An unmanned aerial vehicle autonomous landing method and system based on solar panel inclination recognition

Abstract

This invention discloses a method and system for autonomous landing of unmanned aerial vehicles (UAVs) based on solar panel tilt angle recognition, relating to the fields of autonomous UAV landing and photovoltaic operation and maintenance technology. The method includes: preprocessing and mapping multimodal moment feature vectors to a manifold; constructing a manifold energy function based on the mapping result and minimizing the manifold energy function to obtain pitch and roll angles; establishing a target attitude matrix based on the pitch, roll, and current heading angles to generate initial attitude control torques; calculating a risk value and comparing it with a preset safety threshold; performing corresponding processing on the initial attitude control torques based on the comparison result to obtain the final attitude control torques; and inputting the attitude control torques into a preset power distribution and hybrid control model to generate control commands for each rotor. Even when the visual component fails, this invention can maintain the continuity and robustness of attitude estimation by relying on inertial features and manifold mapping.

Description

Technical Field

[0001] This invention relates to the fields of autonomous drone landing and photovoltaic operation and maintenance technology, specifically to a method and system for autonomous drone landing based on solar panel tilt angle recognition. Background Technology

[0002] With the rapid development of photovoltaic power generation technology, large-scale photovoltaic power plants have been widely built around the world. Photovoltaic power plants typically consist of a large number of solar panels, which are installed on support structures at a certain angle to obtain the optimal angle for receiving solar radiation.

[0003] In recent years, drone technology has been increasingly applied to the inspection and maintenance of photovoltaic power plants. Drones can quickly complete large-scale inspections, effectively improving operation and maintenance efficiency. In some application scenarios, drones also need to dock or land near solar panels to perform tasks such as close-range inspection or charging.

[0004] However, due to the strong reflective properties and limited texture information of solar panels, coupled with their typically tilted angles, traditional visual recognition methods struggle to reliably and accurately acquire their attitude information. Furthermore, most existing UAV landing control methods are based on fixed control strategies. When environmental disturbances such as wind or changes in solar panel attitude occur, the UAV's attitude adjustment can become unstable, impacting the safety and reliability of landing.

[0005] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0006] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an autonomous landing method and system for unmanned aerial vehicles (UAVs) based on solar panel tilt angle recognition. This method and system have the advantages of high robustness and strong anti-interference capability, thereby solving the problem that existing technologies struggle to stably identify solar panel tilt angles and achieve safe landings in environments with strong reflection, low texture, and complex wind disturbances.

[0007] (II) Technical Solution To achieve the aforementioned advantages of high robustness and strong anti-interference capability, the specific technical solution adopted in this invention is as follows: According to one aspect of the present invention, an autonomous landing method for unmanned aerial vehicles (UAVs) based on solar panel tilt angle recognition is provided, the method comprising: S1. Perform time synchronization processing on the acquired visual data and inertial data, and generate multimodal moment feature vectors based on the time synchronization processing results; S2. Preprocess the multimodal moment eigenvectors and map them to a manifold. Construct a manifold energy function based on the mapping result, and minimize the manifold energy function to obtain the pitch and roll angles of the solar panel relative to the UAV. S3. Based on the pitch angle, roll angle and current heading angle, establish the target attitude matrix, and construct a virtual potential energy function by calculating the attitude error to generate the initial attitude control torque; S4. Calculate the risk value by combining the current linear velocity, angular velocity and attitude error, and compare the risk value with the preset safety threshold. Based on the comparison result, perform corresponding processing on the initial attitude control torque to obtain the final attitude control torque. S5. Input the attitude control torque into the preset power distribution and hybrid control model to generate control commands for each rotor, so as to drive the UAV to achieve autonomous landing.

[0008] Furthermore, time synchronization processing is performed on the acquired visual data and inertial data. Based on the time synchronization processing results, multimodal moment feature vectors are generated, including: S11. Use the onboard visual sensor of the UAV to obtain regional images of the solar panel as visual data, and use the inertial measurement unit to obtain the three-axis acceleration data and three-axis angular velocity data of the UAV as inertial data. S12. Based on hardware timestamps and spherical linear interpolation algorithms, perform time synchronization processing on visual data and inertial data; S13. Based on the time-synchronized visual data and inertial data, construct the image brightness moment and inertial data moment respectively; the inertial data moment includes acceleration moment and angular velocity moment; S14. Combine the image brightness moment and the inertial data moment to generate a multimodal moment feature vector.

[0009] Furthermore, the multimodal moment eigenvectors are preprocessed and mapped to a manifold. Based on the mapping result, a manifold energy function is constructed, and the manifold energy function is minimized to obtain the pitch and roll angles of the solar panel relative to the UAV, including: S21. Preprocess the multimodal moment eigenvectors and rearrange the matrix of the preprocessed multimodal moment eigenvectors to generate the feature matrix; the preprocessing includes zero mean and normalization. S22. Based on the preset formula for constructing a symmetric positive definite matrix, the characteristic matrix is ​​mapped to a symmetric positive definite matrix, and the symmetric positive definite matrix is ​​decomposed into eigenvalues. Based on the decomposition result, the symmetric positive definite matrix is ​​transformed to the tangent space using the logarithmic Euclidean mapping to obtain the manifold mapping result as the manifold matrix. S23. Based on the manifold matrix, construct the manifold energy function, and use the Riemann gradient descent method to minimize the manifold energy function. Based on the solution, obtain the pitch angle and roll angle of the solar panel relative to the UAV reference coordinate system.

[0010] Furthermore, based on the manifold matrix, a manifold energy function is constructed, and the Riemann gradient descent method is used to minimize the manifold energy function. Based on the solution, the pitch and roll angles of the solar panel relative to the UAV reference coordinate system are obtained, including: S231. Obtain the candidate normal vectors of the current iteration, establish a candidate symmetric positive definite matrix as the attitude template, and combine it with the manifold matrix to construct the manifold energy function; S232. Calculate the Euclidean gradient of the energy function of the rectangular manifold with respect to the candidate normal vector, and use the attitude template to convert the Euclidean gradient into the Riemann gradient. S233. Based on the Riemann gradient, iteratively update the candidate normal vector and determine whether the update result satisfies any of the preset termination conditions. If yes, stop the iteration and output the initial normal vector; otherwise, return to step S231 until any of the preset termination conditions are satisfied and output the initial normal vector. S234. Based on the preset unit normal vector constraint tolerance threshold, verify the validity of the initial normal vector, and use the verified initial normal vector as the normal vector of the solar panel surface. S235. Using the normal vector as input, calculate the pitch and roll angles of the solar panel relative to the UAV reference coordinate system according to the preset correspondence between the normal vector components and attitude angles in the UAV reference coordinate system.

[0011] Furthermore, the termination conditions include: the difference between the energy function values ​​of the manifold in two adjacent iterations is less than a preset energy threshold; the change in the candidate normal vector in two adjacent iterations is less than a preset normal vector threshold; and the update step size is less than a preset step size threshold.

[0012] Furthermore, the expressions for pitch angle and roll angle are: ; ; In the formula, The pitch angle of the solar panel relative to the UAV's reference coordinate system; The roll angle of the solar panel relative to the UAV's reference coordinate system; The normal vector in the UAV reference coordinate system Components along the axial direction; The normal vector in the UAV reference coordinate system Components along the axial direction; The normal vector in the UAV reference coordinate system The component along the axial direction.

[0013] Furthermore, based on the pitch angle, roll angle, and current heading angle, a target attitude matrix is ​​established, and a virtual potential energy function is constructed by calculating the attitude error to generate the initial attitude control torque, including: S31. Collect the current heading angle of the UAV, and combine it with the pitch angle and roll angle to construct a target attitude matrix that represents the alignment of the UAV with the normal of the solar panel surface. S32. Obtain the current attitude matrix of the UAV, calculate the attitude error between the current attitude matrix and the target attitude matrix through rotation matrix logarithmic mapping, and construct a virtual potential energy function based on the attitude error; S33. Calculate the gradient of the virtual potential energy function, generate the attitude control force field, and map the attitude control force field to the control space of the UAV through the Jacobian matrix to obtain the initial attitude control torque.

[0014] Furthermore, by combining the current linear velocity, angular velocity, and attitude error, a risk value is calculated and compared with a preset safety threshold. Based on the comparison result, the initial attitude control torque is processed accordingly to obtain the final attitude control torque, which includes: S41. Obtain the current linear velocity and angular velocity of the UAV, and decompose the linear velocity into the target velocity along the normal direction of the solar panel surface and the tangential sliding velocity along the solar panel surface based on the normal vector. S42. Combining the current angular velocity, target velocity, tangential slip velocity, and attitude error of the UAV, construct a risk function and calculate the risk value corresponding to the current control action; S43. Compare the risk value with the preset safety threshold. If the risk value does not exceed the safety threshold, the initial attitude control torque is used as the final attitude control torque. If the risk value exceeds the safety threshold, the initial attitude control torque is masked and limited to obtain the final attitude control torque.

[0015] Furthermore, the initial attitude control torque is masked and limited to obtain the final attitude control torque, which includes: Based on the preset control action masking matrix, the diagonal elements corresponding to risk values ​​that exceed the safety threshold are set as the preset first threshold, and the diagonal elements corresponding to risk values ​​that do not exceed the safety threshold are set as the preset second threshold. Candidate attitude control torques are generated based on the setting results. Based on the preset safe contact speed threshold and angular velocity safety threshold, the candidate attitude control torque is limited to obtain the final attitude control torque.

[0016] According to another aspect of the present invention, an autonomous landing system for unmanned aerial vehicles based on solar panel tilt angle recognition is also provided, the system comprising: The feature vector generation module is used to perform time synchronization processing on the acquired visual data and inertial data, and generate multimodal moment feature vectors based on the time synchronization processing results. The attitude acquisition module is used to preprocess and map the multimodal moment feature vectors to a manifold, construct a manifold energy function based on the mapping result, and minimize the manifold energy function to obtain the pitch and roll angles of the solar panel relative to the UAV. The initial attitude control module is used to establish the target attitude matrix based on the pitch angle, roll angle and current heading angle, and to construct a virtual potential energy function by calculating the attitude error in order to generate the initial attitude control torque. The final attitude control module is used to calculate the risk value by combining the current linear velocity, angular velocity and attitude error, compare the risk value with the preset safety threshold, and perform corresponding processing on the initial attitude control torque based on the comparison result to obtain the final attitude control torque. The UAV control module is used to input attitude control torque into a preset power distribution and hybrid control model, generate control commands for each rotor, and drive the UAV to achieve autonomous landing.

[0017] (III) Beneficial Effects Compared with existing technologies, this invention provides a method and system for autonomous landing of unmanned aerial vehicles based on solar panel tilt angle recognition, which has the following advantages: (1) This invention introduces multimodal moment feature fusion, which combines the global brightness moment of vision with the dynamic moment of the UAV body state, namely acceleration moment and angular velocity moment. Even when the vision part fails, the continuity and robustness of attitude estimation can be maintained by relying on inertial features and manifold mapping.

[0018] (2) This invention generates a dynamic self-balancing attitude control force field by constructing a nonlinear virtual potential energy function in the attitude space of the UAV, which effectively overcomes the influence of complex wind disturbance on attitude; at the same time, it introduces a control action feasibility shielding mechanism, and evaluates the danger of actions in real time through a risk function, thereby eliminating control commands that may cause the aircraft to become unstable and overturn from the bottom layer, and ensuring the safe landing of the UAV on a hard glass surface with an inclination angle. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1This is a flowchart of an autonomous landing method for unmanned aerial vehicles based on solar panel tilt angle recognition according to an embodiment of the present invention; Figure 2 This is a schematic diagram of an autonomous landing system for unmanned aerial vehicles based on solar panel tilt angle recognition according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the overall process of an autonomous landing method for unmanned aerial vehicles based on solar panel tilt angle recognition according to an embodiment of the present invention.

[0021] In the picture: 1. Feature vector generation module; 2. Attitude acquisition module; 3. Initial attitude control module; 4. Final attitude control module; 5. UAV control module. Detailed Implementation

[0022] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0023] According to an embodiment of the present invention, an autonomous landing method and system for unmanned aerial vehicles based on solar panel tilt angle recognition is provided.

[0024] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, an autonomous landing method for unmanned aerial vehicles based on solar panel tilt angle recognition includes: S1. Perform time synchronization processing on the acquired visual data and inertial data, and generate multimodal moment feature vectors based on the time synchronization processing results; In this optional embodiment, time synchronization processing is performed on the acquired visual data and inertial data. Based on the time synchronization processing result, a multimodal moment feature vector is generated, including: S11. Use the onboard visual sensor of the UAV to obtain regional images of the solar panel as visual data, and use the inertial measurement unit to obtain the three-axis acceleration data and three-axis angular velocity data of the UAV as inertial data. S12. Based on hardware timestamps and spherical linear interpolation algorithms, perform time synchronization processing on visual data and inertial data; S13. Based on the time-synchronized visual data and inertial data, construct the image brightness moment and inertial data moment respectively; the inertial data moment includes acceleration moment and angular velocity moment; It should be noted that the image brightness moment is to satisfy The two-dimensional geometric brightness moment is calculated according to the following formula: ,in, for order and The image brightness moment corresponding to the order, Image of solar panel area in coordinates The pixel grayscale value at that location, and It is a non-negative integer; the image brightness moment includes those that satisfy... The two-dimensional geometric brightness moment includes at least: The acceleration moment is determined by the three-axis acceleration vector of the UAV. Construct a second-order matrix that satisfies: ,in, For acceleration torque, respectively drones in axis, shaft and The acceleration components along the axes are expanded into a 9-dimensional acceleration feature by rows. The angular velocity moment is derived from the UAV's three-axis angular velocity vectors. Construct it as a third-order tensor, satisfying: ,in, The first of the angular velocity moments One portion, respectively drones in axis, shaft and The angular velocity components along the axial direction, and the angular velocity moments are expanded in a fixed order to form 27-dimensional angular velocity characteristics.

[0025] The multimodal moment feature vector is formed by concatenating the image brightness moment, acceleration moment expansion feature, and angular velocity moment expansion feature in a preset order, satisfying: ,in, For multimodal moment eigenvectors, to The 9-dimensional acceleration feature is obtained by expanding the acceleration moment by rows. to The 27-dimensional angular velocity feature is obtained by expanding the angular velocity moment in a fixed order; the dimensions of the multimodal moment feature vector are: 15 dimensions for image brightness moment, 9 dimensions for acceleration moment, and 27 dimensions for angular velocity moment, with a total dimension of 51 dimensions.

[0026] S14. Combine the image brightness moment and the inertial data moment to generate a multimodal moment feature vector.

[0027] S2. Preprocess the multimodal moment eigenvectors and map them to a manifold. Construct a manifold energy function based on the mapping result, and minimize the manifold energy function to obtain the pitch and roll angles of the solar panel relative to the UAV. In this optional embodiment, the multimodal moment eigenvectors are preprocessed and mapped to a manifold. Based on the mapping result, a manifold energy function is constructed, and the manifold energy function is minimized to obtain the pitch and roll angles of the solar panel relative to the UAV. S21. Preprocess the multimodal moment eigenvectors and rearrange the matrix of the preprocessed multimodal moment eigenvectors to generate the feature matrix; the preprocessing includes zero mean and normalization. It should be noted that, in order to eliminate the difference in magnitude between different moment components, the multimodal moment eigenvectors are normalized to zero mean and unit variance before the moment manifold mapping.

[0028] S22. Based on the preset formula for constructing a symmetric positive definite matrix, the characteristic matrix is ​​mapped to a symmetric positive definite matrix, and the symmetric positive definite matrix is ​​decomposed into eigenvalues. Based on the decomposition result, the symmetric positive definite matrix is ​​transformed to the tangent space using the logarithmic Euclidean mapping to obtain the manifold mapping result as the manifold matrix. S23. Based on the manifold matrix, construct the manifold energy function, and use the Riemann gradient descent method to minimize the manifold energy function. Based on the solution, obtain the pitch angle and roll angle of the solar panel relative to the UAV reference coordinate system.

[0029] In this optional embodiment, a manifold energy function is constructed based on the manifold matrix, and the Riemann gradient descent method is used to minimize the manifold energy function. Based on the solution results, the pitch and roll angles of the solar panel relative to the UAV reference coordinate system are obtained, including: S231. Obtain the candidate normal vectors of the current iteration, establish a candidate symmetric positive definite matrix as the attitude template, and combine it with the manifold matrix to construct the manifold energy function; S232. Calculate the Euclidean gradient of the energy function of the rectangular manifold with respect to the candidate normal vector, and use the attitude template to convert the Euclidean gradient into the Riemann gradient. S233. Based on the Riemann gradient, iteratively update the candidate normal vector and determine whether the update result satisfies any of the preset termination conditions. If yes, stop the iteration and output the initial normal vector; otherwise, return to step S231 until any of the preset termination conditions are satisfied and output the initial normal vector. In this optional embodiment, the termination conditions include: the difference between the energy function values ​​of two adjacent iterations of the rectangular manifold is less than a preset energy threshold; the change in the candidate normal vector between two adjacent iterations is less than a preset normal vector threshold; and the update step size is less than a preset step size threshold.

[0030] S234. Based on the preset unit normal vector constraint tolerance threshold, verify the validity of the initial normal vector, and use the verified initial normal vector as the normal vector of the solar panel surface. S235. Using the normal vector as input, calculate the pitch and roll angles of the solar panel relative to the UAV reference coordinate system according to the preset correspondence between the normal vector components and attitude angles in the UAV reference coordinate system.

[0031] In this optional embodiment, the expressions for pitch angle and roll angle are: ; ; In the formula, The pitch angle of the solar panel relative to the UAV's reference coordinate system; The roll angle of the solar panel relative to the UAV's reference coordinate system; The normal vector in the UAV reference coordinate system Components along the axial direction; The normal vector in the UAV reference coordinate system Components along the axial direction; The normal vector in the UAV reference coordinate system The component along the axial direction.

[0032] It should be further noted that the normalized multimodal moment eigenvectors are rearranged to form the feature matrix. And mapped to a symmetric positive definite matrix according to the following formula: in, It is a real matrix obtained by rearranging the eigenvectors of the multimodal moments. The positive definiteness coefficients are greater than 0. It is the identity matrix. The symmetric positive definite matrix... By transforming to the tangent space using a logarithmic Euclidean mapping, we obtain the rectangular manifold representation matrix. The calculation formula is as follows: ,in, It is a symmetric positive definite matrix eigenvalue decomposition, The eigenvector matrix, A diagonal matrix composed of eigenvalues. Represents a diagonal matrix Each diagonal element is taken as its natural logarithm. The energy function of the rectangular manifold is defined as: ,in, Let the energy function be a rectangular manifold. It is a symmetric positive definite matrix. To determine the candidate normal vector Constructed attitude template matrix, Represents the logarithmic Euclidean mapping. Let Frobenius norm be represented. The rectangular manifold energy function is minimized using the Riemann gradient descent method, and the candidate normal vector is updated as follows: ,in, For the first Candidate normal vectors for the next iteration. For the first The candidate normal vector updated in the next iteration. The iteration step size, For the energy function of the rectangular manifold in The Riemann gradient at that point. The attitude template. From candidate normal vectors Construct a symmetric positive definite matrix: ,in, It is the identity matrix. The positive definiteness coefficient is greater than 0. The gradient of the energy function of the rectangular manifold is obtained by the following formula: ,in, For the rectangular manifold energy function with respect to the candidate normal vector The Euclidean gradient, It is a symmetric positive definite matrix. The pose template matrix, The attitude template matrix, after logarithmic mapping, is relative to the candidate normal vector. The derivative term of the energy function. The Riemann gradient of the energy function is calculated by the following equation: ,in, Let Riemann gradient be the energy function of the rectangular manifold over the rectangular manifold. Attitude template matrix, For the rectangular manifold energy function with respect to the candidate normal vector The Euclidean gradient.

[0033] The normal vector on the unit spherical manifold is updated as follows: ,in, For the first Candidate normal vectors for the next iteration. For the first The candidate normal vector updated in the next iteration. This is the increment of the normal vector in this iteration. The iteration step size, Let be the Riemann gradient of the energy function of the rectangular manifold, and Optimization stops when one of the following conditions is met: the difference between the energy function values ​​of two consecutive iterations is less than a preset energy threshold; the change in the candidate normal vector between two consecutive iterations is less than a preset normal vector threshold; or the update step size is less than a preset step size threshold. Normal vector estimation stops when: The time is considered a valid attitude solution; among which, A preset unit normal vector constraint tolerance threshold is used to determine whether the currently estimated normal vector satisfies the unit length constraint. The tilt angle of the solar panel is determined by the normal vector. The analysis yielded results including the pitch angle. and roll angle .

[0034] S3. Based on the pitch angle, roll angle and current heading angle, establish the target attitude matrix, and construct a virtual potential energy function by calculating the attitude error to generate the initial attitude control torque; In this optional embodiment, a target attitude matrix is ​​established based on the pitch angle, roll angle, and current heading angle, and a virtual potential energy function is constructed by calculating the attitude error to generate the initial attitude control torque, including: S31. Collect the current heading angle of the UAV, and combine it with the pitch angle and roll angle to construct a target attitude matrix that represents the alignment of the UAV with the normal of the solar panel surface. S32. Obtain the current attitude matrix of the UAV, calculate the attitude error between the current attitude matrix and the target attitude matrix through rotation matrix logarithmic mapping, and construct a virtual potential energy function based on the attitude error; S33. Calculate the gradient of the virtual potential energy function, generate the attitude control force field, and map the attitude control force field to the control space of the UAV through the Jacobian matrix to obtain the initial attitude control torque.

[0035] It should be noted that the attitude error is defined using the logarithmic mapping of the rotation matrix: ,in, Let be the target attitude matrix. This is the current attitude matrix of the UAV. This represents the logarithmic mapping from the rotation matrix to the Lie algebra space. The virtual potential energy function is defined as: ,in, For attitude error, It is a positive definite diagonal gain matrix. This represents the virtual potential energy function value. The attitude control force field is obtained from the gradient of the virtual potential energy function. in, For attitude control force field, For virtual potential energy function, Here is the attitude force field gain matrix. This represents the attitude error. The attitude control torque satisfies: ,in, For attitude control torque, For attitude control force field, This is a Jacobian matrix that maps rotational errors to the control space. The attitude control torque further includes an angular velocity feedback term, satisfying: in, For the final attitude control torque, For attitude control torque, This is the angular velocity feedback gain matrix. The current angular velocity of the drone. This represents the desired angular velocity.

[0036] S4. Calculate the risk value by combining the current linear velocity, angular velocity and attitude error, and compare the risk value with the preset safety threshold. Based on the comparison result, perform corresponding processing on the initial attitude control torque to obtain the final attitude control torque. In this optional embodiment, a risk value is calculated by combining the current linear velocity, angular velocity, and attitude error, and then compared with a preset safety threshold. Based on the comparison result, the initial attitude control torque is processed accordingly to obtain the final attitude control torque, which includes: S41. Obtain the current linear velocity and angular velocity of the UAV, and decompose the linear velocity into the target velocity along the normal direction of the solar panel surface and the tangential sliding velocity along the solar panel surface based on the normal vector. S42. Combining the current angular velocity, target velocity, tangential slip velocity, and attitude error of the UAV, construct a risk function and calculate the risk value corresponding to the current control action; S43. Compare the risk value with the preset safety threshold. If the risk value does not exceed the safety threshold, the initial attitude control torque is used as the final attitude control torque. If the risk value exceeds the safety threshold, the initial attitude control torque is masked and limited to obtain the final attitude control torque.

[0037] In this optional embodiment, the initial attitude control torque is shielded and limited to obtain the final attitude control torque, including: Based on the preset control action masking matrix, the diagonal elements corresponding to risk values ​​that exceed the safety threshold are set as the preset first threshold, and the diagonal elements corresponding to risk values ​​that do not exceed the safety threshold are set as the preset second threshold. Candidate attitude control torques are generated based on the setting results. Based on the preset safe contact speed threshold and angular velocity safety threshold, the candidate attitude control torque is limited to obtain the final attitude control torque.

[0038] It should be noted that the risk function is defined as follows: ,in, The current angular velocity of the drone. The approach velocity is along the normal direction of the solar panel surface. The tangential sliding velocity along the surface of the solar panel. For attitude error, Here are the risk weighting coefficients. Only if the risk function satisfies... The corresponding control action can be executed at any time; among which, To preset a safety threshold, construct a control action masking matrix S, and reset the weights of control components that do not meet the safety constraints to zero, satisfying: ,in, For the final attitude control torque, The control torque after shielding treatment. It is a diagonal masking matrix, with diagonal elements corresponding to unsafe control components set to 0 and diagonal elements corresponding to safe control components set to 1. The final control action is processed by a limiting function and satisfies: ,in, For the control action after amplitude limiting, The control torque after shielding treatment. For the amplitude limiting function, and These are the lower and upper limits of the control action, respectively. At the end of descent, the following conditions must be met: ,in, The approach velocity is along the normal direction of the solar panel surface. The safe contact speed threshold must be met at the instant of contact with the solar panel: ,in, The current angular velocity of the drone. The safe threshold for the permissible angular velocity at the moment of contact.

[0039] S5. Input the attitude control torque into the preset power distribution and hybrid control model to generate control commands for each rotor, so as to drive the UAV to achieve autonomous landing.

[0040] It should be added that, such as Figure 3 As shown, this invention is executed by an onboard computing device of the UAV or a computing device that is communicatively connected to the UAV, and the specific steps are as follows: Step 1: Acquire images of the solar panel area using the UAV's onboard visual sensor, and simultaneously acquire the UAV's three-axis acceleration and three-axis angular velocity data using the inertial measurement unit, and perform time synchronization processing on the visual data and inertial data; Step 2: Calculate the image brightness moment based on the image grayscale distribution, construct the acceleration moment based on the acceleration data, and construct the angular velocity moment based on the angular velocity data; Step 3: Combine the image brightness moment, acceleration moment, and angular velocity moment to form a multimodal moment feature vector; Step 4: Map the multimodal moment feature vectors using a moment manifold mapping function to obtain a moment manifold feature representation in a unified feature space; Step 5: Construct a rectangular manifold energy function based on the rectangular manifold feature representation, and solve for the solar panel surface normal vector by minimizing the rectangular manifold energy function; calculate the attitude parameters of the normal vector to obtain the pitch angle and roll angle of the solar panel relative to the UAV reference coordinate system; Step 6: Based on the pitch angle, roll angle, and the UAV's current heading angle, construct a target attitude matrix aligned with the normal to the solar panel surface; wherein, the current heading angle is obtained by the UAV's inertial measurement unit; Step 7: Calculate the attitude error based on the difference between the current attitude matrix of the UAV and the target attitude matrix, and construct a virtual potential energy function based on the attitude error to obtain the attitude control torque; Step 8: Obtain the current linear velocity and angular velocity information of the UAV, and decompose the linear velocity into the approach velocity along the normal direction of the solar panel surface and the tangential sliding velocity along the solar panel surface; construct a control action risk function based on the approach velocity, tangential sliding velocity, angular velocity and attitude error, and shield control actions that exceed the safety threshold. Step 9: Input the shielded attitude control torque and normal descent thrust into the UAV power distribution and hybrid control model to generate control commands for each rotor, so that the UAV can complete alignment, deceleration, docking and stable landing.

[0041] In step 1, the visual acquisition frequency and the IMU acquisition frequency are usually different (e.g., the visual camera is 30Hz and the IMU is 200Hz). In order to ensure the alignment of multimodal data, this invention adopts a time synchronization processing method based on a combination of hardware time-stamping and spherical linear interpolation (Slerp) to eliminate the phase difference caused by time delay to subsequent moment feature fusion.

[0042] In step 2, the construction of multimodal moments is the core foundation of this method; for visual images, this invention calculates the brightness moments of each order of the image, and the image brightness moments are calculated according to the formula... Calculate, where I(x,y) is the pixel gray value of the solar panel region image at coordinates (x,y), and p and q are non-negative integers, satisfying In this embodiment, the first to fourth order luminance moments of the image are extracted, including at least... A total of 15 feature dimensions are used to describe the overall energy distribution, tilt trend, and shape changes of the solar panel in the image, effectively resisting interference from local strong reflective patches. For inertial data, the three-axis acceleration components of the UAV are set as follows: The acceleration moments are constructed in the form of a second-order matrix, satisfying... The acceleration moment, expanded in rows, is a 9-dimensional acceleration characteristic. Let the three-axis angular velocity components be... The angular velocity moment is constructed as a third-order tensor, satisfying The angular velocity moments are expanded in a fixed order to form 27-dimensional angular velocity features; this construction of high-order dynamic moments can capture the energy characteristics of nonlinear high-frequency vibrations of UAVs under wind disturbance.

[0043] In steps 3 and 4, the aforementioned heterogeneous features are combined into multimodal moment feature vectors. Because the dimensions and statistical distributions of visual moments, acceleration moments, and angular velocity moments differ greatly, direct splicing cannot be used for distance measurement; therefore, this invention uses a rectangular manifold mapping function. Projecting the features, specifically, mapping the feature vectors onto a symmetric positive definite (SPD) manifold space, defining a covariance matrix representative, and projecting it onto a unified Riemannian manifold: ,in For the logarithmic mapping operation on the manifold, the mapping result is normalized using the Frobenius norm on the manifold to obtain the moment manifold feature representation T in the unified feature space; the above heterogeneous features are combined into a multimodal moment feature vector, and the brightness geometric moments, acceleration geometric moments, and angular velocity geometric moments are concatenated in a preset order to form the multimodal moment feature vector: The brightness moment has 15 dimensions, the acceleration moment has 9 dimensions, and the angular velocity moment has 27 dimensions, thus forming a unified feature vector with a total dimension of 51. To eliminate the difference in order of magnitude between different moment components, the multimodal moment feature vector is normalized to zero mean and unit variance before the moment manifold mapping. The normalized multimodal moment feature vector is then processed by... The mapping method is used to transform P into a symmetric positive definite matrix to ensure that P belongs to the symmetric positive definite matrix manifold SPD(n). Subsequently, the symmetric positive definite matrix P is transformed by the following logarithmic mapping to obtain the feature representation of the rectangular manifold in the unified feature space: ,in This is an eigenvalue decomposition.

[0044] In step 5, the solar panel normal vector is solved, and the constructed rectangular manifold energy function is: Where T is the multimodal moment characteristic representation, Let n be the mapping function of the rectangular manifold, and n be the three-dimensional normal vector of the solar panel to be solved. The posture template From candidate normal vectors Construct a symmetric positive definite matrix: Where I is the identity matrix, The positive integer is used to ensure the positive definiteness of the matrix. The energy function of the rectangular manifold is defined as follows: The gradient of the rectangular manifold energy function is given by The Riemann gradient of the energy function is obtained as follows: The normal vector is updated on the rotating manifold in the following way: ,in During the optimization process, optimization stops when one of the following conditions is met: (1) (2) (3) The update step size is less than the preset threshold; when the normal vector estimation satisfies The time is considered an effective attitude solution. After obtaining the normal vector n, the pitch and roll angles of the solar panel can be analytically calculated. To solve this optimization problem, the gradient descent method on the Riemannian manifold is used to update n: ,in After obtaining the normal vector n (where n is the learning rate), the pitch angle of the solar panel can be analytically determined. ) and roll angle (Roll, ): , .

[0045] In steps 6 and 7, automatic attitude adjustment is achieved. Traditional PID control suffers from overshoot when landing on an inclined surface. This invention addresses this issue using a special orthogonal group. Construct a virtual potential energy function: Where R is the current attitude matrix of the UAV, Let n be the target attitude matrix determined by the solar panel normal vector n. represent Logarithmic mapping to its Lie algebra so(3); Differentiating the above virtual potential energy function to generate the attitude control force field (control torque). ): ,in As a positive definite diagonal gain matrix, this force field ensures the gradual dissipation of the system's potential energy, thereby driving the UAV's attitude to converge globally asymptotically to the target attitude. This is based on the UAV's current attitude R and the target attitude. The attitude error is defined using a rotation matrix logarithmic mapping. The virtual potential energy function is defined as follows: The attitude force field is obtained from the potential energy gradient. In practical applications, the attitude torque must meet the following requirements. , where J is the Jacobian matrix that maps the rotational error to the control space; to improve system stability, the control torque further includes an angular velocity feedback term: The desired angular velocity This force field ensures the gradual dissipation of the system's potential energy, thereby driving the UAV's attitude to converge globally asymptotically to the target attitude.

[0046] The solar panel normal vector n determines the target attitude matrix. Specific implementation and calculation steps: Given that the unit normal vector of the solar panel surface obtained from the aforementioned steps is... To ensure a smooth landing of the drone, the target Z-axis of the drone's coordinate system needs to be aligned (…). Align it with the normal vector, that is, let =n. Simultaneously, the current heading angle (Yaw) of the UAV is obtained through the UAV's inertial measurement unit and set as n. Target attitude matrix The specific construction steps are as follows: Step 1: Based on the current heading angle of the UAV Construct a reference nose pointing vector for the UAV in the horizontal plane. ,satisfy: Step 2: Using the target Z-axis vector With reference nose pointing vector The cross product is used to calculate and normalize the Y-axis vector of the UAV body target. : Step 3: Through the target Y-axis vector With the target Z-axis vector The cross product is used to calculate the fully orthogonalized X-axis vector of the target organism. : Step 4: Concatenate the orthogonal unit vectors obtained above column by column to synthesize a three-dimensional target attitude matrix that satisfies the special orthogonal group SO(3) constraint. : Through the above calculation steps, the three-dimensional normal vector n of the solar panel surface and the current heading state can be transformed into a complete rotation matrix form, which can then be used to calculate the attitude error. The effective input to the virtual potential function U.

[0047] In step 8, a control action feasibility shielding mechanism is set up. Under near-ground turbulence (ground effect) and complex wind fields, excessive control commands may cause the drone to collide with the solar panel. Therefore, a risk function is constructed: ,in v is the angular velocity, and v is the linear velocity. and The weighting coefficients are dynamically adjusted based on the solar panel material (such as the brittleness of glass); when Risk > Threshold (a set safety threshold), the control action shielding mechanism is activated to control the torque generated in step 7. Apply saturation limiting: This allows for the shielding of control actions that do not meet safety constraints; a control action feasibility shielding mechanism is set up, with the risk function defined as follows: It is used to measure the instability that may be caused by control actions and to construct a masking matrix. Only when the following conditions are met Control actions are allowed to be executed at certain times, thus ensuring that the weight of unsafe actions is zero. The final control action is then processed by a limiting function. .

[0048] In step 9, the hybrid control model is used to... The descent thrust along the normal vector -n is distributed to the four rotors of the UAV, gradually aligning the UAV with the normal direction of the solar panel and completing a stable and flexible landing. Through a hybrid control model, the descent thrust is distributed to each rotor of the UAV, gradually aligning the UAV with the normal direction of the solar panel and completing a stable and flexible landing. At the end of the landing, the following conditions must be met: ,in For safe contact speed, and simultaneously, the following conditions must be met at the moment of contact with the solar panel. This ensures a safe final landing. Through a hybrid control model, the safe attitude control torque, after risk shielding and amplitude limiting, is... The descent thrust along the solar panel normal vector -n is distributed to the four rotors of the UAV, enabling the UAV to gradually align with the solar panel normal direction and complete a stable and flexible landing. The construction process and specific implementation steps of the hybrid control model are as follows: Step 1: Construct the control allocation matrix (hybrid control matrix) for the multi-rotor UAV. Taking a typical "X"-shaped quadcopter UAV as an example, let the rotational speeds of the four rotors be... The lift coefficient generated by the rotor is The reverse torque coefficient is The lever arm length from the drone's center of mass to the rotation centers of each rotor is l. The desired total thrust acting on the drone is... With the desired three-axis attitude control torque The combination forms a generalized control force vector, and its mathematical mapping relationship with the square of the rotational speed of each rotor can be constructed as follows: ; Where M is the constructed Dimensional mixed control matrix.

[0049] Step 2: Calculate the total descent thrust required based on the UAV's preset desired descent acceleration along the solar panel normal vector -n and the UAV's current mass. The safety torque, after risk function processing and amplitude limiting in the aforementioned steps, is then applied. Combine these as input items to the left side of the above equation.

[0050] Step 3: Invert the control matrix M. Perform the reverse calculation to calculate the square of the independent target rotational speed assigned to each rotor: ; The target rotational speed of the four rotors can be obtained by taking the square root of the solution. (i=1,2,3,4), the flight control module converts the target speed into a PWM (Pulse Width Modulation) signal command and sends it to each electronic speed controller (ESC), thereby precisely driving each rotor motor to generate the corresponding lift and torque.

[0051] After the allocation is implemented through the above hybrid control model, the UAV will approach the solar panel surface in a predetermined attitude, and must meet the approach speed constraint at the end of the descent. ,in The approach velocity is along the normal direction of the solar panel surface. A safe contact speed threshold must be set; simultaneously, angular velocity constraints must be strictly met at the instant of contact with the solar panel. ,in To ensure a safe landing and prevent impact damage to the solar panel glass, the angular velocity threshold allowed at the moment of contact is set.

[0052] The present invention also provides a cloud-deployed solar panel tilt angle recognition and UAV landing control service. The service runs on the cloud server side, communicates with the UAV through a wireless network, transmits visual and inertial data in real time by the UAV, and executes all the calculation steps of multimodal moment feature fusion, moment manifold mapping and attitude force field generation by the server. This cloud deployment method effectively reduces the onboard computing load of the UAV and improves the overall cruise efficiency.

[0053] Compared to traditional ArUco code recognition and PID control schemes, the hardware and software collaboration mechanism of this invention significantly improves the landing success rate under extreme light and wind disturbances, and does not require any additional markers to be affixed to the photovoltaic panels, making it perfectly suitable for photovoltaic inspection, photovoltaic cleaning and photovoltaic operation and maintenance scenarios.

[0054] like Figure 2 As shown, according to another embodiment of the present invention, an autonomous landing system for unmanned aerial vehicles based on solar panel tilt angle recognition is also provided, the system comprising: Feature vector generation module 1 is used to perform time synchronization processing on the acquired visual data and inertial data, and generate multimodal moment feature vectors based on the time synchronization processing results; Attitude acquisition module 2 is used to preprocess and map the multimodal moment feature vectors to a manifold, construct a manifold energy function based on the mapping result, and minimize the manifold energy function to obtain the pitch and roll angles of the solar panel relative to the UAV. The initial attitude control module 3 is used to establish the target attitude matrix based on the pitch angle, roll angle and current heading angle, and to construct a virtual potential energy function by calculating the attitude error in order to generate the initial attitude control torque. The final attitude control module 4 is used to calculate the risk value by combining the current linear velocity, angular velocity and attitude error, compare the risk value with the preset safety threshold, and perform corresponding processing on the initial attitude control torque based on the comparison result to obtain the final attitude control torque. The UAV control module 5 is used to input the attitude control torque into the preset power distribution and hybrid control model, generate control commands for each rotor, and drive the UAV to achieve autonomous landing.

[0055] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for autonomous landing of unmanned aerial vehicles (UAVs) based on solar panel tilt angle recognition, characterized in that, The method includes: S1. Perform time synchronization processing on the acquired visual data and inertial data, and generate multimodal moment feature vectors based on the time synchronization processing results; S2. Preprocess the multimodal moment eigenvectors and map them to a manifold. Construct a manifold energy function based on the mapping result, and minimize the manifold energy function to obtain the pitch and roll angles of the solar panel relative to the UAV. S3. Based on the pitch angle, roll angle and current heading angle, establish the target attitude matrix, and construct a virtual potential energy function by calculating the attitude error to generate the initial attitude control torque; S4. Calculate the risk value by combining the current linear velocity, angular velocity and attitude error, and compare the risk value with the preset safety threshold. Based on the comparison result, perform corresponding processing on the initial attitude control torque to obtain the final attitude control torque. S5. Input the attitude control torque into the preset power distribution and hybrid control model to generate control commands for each rotor, so as to drive the UAV to achieve autonomous landing.

2. The method of claim 1, wherein, The step of performing time synchronization processing on the acquired visual data and inertial data, and generating multimodal moment feature vectors based on the time synchronization processing results, includes: S11. Use the onboard visual sensor of the UAV to obtain regional images of the solar panel as visual data, and use the inertial measurement unit to obtain the three-axis acceleration data and three-axis angular velocity data of the UAV as inertial data. S12. Based on hardware timestamps and spherical linear interpolation algorithms, perform time synchronization processing on visual data and inertial data; S13. Based on the time-synchronized visual data and inertial data, construct image brightness moments and inertial data moments respectively; the inertial data moments include acceleration moments and angular velocity moments; S14. Combine the image brightness moment and the inertial data moment to generate a multimodal moment feature vector.

3. The method of claim 1, wherein, The preprocessing and manifold mapping of the multimodal moment eigenvectors, the construction of the manifold energy function based on the mapping result, and the minimization of the manifold energy function to obtain the pitch and roll angles of the solar panel relative to the UAV include: S21. Preprocess the multimodal moment eigenvectors and rearrange the matrix of the preprocessed multimodal moment eigenvectors to generate a feature matrix; the preprocessing includes zero-mean and normalization processing. S22. Based on the preset formula for constructing a symmetric positive definite matrix, the characteristic matrix is ​​mapped to a symmetric positive definite matrix, and the symmetric positive definite matrix is ​​decomposed into eigenvalues. Based on the decomposition result, the symmetric positive definite matrix is ​​transformed to the tangent space using the logarithmic Euclidean mapping to obtain the manifold mapping result as the manifold matrix. S23. Based on the manifold matrix, construct the manifold energy function, and use the Riemann gradient descent method to minimize the manifold energy function. Based on the solution, obtain the pitch angle and roll angle of the solar panel relative to the UAV reference coordinate system.

4. The method of claim 3, wherein, The process involves constructing a manifold energy function based on a manifold matrix, and minimizing the manifold energy function using the Riemann gradient descent method. Based on the solution, the pitch and roll angles of the solar panel relative to the UAV reference coordinate system are obtained, including: S231. Obtain the candidate normal vectors of the current iteration, establish a candidate symmetric positive definite matrix as the attitude template, and combine it with the manifold matrix to construct the manifold energy function; S232. Calculate the Euclidean gradient of the energy function of the rectangular manifold with respect to the candidate normal vector, and use the attitude template to convert the Euclidean gradient into the Riemann gradient. S233. Based on the Riemann gradient, iteratively update the candidate normal vector and determine whether the update result satisfies any of the preset termination conditions. If yes, stop the iteration and output the initial normal vector; otherwise, return to step S231 until any of the preset termination conditions are satisfied and output the initial normal vector. S234. Based on the preset unit normal vector constraint tolerance threshold, verify the validity of the initial normal vector, and use the verified initial normal vector as the normal vector of the solar panel surface. S235. Using the normal vector as input, calculate the pitch and roll angles of the solar panel relative to the UAV reference coordinate system according to the preset correspondence between the normal vector components and attitude angles in the UAV reference coordinate system.

5. The method of claim 4, wherein, The termination conditions include: the difference between the energy function values ​​of two adjacent iterations of the rectangular manifold is less than a preset energy threshold; the change in the candidate normal vector between two adjacent iterations is less than a preset normal vector threshold; and the update step size is less than a preset step size threshold.

6. The method of claim 5, wherein, The expressions for the pitch angle and roll angle are: ； ； In the formula, The pitch angle of the solar panel relative to the UAV's reference coordinate system; The roll angle of the solar panel relative to the UAV's reference coordinate system; The normal vector in the UAV reference coordinate system Components along the axial direction; The normal vector in the UAV reference coordinate system Components along the axial direction; The normal vector in the UAV reference coordinate system The component along the axial direction.

7. The method of claim 1, wherein, The step of establishing a target attitude matrix based on pitch angle, roll angle, and current heading angle, and constructing a virtual potential energy function by calculating attitude error to generate initial attitude control torque includes: S31. Collect the current heading angle of the UAV, and combine it with the pitch angle and roll angle to construct a target attitude matrix that represents the alignment of the UAV with the normal of the solar panel surface. S32. Obtain the current attitude matrix of the UAV, calculate the attitude error between the current attitude matrix and the target attitude matrix through rotation matrix logarithmic mapping, and construct a virtual potential energy function based on the attitude error; S33. Calculate the gradient of the virtual potential energy function, generate the attitude control force field, and map the attitude control force field to the control space of the UAV through the Jacobian matrix to obtain the initial attitude control torque.

8. The method for autonomous landing of a UAV based on solar panel tilt angle recognition according to claim 1, characterized in that, The risk value is calculated by combining the current linear velocity, angular velocity, and attitude error, and then compared with a preset safety threshold. Based on the comparison result, the initial attitude control torque is processed accordingly to obtain the final attitude control torque, which includes: S41. Obtain the current linear velocity and angular velocity of the UAV, and decompose the linear velocity into the target velocity along the normal direction of the solar panel surface and the tangential sliding velocity along the solar panel surface based on the normal vector. S42. Combining the current angular velocity, target velocity, tangential slip velocity, and attitude error of the UAV, construct a risk function and calculate the risk value corresponding to the current control action; S43. Compare the risk value with the preset safety threshold. If the risk value does not exceed the safety threshold, the initial attitude control torque is used as the final attitude control torque. If the risk value exceeds the safety threshold, the initial attitude control torque is masked and limited to obtain the final attitude control torque.

9. The method of claim 8, wherein, The process of masking and limiting the initial attitude control torque to obtain the final attitude control torque includes: Based on the preset control action masking matrix, the diagonal elements corresponding to risk values ​​that exceed the safety threshold are set as the preset first threshold, and the diagonal elements corresponding to risk values ​​that do not exceed the safety threshold are set as the preset second threshold. Candidate attitude control torques are generated based on the setting results. Based on the preset safe contact speed threshold and angular velocity safety threshold, the candidate attitude control torque is limited to obtain the final attitude control torque.

10. A solar panel inclination recognition based autonomous landing system for a UAV for implementing the solar panel inclination recognition based autonomous landing method of any one of claims 1-9, characterized in that, The system includes: The feature vector generation module is used to perform time synchronization processing on the acquired visual data and inertial data, and generate multimodal moment feature vectors based on the time synchronization processing results. The attitude acquisition module is used to preprocess and map the multimodal moment feature vectors to a manifold, construct a manifold energy function based on the mapping result, and minimize the manifold energy function to obtain the pitch and roll angles of the solar panel relative to the UAV. The initial attitude control module is used to establish the target attitude matrix based on the pitch angle, roll angle and current heading angle, and to construct a virtual potential energy function by calculating the attitude error in order to generate the initial attitude control torque. The final attitude control module is used to calculate the risk value by combining the current linear velocity, angular velocity and attitude error, compare the risk value with the preset safety threshold, and perform corresponding processing on the initial attitude control torque based on the comparison result to obtain the final attitude control torque. The UAV control module is used to input attitude control torque into a preset power distribution and hybrid control model, generate control commands for each rotor, and drive the UAV to achieve autonomous landing.

Images