The following constructs a local extraction region, where the local extraction region is the side length. Align the cube with its axis and Uniform sampling along three axes Each sampling location is thus obtained There are 1 sampling point, and the sampling step size is set to 1. and according to The lexicographical order of the sampling points is used to ensure that the features of the local geometric input data are arranged consistently in the time dimension;

[0089] To ensure that the local extraction region is located within the interactive operation area, the airborne computing unit performs constraint pruning on each original sampling point. The constraint pruning is based on the allowable flight area given in step S1. With restricted areas Proceed, where the conditions are not met. The sampling points are projected onto each surface according to the half-space constraint. Points on the boundary and projected are still in units of m. The following expression, for satisfying The sampling points are translated along the out-of-bounds normal direction of the boundary surface with the minimum constraint margin in the restricted area boundary, and the distance is fixed. To ensure that the sampling points do not fall into ;

[0090] After obtaining the cropped sampling points Then, the airborne computing unit processes each Perform implicit and geometric feature queries, where trilinear interpolation is used to retrieve implicit geometric values ​​from the voxel hash table. And generated from step S3 and with Surface normal vector field stored in a common index With curvature field Obtained by trilinear interpolation and ,in It is a unit normal vector and For dimensionless components, Scalar curvature and unit is ;

[0091] The airborne computing unit assembles the spatial location, implicit geometric values, surface normals, and curvature of each sampling point into the local geometric input data in a fixed field order. Record:

[0092] ;

[0093] in Sampling points exist The coordinate components below and the unit is for The numerical value is in meters. They are respectively The three dimensionless components, for The numerical value and the unit is and all Records are stacked in numerical order to form local geometric input data. .

[0094] In this specific embodiment, S5 includes:

[0095] Step S5 is performed by the onboard computing unit in each control cycle. The internal execution outputs aerodynamic disturbance predictions, prediction uncertainty terms, and disturbance sensitivity; the airborne computing unit reads local geometric input data from the shared buffer. and the position in the drone state variables with posture And read the total thrust command from the flight controller within the same control cycle. ,in To be according to The tensor formed by regular sampling stacking and each sampling point contains Fields, World coordinate system The position vector of the UAV is given in units of A unit quaternion that is dimensionless and characterizes the attitude of a UAV. This is the total thrust command for the quadcopter, expressed in Newtons (N), and output by the flight control distribution module.

[0096] Airborne computing unit Perform geometric condition input preprocessing to form geometric condition input Preprocessing includes converting the spatial location components of each sampling point into... The relative coordinates of the origin are used to extract the side length of the local region. Scale normalization is performed so that the components of the relative coordinates fall within the range of Interval, and the sign distance value According to the cutoff distance Normalize and truncate to Interval, curvature according to Perform amplitude truncation and normalization Interval, while preserving the normal vector components Dimensionless input is used to preserve surface orientation information;

[0097] The onboard computing unit generates state inputs for the UAV's state variables. And perform normalization processing, where for The three-dimensional vector normalized to a 1 m position scale. for Quaternion components and maintain unit length constraints to avoid attitude input drift. for Based on the maximum total thrust of the drone Normalized scalar;

[0098] Based on the aforementioned geometric and state inputs, the airborne computing unit invokes a geometrically-based neural operator aerodynamic prediction model for inference. This neural operator aerodynamic prediction model is denoted as... And the parameter set is Its network structure is a three-dimensional Fourier neural operator containing an input upscaling linear layer, four Fourier spectral convolutional layers, and an output projection layer. The input upscaling linear layer maps the 8-dimensional geometric features of each grid point to a feature representation with a channel width of 64. The Fourier spectral convolutional layers have 16 frequency domain modes, and each layer uses gated linear units as activations to maintain differentiability. The state input... The two fully connected embedding network is mapped to a 64-dimensional conditional vector and injected into the feature channel of each Fourier spectral convolutional layer in a channel-by-channel additive modulation manner to achieve geometric conditionalization. The output projection layer performs global max pooling on the spatial features and outputs the scalar aerodynamic disturbance intensity through a two-layer fully connected regression head.

[0099] The model was trained and solidified offline using a dataset generated by a computational fluid dynamics solver. Each sample in the dataset contains With corresponding tags ,in The maximum value of the dynamic pressure at the target surface, expressed in Pa, is obtained by the solver after outputting the dynamic pressure field on the target surface mesh and taking the maximum value. This allows the inference stage to directly output the value with a low interference threshold. Intensity of aerodynamic interference of the same dimension;

[0100] The model output during the inference phase is given by the following formula:

[0101] ;

[0102] in The value represents the predicted aerodynamic disturbance and is expressed as the predicted result of the maximum dynamic pressure at the target surface, expressed in Pa. The prediction uncertainty term is used to characterize the uncertainty. The confidence margin tightened. Let be the perturbation sensitivity vector and its components be right and The partial derivatives constitute, For a neural operator aerodynamic prediction model with fixed parameters, For the set of model parameters, Input the normalized geometric condition tensor. The normalized state input vector, For system time under a unified time base;

[0103] Prediction uncertainty term Based on Monte Carlo random deactivation inference, the onboard computing unit maintains a random deactivation probability during inference. And execute on the same input The forward calculation yielded 16 predicted aerodynamic disturbance values, and their sample mean was taken as... And take twice its sample standard deviation as To form a conservative uncertainty measure consistent with subsequent threshold tightening;

[0104] Perturbation sensitivity vector After turning off random inactivation and using When the deterministic forward computation graph is used as a basis, the result is calculated in one go by the automatic differentiation module and written to the shared buffer.

[0105] In this specific embodiment, S6 includes:

[0106] Step S6 is performed by the onboard computing unit in each control cycle. Internal execution constructs control barrier function constraints and performs rolling solutions for the control sequence of nonlinear model predictive control. The airborne computing unit reads low-disturbance thresholds from the shared buffer. Predicted values ​​of aerodynamic disturbances Prediction uncertainty term Geometric uncertainty term and disturbance sensitivity And calculate the disturbance barrier function value. Its definition is:

[0107] ;

[0108] in To provide system time under a unified time base, For a moment The disturbance barrier function value is given in Pa. Low interference threshold and unit is The predicted values ​​are aerodynamic disturbances expressed in Pa, and represent the predicted maximum dynamic pressure at the target surface. The prediction uncertainty term is expressed in units of 1 / 2. Geometric uncertainty diagram at the current position of the UAV The geometric uncertainty term obtained from the query is in units of For UAVs in the world coordinate system The position vector below, with units of m;

[0109] The onboard computing unit will constrain The constraint is constructed as a control barrier function and tightened in the prediction time domain to maintain consistency, i.e., for the prediction step... Predicted state and The query returned the corresponding And compare it with the same prediction step This is written as a known tightening amount on the right-hand side of the constraint, so that the control solution is always executed against the hard upper limit of the dynamic pressure index at the target surface.

[0110] The system state of nonlinear model predictive control is defined as follows:

[0111] ;

[0112] in For the UAV output by the error state extended Kalman filter in step S2, Downward velocity vector and unit is The control quantity is defined as ,in The total thrust command is in N and is used in step S5. Maintain consistency The system angular velocity command vector and its unit is . ;

[0113] The prediction model employs discrete-time quadrotor point mass dynamics and quaternion attitude analysis, and uses the fourth-order Runge-Kutta method with step size... Discretize the model and fix the model parameters as mass. With gravitational acceleration Furthermore, quaternions are normalized after each discrete propagation to suppress numerical drift;

[0114] The constraint of the actuator is fixed as and And meet the maximum thrust capability of the quadcopter. and Used to limit the rate of attitude change;

[0115] The cost function includes a trajectory tracking error term and an aerodynamic disturbance cost term, wherein the trajectory tracking error term is provided by the position reference sequence in the prediction time domain by the upper-level task module. With attitude reference sequence And use weights and Weighting is applied to the aerodynamic disturbance cost term in the prediction time domain. Apply secondary penalties and use weights A control strategy is employed to suppress approaches to a threshold, while weighting the control increment. To balance traceability with the smoothness of the actuator;

[0116] The sensitivity of the airborne computing unit to disturbances during each solution process According to its opposite and The partial derivative components are divided into blocks and propagated along the hot-start trajectory of the previous control cycle in the prediction time domain to obtain... Then about and First-order approximation is written into the constraint to thus Transform into decision variables The affine inequality constraints, together with the dynamic constraints and the actuator constraints, constitute a sequential quadratic programming subproblem.

[0117] The prediction time domain length is set to And the coverage duration is The solver employs sequential quadratic programming and performs a fixed three outer iterations. Each outer iteration's quadratic programming subproblem is solved using the interior-point method, and a shift-based hot-start of the optimal solution from the previous control cycle is enabled to ensure... The onboard computing unit completes the solution internally and then calculates the control sequence. The first control quantity in As a control command output.

[0118] In this specific embodiment, S7 includes:

[0119] The airborne computing unit at time The first control quantity As control commands, they are sent to the flight controller, where For the first The start time of each control cycle and It is a non-negative integer. Total thrust command and unit is The system angular velocity command vector and its unit is . Flight control will and The attitude inner loop and thrust distribution module are input to generate speed commands for each motor and drive the UAV to fly within the interactive operation area. At the same time, the airborne operation load control module performs close-range interactive operations based on the fixed operation mode parameters in the mission configuration file and runs in parallel with the flight control. The operation mode parameters include load start and stop conditions, duration and intensity settings, and use an independent communication channel with the flight control commands to avoid mutual overwriting.

[0120] exist to During the control cycle interval, the airborne computing unit continuously receives inertial measurement data and lidar point cloud data and writes them into a timestamped circular buffer. At the next control cycle time... The onboard computing unit then re-executes steps S2 to S6 using the synchronization data cached in the most recent control cycle as input for environmental measurement data. In step S2, the UAV state variables are updated through multi-sensor fusion state estimation. and And write it to the shared buffer, and in step S3 use the updated... Point cloud observations for implicit scene models The fusion update is completed, and the surface geometric feature field and geometric uncertainty map are updated synchronously to obtain a new geometric uncertainty term. In step S4 Reconstruct the local extracted region centered on the local geometric input data and resample it. In step S5, New aerodynamic disturbance prediction values ​​are obtained by inputting the UAV state variables into the neural operator aerodynamic prediction model. Prediction uncertainty term With disturbance sensitivity And write it to the shared buffer, based on step S6 and Update the disturbance barrier function values ​​and re-roll the solution of the nonlinear model predictive control to output a new value. This enables the drone to continuously perform close-range interactive operations throughout the entire process. It periodically updates the implicit scene model, local geometric input data, and nonlinear model predictive control solution results, and outputs the updated control commands.

[0121] In this specific embodiment, the control barrier function constraint is in discrete-time form and embedded in the nonlinear model predictive control solution process in step S6. The airborne computing unit writes the disturbance barrier function value as a function of the discrete system state. ,in For the first Each control cycle corresponds to a state The disturbance barrier function value below and the unit is For the first The system state vector for each control cycle, consistent with step S6, is defined as follows:

[0122] ;

[0123] in For the first The start time of each control cycle, and the unit is... World coordinate system The position vector of the UAV is given in units of World coordinate system The velocity vector of the UAV is given in units of A unit quaternion that characterizes the attitude of a UAV and is dimensionless;

[0124] The airborne computing unit evaluates the state pairs of adjacent prediction steps in the prediction time domain of nonlinear model predictive control. Apply discrete-time control barrier function constraints:

[0125] ;

[0126] in The discrete dynamics model used in step S6 during the control period Internally Controlled quantity The system state vector obtained in the next control cycle is obtained through propagation. For preset parameters and take and satisfy With limitation Minimum convergence rate in discrete time, thus guaranteeing that when The system state remains unchanged as the control cycle progresses. Within the safe set;

[0127] The onboard computing unit applies the above constraints to the prediction step during each rolling solution. arrive Gradually applied, among which The prediction time domain length is set for step S6 and taken as follows: and using the perturbation sensitivity in step S6 The obtained first-order linearization result will Regarding decision variables The constraints are transformed into affine inequalities to be solved together with dynamic constraints and actuator constraints, so that the discrete-time control barrier function constraints can be directly verified by the solver in each control cycle and acted on the control sequence output in the form of hard constraints.

[0128] In this specific embodiment, when the rolling control is executed in step S7, the onboard calculation unit at the beginning of each control cycle... A threshold determination is performed on the geometric uncertainty term and the predicted uncertainty term, and a maneuvering strategy that switches to active sensing upon triggering is adopted to reduce the uncertainty in the geometric uncertainty diagram. The threshold determination is based on the uncertainty trigger amount. Completed and satisfied:

[0129] ;

[0130] in For the first The start time of each control cycle and It is a non-negative integer. For dimensionless uncertainty triggering quantity, To find the maximum value operator, Geometric uncertainty diagram at the current position of the UAV The geometric uncertainty term obtained from the query is in units of World coordinate system The position vector of the UAV is given in units of The preset geometric uncertainty threshold is given in Pa and set to [value]. The prediction uncertainty term is expressed in units of 1 / 2. The preset prediction uncertainty threshold is set in Pa and is set to... ;

[0131] when When the onboard computing unit enters the active perception mode and freezes the interactive operation payload output to keep it at zero, the disturbance control barrier function constraint and the allowable flight area constraint remain effective to ensure that the active perception maneuver process still meets the low interference and area boundary requirements.

[0132] The airborne computing unit, within the same control cycle after entering active sensing mode, uses the geometric uncertainty diagram to... The voxel center with the largest geometric uncertainty is obtained by searching within a local cube centered at a point with a side length of 0.6 m, and is used as the observation center for high uncertainty. and exist The following expression is in m units, and at the same time... Implicit Scenario Model for Querying The gradient is calculated and normalized to obtain the corresponding surface normal vector. and It is a dimensionless unit vector;

[0133] The airborne computing unit will subsequently The nonlinear model predictive control reference trajectory is switched to orbit around A fixed-radius orbital observation trajectory and setting the orbital radius. With the orbital linear velocity The initial reference position of the reference trajectory is defined as Ensure that the drone first follows The orientation is adjusted to distance from the high uncertainty region and obtain a larger field of view coverage. During the orbiting phase, the onboard computing unit sets the attitude reference so that the line of sight of the onboard geometry sensor always points to... This creates a continuously changing observation direction in the azimuth angle and improves the angular coverage. In this embodiment, angular coverage is defined as the number of sectors covered by observation rays from different azimuth sectors within a unit time window for the same voxel, and the azimuth sectors are arranged according to... The voxel is divided into 12 sectors, and a sector hit count is maintained for each voxel during fusion in step S3 for updating the geometric uncertainty map.

[0134] The onboard computing unit still operates in active sensing mode according to the control cycle. Repeat steps S2 to S6 to continuously update the UAV state variables, implicit scene model, surface geometric features, and geometric uncertainty map, and recalculate. and When continuous Each control cycle satisfies When the onboard computing unit exits the active sensing mode and resumes the upper-level interactive operation trajectory reference and load output, it increases the view coverage and reduces the uncertainty in the geometric uncertainty map by using a determined orbital observation maneuver in reflective and weak texture areas where geometric reconstruction is unreliable.

[0135] 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.

[0136] This invention employs a closed-loop combination of "implicit scene modeling, aerodynamic interference prediction, and constrained trajectory optimization control" to computably and constrain aerodynamic interferences such as rotor downwash and ground effects during close-range operations. First, an implicit scene model is constructed or updated based on environmental measurement data and state variables to obtain the surface geometric features within the interactive operation area. Then, local geometric inputs and UAV state variables are input into the aerodynamic prediction model to obtain predicted values ​​of aerodynamic interference intensity at the target surface. Finally, the low-interference threshold and the prediction results are combined to construct a control barrier function constraint and introduced into a nonlinear model predictive control to solve the control sequence. This allows the UAV to continuously satisfy low-interference constraints and continuously update them while meeting dynamic and actuator constraints, thereby achieving low-interference close-range interactive operation oriented towards the target surface, reducing the risk of damage to the target surface, and improving operational stability.

[0137] This invention addresses the aforementioned technical problems by making targeted improvements to the algorithm structure: First, it further generates a geometric uncertainty map and forms a geometric uncertainty term in the implicit scene model to characterize mapping errors caused by reflective or weakly textured areas. This term, along with the uncertainty term of aerodynamic prediction, is introduced into the control barrier function to tighten the confidence margin for low-interference thresholds, thus maintaining a conservative and safe approach even when perception is unreliable. Second, it employs a geometrically conditional neural operator aerodynamic prediction structure, using local boundary geometry such as symbolic distance or volume density values, surface normals, and curvature as conditional inputs. This makes the model more sensitive to differences in backflow and ground effects caused by corners, grooves, slits, and curvature abrupt changes, improving prediction accuracy and generalization ability. Third, it outputs the sensitivity of aerodynamic disturbances to state variables, which is used to linearize the control barrier function constraints to the first order and accelerate the model's predictive control solution. This is more conducive to achieving real-time rolling optimization within a preset control cycle and obtaining better lossless operation results.

Claims

1. A non-destructive operation method for drones based on AI and low-interference short-range interactive trajectory planning, comprising: S1. Obtain job task parameters, including interactive job area parameters and low interference threshold; S2. Based on the task parameters, control the UAV sensors to collect environmental measurement data and estimate the UAV state variables; S3. Based on the environmental measurement data and the UAV state variables, construct or update the implicit scene model, generate the surface geometric features within the interactive operation area, and generate a geometric uncertainty map to determine the geometric uncertainty term. S4. Based on the implicit scene model, surface geometric features, and UAV state variables, extract the local geometric input data corresponding to the interactive operation area; S5. Input the local geometric input data and UAV state variables into the neural operator aerodynamic prediction model based on geometric conditions to obtain the predicted aerodynamic disturbance value, prediction uncertainty term, and disturbance sensitivity; S6. Based on the low disturbance threshold, predicted aerodynamic disturbance value, prediction uncertainty term, and geometric uncertainty term, calculate the disturbance barrier function value, construct the control barrier function constraint with the disturbance barrier function value not less than zero, and introduce it into the nonlinear model predictive control. Linearize the control barrier function constraint based on the disturbance sensitivity and solve it in the prediction time domain to obtain the control sequence. Output the first control quantity of the control sequence as the control command; S7. Based on control commands, control the UAV to perform close-range interactive operations within the interactive operation area, and repeat steps S2 to S6 with a preset control cycle to continuously update the implicit scene model, local geometric input data, and the solution results of nonlinear model predictive control.

2. The method for non-destructive operation of a drone based on AI and low-interference close-range interactive trajectory planning according to claim 1, S1 includes: Obtain the target surface position parameters of the target to be operated on, wherein the target surface position parameters are used to characterize the position and orientation of the target surface in a preset coordinate system; Obtain the interactive operation area parameters, which are used to characterize the boundary of the permitted flight area and the boundary of the prohibited area surrounding the target surface; The low interference threshold is obtained, which is a threshold characterizing the upper limit of the aerodynamic interference intensity that the target surface can withstand. The target surface position parameters, the interactive operation area parameters, and the low interference threshold are then output as the operation task parameters.

3. The method for non-destructive operation of a drone based on AI and low-interference close-range interactive trajectory planning according to claim 1, S2 includes: Based on the task parameters, the UAV's onboard sensors are controlled to collect environmental measurement data within the interactive work area. The environmental measurement data includes inertial measurement data and measurement data used to characterize the environmental geometry. The environmental measurement data are synchronized in time and transformed in coordinates. Based on the synchronized environmental measurement data, the UAV state variables are obtained through multi-sensor fusion state estimation, and the UAV state variables include position and attitude.

4. The method for non-destructive operation of a drone based on AI and low-interference close-range interactive trajectory planning according to claim 1, S3 includes: Based on the environmental measurement data and the UAV state variables, the measurement data used to characterize the environmental geometry is transformed into three-dimensional spatial observation data in a preset coordinate system through coordinate transformation. The three-dimensional spatial observation data is fused into an implicit scene model to construct or update the implicit scene model; the interactive operation area is sampled in the updated implicit scene model, and for each sampling point, the implicit geometric quantities output by the implicit scene model are queried and obtained, including the symbolic distance value and / or the volume density value. The surface geometric features are generated by calculating the gradient and / or performing numerical difference on the implicit geometric quantities to determine the surface normal vector, calculating the second derivative of the implicit geometric quantities and / or performing numerical approximation based on the rate of change of the normal vector to determine the curvature. A geometric uncertainty map is generated based on at least one of the observation count, reprojection error, measurement residual, photometric consistency, and view coverage of the model unit in the implicit scene model. The geometric uncertainty term is then obtained by querying the location corresponding to the UAV state variable based on the geometric uncertainty map.

5. The method for non-destructive operation of a drone based on AI and low-interference close-range interactive trajectory planning according to claim 1, S4 includes: A local extraction region is determined based on the UAV state variables. The local extraction region is located within the interactive operation area and centered on the position corresponding to the UAV state variables, and has a preset spatial range. Multiple sampling points are selected within the local extraction region according to preset sampling rules. For each sampling point, the implicit geometric values ​​of the sampling point are obtained by querying based on the implicit scene model. The implicit geometric values ​​include the symbolic distance value and / or the volume density value. The surface normal vector and curvature corresponding to the sampling point are obtained based on the surface geometric features. The spatial location, implicit geometric values, surface normal vectors, and curvature of the multiple sampling points are combined according to a preset data structure to form the local geometric input data.

6. The method for non-destructive operation of a drone based on AI and low-interference close-range interactive trajectory planning according to claim 1, S5 includes: The local geometric input data is normalized and converted into geometric condition input for the neural operator aerodynamic prediction model. Simultaneously, the UAV state variables are normalized and converted into state input for the neural operator aerodynamic prediction model. The geometric condition input and the state input are then input into the geometric condition-based neural operator aerodynamic prediction model to obtain aerodynamic interference prediction values ​​characterizing the aerodynamic interference intensity caused by the rotor downwash and ground effect on the target surface. The aerodynamic interference intensity is at least one of the following indicators at the target surface: wind speed, dynamic pressure / wind pressure, shear stress, pulsation intensity, vorticity, force, and torque. Furthermore, the aerodynamic interference intensity is at least one of the maximum value, peak value, root mean square value, average value, and integral value of the index on the target surface; based on at least one of the variance output of the geometrically based neural operator aerodynamic prediction model for the same input, the dispersion of the integrated inference output, and the dispersion of the Monte Carlo random inactivation inference output, the prediction uncertainty term is determined; based on the differentiable calculation of the UAV state variables by the geometrically based neural operator aerodynamic prediction model, the partial derivative of the aerodynamic interference prediction value with respect to the UAV state variables is obtained, and the disturbance sensitivity is obtained.

7. The method for non-destructive operation of a drone based on AI and low-interference close-range interactive trajectory planning according to claim 1, S6 includes: Based on the low disturbance threshold, the predicted aerodynamic disturbance value, the prediction uncertainty term, and the geometric uncertainty term, a disturbance barrier function value is calculated. The disturbance barrier function value is equal to the low disturbance threshold minus the sum of the predicted aerodynamic disturbance value, the geometric uncertainty term, and the prediction uncertainty term. The disturbance barrier function value is not less than zero to construct a control barrier function constraint, and the control barrier function constraint is linearized to the first order based on the disturbance sensitivity to obtain a linearized constraint. The linearization constraints, along with the UAV dynamics constraints and actuator constraints, are input into the nonlinear model predictive control. The control sequence is then solved in the prediction time domain, so that the cost function includes a trajectory tracking error term and an aerodynamic disturbance cost term. The first control quantity of the control sequence is output as the control instruction.

8. The method for non-destructive operation of a drone based on AI and low-interference close-range interactive trajectory planning according to claim 1, S7 includes: Based on the control commands, the drone is controlled to fly within the interactive work area and perform the close-range interactive operation; During the execution of the close-range interaction operation, steps S2 to S6 are repeated with a preset control cycle to reacquire the environmental measurement data and update the UAV state variables, thereby updating the implicit scene model, the surface geometric features and the geometric uncertainty term, updating the local geometric input data, recalculating the aerodynamic disturbance prediction value, the prediction uncertainty term and the disturbance sensitivity, and resolving the nonlinear model predictive control to output the updated control command.

9. The method for non-destructive operation of a drone based on AI and low-interference short-range interactive trajectory planning according to claim 7, characterized in that, The control barrier function constraint includes at least one of the following: control barrier function constraints based on continuous-time form: ； in, For extended class K functions; and / or control barrier function constraints based on discrete-time form: ； in, These are preset parameters. and These represent the system states for adjacent control cycles.

10. A method for non-destructive operation of a drone based on AI and low-interference short-range interactive trajectory planning according to claim 8, characterized in that, Within the preset control period, when the geometric uncertainty term and / or the prediction uncertainty term are greater than the preset uncertainty threshold, the UAV is controlled to execute an active perception maneuvering strategy to increase the spectral coverage of reflective areas and / or weak texture areas and reduce the uncertainty in the geometric uncertainty map.

Images