An artificial intelligence-based unmanned aerial vehicle automatic airport inspection system for glass curtain walls

Abstract

The application discloses an unmanned aerial vehicle automatic airport inspection system based on artificial intelligence, and particularly relates to the field of unmanned aerial vehicle inspection, and is used for solving the problem of lack of cooperation between curvature optimization and dynamic airspace management in airport glass curtain wall inspection. The system is characterized in that curvature grid and clearance profile are fused in a unified coordinate frame, real-time airspace information is immediately inherited in the process of flight path generation, a pause signal is triggered by distance and attitude double features in the process of subsequent flight stage safety monitoring, only the unexecuted section is implemented with normal translation and the tower station is simultaneously given in the process of trajectory correction, global recalculation is avoided, experience iteration is continuously written into a learning library, safety bandwidth is adaptively converged with task accumulation, and a cooperative closed chain of wall inspection and airspace scheduling is formed by multi-unit head-to-tail connection, the camera visual axis is kept adhering to the curtain wall, dynamic corridors are avoided in real time, the inspection is not interrupted, scheduling has zero interference, and the algorithm load is stably in the millisecond level, the flight path pause time of one-time inspection of the airport curtain wall is significantly shortened, and the frequency of manual intervention is compressed.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of unmanned aerial vehicle inspection, more particularly, the present application relates to a glass curtain wall unmanned aerial vehicle automatic airport inspection system based on artificial intelligence. BACKGROUND

[0002] On the glass curtain wall outer facade of the airport terminal, the inspection operation relies on the unmanned aerial vehicle to fly along the double curvature wall, and the heading is output in real time by the curvature smoothing algorithm, so that the camera visual axis is always perpendicular to the glass surface. At the same time, the airport operation command center will repeatedly adjust the clearance corridor around the curtain wall according to the flight in and out and the ground vehicle walking condition, and the corridor boundary will shrink or expand outward according to the dispatch rhythm, especially near the corner and concave-convex nodes, which is extremely close to the inspection route, forming a dynamic interleaved airspace situation.

[0003] The current automatic inspection method only pays attention to curvature smoothing and endurance consumption when planning the route, and does not take the real-time change of the clearance corridor into account. Once the wall-attached trajectory generated by the algorithm is fixed, it will not be updated, and when the clearance boundary approaches the curtain wall, the unmanned aerial vehicle still advances according to the original trajectory, and finally crosses into the controlled airspace, causing the tower to interrupt the inspection and return to the original position, and the flight scheduling is also forced to be rearranged. The root cause of the problem is that the curvature optimization and dynamic airspace management lack coordination, and simply expanding the safety distance cannot fundamentally solve the conflict. It is necessary to integrate the dynamic clearance model in the route generation stage to enable the curtain wall inspection and airport operation to coexist for a long time.

[0004] In order to solve the above problems, a technical scheme is provided. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide a glass curtain wall unmanned aerial vehicle automatic airport inspection system based on artificial intelligence, which solves the problem of lack of coordination between curvature optimization and dynamic airspace management in airport glass curtain wall inspection. This is achieved by first integrating the curvature grid and the clearance profile in a unified coordinate frame, so that the flight path generation immediately inherits real-time airspace information, the safety monitoring in the subsequent flight stage triggers a pause signal with distance-attitude dual features, the trajectory correction only implements normal translation for the unexecuted segment and synchronously gives the tower, avoiding global recalculation, and the experience iteration continuously writes into the learning library, so that the safety bandwidth adaptsively converges with the task accumulation; the multi-unit head-to-tail connection forms a coordinated closed loop of wall-attached inspection-airspace scheduling, which not only keeps the camera visual axis attached to the curtain wall, but also avoids the dynamic corridor in real time, so as to solve the problems raised in the background art.

[0006] To achieve the above purpose, the present application provides the following technical scheme:

[0007] Coordinate fusion unit: superimpose the curtain wall curvature grid and the clearance corridor profile under the unified coordinate reference before takeoff, generate a shared coordinate table and load it to the flight terminal;

[0008] Trajectory generation unit: read shared coordinate table, generate straight-line and arc hybrid trajectory according to minimum curvature principle, and write minimum safety zone at each node;

[0009] Safety monitoring unit: during flight, output residual distance curve and angular velocity signal synchronously with radar and inertial navigation, extract key features to perform close-to-wall safety judgment, if the judgment result exceeds the standard, pause and request the latest corridor profile;

[0010] Trajectory correction unit: after the ground station pushes the latest corridor profile, the path planner translates the unexecuted trajectory along the normal direction and broadcasts the correction result for tower monitoring, and the flight control then resumes execution of the updated trajectory;

[0011] Experience iteration unit: after landing, write flight records into the learning library, preset safety bandwidth for the next task and reduce online downtime.

[0012] In a preferred embodiment, the coordinate fusion unit includes the following:

[0013] Before takeoff, extract the curtain wall curvature grid containing position coordinates and local curvature values from the airport building database, obtain the clearance corridor profile composed of polygon boundary line segments from the airport operation command center, verify point cloud density consistency, select a reference point set to calculate the transformation matrix to minimize the sum of Euclidean distances to establish a unified coordinate reference, calculate the projection distance of each curtain wall curvature grid point to the clearance corridor profile to identify the overlapping area and generate a fusion grid by superimposition, extract the shared coordinate table organized by the fusion grid points, and verify and load it into the flight terminal storage module through a wireless transmission protocol.

[0014] In a preferred embodiment, the trajectory generation unit includes the following:

[0015] The trajectory generation unit extracts the shared coordinate table from the storage module, verifies the integrity of the matrix row number, filters points with curvature values exceeding the preset threshold as curvature key points, calculates the minimum curvature path between adjacent curvature key points to generate a close-to-wall straight-line and arc hybrid trajectory, calculates the shortest normal distance of each path point to the clearance corridor profile to quantify the clearance and obtain the minimum safety bandwidth value, and appends the minimum safety bandwidth value to the path point sequence as a floating-point number format attribute field.

[0016] In a preferred embodiment, the shared coordinate table is a matrix, each row corresponding to a point, including unified coordinates, curvature values, and profile flags;

[0017] Unified coordinates refer to the three-dimensional position coordinate values of each point in the fusion grid under the unified coordinate reference, including x component, y component, and z component;

[0018] Curvature value refers to the local curvature value derived from the curtain wall curvature mesh, which directly inherits the original local curvature value for the point derived from the curtain wall curvature mesh, and fills in zero for the point derived from the clearance corridor profile to represent no curvature information;

[0019] Profile flag refers to a binary identification value, which is assigned a value of one for the point belonging to the clearance corridor profile boundary, and a value of zero for the point not belonging to the clearance corridor profile boundary.

[0020] In a preferred embodiment, the safety monitoring unit includes the following:

[0021] During flight, the residual distance readings collected by the radar form a residual distance curve, and the inertial navigation sensors record the three-axis angular velocity signals. The clock synchronizes the alignment sequence data, extracts key features, including the margin contraction speed ratio and the attitude amplitude density, and the pre-trained lightweight random forest inputs the margin contraction speed ratio and the attitude amplitude density to output the attitude stability adjustment coefficient and determine whether it is over-standard. If so, pause and request the latest clearance corridor profile and lock the high-precision coordinates.

[0022] In a preferred embodiment, the margin contraction speed ratio calculates the slope difference between adjacent time points based on the residual distance curve, and selects the maximum negative slope value to quantify the clearance margin reduction rate.

[0023] In a preferred embodiment, the attitude amplitude density applies a short window fast Fourier transform to the three-axis angular velocity signals, integrates the part exceeding the medium frequency threshold after obtaining the frequency spectrum amplitude, to evaluate the attitude oscillation intensity.

[0024] In a preferred embodiment, the trajectory correction unit includes the following:

[0025] After the ground station pushes the latest clearance corridor profile, it checks the integrity of the boundary line segment, extracts the unexecuted flight path sub-sequence based on the current UAV position binary search matching index, calculates the normal distance from each point to the latest clearance corridor profile to translate the point position and generate the corrected unexecuted flight path sub-sequence, sends the point position of the corrected unexecuted flight path sub-sequence and the minimum safety bandwidth value, and the tower station confirms and replaces the original unexecuted flight path sub-sequence and removes the pause to start the forward movement.

[0026] In a preferred embodiment, the experience iteration unit includes the following:

[0027] After landing, the flight record is extracted and the sequence integrity is checked, the flight record is appended to the learning library and the structured field is defined, the multi-task flight record is queried to calculate the preset safety bandwidth intention index decay, the minimum safety bandwidth value of the shared coordinate table is updated and version marking control is performed, and the buffer threshold limit is expanded during flight.

[0028] The technical effect and advantages of the glass curtain wall unmanned aerial vehicle automatic airport inspection system based on artificial intelligence of the present application are as follows:

[0029] The present application fuses the curvature grid and the clearance profile in the unified coordinate frame in advance, the flight path generation immediately inherits the real-time airspace information, the subsequent flight stage safety monitoring triggers the pause signal with the distance-attitude double features, the trajectory correction only implements the normal translation for the unexecuted segment and synchronously gives the tower station, the global recalculation is avoided, the experience iteration continuously writes into the learning library, the safety bandwidth is adaptively converged with the task accumulation, the multi-unit head-tail connection forms the cooperative closed chain of the wall-adhesion inspection-airspace scheduling, the camera visual axis is adhered to the curtain wall and the dynamic corridor is avoided in real time, the inspection is not interrupted, the scheduling is zero-interference, the algorithm load is stably in the millisecond level, the overall flight path pause time of the airport curtain wall one-time inspection is significantly shortened and the artificial intervention frequency is compressed. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 The structure diagram of the glass curtain wall unmanned aerial vehicle automatic airport inspection system based on artificial intelligence of the present application is shown. DETAILED DESCRIPTION

[0031] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0032] Embodiment 1: Figure 1 The glass curtain wall unmanned aerial vehicle automatic airport inspection system based on artificial intelligence of the present application is given, which comprises:

[0033] The coordinate fusion unit: before taking off, the curtain wall curvature grid and the clearance corridor profile are superimposed under the unified coordinate reference, the shared coordinate table is generated and loaded to the flight terminal.

[0034] The flight path generation unit: the shared coordinate table is read, the wall-adhesion straight line and circular arc hybrid flight path is generated according to the minimum curvature principle, and the minimum safety belt is written at each node.

[0035] The safety monitoring unit: during the flight, the residual distance curve and the angular velocity signal are synchronously output from the distance radar and the inertial navigation, the key features are extracted to perform the wall-adhesion safety judgment, if the judgment result is out of standard, the progress is paused and the latest corridor profile is requested.

[0036] The trajectory correction unit: after the ground station pushes the latest corridor profile, the path planner corrects the unexecuted flight path along the normal direction, and broadcasts the correction result for the tower station monitoring, and the flight control subsequently releases the pause to execute the updated flight path.

[0037] Experience iteration unit: after landing, write flight records into the learning library, preset safety bandwidth for the next task and reduce online downtime.

[0038] In the inspection operation of the glass curtain wall of the airport terminal building, the UAV needs to fly along the hyperbolic curvature surface to ensure that the camera's visual axis is always perpendicular to the glass surface. At the same time, the airport operation command center adjusts the clearance corridor boundary around the curtain wall in real time according to the arrival and departure of flights and the movement of ground vehicles. These boundaries are extremely close to the inspection route at the corner and concave-convex nodes, forming a dynamic and interlaced airspace environment. The current inspection method only focuses on smooth curvature and endurance consumption, and does not take into account the real-time changes of the clearance corridor in the route planning constraints, resulting in the generated wall-clinging trajectory being fixed and unable to adapt to the approaching boundary, causing the UAV to enter the controlled airspace and interrupt the inspection.

[0039] To solve the problem of lack of coordination between curvature optimization and airspace management, the coordinate fusion unit superimposes the curtain wall curvature grid and the clearance corridor profile under the unified coordinate reference before takeoff, generates a shared coordinate table and loads it to the flight terminal, thereby integrating the dynamic clearance model at the route generation stage and ensuring that the subsequent inspection route inherits real-time airspace information from the source, achieving long-term coexistence of inspection and airport operation.

[0040] The processing method of this scheme is to superimpose the curtain wall curvature grid and the clearance corridor profile in the same coordinate frame to avoid deviation caused by independent coordinates, and generate a shared coordinate table that can be directly called by downstream units, thereby providing a unified basis for curvature and airspace constraints for route generation.

[0041] The processing of the coordinate fusion unit is divided into the following steps:

[0042] 1-1. The process of collecting the curtain wall curvature grid and the clearance corridor profile is to obtain the surface geometry information and airspace boundary data required for inspection from the source, thereby laying a accurate input basis before fusion.

[0043] Before takeoff, the curtain wall curvature grid is extracted from the airport building database. The grid is represented by a three-dimensional point cloud, and each point contains position coordinates and local curvature value. At the same time, the clearance corridor profile is obtained from the airport operation command center in real time. The profile is composed of polygon boundary line segments, and each line segment is labeled with start and end point coordinates. When extracting, verify the consistency of point cloud density by comparing the average distance between adjacent points to ensure that the grid resolution matches the length of the contour line segment, avoiding scale deviation during fusion.

[0044] This step outputs the original data set of the curtain wall curvature grid and the clearance corridor profile, which facilitates the direct reference of these point information by subsequent mapping processing, achieving reliable preparation of input data.

[0045] 1-2. Establish a unified coordinate reference, eliminate coordinate deviations between data from different sources, and ensure that the overlay operation is performed within the same framework, thereby improving the geometric consistency of the fusion.

[0046] Map the curtain wall curvature grid and the clearance corridor profile into the airport global coordinate framework, and align the reference points by the least square method. Specifically, select a set of reference points in the curtain wall curvature grid and a corresponding set of reference points in the clearance corridor profile, calculate the transformation matrix to minimize the sum of the Euclidean distances between the point sets; in the operation, first initialize the transformation matrix as the unit matrix, then iteratively adjust the matrix parameters to minimize the sum of the squared distances between each reference point after transformation and the corresponding point; the summing process involves accumulating the distance calculations for all pairs of reference points, and gradually approaching the optimal solution through gradient descent, where the distance calculation is the square root of the sum of the squares of the transformed coordinates minus the target coordinates.

[0047] This step outputs the curtain wall curvature grid and the clearance corridor profile under the unified coordinate reference, ensuring that the downstream overlay directly uses these aligned point positions, avoiding airspace misjudgment caused by deviation accumulation.

[0048] 1-3. Overlay the curtain wall curvature grid and the clearance corridor profile, identify the geometric overlapping area, and label potential conflict points in the shared table to achieve the preliminary integration of curvature and airspace.

[0049] Based on the curtain wall curvature grid and the clearance corridor profile under the unified coordinate reference, perform geometric overlay in three-dimensional space to identify overlapping areas. Specifically, calculate the projection distance of each curtain wall curvature grid point within the clearance corridor profile, which is obtained by first projecting the point onto each profile segment, then calculating the straight-line distance from the grid point to the projection point, and taking the minimum distance value among all segments; if the minimum distance value is negative or zero, it is determined that the point is located within the clearance corridor profile.

[0050] This step outputs the fused grid after overlay, and labels the overlapping state of each point, which facilitates the inheritance of these labeling information by the subsequent table generation, and improves the response accuracy of the inspection path to dynamic boundaries.

[0051] 1-4. Generate a shared coordinate table to organize the fused information into a structured format, thereby facilitating efficient reading of curvature and profile constraints by flight terminals, and realizing the standardization of data sharing.

[0052] Based on the superimposed fusion grid, the coordinates, curvature values and contour boundary information of all points are extracted and organized into a table structure. Specifically, the shared coordinate table is a matrix, each row corresponds to a point, including uniform coordinates, curvature values (if derived from the curtain wall curvature grid, the original value is directly inherited, otherwise fill in zero to represent no curvature information) and contour flag (assigned to one to represent the clear corridor contour boundary, assigned to zero to represent non-boundary points); the filling rule ensures that the curvature value is only valid at the grid source point, and the flag bit is binary based on the overlap state.

[0053] This step ensures that the shared coordinate table contains complete fusion information, making it easy for downstream units to directly parse the matrix row when called, achieving seamless information transfer.

[0054] 1-5. Load the shared coordinate table to the flight terminal and plant the fusion data in real time into the unmanned aerial vehicle control module, thereby supporting immediate flight path adjustment after takeoff and achieving synchronization of ground and air data.

[0055] Based on the generated shared coordinate table, upload it to the storage module of the unmanned aerial vehicle flight terminal through the wireless transmission protocol, and add a check code during transmission for verification, that is, calculate the hash value of the table data and send it with the hash value, and the receiving end recalculates the hash to confirm the integrity.

[0056] The coordinate fusion unit realizes the unified superposition of the curtain wall curvature grid and the clear corridor contour before takeoff, ensuring that the generated shared coordinate table inherits the curvature constraints and dynamic airspace boundary information under the airport glass curtain wall inspection scenario, thereby providing a fusion basis for subsequent flight path generation and avoiding airspace conflicts caused by independent coordinates.

[0057] The coordinate fusion unit realizes the unified superposition of the curtain wall curvature grid and the clear corridor contour before takeoff, and generates a shared coordinate table loaded to the flight terminal, thereby providing a unified data basis for the fusion of curvature constraints and dynamic airspace boundaries for inspection route planning. However, the current method ignores the clear change during the planning stage, leading to frequent conflicts after trajectory fixation. Therefore, the flight path generation unit generates a straight line and circular arc mixed flight path based on the shared coordinate table and the minimum curvature principle, and writes the minimum safety zone at the nodes, ensuring that the generated flight path inherits real-time airspace information from the source, achieving preliminary coordination of camera visual axis vertical adhesion and dynamic clear corridor avoidance in glass curtain wall inspection.

[0058] The processing of the flight path generation unit is divided into the following steps:

[0059] 2-1. Read the shared coordinate table to obtain the fused curvature and contour information, thereby directly incorporating the curtain wall geometry and airspace constraints in the flight path planning.

[0060] The flight terminal extracts the shared coordinate table from the storage module. The shared coordinate table is in a matrix structure, each row containing a uniform coordinate, a curvature value, and a contour flag. The uniform coordinate is a three-dimensional position coordinate value, the curvature value is a local curvature numerical value or zero padding, and the contour flag is a binary identification value. When extracting, the number of matrix rows is checked to match the expected total number to ensure data integrity.

[0061] 2-2. Extract the curvature key points and identify the high curvature area to guide the selection of the segmentation points of the straight line and arc hybrid flight path.

[0062] Based on the complete matrix of the shared coordinate table, the curvature values of each row are traversed, and the points with curvature values exceeding the preset threshold are selected as curvature key points. These points are used as the turning positions of the flight path in the downstream generation to avoid the deviation of the visual axis caused by insufficient smoothing.

[0063] 2-3. Generate a wall-adhesive straight line and arc hybrid flight path to minimize the cumulative curvature and ensure that the camera visual axis is always perpendicular to the glass surface during the flight of the unmanned aerial vehicle.

[0064] Based on the set of curvature key points, the adjacent points are connected to form a hybrid trajectory according to the minimum curvature principle. The minimum curvature path between adjacent curvature key points is calculated, and the total length is obtained by adding the length of the straight line segment and the length of the arc segment. The length of the straight line segment is obtained by dividing the path into small segments along one coordinate axis, calculating the square of the slope between two points for each small segment, taking the square root of the sum of all small segments, and accumulating all small segments to approximate the integral value. The slope is obtained by dividing the difference between the adjacent points on the other two coordinate axes by the difference in the axis. The length of the arc segment is obtained by multiplying the radius of each arc by the angular radian. The radius and angular radian are iteratively adjusted to minimize the total curvature integral, i.e., repeatedly calculating the path curvature with different parameters and selecting the minimum one.

[0065] This calculation ensures that the straight line segment dominates in the low curvature area and the arc segment smoothly transitions in the high curvature area, with the unit being length. This step outputs the path point sequence of the wall-adhesive straight line and arc hybrid flight path, which facilitates the direct insertion of safety information along these points in the subsequent node processing, and improves the adhesion of the trajectory to the hyperbolic surface of the curtain wall.

[0066] 2-4. Calculate the minimum safety zone to quantify the clearance at the flight path nodes to prevent boundary intrusion when the dynamic corridor shrinks.

[0067] Based on the path point sequence of the mixed straight line and arc flight path along the wall, the shortest distance to the clearance corridor contour is calculated for each path point. Specifically, the boundary points are identified by using the contour markers in the shared coordinate table, all contour markers with a value of 1 are traversed, the Euclidean distance is calculated by taking the square root of the sum of the squares of the difference between the uniform coordinate components of each path point and the uniform coordinate components of the boundary points, the minimum Euclidean distance is selected, and then multiplied by the cosine value of the angle between the path point normal vector and the distance vector, which is obtained by dividing the product of the corresponding components of the two vectors by the product of their respective modules, to ensure the projection along the normal.

[0068] The calculation ensures that the safety belt is projected along the normal, and the unit is length. The minimum safety belt width value of each path point is output, which facilitates the direct addition of these values in the write operation, and realizes the buffer protection of the space intersection near the corner concave-convex nodes.

[0069] The minimum safety belt refers to the shortest normal distance from each path point in the mixed straight line and arc flight path along the wall to the clearance corridor contour, which is used to quantify the buffer margin between the UAV and the dynamic airspace boundary. Specifically, the minimum safety belt is determined by calculating the perpendicular distance from the uniform coordinates of the path point to the nearest clearance corridor boundary line, the distance is projected along the normal of the curtain wall surface, and the unit is length, which ensures that enough space is provided during the inspection to avoid trajectory conflicts caused by the contraction or expansion of the clearance corridor, thereby maintaining the vertical fit of the camera visual axis and flight safety.

[0070] 2-5. Write the minimum safety belt to the node, embed the safety information into the flight path structure, so as to provide real-time judgment of the margin for the downstream monitoring unit. Based on the minimum safety belt width value of each path point, these values are appended to the path point sequence of the mixed straight line and arc flight path along the wall to form extended path data. Specifically, the minimum safety belt width value is appended as an attribute field after each node in the path point sequence, and the field is defined as a floating point number format for unified storage and reading. This step ensures the completeness of the extended path data, which facilitates the direct reading of these fields during safety monitoring, and realizes the dynamic adaptability of the inspection flight path.

[0071] The flight path generation unit realizes the generation of the mixed straight line and arc flight path along the wall based on the shared coordinate table, and writes the minimum safety belt to the node, which ensures the comprehensive integration of flight path planning from curvature minimization to airspace margin buffer in the airport glass curtain wall inspection scene, thereby avoiding trajectory conflicts caused by changes in the clearance corridor.

[0072] The flight path generation unit reads the shared coordinate table, generates a straight-line and arc hybrid flight path that adheres to the wall, and writes it into the minimum safety zone, thereby providing an initial path for the flight phase that inherits the curvature and airspace constraints. However, the current inspection ignores real-time clearance changes during flight, leading to frequent trajectory conflicts. To address this, the safety monitoring unit outputs the residual distance curve and angular velocity signal synchronously and extracts key features to perform judgment, ensuring stable UAV attitude and immediate response to dynamic corridor avoidance during glass curtain wall inspection.

[0073] The key features include the excess shrinkage speed ratio and the attitude amplitude density.

[0074] The safety monitoring unit processing is divided into the following steps:

[0075] 3-1. Synchronously output the residual distance curve and angular velocity signal to capture the instantaneous distance change of the UAV to the clearance corridor boundary and the attitude dynamics during flight, thereby integrating distance radar and inertial navigation data in monitoring.

[0076] During flight, the distance radar continuously collects residual distance readings to the clearance corridor boundary, forming a residual distance curve composed of time series distance values; at the same time, the inertial navigation sensor records real-time three-axis angular velocity signals, including roll, pitch, and yaw axis angular velocity values. During collection, the clock is synchronized to align the residual distance readings and three-axis angular velocity signals to the millisecond level, avoiding timing deviation affecting feature extraction accuracy. Output the residual distance curve and three-axis angular velocity signal for real-time tracking of dynamic airspace intersection.

[0077] 3-2. Calculate the excess shrinkage speed ratio to quantify the rate of clearance reduction, thereby identifying the risk of boundary approaching. Based on the residual distance curve, calculate the slope difference of adjacent time points, and select the maximum negative slope as the excess shrinkage speed ratio. The specific calculation is to first take the derivative of the residual distance curve point by point, i.e., each time point uses the next distance value minus the current distance value divided by the time interval to get the slope, then calculates the difference value of adjacent slopes, only retains the negative difference value sequence, and selects the maximum value as the excess shrinkage speed ratio, where the negative difference value represents accelerated shrinkage, ensuring to capture accelerated shrinkage. Output the excess shrinkage speed ratio for downstream feature integration to use the excess shrinkage speed ratio value, improving the sensitivity to corridor shrinkage near corner nodes.

[0078] 3-3. Calculate the attitude amplitude density to assess the strength of attitude oscillation, thereby detecting stability deviation during wall-attached flight.

[0079] Based on the three-axis angular velocity signals, a short-window Fast Fourier Transform (FFT) is applied to each axis signal to obtain the spectral amplitude. Then, the amplitudes exceeding the intermediate frequency (IF) threshold are integrated to obtain the attitude amplitude density. First, the three-axis angular velocity signals are divided into short time windows. Within each window, a FFT is performed to obtain the frequency domain amplitude spectrum. Then, the portions of the amplitude spectrum with frequencies higher than the IF threshold are selected. These amplitudes are integrated and summed from the IF threshold to infinity. The average of the three-axis integrals is used as the attitude amplitude density to ensure the quantification of high-frequency disturbances.

[0080] Output attitude amplitude density to facilitate the direct input of attitude amplitude density values, enabling precise monitoring of flight attitude on hyperbolic curvature surfaces.

[0081] 3-4. Perform a wall-hugging safety judgment, comprehensively assess the posture stability status, and thus decide whether to trigger a pause.

[0082] The residual shrinkage ratio and attitude amplitude density are used as inputs to a lightweight random forest model, which outputs attitude stabilization adjustment coefficients. If the attitude stabilization adjustment coefficients exceed a preset limit, the model is considered to have exceeded the limit.

[0083] The lightweight random forest construction process is based on ensemble learning, combining multiple decision trees to form an efficient model to meet real-time response requirements, such as millisecond-level computation. In the initial construction phase, guide samples are generated from the training dataset. This involves sampling with replacement from the original dataset to obtain a subset of the same size, allowing some data points to appear repeatedly. Approximately one-third of the data is reserved as out-of-bag samples for subsequent validation. Each decision tree is built independently, first randomly selecting a subset of features from the guide samples to reduce inter-tree correlation and improve generalization ability. The decision tree construction starts from the root node, recursively splitting based on the input features—namely, the residual shrinkage rate ratio and attitude amplitude density—using a preset branch threshold. Specifically, for each node, it is evaluated whether the feature value exceeds the preset threshold. If so, the split continues along the "yes" branch; otherwise, it follows the "no" branch. The splitting criterion uses mean squared error as a metric to ensure that each split reduces the variance of the data within the node. The recursive process continues until leaf node conditions are reached, such as the number of node samples falling below the minimum sample splitting threshold or reaching the maximum depth limit, thus forming a shallow tree structure to maintain the model's lightweight nature. After all decision trees are built, they are integrated into a random forest through an aggregation mechanism. The number of trees is kept within a small range to optimize computational efficiency and ensure that the overall model can quickly output values ​​in attitude stability evaluation.

[0084] The lightweight random forest training algorithm employs a guided aggregation method combined with feature randomness to achieve efficient learning and minimize the risk of overfitting. Training begins with dataset partitioning, where the original dataset is used to generate multiple guide samples, each created through sampling with replacement to introduce diversity. For each decision tree, a classification and regression tree algorithm is used for training: starting from the root node, the reduction in mean squared error of each possible branch threshold is calculated, and the branch threshold that results in the largest reduction is selected for splitting; for the two input features, residual shrinkage ratio and attitude amplitude density, a preset branch threshold is used as the split point, and data subsets are recursively assigned to child nodes until leaf nodes, where leaf nodes store the average value as the output. During training, each tree only considers a randomly selected subset of features, typically the square root of the total number of features, to reduce computational complexity and inter-tree correlation. Out-of-bag samples are used for internal cross-validation: for each tree, its out-of-bag samples that were not used in training are input into the model, and the prediction error is calculated to evaluate performance. After all trees are trained, the prediction results are aggregated by averaging the outputs of all trees to form the attitude stability adjustment coefficient as the final output. This training mechanism ensures that the model converges quickly with limited resources, making it suitable for real-time processing of distance and attitude features in flight monitoring scenarios.

[0085] The optimization of lightweight random forests focuses on improving prediction accuracy and computational efficiency, reducing overfitting and highlighting key features through various techniques. First, variance is reduced through aggregation of uncorrelated trees: each tree is trained on independently guided samples, and the average output smooths the noise of individual trees, ensuring the overall model robustness. Second, feature importance is evaluated to streamline input: the mean-reduction impurity method is used to calculate the decrease in model accuracy after excluding residual shrinkage ratios or attitude amplitude density; or permutation importance is employed, quantifying the contribution by randomly permuting feature values ​​in out-of-bag samples and observing the average decrease in accuracy, thus prioritizing the retention of highly important features. Optimization also includes handling missing values: for potentially missing inputs such as attitude amplitude density, feature randomness is used to estimate imputation values, maintaining model integrity. Furthermore, hyperparameter tuning techniques are employed, such as grid search systematically traversing combinations of the number of trees and maximum depth, evaluating performance metrics such as mean squared error on the validation set, and selecting the parameter set that minimizes the error; or random search sampling in the parameter space to efficiently explore a wide range of settings. This optimization process ensures that the model maintains high accuracy in millisecond-level responses while adapting to the real-time requirements of dynamic airspace monitoring.

[0086] The parameter settings of the light-weight random forest are aimed at balancing the model complexity and computational overhead, and all parameters are fixed before training to support efficient deployment. The main parameters include the node size, i.e., the minimum number of samples in a leaf node, which is used to control the tree depth and prevent overfitting, and is usually set to a small value to allow a shallow tree structure; the number of trees, i.e., the total number of decision trees in the ensemble, which is set to a limited number such as tens of trees to achieve light weight and ensure millisecond-level inference time; the number of sampled features, i.e., the size of the feature subset considered when splitting each node, which is usually set to the square root of the total number of features or all for models with only two input features to maintain diversity. In addition, the maximum depth parameter limits the number of splitting layers of the tree, and is set to a shallow value to reduce the computational load; the minimum sample split threshold defines the minimum number of samples required for node splitting, and is preset to be higher than the default to speed up training; the preset branch threshold is specifically set for the residual contraction speed ratio and the attitude amplitude density, such as a fixed threshold range based on domain knowledge, which is used for recursive splitting rather than data-driven learning to simplify the model. Other parameters such as the proportion of bootstrap samples are set to the size of the dataset by default, and out-of-bag samples are used for validation and not involved in the setting. These parameters are finally determined through validation set tuning to ensure that the dimensionless output of the attitude stability adjustment coefficient accurately triggers the judgment when it exceeds the limit value, optimizing the response efficiency in the inspection scenario.

[0087] 3-5. In response to the judgment that the threshold is exceeded, the progress is interrupted before the risk and the information is updated, thereby maintaining the continuity of the inspection. Based on the judgment result, if the threshold is exceeded, the progress of the UAV is paused, and the ground station is requested for the latest clear corridor profile. Specifically, the pause command locks the current position, and high-precision coordinates are recorded to ensure that the request includes the accurate position, and the current coordinates are sent through the wireless link to obtain the updated profile. This step ensures flight safety and facilitates subsequent processing by the trajectory correction unit, thereby reducing the intervention of the tower station.

[0088] The safety monitoring unit realizes the extraction and judgment of distance-attitude features during flight, ensuring that the UAV responds to changes in the clear corridor in real time in the scenario of airport glass curtain wall inspection, thereby minimizing interruptions caused by trajectory conflicts.

[0089] The safety monitoring unit judges the threshold by extracting key features and pauses the progress to request the latest clear corridor profile, thereby providing dynamic updated airspace boundary information for the correction stage. However, the current inspection needs to be globally re-planned when the boundary changes, resulting in a sharp increase in computational load. Therefore, the trajectory correction unit corrects the unexecuted flight path along the normal direction and broadcasts the results, ensuring local efficiency of flight path adjustment in glass curtain wall inspection and synchronization of tower monitoring, and achieving rapid recovery after interruption.

[0090] The processing of the trajectory correction unit is divided into the following steps:

[0091] 4-1. Receive the latest clearance corridor profile, obtain the updated boundary data pushed by the ground station, and integrate the real-time spatial changes in the correction. The ground station pushes the latest clearance corridor profile in response to the request, and the latest clearance corridor profile is composed of polygon boundary segments, each of which is marked with the starting point and ending point coordinates. When pushing, check the integrity of the boundary segment, that is, check whether the starting point and ending point coordinates are continuous, to avoid transmission errors affecting the accuracy of the correction.

[0092] 4-2. Identify the unexecuted flight path, isolate the flown part, and only adjust the remaining path.

[0093] Based on the boundary data set of the latest clearance corridor profile and the current UAV position, query the path point sequence of the mixed straight line and circular arc wall-following flight path, and extract the remaining points from the current position as the unexecuted flight path sub-sequence. Specifically, the index of the current position to the path point sequence is matched by the binary search algorithm, that is, the sequence is repeatedly divided and compared with the coordinate value until the index is located, and all points after the located index are extracted to form the unexecuted flight path sub-sequence. The unexecuted flight path sub-sequence is directly operated by the downstream correction to avoid global recalculation and improve computational efficiency.

[0094] 4-3. Perform normal translation correction to offset the unexecuted flight path to avoid new boundaries and maintain the perpendicularity of the camera visual axis to the glass surface.

[0095] Based on the unexecuted flight path sub-sequence, the normal distance to the latest clearance corridor profile is calculated for each point, and the point is translated along the normal. The translation vector is calculated, the normal distance of the point to the nearest boundary line is calculated, that is, the minimum vertical distance of the projection point to the polygon boundary segment, and the positive and negative signs are assigned according to the position inside and outside, then the normal distance is multiplied by the unit normal vector of the curtain wall surface at the point to obtain the vector component, ensuring that the offset is perpendicular to the surface. The corrected unexecuted flight path sub-sequence is output to facilitate the broadcast to directly use these offset points to achieve accurate avoidance of the overlapping airspace near the corner concave-convex nodes.

[0096] 4-4. Broadcast the correction results to inform the tower monitoring to update the path and coordinate the airport scheduling to avoid interference.

[0097] Based on the corrected unexecuted flight path sub-sequence, the point position and minimum safety bandwidth value of the unexecuted flight path sub-sequence are sent to the tower via a wireless link. Ensure that the tower receives complete correction data to facilitate subsequent execution of monitoring these points to achieve zero interference synchronization between patrol and flight operation.

[0098] 4-5. Unpause the execution of the updated flight path, resume the flight and apply the revised path, complete the continuous inspection after the interruption. Based on the confirmation after broadcasting, the flight control module replaces the original unexecuted flight path sub-sequence with the revised unexecuted flight path sub-sequence, and unpauses the command to start moving forward. The tower confirms the synchronization timing, that is, waits for the response signal within a fixed time window to align the update, ensuring no delay integration. Ensuring seamless integration of the updated flight path, facilitating subsequent recording by the experience iteration unit, significantly shortening the flight path pause time.

[0099] The trajectory revision unit implements normal translation revision of the unexecuted flight path and broadcasts it for tower monitoring, ensuring local adjustment of dynamic clearance changes in the airport glass curtain wall inspection scenario, thereby minimizing computational load and quickly resuming flight.

[0100] The trajectory revision unit implements normal translation revision of the unexecuted flight path and broadcasts it for tower monitoring, thereby achieving rapid resumption of flight after interruption. However, the current inspection lacks task-based experience accumulation, resulting in fixed safety parameters. Therefore, the experience iteration unit writes flight records to the learning library and presets a safety bandwidth, ensuring adaptive optimization of safety parameters in glass curtain wall inspection, and achieving reduced pauses under multi-task accumulation.

[0101] 5-1. Collect flight records to capture all data during this inspection, and accumulate distance, attitude and revision event information in iterations.

[0102] After landing, extract the flight records from the flight terminal, including the residual distance curve, three-axis angular velocity signal, attitude stabilization adjustment coefficient, revised unexecuted flight path sub-sequence and pause count. When extracting, check the integrity of the record sequence, that is, check whether the number of points in the residual distance curve, the number of axes in the three-axis angular velocity signal, the value of the attitude stabilization adjustment coefficient, the length of the revised unexecuted flight path sub-sequence and the count of the pause count match the expected value, to avoid incomplete data affecting the accuracy of iteration. Achieve comprehensive archiving of dynamic airspace interleaving events.

[0103] 5-2. Write to the learning library, persist the flight records for cross-task analysis.

[0104] Based on the complete data set of flight records, append the data set to the learning library, which is a time series database structure, indexed and stored by task number. Create a new entry to insert the residual distance curve, three-axis angular velocity signal, attitude stabilization adjustment coefficient, revised unexecuted flight path sub-sequence and pause count, and define the entry as a structured field in a unified format to ensure query consistency. Output the updated learning library to facilitate downstream analysis to directly query these historical entries, improving the continuity of experience accumulation.

[0105] 5-3. Analyze historical records to extract patterns to guide parameter adjustment.

[0106] Based on the updated learning library, the multi-task flight record is queried, the average pause duration and the margin contraction frequency are calculated, the preset safety bandwidth is calculated, the ratio of the total number of historical pauses to the total number of tasks is calculated, the negative ratio is taken as an index, the natural base is taken as an exponent to obtain a decay factor, and finally the initial safety bandwidth is multiplied by the decay factor to obtain the preset safety bandwidth, wherein the initial safety bandwidth is a fixed reference length, the total number of historical pauses is the cumulative number of all task pauses, and the total number of tasks is the number of executed tasks, and the calculation ensures that the bandwidth decays exponentially with experience. The preset safety bandwidth is output, which is convenient for downstream applications to directly use the preset safety bandwidth value to realize adaptive buffering of the change of the clear corridor.

[0107] 5-4. The preset safety bandwidth is initialized for the next task to reduce the need for online adjustment.

[0108] Based on the preset safety bandwidth, the minimum safety bandwidth value in the shared coordinate table of the flight path generation unit is updated to the preset safety bandwidth. The minimum safety bandwidth value of each node is covered, and version mark control update is adopted, that is, a task serial number label is attached to prevent multiple version conflicts. Ensure that the next task inherits the optimized parameters, which facilitates the safety monitoring unit to reduce the judgment of over-standard frequency and realizes the cumulative compression of the inspection pause.

[0109] 5-5. Reduce online pause, apply iterative result to optimize flight logic, and improve overall efficiency.

[0110] Based on the updated minimum safety bandwidth value, the buffer threshold limit is expanded during flight, and the preset threshold value of the attitude stabilization adjustment coefficient is adjusted by a proportion, and the proportion is calculated by the inverse ratio of the total number of historical pauses to ensure that the threshold limit is dynamically increased to reduce the triggering frequency. Ensure that the online pause is reduced, which facilitates the formation of a closed loop to realize the significant shortening of the flight path pause time.

[0111] The above formulas are all dimensionless numerical calculations, and the formulas are obtained by software simulation of a large amount of data to obtain a formula of the most recent real situation, and the preset parameters in the formula are set by a person skilled in the art according to the actual situation.

[0112] It should be noted that the system of the present application can be deployed on the device itself to realize embedded application, or can be run on PC or other terminal with user interface, so as to meet various hardware environments and use requirements.

[0113] The above only describes some exemplary embodiments of the present application by way of illustration, and it is needless to say that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present application. Therefore, the above figures and descriptions are illustrative in nature and should not be understood as limiting the scope of protection of the claims of the present application.

[0114] It has to be noted that, in the present document, the terms "first", "second", etc. merely serve the purpose of differentiation and do not require or imply any kind of ordering or sequence of the entities or actions associated therewith. Furthermore, the terms "comprising", "containing", etc. are to be interpreted as non- exclusive in the sense that a process, method, article, or apparatus that comprises a list of elements is not necessarily limited to those elements, but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises... a", "contains... a", etc. does not, without further restriction, exclude the presence of additional identical elements in the process, method, article, or apparatus.

[0115] The above description is only specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An artificial intelligence-based unmanned aerial vehicle automatic airport inspection system for glass curtain walls, characterized in that, include: Coordinate fusion unit: Before takeoff, the curtain wall curvature grid and the airspace corridor outline are superimposed under a unified coordinate reference to generate a shared coordinate table and load it into the flight terminal; Track generation unit: Reads the shared coordinate table, generates a hybrid straight-circular-arc track that adheres to the wall according to the principle of minimum curvature, and writes the minimum safety band at each node; Safety monitoring unit: During flight, the distance radar and inertial navigation synchronously output residual distance curves and angular velocity signals, extract key features to perform wall-hugging safety judgments, and if the judgment results exceed the limits, pause the advance and request the latest corridor outline; Track correction unit: After the ground station pushes the latest corridor outline, the path planner corrects the unexecuted trajectory along the normal direction and broadcasts the correction results for tower monitoring. The flight control then releases the pause on updating the trajectory. Experience Iteration Unit: After landing, the flight record is written to the learning library to preset a safe bandwidth for the next mission and reduce online downtime. 2.The artificial intelligence-based unmanned aerial vehicle automatic airport inspection system for glass curtain walls according to claim 1, characterized in that, The coordinate fusion unit includes the following: Before takeoff, the curtain wall curvature mesh containing location coordinates and local curvature values ​​is extracted from the airport building database. The airspace corridor outline composed of polygonal boundary segments is obtained from the airport operations command center. The consistency of point cloud density is verified. A set of reference points is selected to calculate the transformation matrix to minimize the sum of Euclidean distances and thus establish a unified coordinate reference. The projection distance from each curtain wall curvature mesh point to the airspace corridor outline is calculated to identify overlapping areas and generate a fused mesh. A shared coordinate table organized by the points of the fused mesh is extracted. A check code is added for verification via wireless transmission protocol and loaded into the flight terminal storage module. 3.The artificial intelligence-based glass curtain wall unmanned aerial vehicle automatic airport inspection system according to claim 2, characterized in that, The trajectory generation unit includes the following: The flight terminal of the trajectory generation unit extracts the shared coordinate table from the storage module, verifies the integrity of the matrix rows, traverses the curvature values ​​to filter points that exceed the preset threshold as curvature key points, calculates the minimum curvature path between adjacent curvature key points to minimize the overall curvature accumulation and generate a wall-hugging straight-circular-arc hybrid trajectory, calculates the shortest normal distance from each path point to the clearance corridor outline to quantify the clearance margin and obtain the minimum safety belt width value, and appends the minimum safety belt width value to the path point sequence as a floating-point format attribute field.

4. The artificial intelligence-based glass curtain wall unmanned aerial vehicle automatic airport inspection system according to claim 2, Its features are: The shared coordinate table is a matrix, with each row corresponding to a point, including uniform coordinates, curvature values, and contour markers; Unified coordinates refer to the three-dimensional position coordinates of each point in the fused mesh under a unified coordinate reference, including the x, y and z components; The curvature value refers to the local curvature value derived from the curtain wall curvature grid. Points derived from the curtain wall curvature grid directly inherit the original local curvature value, while points derived from the clear corridor outline are filled with zero to indicate no curvature information. The outline marker refers to a binary identifier value, which is assigned a value of one for points belonging to the boundary of the clearance corridor outline and a value of zero for points not belonging to the boundary of the clearance corridor outline. 5.The artificial intelligence-based unmanned aerial vehicle automatic airport inspection system for glass curtain walls of claim 3, characterized in that, The security monitoring unit includes the following components: Residual distance readings are collected by the radar during the flight to form a residual distance curve, and three-axis angular velocity signals are recorded by the inertial navigation sensor. The sequence data are aligned in time and synchronized, and key features are extracted, including the margin contraction speed ratio and the attitude amplitude density. A pre-trained lightweight random forest is used to input the margin contraction speed ratio and the attitude amplitude density to output the attitude stabilization adjustment coefficient and determine whether it exceeds the standard. If it does, the forward request is suspended, the latest clearance corridor profile is requested, and the high-precision coordinates are locked.

6. The artificial intelligence-based unmanned aerial vehicle automatic airport inspection system for glass curtain walls according to claim 5, characterized in that: The margin contraction speed ratio is calculated based on the residual distance curve to obtain the slope difference at adjacent time points, and the maximum negative slope value is selected to quantify the clearance margin reduction rate.

7. The artificial intelligence-based unmanned aerial vehicle automatic airport inspection system for glass curtain walls according to claim 5, characterized in that: The attitude amplitude density applies a short window fast Fourier transform to the three-axis angular velocity signals, integrates the part exceeding the intermediate frequency threshold after obtaining the frequency spectrum amplitude, and evaluates the attitude oscillation strength. 8.The artificial intelligence-based glass curtain wall unmanned aerial vehicle automatic airport inspection system according to claim 5, characterized in that, The trajectory correction unit includes the following: After the ground station pushes the latest clearance corridor profile, the boundary line segment integrity is verified, the unexecuted flight path sub-sequence is extracted based on the current unmanned aerial vehicle position binary search matching index, the normal distance from each point to the latest clearance corridor profile is calculated to translate the point position to generate the corrected unexecuted flight path sub-sequence, and the corrected unexecuted flight path sub-sequence point position and the minimum safety bandwidth value are sent. After the tower station confirms, the original unexecuted flight path sub-sequence is replaced and the suspension is started. 9.The artificial intelligence-based glass curtain wall unmanned aerial vehicle automatic airport inspection system according to claim 8, characterized in that, The experience iteration unit includes the following: After landing, the flight record is extracted and the sequence integrity is verified, the flight record is appended to the learning library and the structured field is defined, the multi-task flight record is queried to calculate the preset safety bandwidth intention index decay, the minimum safety bandwidth value of the shared coordinate table is updated and version marking control is performed, and the buffer threshold limit is expanded during the flight.

Images