An unmanned aerial vehicle hanging line self-adaptive alignment system and method based on AI vision

Abstract

The application relates to the field of AI vision technology and discloses an unmanned aerial vehicle wire hanging self-adaptive alignment system and method based on AI vision, wherein the unmanned aerial vehicle wire hanging self-adaptive alignment system based on AI vision comprises the following steps: target guide wire extraction; normal deviation calculation; predicted tangential difference calculation; capture window shrinkage coefficient calculation; target wire hanging execution point position and target yaw angle updating; and outputting a control instruction to complete alignment. The application establishes an integrated closed loop of perception, modeling, prediction, constraint and control around the mechanical hangability of the wire hanging tail section, identifies the target guide wire, the wire hanging soft wire and the wire hanging port at the same time, restores local three-dimensional geometry, constructs a dynamic tangential capture window, and generates the target wire hanging execution point position, the target yaw angle and the final control instruction according to the dynamic tangential capture window, so that wire hanging control is transformed from pure image alignment into geometric alignment facing the line entry channel, and the application is suitable for the automatic implementation of unmanned aerial vehicle wire hanging operation of a power transmission line.

Description

Technical Field

[0001] This invention relates to the field of AI vision technology, and more specifically, to an AI vision-based adaptive alignment system and method for drones with attached wires. Background Technology

[0002] With the continuous expansion of power transmission lines, power outages for maintenance and grounding line installation are becoming increasingly frequent in complex scenarios such as mountainous areas, river crossings, and valley crossings. Traditional line installation methods typically rely on manual observation, adjustment, and connection by workers at high or near-line locations, which presents problems such as high labor intensity, high operational risks, and poor environmental adaptability. In recent years, drones have gradually entered the fields of power transmission line inspection, installation, and auxiliary construction, making it possible to move near-line operations from ground-based manual work to a combination of airborne vision and flight control.

[0003] Existing UAV vision solutions related to power transmission lines primarily focus on conductor detection, line tracking, defect identification, or position maintenance. Their control is typically based on center deviation, slope deviation, lateral displacement, or attitude error in the image. While these solutions are suitable for inspection and line-following flights, they still have significant limitations in scenarios involving the final stages of line installation. This is because the line installation action is not simply about approaching a conductor; it requires the installation opening to enter a local entry channel formed by the target conductor, the installation cable, and the mechanical opening in a suitable direction. If only image overlap or minimizing positional deviation is used as the control objective, it is difficult to simultaneously input the relationship between the local tangential direction of the target conductor, the local curvature of the target conductor, and the entry direction of the installation cable.

[0004] In the actual hanging section, the target conductor exhibits local sag and curvature, while the hanging wire is a flexible material. Both undergo asynchronous changes under the combined effects of crosswinds, downwash, and control delays. Even if the hanging point is close to the target conductor's center in the image, the actual hanging window in three-dimensional space may continue to shrink due to orientation mismatch, local bending, and future drift. Therefore, existing solutions do not explicitly quantify the mechanical hanging capability of the hanging section, which can easily lead to situations where the wire appears aligned in the image but mechanically fails to hang successfully on the target conductor. Summary of the Invention

[0005] This invention provides an AI vision-based adaptive alignment system and method for drones with attached wires, solving the technical problems mentioned in the background.

[0006] This invention provides an AI vision-based adaptive alignment system for drones with attached wires, comprising:

[0007] The first module uses binocular vision to extract the centerline point set of the target wire, the hanging wire, and the hanging point.

[0008] The second module reconstructs the three-dimensional geometry based on the centerline point set within the local hanging plane, fits the catenary to obtain the local tangential angle and local curvature of the target traverse, fits the equivalent tangent to obtain the equivalent tangential angle, and calculates the normal deviation of the hanging execution point relative to the target traverse.

[0009] The third module calculates the current tangential difference and the rate of change of tangential difference using the equivalent tangential angle and the local tangential angle of the target traverse, and calculates the predicted tangential difference by combining the current tangential difference and the rate of change of tangential difference.

[0010] The fourth module calculates the dynamic tangential capture window based on the nominal mechanical opening width, predicted tangential difference, tangential difference change rate, and local curvature of the target conductor, and uses the dynamic tangential capture window and the nominal mechanical opening width to calculate the capture window contraction coefficient.

[0011] The fifth module calculates the normal compensation, vertical compensation, and yaw compensation based on the normal deviation, the local curvature of the target traverse, and the predicted tangential difference, and then updates the target traverse execution point position and the target yaw angle.

[0012] The sixth module generates the desired propulsion speed under the constraint of the capture window contraction coefficient. It then generates the final control command by combining the desired propulsion speed, normal compensation, vertical compensation, and yaw compensation. The final pose result, consisting of the target hook execution point position and the target yaw angle, and the final control command are output to complete the alignment.

[0013] This invention provides an AI vision-based adaptive alignment method for drone wiring, comprising the following steps:

[0014] Step S1: Use binocular vision to extract the centerline point set of the target wire, the hanging wire, and the hanging point;

[0015] Step S2: Reconstruct the three-dimensional geometry based on the centerline point set in the local hanging plane, fit the catenary to obtain the local tangential angle and local curvature of the target traverse, fit the equivalent tangent to obtain the equivalent tangential angle, and calculate the normal deviation of the hanging execution point relative to the target traverse.

[0016] Step S3: Calculate the current tangential difference and the rate of change of tangential difference using the equivalent tangential angle and the local tangential angle of the target traverse; and calculate the predicted tangential difference by combining the current tangential difference and the rate of change of tangential difference.

[0017] Step S4: Calculate the dynamic tangential capture window based on the nominal mechanical opening width, predicted tangential difference, tangential difference change rate, and local curvature of the target conductor; and calculate the capture window contraction coefficient using the dynamic tangential capture window and the nominal mechanical opening width.

[0018] Step S5: Calculate the normal compensation, vertical compensation, and yaw compensation based on the normal deviation, local curvature of the target traverse, and predicted tangential difference, and then update the target traverse execution point position and target yaw angle.

[0019] Step S6: Generate the desired propulsion speed under the constraint of the capture window contraction coefficient. Combine the desired propulsion speed, normal compensation, vertical compensation and yaw compensation to generate the final control command. Output the final pose result consisting of the target hook execution point position and the target yaw angle, as well as the final control command to complete the alignment.

[0020] The beneficial effects of this invention are as follows: This invention establishes an integrated closed loop of perception, modeling, prediction, constraint, and control around the mechanical attachability of the end section of the wire. By simultaneously identifying the target conductor, the wire attaching cable, and the wire attachment point, it recovers the local three-dimensional geometry, constructs a dynamic tangential capture window, and generates the target wire attachment execution point position, target yaw angle, and final control command accordingly. This transforms wire attachment control from simple image alignment to geometric alignment oriented towards the wire insertion channel. Therefore, it can maintain consistency between the propulsion strategy and window margin under conditions of orientation mismatch, local bending, and control delay, and ensures that the pose and control results are simultaneously traceable and reproducible. This invention is suitable for the automated implementation of UAV wire attachment operations on power transmission lines. Attached Figure Description

[0021] Figure 1 This is a calculation flowchart of an AI vision-based drone wire-attached adaptive alignment system according to the present invention.

[0022] Figure 2 This is a schematic diagram of a computational scenario for an AI vision-based drone wire-attached adaptive alignment system according to the present invention. Detailed Implementation

[0023] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0024] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of the present invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of the present invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" indicate that the element or object preceding the term encompasses the elements or objects listed following the term and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0025] like Figures 1-2 As shown, an AI vision-based adaptive alignment system for drones with attached wires includes:

[0026] The first module uses binocular vision to extract the centerline point set of the target wire, the hanging wire, and the hanging point.

[0027] The second module reconstructs the three-dimensional geometry based on the centerline point set within the local hanging plane, fits the catenary to obtain the local tangential angle and local curvature of the target traverse, fits the equivalent tangent to obtain the equivalent tangential angle, and calculates the normal deviation of the hanging execution point relative to the target traverse.

[0028] The third module calculates the current tangential difference and the rate of change of tangential difference using the equivalent tangential angle and the local tangential angle of the target traverse, and calculates the predicted tangential difference by combining the current tangential difference and the rate of change of tangential difference.

[0029] The fourth module calculates the dynamic tangential capture window based on the nominal mechanical opening width, predicted tangential difference, tangential difference change rate, and local curvature of the target conductor, and uses the dynamic tangential capture window and the nominal mechanical opening width to calculate the capture window contraction coefficient.

[0030] The fifth module calculates the normal compensation, vertical compensation, and yaw compensation based on the normal deviation, the local curvature of the target traverse, and the predicted tangential difference, and then updates the target traverse execution point position and the target yaw angle.

[0031] The sixth module generates the desired propulsion speed under the constraint of the capture window contraction coefficient. It then generates the final control command by combining the desired propulsion speed, normal compensation, vertical compensation, and yaw compensation. The final pose result, consisting of the target hook execution point position and the target yaw angle, and the final control command are output to complete the alignment.

[0032] In one embodiment of the present invention, the extraction of the centerline point set of the target conductor, the hanging wire, and the hanging opening using binocular vision includes:

[0033] The left and right camera images are input into a joint segmentation network to segment and obtain the target wire mask, the hanging wire mask, and the hanging opening mask. The calculation formula is as follows:

[0034] ;

[0035] In the formula, Image from the left camera. Image from the right camera. For joint segmentation networks, This is the set of training parameters for the joint segmentation network. The target wire mask corresponding to the left camera image. The soft wire mask corresponding to the left camera image. The mask for the hanging hole corresponding to the left camera image. The target guideline mask corresponding to the right camera image. The soft-wire mask corresponding to the right camera image. The mask for the hanging hole corresponding to the image from the right camera;

[0036] The target conductor mask, the hanging wire mask, and the hanging opening mask are refined separately to extract the target conductor centerline point set. Center point set of hanging wire and the center line point set of the hanging port .

[0037] It should be noted that the left camera image is the raw operational view data output by the left imaging unit at the current sampling time. It provides the texture, edges, and relative positional relationships of the target conductor, hanging cable, and hanging point from the left perspective. This data can be continuously acquired by a left industrial camera installed near the hanging mechanism and whose intrinsic parameters are calibrated and time-synchronized. The right camera image is the raw operational view data output by the right imaging unit at the current sampling time. This data can be continuously acquired by a right industrial camera installed in pair with the left industrial camera and whose time is synchronized. The joint segmentation network is a visual recognition model that simultaneously receives images from both left and right perspectives and outputs multi-target pixel-level classification results. It is used to separate the target conductor, hanging cable, and hanging point from complex backgrounds. The training parameter set of the joint segmentation network is the set of weight parameters, bias parameters, and normalization parameters that are fixed after offline training. The target conductor mask corresponding to the left camera image is the set of pixel regions in the left camera image that are identified as the target conductor. The hanging cable mask corresponding to the left camera image is the set of pixel regions in the left camera image that are identified as the hanging cable. The mask for the wire hanging opening in the left camera image is the set of pixel regions in the left camera image that are identified as wire hanging openings. The mask for the target conductor in the right camera image is the set of pixel regions in the right camera image that are identified as target conductors. The mask for the flexible wire hanging in the right camera image is the set of pixel regions in the right camera image that are identified as flexible wire hanging. The mask for the wire hanging opening in the right camera image is the set of pixel regions in the right camera image that are identified as wire hanging openings. The set of centerline points for the target conductor is a continuous sequence of center pixels obtained after thinning or center extraction of the corresponding mask. The set of centerline points for the flexible wire hanging is a continuous sequence of center pixels obtained after thinning or center extraction of the corresponding mask. The set of centerline points for the wire hanging opening is a continuous sequence of center pixels obtained after thinning or center extraction of the corresponding mask.

[0038] It should be noted that the joint segmentation network is constructed as follows: a segmentation structure using a dual-input shared encoder and a three-branch decoder is employed. First, time synchronization, distortion correction, and epipolar constraint preparation are performed on the left and right camera images. Then, the two images are input into a shared backbone network to extract shallow texture features and deep semantic features. Subsequently, left and right view features are aligned in the intermediate fusion layer. Finally, three types of probability maps are output from the target conductor branch, the hanging wire branch, and the hanging point branch, respectively. The backbone network can use a lightweight residual network, and the decoder can use a stepwise upsampling and skip connection structure. The joint segmentation network is trained as follows: paired binocular images are acquired in a ground test field and a real transmission line environment. The target conductor region, hanging wire region, and hanging point region are labeled frame by frame to form training, validation, and test sets. During training, class loss, boundary loss, and left-right consistency loss are jointly optimized, and brightness perturbation, blur perturbation, background replacement, local occlusion, and wind-induced deformation enhancement are added to compensate for data biases caused by mountain backlighting, cloud backgrounds, and tower material interference.

[0039] It should be noted that the refinement process is as follows: First, connected component filtering is performed on the segmentation results, retaining the main connected regions that meet the constraints of length, area, and spatial position. Then, morphological refinement is performed on the target wire mask and the hanging wire mask to obtain the single-pixel width centerline. Next, spur removal, short branch removal, and breakpoint bridging are performed on the centerline to ensure the continuity of the point sequence. For the hanging opening mask, the two sides of the opening are extracted first, and then the centerline of the hanging opening is obtained based on the centerlines of the two sides of the boundary. The definition of the hanging opening centerline point set is as follows: when the hanging opening is a long and narrow groove in the current view, the equidistant centerlines of the left and right boundaries of the opening are used as the hanging opening centerline point set; when the hanging opening degenerates into an approximate point shape due to view or distance, several equidistant sampling points are constructed in the direction of the opening, starting from the geometric center of the opening, to form a short centerline point set. This can cover both frontal view and oblique view and partial occlusion conditions.

[0040] In one embodiment of the present invention, three-dimensional geometry is reconstructed based on a centerline point set within a local catenary plane; a catenary is fitted to obtain the local tangential angle and local curvature of the target traverse; an equivalent tangent is fitted to obtain the equivalent tangential angle; and the normal deviation of the catenary execution point relative to the target traverse is calculated, including:

[0041] The formula for calculating the coordinates of a local 3D point is:

[0042] ;

[0043] ;

[0044] ;

[0045] In the formula, To match point disparity, For the focal length of a binocular camera, The baseline length of the binocular camera. For depth, The horizontal pixel coordinates are Vertical pixel coordinates The coordinates of the horizontal principal point are... The coordinates of the vertical principal point are... The equivalent focal length is horizontal. For the longitudinal equivalent focal length, The horizontal position of the camera coordinate system. The vertical position of the camera coordinate system;

[0046] The formula for calculating the catenary height function by fitting a catenary is as follows:

[0047] ;

[0048] In the formula, This represents the horizontal coordinate of the local hanging line plane. It is a function of the catenary height. For the sag parameters of the catenary, The parameter represents the center position of the catenary. These are the vertical translation parameters;

[0049] The formulas for calculating the local tangential angle and local curvature of the target traverse are as follows:

[0050] ;

[0051] ;

[0052] In the formula, The horizontal coordinate of the target traverse corresponding to the nearest neighbor of the hanging execution point. The local tangential angle of the target traverse. Let the local curvature of the target conductor be denoted by ; the formulas for calculating the local linear approximation function and equivalent tangential angle of the fitted wire are as follows:

[0053] ;

[0054] ;

[0055] In the formula, This is a local linear approximation function for the hanging wire. The local equivalent slope of the hanging wire. The bias is used for linear fitting of the hanging wire. This is the equivalent tangential angle;

[0056] The formula for calculating the normal deviation is:

[0057] ;

[0058] In the formula, This is the current location of the wire execution point. The location of the nearest point on the target traverse. Let the local normal unit vector of the target traverse be denoted as . This is the normal deviation.

[0059] It should be noted that the matching point disparity is the pixel position difference of the same spatial point along the epipolar direction in the left and right views, reflecting the distance of the spatial point from the camera. The binocular camera focal length is an intrinsic parameter in the binocular imaging model, characterizing the imaging scale relationship, used to establish a definite mapping between disparity and spatial depth, and can be obtained through offline calibration using a binocular calibration board. The binocular camera baseline length is a fixed distance between the optical centers of the left and right cameras, used to construct the geometric scale for solid triangulation, and can be obtained through a combination of mechanical measurement and binocular calibration. Depth is the spatial distance of the matching point along the camera's optical axis, used to characterize how far the target point is from the camera. The lateral pixel coordinates are the horizontal position of the matching point in the calibrated image, used in calculating the lateral spatial position. The longitudinal pixel coordinates are the vertical position of the matching point in the calibrated image, used in calculating the longitudinal spatial position. The lateral principal point coordinates are the horizontal reference position of the camera's optical axis after it is projected onto the image plane, used to achieve center alignment between pixel coordinates and camera coordinates, and can be obtained through binocular calibration. The vertical principal point coordinates are the vertical reference position of the camera's optical axis after it is projected onto the image plane. They are used to center-align the pixel coordinates to the camera coordinates and can be obtained through binarized image mapping. The horizontal equivalent focal length is the camera's focal length parameter at the horizontal pixel scale, used to perform horizontal spatial position conversions and can be obtained through binarized image mapping. The vertical equivalent focal length is the camera's focal length parameter at the vertical pixel scale, used to perform vertical spatial position conversions and can be obtained through binarized image mapping. The horizontal position in the camera coordinate system is the horizontal spatial position of the matching point within the camera coordinate system. The vertical position in the camera coordinate system is the vertical spatial position of the matching point within the camera coordinate system.

[0060] It should be noted that the local 3D point coordinates are spatial points composed of depth, lateral position in the camera coordinate system, and longitudinal position in the camera coordinate system, used to reconstruct the true local geometry of the target traverse and the hanging line. The lateral coordinates of the local hanging plane are independent variables describing the lateral position change of the traverse within the local hanging plane, used as the independent variable of the catenary function. The catenary height function describes the relationship between the height of the target traverse and its lateral position within the local hanging plane, used to express the local sag shape. The catenary sag parameter is a fitting parameter controlling the local sag and flare of the target traverse, used to determine the shape of the catenary. The catenary center position parameter is a lateral parameter characterizing the local center position of the catenary, used to determine the lateral translation of the local curve in the plane. The vertical translation parameter is a fitting parameter for adjusting the vertical position of the overall catenary height function, used to ensure that the fitted curve is consistent with the actual traverse height. The lateral coordinates of the target traverse corresponding to the nearest neighbor of the hanging execution point are the lateral calculation positions on the target traverse with the smallest distance to the current hanging execution point, used to calculate the local tangential angle and local curvature most relevant to the current hanging action. The local tangential angle of the target traverse is the local orientation angle of the target traverse at its nearest neighbor evaluation position, used to characterize the current extension direction of the target traverse. The local curvature of the target traverse is the degree of bending of the target traverse at its nearest neighbor evaluation position, used to characterize the degree to which a local arc segment occupies the mechanical channel. The local linear approximation function of the hanging wire is a function that linearly expresses the geometry of the hanging wire in a short interval near the hanging point, used to determine the current entry direction. The local equivalent slope of the hanging wire is the slope parameter of the local linear approximation function of the hanging wire, used to characterize the rate of change of direction in the short interval of the hanging wire. The linear fitting bias of the hanging wire is the intercept parameter of the local linear approximation function of the hanging wire, used to compensate for the overall position of the fitted line in the plane. The equivalent tangential angle is the orientation angle obtained by converting the local equivalent slope of the hanging wire, used to characterize the current entry direction of the hanging wire. The spatial difference vector is the three-dimensional position difference vector between the current hanging execution point position and the nearest point position of the target traverse, used for subsequent projection to obtain the normal deviation. The current wire-attaching execution point position is the spatial position of the current wire-attaching point center in the unified working coordinate system, used to characterize the actual execution point location. The target traverse nearest point position is the spatial point on the target traverse with the smallest distance from the current wire-attaching execution point position, used as a reference point for normal deviation. The target traverse local normal unit vector is a normalized normal vector orthogonal to the local tangential direction of the target traverse, used to define the lateral correction direction. Normal deviation is the signed deviation of the current wire-attaching execution point position relative to the target traverse nearest point position in the direction of the target traverse local normal unit vector, used to measure the degree of lateral misalignment.

[0061] It should be noted that the local hanging line plane is established as follows: First, the global vertical reference direction is determined based on the attitude calculation results and gravity direction output by the flight control. Then, the local tangential direction at the nearest point of the target conductor is used as the first basis vector of the local hanging line plane. The global vertical reference direction is projected onto a plane orthogonal to this tangential direction and normalized to obtain the local plane vertical direction unit vector. Finally, the local hanging line plane is spanned by the local tangential direction and the local plane vertical direction unit vector. The matching method for the centerline point set is as follows: First, epipolar correction is performed on the left and right images so that corresponding points are mainly distributed near the same image row. Then, using the target conductor centerline point set and the hanging line centerline point set as constraints, matching is performed only within a narrow epipolar search band. The matching cost can be jointly scored using local grayscale similarity, gradient similarity, and semantic consistency. For multiple candidate points, matching results with tangential continuity and smooth disparity changes are prioritized. The catenary fitting method is as follows: first, outlier removal and scale normalization are performed on the local 3D point coordinates; then, the initial values ​​of the catenary sag parameters, catenary center position parameters, and vertical translation parameters are solved using the least squares objective function; subsequently, robust iterative optimization is used to suppress the influence of occlusion points and mismatched points. When the amount of data is small, the initial values ​​can be estimated using local quadratic curves first, and then the model can be switched to refine the catenary model. The determination of the target traverse lateral coordinates corresponding to the nearest neighbor of the catenary execution point is as follows: a local search interval is constructed within the domain of the catenary height function; after projecting the current catenary execution point position onto the local catenary plane, the Euclidean distance from it to the discrete sampling point or analytical function point of the catenary is calculated, and the lateral coordinates corresponding to the minimum distance are taken as the nearest neighbor lateral coordinates; to suppress jitter, continuity constraints can be applied to the evaluation results of adjacent frames.

[0062] It should be noted that the window selection method for local linear fitting of the suspended wire is as follows: a fixed-length window is selected outward from the center of the suspended wire opening along the centerline point set of the suspended wire, prioritizing the retention of the nearest effective visible wire point outside the opening; the window length is adaptively adjusted according to the resolution, shooting distance, and visible wire length to prevent slope jitter caused by an excessively short window and to avoid introducing non-entry segment oscillations into the fitting due to excessively long sampling. The method for obtaining the local normal unit vector of the target traverse is as follows: first, a unit tangential vector is constructed based on the local tangential angle of the target traverse; then, this unit tangential vector is rotated 90 degrees and normalized in the local suspended wire plane to obtain the local normal unit vector of the target traverse; to maintain directional consistency, the normal direction can be specified to always point to the side where the current suspended wire execution point is located. The current wire-attaching execution point position is obtained as follows: First, the aircraft's position in the world coordinate system is estimated from the flight control status. Then, the fixed offset of the wire-attaching mechanism's center of mass relative to the aircraft's center of mass is calculated by combining the rigid external parameters of the wire-attaching mechanism. Finally, this offset is transformed to the world coordinate system and superimposed on the aircraft's position to obtain the current wire-attaching execution point position. The external parameters can be pre-measured on the ground using a calibration fixture. The nearest point position of the target traverse is obtained as follows: After obtaining the nearest neighbor's lateral coordinates, these lateral coordinates are substituted into the catenary height function to obtain the target point in the local wire-attaching plane. Then, the basis vectors of the local wire-attaching plane are combined to restore the three-dimensional position in the world coordinate system, which is taken as the nearest point position of the target traverse.

[0063] In one embodiment of the present invention, the current tangential difference and the rate of change of tangential difference are calculated using the equivalent tangential angle and the local tangential angle of the target traverse, and the predicted tangential difference is calculated by combining the current tangential difference and the rate of change of tangential difference, including:

[0064] The formula for calculating the current tangential difference is:

[0065] ;

[0066] In the formula, For the equivalent tangential angle, The local tangential angle of the target traverse. This represents the current tangential difference;

[0067] The formula for calculating the rate of change of tangential difference is:

[0068] ;

[0069] In the formula, The difference in tangential direction at the previous moment. The sampling time interval, This represents the rate of change of tangential difference.

[0070] The formula for calculating the predicted tangential difference is:

[0071] ;

[0072] In the formula, To control the total latency, To predict the tangential difference.

[0073] It should be noted that the current tangential difference is the instantaneous directional deviation of the equivalent tangential angle relative to the local tangential angle of the target guideline, used to characterize whether the current incoming direction is consistent with the target guideline direction. The sampling time interval is the time interval between two consecutive updates of the tangential difference by the system, used to convert the change in directional difference into a change per unit time. A preferred value is 0.02 to 0.05 seconds, a range that can accommodate the refresh rate of airborne vision inference, the flight control cycle, and the dynamic response requirements of the final stage of the guideline connection. The tangential difference change rate is the change per unit time of the current tangential difference relative to the tangential difference at the previous moment, used to reflect whether the directional error is currently increasing or decreasing. The total control delay is the total time from the completion of image acquisition to the actual application of the corresponding control command to the UAV, used for delay compensation of the directional error. The time-delay tangential drift is the additional directional deviation formed by the continued accumulation of the tangential difference change rate within the total control delay, used to incorporate future drift into the current decision. The predicted tangential difference is the estimated future tangential difference value used for the actual control decision after considering the total control delay.

[0074] It should be noted that the cache update method for the tangential difference at the previous moment is as follows: before the end of each control cycle, the filtered current tangential difference and its timestamp are written into the circular buffer. In the next cycle, the most recent valid value is read as the tangential difference at the previous moment. When the current frame identification or matching fails, the previous frame cache is retained and an effective flag is used to prohibit abnormal differences, preventing false jumps in the direction change rate. The method for measuring the total control delay is as follows: a unified clock timestamp is recorded at five nodes: image exposure completion, image inference, inference output completion, control command issuance, and flight control execution receipt. After continuously collecting multiple sets of samples, the average value or high quantile value is calculated as the total control delay. If the airborne computing load changes significantly, a delay lookup table can be established according to resolution and flight mode. The limiting and outlier suppression methods for predicted tangential difference are as follows: First, median filtering or first-order low-pass filtering is applied to the current tangential difference and the rate of change of tangential difference. Then, a physically acceptable upper limit is set for the predicted tangential difference. If the value exceeds the limit, the output is based on the limiting value. For single-frame spikes caused by instantaneous misidentification, a substantial update can be made only after two or more consecutive frames of change in the same direction. The equivalent tangential angle represents the current entry direction of the suspended wire, while the local tangential angle of the target conductor represents the current extension direction of the target conductor. Therefore, the equivalent tangential angle minus the local tangential angle of the target conductor is defined as the current tangential difference, which can immediately determine whether the current entry direction is approaching the direction of the target conductor. For example, if the center position is already close but the current tangential difference is still large, the wire insertion point will quickly brush past the edge of the conductor instead of entering along the channel. In addition, the directional error of the end section of the suspended wire will continue to evolve with wind disturbance, wire sway, and machine response. Therefore, the rate of change of tangential difference is obtained by discretizing the current tangential difference with the tangential difference at the previous moment, which can directly indicate whether the directional relationship is deteriorating or has converged.

[0075] In one embodiment of the present invention, a dynamic tangential capture window is calculated based on the nominal mechanical aperture width, the predicted tangential difference, the rate of change of the tangential difference, and the local curvature of the target conductor. The capture window contraction coefficient is then calculated using the dynamic tangential capture window and the nominal mechanical aperture width, including:

[0076] The formula for calculating the dynamic tangential capture window is as follows:

[0077] ;

[0078] In the formula, The nominal mechanical opening width, For effective entry depth, To predict the tangential difference, For the local curvature of the target conductor, The current visible arc length, To control the total latency, The rate of change of tangential difference. For dynamic tangential capture window;

[0079] The formula for calculating the capture window shrinkage coefficient is as follows:

[0080] ;

[0081] In the formula, This is the capture window shrinkage factor.

[0082] It should be noted that the nominal mechanical opening width is the nominal opening width of the wire-attaching mechanism in its static design state. It serves as the initial upper limit of the dynamic tangential capture window, and is preferably set to 1.3 to 1.8 times the outer diameter of the target conductor. A smaller opening will make simultaneous positioning difficult, while a larger opening will reduce guiding constraints and mechanical compactness. The effective penetration depth is the equivalent geometric depth that the target conductor needs to travel from initial contact with the wire-attaching opening to reaching a stable attachment position. It measures how much the directional deviation is amplified during penetration. A preferred value is 0.8 to 1.5 times the nominal mechanical opening width, a range that balances guiding stability and structural compactness. The directional misalignment erosion width is the amount of loss to the actual penetration width caused by the projection of the predicted tangential difference onto the effective penetration depth. It represents the mechanical window erosion caused by directional inconsistency. The current visible arc length is the length of the target conductor arc segment reliably identified and involved in local curvature calculation in the current frame, reflecting the observation range upon which the current local curvature calculation depends. Local bending erosion width is the amount of width encroachment caused by the local curvature of the target conductor on the straight entry channel within the currently visible arc length, used to represent the occupancy of the hanging port channel by the bending arc segment. Time-delay drift erosion width is the additional loss to the entry width caused by the continued change of the tangential difference rate within the total control time delay, used to pre-deduct future drift erosion. Dynamic tangential capture window is the remaining effective width that can still guarantee the target conductor's entry into the hanging port at the current moment, used as a core quantity for the allowable advance along the line. Capture window contraction coefficient is the normalized ratio of the dynamic tangential capture window to the nominal mechanical opening width, used to unify the advance constraint scale under different mechanism dimensions.

[0083] It should be noted that the nominal mechanical opening width is calibrated as follows: After the wire-hanging mechanism is assembled, the minimum passage width between the effective boundaries on both sides of the wire-hanging opening is measured using gauge blocks, calipers, or a visual ruler. This measurement is repeated multiple times under no-load and standard working conditions, and the average value is taken as the nominal mechanical opening width. If the wire-hanging opening exhibits elastic deformation, the working opening value should also be measured within the designed clamping force range. The effective wire penetration depth is determined as follows: Using a standard wire sample with the same outer diameter as the target wire, on a ground test bench, starting from the first contact of the wire with the wire-hanging opening boundary, the wire is advanced along the designed wire-hanging direction until it reaches a stable connection position. The equivalent advancement distance of this process is recorded as the effective wire penetration depth. The test can be repeated under different advancement speeds and wire outer diameters, and the safe side value is taken. The current visible arc length is calculated as follows: In the current frame, the distance between adjacent spatial points is calculated and accumulated according to the point order of the target conductor centerline point set. Only the local main arc segments that meet the confidence threshold, are unobstructed, and have continuous curves are retained as the current visible arc length. If there are breaks in the point set, they can be bridged first before calculating the length of the main arc segment. The handling method when the dynamic tangential capture window is less than or equal to zero is as follows: When the dynamic tangential capture window is less than or equal to zero, the non-advancement status flag is immediately set, the advance along the line is stopped, and only normal compensation, vertical compensation, and yaw compensation are performed. If necessary, a small retreat or hovering repositioning is performed. Only after several consecutive frames have recovered to a positive value can advance along the line be allowed again. The saturation constraint method of the capture window contraction coefficient is as follows: First, the original capture window contraction coefficient is calculated according to the ratio of the dynamic tangential capture window to the nominal mechanical opening width, and then it is restricted to between zero and one. When it is less than zero, it is taken as zero, and when it is greater than one, it is taken as one, to ensure that the advance speed calculation will not be unnecessarily amplified due to noise or fitting error.

[0084] In one embodiment of the present invention, the normal compensation, vertical compensation, and yaw compensation are calculated based on the normal deviation, the local curvature of the target traverse, and the predicted tangential difference, respectively, and then the target traverse execution point position and the target yaw angle are updated, including:

[0085] The formulas for calculating the normal compensation, vertical compensation, and yaw compensation are as follows:

[0086] ;

[0087] ;

[0088] ;

[0089] In the formula, For normal deviation, For normal compensation amount, For the local curvature of the target conductor, The current visible arc length, This is the vertical compensation amount. To predict the tangential difference, This is the yaw compensation amount;

[0090] The formula for calculating the updated target hanging point position is as follows:

[0091] ;

[0092] In the formula, This is the current location of the wire execution point. Let the local normal unit vector of the target traverse be denoted as . Let be the unit vector in the vertical direction of the local plane. The target is the location of the execution point;

[0093] The formula for calculating the updated target yaw angle is:

[0094] ;

[0095] In the formula, The current yaw angle, The target yaw angle.

[0096] It should be noted that the normal compensation amount is a correction amount used to offset the normal deviation along the local normal unit vector direction of the target guideline, used to laterally pull the wire-attaching execution point back to the vicinity of the target guideline. The vertical compensation amount is a correction amount used to correct the local bending and encroachment of the guideline along the local plane vertical unit vector direction, used to ensure that the guideline still forms sufficient vertical accommodation space in front of the wire-attaching opening. The yaw compensation amount is an orientation correction amount used to offset the predicted tangential difference, used to achieve consistent entry direction with the guideline while correcting the position. The local plane vertical unit vector is the vertical direction basis vector in the local wire-attaching plane, used to define the application direction of the vertical compensation amount. The target wire-attaching execution point position is the spatial position of the target center of the wire-attaching opening obtained after superimposing the normal compensation amount and the vertical compensation amount, used as the spatial tracking target of the subsequent controller. The current yaw angle is the current orientation angle of the UAV around the vertical axis, used as the superposition reference of the yaw compensation amount, and can be obtained from the inertial measurement results and heading calculation results of flight control fusion. The target yaw angle is the target orientation angle obtained by adding yaw compensation to the current yaw angle, and is used to indicate the angle that the cable hanger should ultimately point to.

[0097] It should be noted that the method for establishing the vertical unit vector of the local plane is as follows: first, obtain the unit vector of the gravity direction in the world coordinate system; then, project this unit vector onto the local hanging plane orthogonal to the local tangential direction of the target traverse; normalize the projection result to obtain the vertical unit vector of the local plane; if the projection result is close to zero, the result of the previous valid frame is used for smooth transition. The method for unifying the coordinates of the target hanging execution point position is as follows: express the current hanging execution point position, the local normal unit vector of the target traverse, and the vertical unit vector of the local plane in the same world coordinate system, and then perform vector superposition to obtain the target hanging execution point position; if the previous result comes from the camera coordinate system or the local hanging plane coordinate system, coordinate transformation should be completed first through the extrinsic parameter matrix. The method for acquiring the current yaw angle and defining the direction is as follows: the flight control unit integrates the inertial measurement unit, magnetic heading information, and positioning heading information to output the current yaw angle, and uniformly stipulates that counterclockwise is the positive direction or uses the existing positive direction of the flight control interface. The limiting method for the target wire attachment point position and the target yaw angle is as follows: based on the allowable geometric travel of the wire attachment mechanism, the safe distance of the UAV to the wire, and the upper limit of the allowable angular velocity of the flight control, the maximum change in the target wire attachment point position and the target yaw angle in a single cycle is set; when the upper limit is exceeded, the limit value is applied in segments to prevent the wire swing from being amplified due to excessive correction at one time.

[0098] In one embodiment of the present invention, a desired propulsion speed is generated under the constraint of the capture window contraction coefficient. The desired propulsion speed, normal compensation, vertical compensation, and yaw compensation are combined to generate a final control command. The final pose result, consisting of the target alignment execution point position and the target yaw angle, and the final control command are output to complete the alignment, including:

[0099] The formula for calculating the desired propulsion speed is:

[0100] ;

[0101] In the formula, To achieve the maximum advance speed along the route, The capture window contraction factor, This is the result of the positive part operator extraction operation performed on the capture window contraction coefficient. To achieve the desired speed of advancement;

[0102] The formulas for calculating the normal commanded velocity, vertical commanded velocity, and desired yaw rate are as follows:

[0103] ;

[0104] ;

[0105] ;

[0106] In the formula, For normal control gain, For normal compensation amount, For normal command speed, For vertical control gain, This is the vertical compensation amount. For vertical command speed, For yaw control gain, This is the yaw compensation amount. The desired yaw rate;

[0107] The formula for calculating the desired velocity vector of a computer system is:

[0108] ;

[0109] ;

[0110] In the formula, Let be the local tangential unit vector of the target traverse. Let the local normal unit vector of the target traverse be denoted as . Let be the unit vector in the vertical direction of the local plane. The desired velocity vector in the world coordinate system. Let be the transpose of the rotation matrix from the body coordinate system to the world coordinate system. The desired velocity vector of the machine system;

[0111] The formula for calculating the final control command is as follows:

[0112] ;

[0113] In the formula, The lateral velocity command component of the machine system. For the longitudinal velocity command component of the machine system, The vertical velocity command component of the machine system. This is the final control command;

[0114] The formula for calculating the final pose result is as follows:

[0115] ;

[0116] In the formula, The horizontal coordinate of the target line. The vertical coordinate of the target line is... The vertical coordinates of the target line are given. For the target yaw angle, This represents the final pose result.

[0117] It should be noted that the non-negative truncation coefficient is a coefficient not less than zero obtained by performing a positive part operator extraction operation on the capture window contraction coefficient. It is used to ensure that the desired advance speed does not exhibit reverse anomalies due to a negative window. The maximum advance speed along the target conductor is the upper limit speed at which the system is allowed to advance along the target conductor direction. It is used to limit the maximum advance intensity at the end of the wire-attaching phase, with a preferred value of 0.1 m / s to 0.4 m / s. This means that during the close-range wire-attaching phase, execution efficiency, airflow disturbance, and flight control speed tracking capabilities need to be considered. The desired advance speed is the advance speed along the local tangential unit vector direction of the target conductor, generated under the constraint of the capture window contraction coefficient. It is used to determine whether to guide the system to continue forward into the wire-attaching point. The normal control gain is a proportional coefficient that maps the normal compensation amount to the normal command speed. It is used to adjust the lateral correction speed, with a preferred value of 0.6 to 1.8. This means that normal correction needs to converge relatively quickly, but excessive gain can easily amplify image jitter and cause lateral oscillations. The normal command speed is the normal speed command obtained by mapping the normal compensation amount to the normal control gain. It is used to drive the wire-attaching execution point to converge towards the normal center of the conductor. The vertical control gain is a proportional coefficient that maps the vertical compensation amount to the vertical command velocity. It is used to adjust the intensity of vertical clearance correction, with a preferred value of 0.4 to 1.5. This means that vertical compensation mainly absorbs curvature encroachment, and its dynamic change is usually slower than normal correction. The vertical command velocity is the vertical velocity command obtained after mapping the vertical compensation amount to the vertical control gain, used to correct the vertical position of the cable hanger. The yaw control gain is a proportional coefficient that maps the yaw compensation amount to the desired yaw angular velocity. It is used to adjust the orientation correction sensitivity, with a preferred value of 0.8 to 2.5. This means that yaw correction must both reduce the predicted tangential difference and prevent overshoot from causing secondary misalignment. The desired yaw angular velocity is the yaw angular velocity command obtained after mapping the yaw compensation amount to the yaw control gain, used to drive the aircraft's orientation to synchronize with the cable hanger's orientation.

[0118] It should be noted that the local tangential unit vector of the target traverse is the normalized direction vector of the target traverse along its extension direction at the current evaluation position, used to represent the reference direction of propulsion along the traverse. The desired velocity vector in the world coordinate system is the velocity vector synthesized from the propulsion velocity along the traverse, the normal command velocity, and the vertical command velocity in the world coordinate system, used as a unified spatial control quantity. The transpose of the rotation matrix from the body coordinate system to the world coordinate system is the coordinate transformation matrix that transforms the velocity vector in the world coordinate system to the body coordinate system, used to convert the environmental reference velocity into a body velocity that the flight control can receive. The desired velocity vector in the body coordinate system is the body velocity vector obtained after coordinate transformation of the desired velocity vector in the world coordinate system, used as an intermediate expression of the flight control velocity command. The lateral velocity command component in the body coordinate system is the velocity component of the desired velocity vector in the body coordinate system on the lateral axis of the body coordinate system, used to achieve left or right movement control of the body. The longitudinal velocity command component in the body coordinate system is the velocity component of the desired velocity vector in the body coordinate system on the longitudinal axis of the body coordinate system, used to achieve forward or backward movement control of the body. The vertical velocity command component of the airframe is the velocity component of the desired velocity vector of the airframe on the vertical axis of the airframe, used to achieve airframe climb or descent control. The final control command is an output quantity composed of the lateral velocity command component, longitudinal velocity command component, vertical velocity command component, and desired yaw rate of the airframe, used to directly drive flight control execution. The lateral coordinate of the target wire is the coordinate component of the target wire execution point position on the lateral axis of the selected coordinate system. The longitudinal coordinate of the target wire is the coordinate component of the target wire execution point position on the longitudinal axis of the selected coordinate system. The vertical coordinate of the target wire is the coordinate component of the target wire execution point position on the vertical axis of the selected coordinate system, used to characterize the vertical position of the target point. The final attitude result is an output quantity composed of the lateral coordinate of the target wire, the longitudinal coordinate of the target wire, the vertical coordinate of the target wire, and the target yaw angle, used to provide the target attitude result of the wire execution point.

[0119] It should be noted that the positive part operator extraction operation is defined as follows: the capture window contraction coefficient is compared with zero; if it is greater than zero, the original value is retained; if it is less than or equal to zero, zero is output, thus obtaining the non-negative truncation coefficient. This process should be performed before the propulsion speed is generated to ensure that the propulsion speed along the line is zero when there is no effective window. The tuning method for the maximum propulsion speed along the line and the normal control gain, vertical control gain, and yaw control gain is as follows: first, initial values ​​are given in the ground simulation environment based on the mass of the UAV, the mass of the cable-attached mechanism, the length of the cable, and the expected control cycle; then, they are gradually corrected through semi-physical experiments and small-scale flight tests. During tuning, the focus is on observing normal overshoot, vertical ripple, and yaw overshoot until all three types of responses meet the cable-attached safety requirements. The construction method for the local tangential unit vector of the target traverse is as follows: the local tangential direction is obtained from the local tangential angle of the target traverse or the first derivative of the catenary function; then, the direction vector is normalized to obtain the local tangential unit vector of the target traverse; to maintain directional consistency, it can be specified that it always points to the preset propulsion direction along the line.

[0120] It should be noted that the transpose of the rotation matrix from the body coordinate system to the world coordinate system is constructed as follows: based on the roll, pitch, and yaw angles output by the flight controller, the rotation matrix from the body coordinate system to the world coordinate system is constructed in a predetermined rotation order. Then, its transpose is obtained for the conversion of the world coordinate system velocity vector to the body coordinate system velocity vector. The final control command output interface and limiting method are as follows: the body coordinate system lateral velocity command components, body coordinate system longitudinal velocity command components, body coordinate system vertical velocity command components, and the desired yaw rate are encapsulated into a unified control message, which is output to the flight controller via serial port, Ethernet, or flight controller software interface. Before output, speed, angular velocity, and acceleration limits should be applied to each component. The final attitude result's coordinate reference and timestamp encapsulation method is as follows: the target line's lateral coordinates, target line's longitudinal coordinates, target line's vertical coordinates, and target yaw angle are uniformly encapsulated under the same world coordinate system reference, and the current control cycle timestamp and validity flag are appended. If the host computer needs to display the relative attitude, it can also simultaneously provide the difference relative to the current body position.

[0121] It should be noted that because the capture window shrinkage coefficient may drop below zero due to future collision risks, performing a positive part operator extraction operation on the capture window shrinkage coefficient to obtain a non-negative truncation coefficient ensures that the desired propulsion speed automatically returns to zero under high-risk conditions, preventing the abnormality of continuing propulsion when there is no effective window; for example, when the dynamic tangential capture window has collapsed, the system will no longer continue to propel into the inside of the guide wire. The final motion of the UAV occurs in three-dimensional space, and the three directions of line-following propulsion, normal correction, and vertical clearance have different physical meanings. Therefore, expanding and synthesizing the desired propulsion speed, normal command speed, and vertical command speed along the local tangential unit vector of the target guide wire, the local normal unit vector of the target guide wire, and the local vertical unit vector of the plane into the desired speed vector in the world coordinate system can avoid the directional distortion caused by simple axial superposition. In addition, the flight control system typically receives speed commands in the body coordinate system and uses the transpose of the rotation matrix from the body coordinate system to the world coordinate system to perform coordinate transformation to generate the desired velocity vector of the body coordinate system. It can correctly translate the environmental reference velocity into the body coordinate system velocity using the current attitude as a bridge. For example, the orientation of the guide wire may not be consistent with the nose of the aircraft. If the world coordinate velocity is directly sent to the flight control system, it will cause obvious lateral errors.

[0122] In one embodiment of the present invention, a drone wire-attaching adaptive alignment method based on AI vision includes the following steps:

[0123] Step S1: Use binocular vision to extract the centerline point set of the target wire, the hanging wire, and the hanging point;

[0124] Step S2: Reconstruct the three-dimensional geometry based on the centerline point set in the local hanging plane, fit the catenary to obtain the local tangential angle and local curvature of the target traverse, fit the equivalent tangent to obtain the equivalent tangential angle, and calculate the normal deviation of the hanging execution point relative to the target traverse.

[0125] Step S3: Calculate the current tangential difference and the rate of change of tangential difference using the equivalent tangential angle and the local tangential angle of the target traverse; and calculate the predicted tangential difference by combining the current tangential difference and the rate of change of tangential difference.

[0126] Step S4: Calculate the dynamic tangential capture window based on the nominal mechanical opening width, predicted tangential difference, tangential difference change rate, and local curvature of the target conductor; and calculate the capture window contraction coefficient using the dynamic tangential capture window and the nominal mechanical opening width.

[0127] Step S5: Calculate the normal compensation, vertical compensation, and yaw compensation based on the normal deviation, local curvature of the target traverse, and predicted tangential difference, and then update the target traverse execution point position and target yaw angle.

[0128] Step S6: Generate the desired propulsion speed under the constraint of the capture window contraction coefficient. Combine the desired propulsion speed, normal compensation, vertical compensation and yaw compensation to generate the final control command. Output the final pose result consisting of the target hook execution point position and the target yaw angle, as well as the final control command to complete the alignment.

[0129] Specifically, the actual deployment and use of this invention can be divided into four stages: ground preparation, airborne deployment, online operation, and result output. The ground preparation stage first completes the calibration of the binocular camera, the calibration of the extrinsic parameters of the wire-attaching mechanism, the calibration of the geometric dimensions of the wire-attaching opening, and the determination of the total control delay. Binocular camera calibration is used to obtain the binocular camera focal length, binocular camera baseline length, lateral principal point coordinates, longitudinal principal point coordinates, lateral equivalent focal length, and longitudinal equivalent focal length. Extrinsic parameter calibration of the wire-attaching mechanism is used to determine the fixed relationship between the current wire-attaching execution point position and the center of the aircraft. Geometric dimension calibration of the wire-attaching opening is used to obtain the nominal mechanical opening width and effective wire insertion depth. The determination of the total control delay is used to establish a unified time budget between image acquisition, inference, control issuance, and flight control execution. After calibration, binocular training samples of the target conductor, the wire-attaching cable, and the wire-attaching opening are collected at the ground test site to train the joint segmentation network, and the recognition continuity and refinement stability of the three types of objects are confirmed on the validation set.

[0130] It should be noted that the airborne deployment phase integrates the binocular camera, airborne computing unit, flight control, positioning unit, inertial measurement unit, and wire-attaching mechanism into a unified system. The binocular camera is mounted on a rigid bracket that allows clear observation of the target conductor and wire-attaching opening. The airborne computing unit is responsible for joint segmentation, centerline extraction, local 3D reconstruction, catenary fitting, tangential difference prediction, dynamic tangential capture window calculation, and control command generation. The flight control unit is responsible for receiving the final control commands and outputting attitude and position state quantities. After the system is powered on, the UAV first performs a near-line approach mission, completing target conductor locking and wire-attaching locking before entering the end of the wire-attaching section, and then switching to the adaptive alignment mode of this invention. At this time, the system continuously acquires images from the left and right cameras in a control cycle, extracts the target conductor centerline point set, wire-attaching centerline point set, and wire-attaching opening centerline point set in real time, and combines the current yaw angle and the current wire-attaching execution point position to complete local geometric modeling. During the online operation phase, it operates in a closed loop according to the above-described AI vision-based UAV wire-attaching adaptive alignment method.

[0131] It should be noted that the final output of this invention includes two types of quantities: the final pose result and the final control command. The final pose result is composed of the target line's lateral coordinates, longitudinal coordinates, vertical coordinates, and yaw angle, used to inform the host computer and flight controller of the target state at the target execution point. The final control command is composed of the aircraft system's lateral velocity command component, longitudinal velocity command component, vertical velocity command component, and desired yaw angle velocity, used to directly drive the UAV to complete the alignment. For example, in a certain control cycle, the system outputs the target line's lateral coordinates as 12.48 meters, longitudinal coordinates as 3.62 meters, vertical coordinates as 18.27 meters, and yaw angle as 6 degrees. Simultaneously, the system outputs a lateral velocity command of 0.12 m / s, a longitudinal velocity command of 0.18 m / s, a vertical velocity command of -0.05 m / s, and a desired yaw rate of -3 degrees / s. This indicates that the UAV should moderately move to the right, forward, and slightly descend while rotating in the opposite direction at a small angle within the same cycle to follow the still-accessible entry channel. For example, in another control cycle, if the dynamic tangential capture window drops below zero, the system can output a desired propulsion speed of zero, maintaining only normal and yaw corrections until the window stabilizes and recovers before resuming propulsion. This will not be elaborated further here.

[0132] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0133] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. An AI vision-based adaptive alignment system for drones with attached wires, characterized in that, include: The first module uses binocular vision to extract the centerline point set of the target wire, the hanging wire, and the hanging point. The second module reconstructs the three-dimensional geometry based on the centerline point set within the local hanging plane, fits the catenary to obtain the local tangential angle and local curvature of the target traverse, fits the equivalent tangent to obtain the equivalent tangential angle, and calculates the normal deviation of the hanging execution point relative to the target traverse. The third module calculates the current tangential difference and the rate of change of tangential difference using the equivalent tangential angle and the local tangential angle of the target traverse, and calculates the predicted tangential difference by combining the current tangential difference and the rate of change of tangential difference. The fourth module calculates the dynamic tangential capture window based on the nominal mechanical opening width, predicted tangential difference, tangential difference change rate, and local curvature of the target conductor, and uses the dynamic tangential capture window and the nominal mechanical opening width to calculate the capture window contraction coefficient. The fifth module calculates the normal compensation, vertical compensation, and yaw compensation based on the normal deviation, the local curvature of the target traverse, and the predicted tangential difference, and then updates the target traverse execution point position and the target yaw angle. The sixth module generates the desired propulsion speed under the constraint of the capture window contraction coefficient. It then generates the final control command by combining the desired propulsion speed, normal compensation, vertical compensation, and yaw compensation. The final pose result, consisting of the target wire execution point position and the target yaw angle, along with the final control command, is output to complete the alignment. The final pose result is composed of the target wire's lateral coordinates, longitudinal coordinates, vertical coordinates, and yaw angle, and is used to inform the host computer and flight controller of the target state of the wire execution point that should be reached. The final control command is composed of the aircraft system's lateral velocity command component, longitudinal velocity command component, vertical velocity command component, and desired yaw angle velocity, and is used to directly drive the UAV to complete the alignment.

2. The AI ​​vision-based drone wire-attaching adaptive alignment system according to claim 1, characterized in that, Acquire images from the left and right cameras, input them into a joint segmentation network, and segment to obtain the target wire mask, hanging wire mask, and hanging opening mask. The target conductor mask, the hanging wire mask, and the hanging port mask are refined respectively to extract the center line point set of the target conductor, the hanging wire, and the hanging port.

3. The AI ​​vision-based drone wire-attaching adaptive alignment system according to claim 1, characterized in that, The depth is generated by calculating the disparity of the matching points of the centerline point set by combining the focal length and baseline length of the binocular camera. The position of the horizontal and vertical pixel coordinates is calculated by combining the depth, the horizontal equivalent focal length, the vertical equivalent focal length, the horizontal principal point coordinates and the vertical principal point coordinates to generate the horizontal position and the vertical position of the camera coordinate system. The local three-dimensional point coordinates are formed by the depth, the horizontal position and the vertical position of the camera coordinate system. Within the local suspension plane, the target traverse is fitted with a catenary based on the horizontal coordinates of the local suspension plane and the coordinates of local three-dimensional points, generating a catenary height function that includes catenary sag parameters, catenary center position parameters, and vertical translation parameters.

4. The AI ​​vision-based drone wire-attaching adaptive alignment system according to claim 3, characterized in that, Combining the sag parameters and center position parameters of the catenary, the local tangential angle and local curvature of the target traverse are calculated based on the lateral coordinates of the nearest neighbor of the target traverse at the suspension point. Local linear fitting is performed on the suspension line within the local suspension plane to generate a local linear approximation function for the suspension line that includes the local equivalent slope and the linear fitting bias. The arctangent value of the local equivalent slope is then calculated to obtain the equivalent tangential angle. Calculate the spatial difference vector between the current wire-hanging execution point and the nearest point of the target traverse, and calculate the dot product of the spatial difference vector and the local normal unit vector of the target traverse to obtain the normal deviation.

5. The AI ​​vision-based drone wire-attaching adaptive alignment system according to claim 1, characterized in that, The current tangential difference is obtained by subtracting the local tangential angle of the target traverse from the equivalent tangential angle; the tangential difference at the previous moment is subtracted from the current tangential difference, and the rate of change of the tangential difference is obtained by dividing the result of the subtraction by the sampling time interval. Multiply the rate of change of tangential difference by the total control delay to obtain the delay tangential drift, and add the delay tangential drift to the current tangential difference to calculate the predicted tangential difference.

6. The AI ​​vision-based drone wire-attaching adaptive alignment system according to claim 1, characterized in that, The directional misalignment erosion width is obtained by multiplying the absolute value of the sine of the predicted tangential difference with the effective penetration depth. The local bending erosion width is obtained by calculating half the product of the absolute value of the local curvature of the target conductor and the square of the current visible arc length. The time delay drift erosion width is obtained by calculating the absolute value of the rate of change of tangential difference, the effective penetration depth, and the total control time delay; The dynamic tangential capture window is calculated by subtracting the directional misalignment erosion width, the local bending erosion width, and the time-delay drift erosion width from the nominal mechanical opening width in sequence. The capture window shrinkage coefficient is calculated by dividing the dynamic tangential capture window by the nominal mechanical opening width.

7. The AI ​​vision-based drone wire-attaching adaptive alignment system according to claim 1, characterized in that, The negative of the normal deviation is extracted as the normal compensation amount; the negative of half the product of the local curvature of the target traverse and the square of the current visible arc length is extracted as the vertical compensation amount; and the negative of the predicted tangential difference is extracted as the yaw compensation amount. Based on the current location of the hanging line execution point, the normal compensation amount extending along the local normal unit vector of the target traverse and the vertical compensation amount extending along the vertical direction of the local plane are superimposed to update and obtain the target hanging line execution point location. The target yaw angle is updated by adding yaw compensation to the current yaw angle.

8. The AI ​​vision-based drone wire-attaching adaptive alignment system according to claim 1, characterized in that, The positive part operator is extracted from the capture window contraction coefficient to obtain the non-negative truncation coefficient. The non-negative truncation coefficient is multiplied by the maximum advance speed along the line to calculate the desired advance speed. The normal compensation is multiplied by the normal control gain to generate the normal command velocity, the vertical compensation is multiplied by the vertical control gain to generate the vertical command velocity, and the yaw compensation is multiplied by the yaw control gain to generate the desired yaw rate. The desired propulsion velocity is expanded along the local tangential unit vector direction of the target traverse, the normal command velocity is expanded along the local normal unit vector direction of the target traverse, and the vertical command velocity is expanded along the local vertical unit vector direction of the plane. The results of the expansion in the three directions are added together to synthesize the desired velocity vector in the world coordinate system.

9. The AI ​​vision-based drone wire-attaching adaptive alignment system according to claim 8, characterized in that, The desired velocity vector in the machine system is generated by performing coordinate transformation on the desired velocity vector in the world coordinate system using the transpose of the rotation matrix from the body coordinate system to the world coordinate system. The desired velocity vector of the aircraft system includes the lateral velocity command component, longitudinal velocity command component, and vertical velocity command component of the aircraft system. The lateral velocity command component, longitudinal velocity command component, and vertical velocity command component of the aircraft system are combined with the desired yaw rate to construct the final control command. Extract the target line's horizontal, vertical, and lateral coordinates from the target line's execution point location. Combine these coordinates with the target line's yaw angle to construct the final pose result.

10. An AI vision-based adaptive alignment method for drone wiring, characterized in that, Implementing an AI vision-based drone wire-attached adaptive alignment system as described in any one of claims 1 to 9 includes the following steps: Step S1: Use binocular vision to extract the centerline point set of the target wire, the hanging wire, and the hanging point; Step S2: Reconstruct the three-dimensional geometry based on the centerline point set in the local hanging plane, fit the catenary to obtain the local tangential angle and local curvature of the target traverse, fit the equivalent tangent to obtain the equivalent tangential angle, and calculate the normal deviation of the hanging execution point relative to the target traverse. Step S3: Calculate the current tangential difference and the rate of change of tangential difference using the equivalent tangential angle and the local tangential angle of the target traverse; and calculate the predicted tangential difference by combining the current tangential difference and the rate of change of tangential difference. Step S4: Calculate the dynamic tangential capture window based on the nominal mechanical opening width, predicted tangential difference, tangential difference change rate, and local curvature of the target conductor; and calculate the capture window contraction coefficient using the dynamic tangential capture window and the nominal mechanical opening width. Step S5: Calculate the normal compensation, vertical compensation, and yaw compensation based on the normal deviation, local curvature of the target traverse, and predicted tangential difference, and then update the target traverse execution point position and target yaw angle. Step S6: Generate the desired propulsion speed under the constraint of the capture window contraction coefficient. Combine the desired propulsion speed, normal compensation, vertical compensation and yaw compensation to generate the final control command. Output the final pose result consisting of the target hook execution point position and the target yaw angle, as well as the final control command to complete the alignment.

Images