Unmanned aerial vehicle folding machine arm structure and working method thereof

Abstract

The application relates to the technical field of unmanned aerial vehicle arms, in particular to an unmanned aerial vehicle folding arm structure and a working method thereof, which comprises the following steps: when the unmanned aerial vehicle is in an unfolded working state, a stress sensor at a hinge obtains real-time strain signals containing flight bending moment and torsional load time history data of the arm, a displacement sensor at a driving motor synchronously collects an instantaneous value sequence of an arm unfolding angle, and the two groups of signals are aligned on a time line; the two groups of signals are input into a pre-trained load-deformation correlation model, arm structure dynamics response is analyzed, and a theoretical deformation distribution cloud diagram of multiple preset characteristic sections is solved out; a main compensation control signal is generated for the driving motor according to the cloud diagram, the arm unfolding angle is fine-tuned, and structural deformation is inhibited. The method can accurately monitor the arm state and actively inhibit deformation, thereby guaranteeing the flight stability of the unmanned aerial vehicle.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) arm technology, and in particular to a folding UAV arm structure and its working method. Background Technology

[0002] The folding arm structure of a drone is a key component for achieving miniaturized storage and stable flight. In its deployed state, it must withstand bending moments and torsional loads generated during flight. These loads can cause structural deformation of the arm, thereby affecting the drone's flight stability and the arm's lifespan. Existing technologies employ various methods: some solutions deploy sensors on the folding arm to collect load or displacement signals for monitoring its status; others use drive motors to adjust the arm's deployment angle to ensure it is fully deployed.

[0003] In existing technologies, sensors mostly acquire single-type signals, either only load-related strain signals or only displacement signals related to the deployment angle. These two types of signals are not aligned on a timeline, making it impossible to achieve a precise correlation between load and boom deformation. Furthermore, existing technologies lack effective means to analyze the boom's structural dynamic response based on real-time acquired data. This makes it difficult to accurately obtain the deformation trends and weak points of various characteristic sections of the boom, and the control of boom deformation is mostly a passive response, unable to proactively adjust based on deformation predictions.

[0004] How to accurately predict the deformation of the boom by precisely collecting relevant signals and establishing effective correlations, and then actively control and suppress the deformation to avoid affecting the flight stability and structural life of the UAV due to boom deformation, has become a problem that needs to be solved in the current design of UAV folding boom structures. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a folding arm structure for unmanned aerial vehicles and its operating method.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for operating a folding arm structure of a drone, comprising:

[0007] When the UAV is in the unfolded operation state, real-time strain signals are obtained by stress sensors deployed at the hinge of the folding arm structure. The real-time strain signals include time history data of the bending moment and torsional load borne by the arm during flight.

[0008] Displacement sensors installed at the drive motor of the folding arm structure synchronously collect the instantaneous value sequence of the arm deployment angle, and the instantaneous value sequence of the arm deployment angle is aligned with the real-time strain signal on the timeline;

[0009] The real-time strain signal and the instantaneous value sequence of the boom deployment angle are input into a pre-trained load-deformation correlation model, which is used to analyze the dynamic response of the boom structure under different load modes.

[0010] Based on the analysis results of the pre-trained load-deformation correlation model, the theoretical deformation distribution cloud map of the folding arm structure at multiple preset feature sections is calculated. The theoretical deformation distribution cloud map reflects the deformation trend and weak points of the arm structure in space.

[0011] Based on the theoretical deformation distribution cloud map, an active compensation control signal is generated for the drive motor of the folding arm structure. The active compensation control signal is used to fine-tune the unfolding angle of the arm to suppress the structural deformation predicted by the theoretical deformation distribution cloud map.

[0012] As a further aspect of the present invention, the instantaneous value sequence of the real-time strain signal and the boom deployment angle is input into a pre-trained load-deformation correlation model, including:

[0013] Modal decomposition is performed on the real-time strain signal to separate the trend component representing the low-frequency principal stress and the detail component representing the high-frequency vibration stress.

[0014] The instantaneous value sequence of the arm deployment angle is calculated by difference to obtain the variation curves of the arm deployment angular velocity and angular acceleration;

[0015] The separated trend component, detail component and the change curves of the arm deployment angular velocity and angular acceleration are aligned and fused according to the timestamps to form a structural load state matrix.

[0016] The structural load state matrix is ​​input into a convolutional neural network model trained using historical flight data, i.e., the pre-trained load-deformation correlation model.

[0017] The forward propagation calculation of the convolutional neural network model outputs the overall stiffness evaluation vector of the boom structure corresponding to the current structural load state matrix.

[0018] As a further aspect of the present invention, based on the analysis results of the pre-trained load-deformation correlation model, the theoretical deformation distribution cloud map of the folding arm structure at multiple preset feature sections is calculated, including:

[0019] Receive the overall stiffness evaluation vector of the boom structure output by the pre-trained load-deformation correlation model;

[0020] Based on the UAV 3D structure database, a 3D solid model and material physical property parameters of the folding arm structure are extracted. The material physical property parameters include at least the elastic modulus and Poisson's ratio.

[0021] On the three-dimensional solid model, multiple preset feature sections are selected along the axis of the boom. The location of the preset feature sections is selected based on the abrupt change in structural geometry or the critical node of the load transfer path.

[0022] Using the overall stiffness evaluation vector of the boom structure as the boundary condition, and in combination with the material physical property parameters, a quasi-static mechanical solution is performed on the simplified finite element model of the boom containing multiple preset feature sections.

[0023] Displacement field data on each preset feature section is extracted from the quasi-static mechanical solution, and the displacement field data of all sections are mapped back to three-dimensional spatial coordinates to generate the theoretical deformation distribution cloud map.

[0024] As a further aspect of the present invention, using the overall stiffness evaluation vector of the boom structure as boundary conditions, and combining the material physical property parameters, a quasi-static mechanical solution is performed on the simplified finite element model of the boom containing multiple preset feature sections, including:

[0025] The overall stiffness evaluation vector of the arm structure is decomposed into force components and moment components corresponding to the local coordinate system of the arm.

[0026] The force components and torque components are used as nodal loads and applied to the corresponding fixed nodes at the root of the boom in the simplified finite element model of the boom.

[0027] The initial displacement and rotation of all nodes in the simplified finite element model of the robotic arm, except for the root fixed node, are set to zero;

[0028] The static solver based on the finite element method is invoked, and the displacement and stress response of all nodes of the simplified finite element model of the boom under the nodal load is calculated using the material physical property parameters.

[0029] From the displacement response output by the statics solver, the displacement data of the grid nodes located on the preset characteristic cross section are selected to form the preliminary displacement field of each cross section.

[0030] As a further aspect of the present invention, based on the theoretical deformation distribution cloud map, an active compensation control signal is generated for the drive motor of the folding arm structure, including:

[0031] Extract the predicted deformation displacement vector of the feature points at the end of the boom from the theoretical deformation distribution cloud map;

[0032] The predicted deformation displacement vector is compared with the theoretical end position vector in the zero deformation state to obtain the position deviation vector of the computer arm end.

[0033] The position deviation vector is converted into the theoretical compensation angle increment required for each joint of the folding arm through inverse kinematics calculation;

[0034] The current angle feedback value of the drive motor of the folding arm structure is queried, and the theoretical compensation angle increment is superimposed on the target angle command of the drive motor to form an angle adjustment command.

[0035] Based on the difference between the angle adjustment command and the current angle feedback value, the active compensation control signal is generated by a proportional-integral-derivative control algorithm. The active compensation control signal directly acts on the torque output of the drive motor.

[0036] As a further aspect of the present invention, the working method further includes a dynamic compensation gain adjustment step based on flight status feedback:

[0037] Continuously receive real-time flight status data from the flight control unit of the UAV, wherein the real-time flight status data includes at least flight speed, flight altitude and attitude angle;

[0038] Calculate the current airspeed vector and angle of attack from the real-time flight status data based on the flight speed and attitude angle;

[0039] Establish a lookup table indexed by airspeed and angle of attack, which stores the compensation gain coefficients corresponding to different flight states;

[0040] Based on the current airspeed vector and airflow angle of attack, a set of real-time compensation gain coefficients are obtained by querying the lookup table.

[0041] The real-time compensation gain coefficient is applied to the gain parameters of the proportional, integral, and derivative terms of the proportional-integral-derivative control algorithm to dynamically adjust the strength and response speed of the active compensation control signal.

[0042] As a further aspect of the present invention, the working method further includes a boom stiffness attenuation early warning step based on structural health:

[0043] After each UAV flight mission is completed, historical data of all real-time strain signals from the mission are retrieved from the onboard storage unit.

[0044] Statistical features are extracted from the historical data to calculate the average value, peak value, and variance rate of change of the strain signal throughout the entire flight mission cycle.

[0045] The calculated average value, peak value, and variance rate of change of the strain signal are input into a pre-trained structural health assessment model, which is used to identify small attenuation trends in the stiffness of the boom structure.

[0046] Obtain the structural health index output by the structural health assessment model. The structural health index is a value between zero and one, and the lower the value, the higher the stiffness decay.

[0047] If the structural health index is lower than the preset warning threshold, a stiffness attenuation warning message containing the specific arm number and suggested inspection items will be generated and recorded in the maintenance log.

[0048] As a further aspect of the present invention, statistical feature extraction is performed on the historical data to calculate the average value, peak value, and variance rate of change of the strain signal throughout the entire flight mission cycle, including:

[0049] From the historical data of the real-time strain signal, extract the signal data of two independent channels representing bending moment and torsional load respectively;

[0050] For the signal data of each channel, the pulse interference segment caused by known maneuvering motion is removed to obtain a smoothed strain signal sequence;

[0051] The arithmetic mean of the smoothed strain signal sequence over the entire mission duration is calculated to obtain the average value of the strain signal.

[0052] Search for the point with the maximum absolute value in the smoothed strain signal sequence, record its value and the time of occurrence, and the value is the peak value of the strain signal.

[0053] The task duration is divided into several equal-length sub-time periods. The variance of the strain signal sequence in each sub-time period is calculated, and the change in variance between adjacent sub-time periods is calculated. The average value of all changes is then taken to obtain the variance change rate.

[0054] As a further aspect of the present invention, the working method further includes a decision-making step for the optimal arm deployment angle for different task modes:

[0055] Obtain the mission mode code for the upcoming flight mission, wherein the mission mode code corresponds to one of the following: cruise mode, high-speed maneuvering mode, hovering observation mode, or wind-resistant attitude stabilization mode;

[0056] Retrieve the historical best flight database corresponding to the mission mode code. The historical best flight database stores the correlation records between different arm deployment angles, flight states and final structural deformation data under the same mission mode.

[0057] Based on the current mission environment prediction parameters, a matching query is performed in the historical best flight database. The mission environment prediction parameters include at least the predicted wind speed and the predicted payload.

[0058] From the results of the matching query, the record with the smallest structural deformation data under the estimated parameters of similar task environment is selected, and the corresponding arm deployment angle in the record is extracted as the initial recommended deployment angle.

[0059] Before the drone takes off, the initial recommended deployment angle is sent as a preset command to the drive motor of the folding arm structure to control the arm to deploy to the initial recommended deployment angle.

[0060] The matching query in the historical best flight database based on the estimated parameters of the current mission environment specifically includes:

[0061] Read the predicted mission environment parameters from the flight mission planning system;

[0062] In the historical best flight database, the similarity distance between the environmental parameters of each historical mission record and the estimated environmental parameters of the current mission is calculated.

[0063] Set a similarity threshold, and select all historical task records whose similarity distance is less than the similarity threshold to form a candidate record set;

[0064] For each record in the candidate record set, sort them according to the size of the final structural deformation data of the record, wherein the final structural deformation data is quantified by the average or peak value of the strain signal in the corresponding historical task;

[0065] From the sorting results, select several records with the smallest final structural deformation data as the optimal matching records, which are used to extract the initial recommended unfolding angle.

[0066] As a further aspect of the present invention, the present invention also includes a folding arm structure for a drone, the structure comprising: a processor; a memory storing a computer program; the processor executing the computer program implements the working method of the folding arm structure for a drone as described above.

[0067] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0068] Stress sensors are deployed at the hinges of the folding arm structure to collect real-time strain signals containing time-history data of bending moments and torsional loads experienced by the arm during flight. Simultaneously, displacement sensors are deployed at the drive motors of the folding arm structure to synchronously collect instantaneous sequences of the arm's deployment angle. The two sets of acquired signals are aligned on a timeline. This collaborative deployment of the two types of sensors comprehensively captures load changes and structural positional changes during arm flight. The alignment of the timelines ensures precise matching of load data with deformation-related data, overcoming the limitations of existing technologies where single-sensor data cannot fully reflect the arm's state and has low data correlation. This effectively avoids biases in arm state analysis caused by data misalignment, making subsequent structural dynamics response analysis more targeted.

[0069] Real-time strain signals aligned with the timeline and instantaneous values ​​of the arm deployment angle are input into a pre-trained load-deformation correlation model. The model analyzes the dynamic response of the arm structure under different load modes, calculating the theoretical deformation distribution cloud map of the folding arm structure at multiple preset characteristic sections. Based on this cloud map, active compensation control signals are generated for the drive motors, suppressing structural deformation by fine-tuning the arm deployment angle. This method can accurately capture the deformation trend and weak points of the arm in space. Compared to existing technologies that cannot accurately locate deformation patterns and can only passively respond to deformation, this method enables early prediction and active suppression of arm deformation, reducing fatigue losses caused by excessive deformation, avoiding adverse effects of deformation on the flight stability of the UAV, and extending the service life of the arm structure. Attached Figure Description

[0070] Figure 1 This is a flowchart of the working method of the folding arm structure of the UAV described in this invention;

[0071] Figure 2 A flowchart for solving the theoretical deformation distribution cloud map;

[0072] Figure 3 A graph showing the incremental compensation angle of the joints of the folding arm of a drone.

[0073] Figure 4 Comparison of deformation compensation effects at the end of a drone's folding arm;

[0074] Figure 5 This is a diagram showing the similarity matching analysis of the UAV mission environment. Detailed Implementation

[0075] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0076] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0077] See Figure 1 When the UAV is in deployment mode, stress sensors deployed at the hinges of the folding arm structure continuously acquire real-time strain signals. These signals contain time-history data on the bending moment and torsional loads experienced by the arm during flight. Simultaneously, displacement sensors deployed at the drive motor of the folding arm structure synchronously collect the instantaneous sequence of the arm's deployment angle, ensuring strict alignment between this sequence and the real-time strain signals through time synchronization technology. The aligned real-time strain signals and the instantaneous sequence of the arm's deployment angle are input into a pre-trained load-deformation correlation model. This model analyzes the dynamic response characteristics of the arm structure under different load modes. Based on the model's analysis, theoretical deformation distribution cloud maps of the folding arm structure at multiple preset characteristic sections can be calculated. These cloud maps visually reflect the deformation trend and potential weak points of the arm structure in space. Based on the calculated theoretical deformation distribution cloud maps, active compensation control signals are generated for the drive motor of the folding arm structure. These control signals can fine-tune the arm's deployment angle, thereby offsetting or suppressing the structural deformation predicted by the model, improving flight stability and structural safety.

[0078] In one embodiment of the invention, when the UAV is hovering while performing an inspection mission, stress sensors deployed at the hinge of the folding arm structure continuously transmit real-time strain signals. These real-time strain signals are time-series sequences containing timestamps, bending moment load values, and torsional load values. In a specific example, a 2-second real-time strain signal shows that the bending moment load fluctuates between -15 N·m and +20 N·m, while the torsional load varies between 5 N·m and 8 N·m. Modal decomposition is performed on this real-time strain signal. The empirical modal decomposition method is used to separate the original signal into several intrinsic mode functions (EMFs). Components with frequencies below 1 Hz are classified as trend components representing low-frequency principal stresses, while components with frequencies above 10 Hz are classified as detail components representing high-frequency vibrational stresses. The trend components primarily reflect the steady-state aerodynamic loads borne by the arm, while the detail components capture the dynamic responses caused by rotor vibrations or gusts of wind. The instantaneous sequence of arm deployment angles acquired synchronously is differentially calculated. The instantaneous sequence is sampled at a frequency of 100Hz. For example, a set of angle values ​​[45.0°, 45.1°, 45.3°, 45.6°] is used to obtain an angular velocity sequence [0.1° / s, 0.2° / s, 0.3° / s] through first-order backward differential calculation. Then, the angular velocity sequence is differentially calculated again to obtain an angular acceleration sequence [0.1° / s², 0.1° / s²]. The separated trend components, detail components, and the calculated curves of arm deployment angular velocity and angular acceleration are aligned and fused based on the precise timestamps carried by their respective data points. In some embodiments, the alignment process uses linear interpolation to unify all data sequences to a common 100Hz time base. After fusion, a multi-row, multi-column structural load state matrix is ​​formed. Each row of the matrix corresponds to a sampling time, and each column represents a physical quantity channel. For example, channel one is the bending moment trend component, channel two is the bending moment detail component, channel three is the torque trend component, and so on.

[0079] It is understandable that the core step is to input the structural load state matrix into a convolutional neural network model trained using historical flight data. The pre-trained load-deformation correlation model is itself a convolutional neural network model. During model training, over 1000 sets of historical flight data covering different flight conditions were used. Each set of historical flight data includes a structural load state matrix and corresponding arm deformation baseline data obtained through high-precision laser displacement sensors. The convolutional neural network model learns the complex mapping relationship between the input matrix and the output deformation features. During inference, the currently generated online structural load state matrix is ​​input into this convolutional neural network model, which performs forward propagation calculations through its multiple convolutional layers, pooling layers, and fully connected layers. The model ultimately outputs a fixed-dimensional overall stiffness evaluation vector for the arm structure, such as a 128-dimensional vector, where each element comprehensively represents the arm's stiffness performance in the corresponding mode or direction. An optional implementation detail is that, when performing modal decomposition on the real-time strain signal, the following mathematical process can be used for component selection:

[0080]

[0081] in: The time-history data representing the separated j-th type component (j=1 represents the trend component, j=2 represents the detail component) This represents the i-th eigenmode function obtained from empirical mode decomposition. Represents the intrinsic mode function The dominant frequency, The preset frequency threshold represents the distinction between trends and details. It's a filtering function; the function value is 1 if the condition is met, and 0 otherwise. This represents the total number of intrinsic mode functions (EMFs). This formula describes the process of classifying and extracting target components from a series of EMFs based on the relationship between their dominant frequencies and thresholds. In some embodiments, the structural load state matrix undergoes standardization preprocessing before being input into the convolutional neural network model; that is, the data for each channel is subtracted from its historical mean and divided by its historical standard deviation to ensure the stability of the model input. The forward propagation computation process of the convolutional neural network model is a completely deterministic matrix and nonlinear transformation operation, without any stochastic processes, ensuring that a consistent overall stiffness evaluation vector output for the boom structure is always produced for the same input matrix.

[0082] In one embodiment of the present invention, the method for solving the theoretical deformation distribution cloud map based on the analysis results of a pre-trained load-deformation correlation model involves the conversion from a digital model to a physical simulation, see below. Figure 2The system receives the overall stiffness evaluation vector of the arm structure output by a pre-trained load-deformation correlation model. Based on the UAV 3D structure database, a precise 3D solid model and material physical property parameters of the folding arm structure are extracted. The material physical property parameters explicitly include the elastic modulus and Poisson's ratio. The 3D solid model is stored in STEP or IGES format and contains complete geometric and topological information. On the 3D solid model, multiple preset feature sections are selected along the arm axis. The location of the preset feature sections is selected based on the geometric abrupt changes in the structure or the critical nodes of the load transfer path. For example, the first preset feature section is located at the root of the hinge connecting the arm and the fuselage, the second preset feature section is located at the middle reinforcing rib of the arm, and the third preset feature section is located at the end near the motor mounting base.

[0083] Using the overall stiffness evaluation vector of the boom structure as boundary conditions and combining material physical property parameters, a quasi-static mechanical solution is performed on a simplified finite element model of the boom containing multiple preset characteristic sections. The simplified finite element model of the boom is discretized from hexahedral solid elements or shell element meshes. In specific implementation, the solution process includes decomposing the overall stiffness evaluation vector of the boom structure into force components and moment components corresponding to the local coordinate system of the boom. The local coordinate system of the boom has its origin at the center of the boom root, with the X-axis along the boom span, the Y-axis vertically upward, and the Z-axis determined by the right-hand rule. The decomposed force components include Fx, Fy, and Fz, and the moment components include Mx, My, and Mz. The force components and moment components are applied as nodal loads to the corresponding fixed nodes at the boom root in the simplified finite element model of the boom. The application of nodal loads follows the input specifications for nodal forces and moments in finite element analysis. The initial displacement and rotation of all nodes in the simplified finite element model of the boom, except for the fixed nodes at the root, are set to zero as the initial conditions for the quasi-static mechanical solution. The statics solver based on the finite element method is invoked, and the displacement and stress responses of all nodes in the simplified finite element model of the boom under nodal loads are calculated using material physical property parameters. The statics solver solves the problem based on the following equilibrium equations:

[0084]

[0085] in: The total stiffness matrix represents the simplified finite element model of the robotic arm, and its elements are determined by the element type, mesh size, and material physical property parameters (elastic modulus, Poisson's ratio). The displacement vector representing all nodes, including linear and angular degrees of freedom; This represents the equivalent nodal load vector composed of the decomposed force and moment components. From the displacement response output by the statics solver, displacement data of mesh nodes located on preset characteristic sections are selected to form the preliminary displacement field for each preset characteristic section. In some embodiments, the selection process is achieved by traversing all mesh nodes and determining whether their spatial coordinates are within the tolerance range of the plane containing the preset characteristic section. The preliminary displacement field of each preset characteristic section contains the Ux, Uy, and Uz displacement component data of hundreds of nodes on that section. All the preliminary displacement field data of the preset characteristic sections are mapped back to the complete 3D solid model coordinate space using a 3D spatial interpolation method, and combined to generate a theoretical deformation distribution cloud map covering the surface of the boom structure. The theoretical deformation distribution cloud map uses color gradients to represent the magnitude of displacement; for example, blue represents regions where the displacement is close to zero, and red represents regions where the displacement reaches its maximum value.

[0086] Optionally, after generating the theoretical deformation distribution cloud map, the cloud map data can be overlaid with the stress response results in the simplified finite element model of the robotic arm to simultaneously observe deformation and stress concentration. The process of extracting displacement field data from each preset feature section from the quasi-static mechanical solution and mapping all the displacement field data of the preset feature sections back to three-dimensional spatial coordinates to generate the theoretical deformation distribution cloud map is automated. It can be understood that the theoretical deformation distribution cloud map visually reveals the spatial deformation morphology of the robotic arm under specific loads, with deformation trends and weak points intuitively displayed. For example, the cloud map shows that the maximum deformation occurs in the region between the second and third preset feature sections. In some embodiments, the mapping process uses linear interpolation or radial basis function interpolation algorithms to ensure that discrete node displacement data can be continuously and smoothly presented on the surface of the three-dimensional model. The final generated theoretical deformation distribution cloud map is a structured dataset that includes vertex color information for rendering, as well as the maximum displacement value, average displacement value, and deformation vector matrix of each preset feature section for numerical analysis. It is understandable that the theoretical deformation distribution cloud map provides a clear basis for deformation quantification for subsequent active compensation control. The solution accuracy of the simplified finite element model of the boom directly depends on the accuracy of the overall stiffness evaluation vector of the boom structure and the authenticity of the material physical property parameters.

[0087] In one embodiment of the present invention, generating an active compensation control signal based on a theoretical deformation distribution cloud map involves a closed-loop control process from deformation prediction to action execution. The predicted deformation displacement vector of a feature point at the end of the arm is extracted from the theoretical deformation distribution cloud map. The feature point is typically chosen as the geometric center of the motor mounting plane at the end of the arm. The predicted deformation displacement vector is compared with the theoretical end-effector position vector in the zero-deformation state. The theoretical end-effector position vector in the zero-deformation state is uniquely determined by the arm's designed unfolding angle and geometric parameters. The comparison operation is vector subtraction, calculating the position deviation vector at the end of the arm, for example, obtaining a position deviation vector of [1.2, -0.3, 0.8] mm. The position deviation vector is converted into the theoretical compensation angle increment required by each joint of the folding arm through inverse kinematics calculation. The inverse kinematics calculation is based on the link parameters and joint constraints of the folding arm. For a two-joint folding arm, the calculation may output a shoulder joint compensation increment of 0.15 degrees and an elbow joint compensation increment of -0.08 degrees. The current angle feedback value of the drive motor of the folding arm structure is queried. This value is provided in real-time by the motor's built-in encoder; for example, the shoulder joint motor feedback angle is 89.95 degrees, and the elbow joint motor feedback angle is 45.02 degrees. The theoretical compensation angle increment is superimposed on the target angle command of the drive motor to form an angle adjustment command. If the original target angle command was 90.00 degrees for the shoulder joint and 45.00 degrees for the elbow joint, the new angle adjustment command becomes 90.15 degrees for the shoulder joint and 44.92 degrees for the elbow joint. Based on the difference between the angle adjustment command and the current angle feedback value, an active compensation control signal is generated using a proportional-integral-derivative (PID) control algorithm. The formula for calculating the torque output command using the PID algorithm is expressed as:

[0088]

[0089] in: This represents the active compensation control signal that is calculated, generated, and ultimately output to the drive motor. Its physical meaning is the motor's desired output torque. This represents the real-time difference between the angle adjustment command and the current angle feedback value. The gain parameter representing the proportional term. The gain parameter representing the integral term. The gain parameter representing the differential term. This represents the integral operation of the following error over time. This represents the rate of change of the following error over time. The active compensation control signal acts directly on the torque output terminal of the drive motor. The drive motor adjusts its current according to the received torque command to achieve precise and rapid fine-tuning of the output shaft angle.

[0090] In practical implementation, the dynamic compensation gain adjustment step based on flight status feedback runs in parallel with the main control loop, continuously receiving real-time flight status data from the UAV's flight control unit. This real-time flight status data is updated at a frequency of 100 Hz and includes at least flight speed, altitude, and attitude angle. From the real-time flight status data, the current airspeed vector and angle of attack are calculated based on the flight speed and attitude angle. For example, if the flight speed is 15 m / s and the pitch angle is 2 degrees, the calculated angle of attack is approximately 5 degrees. A lookup table is established, indexed by airspeed and angle of attack. This lookup table is a two-dimensional table where the row index is the airspeed value and the column index is the angle of attack value. Each cell stores a set of corresponding compensation gain coefficients [Kp, Ki, Kd]. Based on the currently calculated airspeed vector and angle of attack, a set of real-time compensation gain coefficients is retrieved from the lookup table. The lookup process uses bilinear interpolation. For example, if the current airspeed is 12 m / s and the angle of attack is 3 degrees, the real-time Kp, Ki, and Kd are obtained by interpolating from the four closest points in the lookup table. The real-time compensation gain coefficient is applied to the gain parameters of the proportional, integral, and derivative terms of the proportional-integral-derivative control algorithm to dynamically adjust the strength and response speed of the active compensation control signal.

[0091] In some embodiments, the lookup table contains the set of optimal gain parameters tuned using a system identification method under different combinations of airspeed and angle of attack through a series of wind tunnel and flight tests. It can be understood that dynamic gain adjustment allows the compensation system to use a gentler gain to prevent overshoot during hovering (low airspeed, small angle of attack) and a stronger gain to counteract greater aerodynamic deformation during high-speed forward flight (high airspeed, large angle of attack). Optionally, the lookup table may also include additional dimensions related to flight altitude or air density to address more complex environmental changes. The process of calculating and generating the active compensation control signal using the proportional-integral-derivative (PID) control algorithm is a continuous loop. Each control cycle recalculates the torque output command based on the latest angle adjustment command, the current angle feedback value, and the real-time compensation gain coefficient, thereby achieving continuous and adaptive deformation suppression. In some embodiments, the inverse kinematics solution step considers the kinematic limits of the joints to ensure that the calculated theoretical compensation angle increment does not cause the joint command to exceed its mechanical limits. It is understandable that the entire process from deformation prediction to the generation of active compensation control signals constitutes a high-frequency real-time control loop, whose response speed is significantly higher than that of the aircraft's own attitude control loop, in order to achieve active and real-time suppression of structural deformation.

[0092] See Figure 3This is a graph showing the incremental compensation angles of the joints in a drone's folding arm. It illustrates the changes in the shoulder and elbow joint compensation angles over time during the active compensation control phase, representing core control data for suppressing arm structural deformation. The shoulder joint compensation increment fluctuates between approximately -0.20° and +0.20°, with a larger amplitude and a frequency of approximately 0.4Hz, reflecting a more sensitive compensation response to deformation. The elbow joint compensation increment fluctuates between approximately -0.10° and +0.10°, with an amplitude approximately half that of the shoulder joint, exhibiting an inverse phase change compared to the shoulder joint curve, consistent with the kinematic coupling characteristics of the two-joint folding arm. Both curves are superimposed with high-frequency noise to simulate disturbances in real sensor data acquisition and control signals. The compensation increment fluctuates around 0°, indicating that the system continuously fine-tunes the joint angles to suppress arm deformation. The inverse phase change reflects the inverse kinematics calculation results, demonstrating how the coordinated movement of the two joints corrects the end-effector position deviation.

[0093] In one embodiment of the present invention, the arm stiffness attenuation early warning step based on structural health is automatically activated after each flight mission. After the UAV completes a 1200-second inspection flight mission and lands with its power shut off, the maintenance system automatically retrieves the historical data of real-time strain signals recorded throughout the mission from the onboard non-volatile storage unit. The historical data is a complete dataset containing timestamps, bending moment load channels, and torsional load channels. Statistical feature extraction is performed on the historical data to calculate the average value, peak value, and variance rate of change of the strain signals throughout the entire flight mission cycle. The specific implementation process is as follows: From the historical data of real-time strain signals, signal data representing two independent channels of bending moment load and torsional load are extracted respectively. For example, the bending moment channel signal is an array M(t) with a length of 120,000 sampling points (100Hz sampling rate), and the torque channel signal is an array T(t) of the same length. For the signal data of each channel, pulse interference segments caused by known maneuvers are removed. These known maneuvers include sharp turns and rapid ascents and descents. The start and end times of these maneuvers have been marked by the flight control system. The removal method involves replacing the data within the marked time period with linear interpolation of adjacent data, resulting in a smoothed strain signal sequence. The arithmetic mean of the smoothed strain signal sequence is calculated over the entire mission duration to obtain the average strain signal. The calculation formula is as follows:

[0094]

[0095] in: represents the k-th data point in the smoothed strain signal sequence, and N represents the total number of data points in the sequence. The absolute value represents the calculated average strain signal. This value is used to calculate the average load amplitude. The point with the largest absolute value in the smoothed strain signal sequence is searched, and its value and occurrence time are recorded; this value is the peak value of the strain signal. The total task duration of 1200 seconds is divided into 24 sub-time periods of 50 seconds each. The variance of the strain signal sequence in each sub-time period is calculated using the standard unbiased estimation formula. The change in variance between adjacent sub-time periods is also calculated, and the average of all changes is then taken to obtain the variance change rate. Table 1 below shows an example of the statistical characteristics calculated for the bending moment channel of the left front boom after one task:

[0096] Table 1: Examples of Statistical Characteristics of Single Task in the Bending Moment Channel of the Left Front Arm

[0097]

[0098] In some embodiments, smoothing can also be implemented using a low-pass digital filter, with the cutoff frequency set to half of the highest known frequency of maneuvering. It can be understood that the average value reflects the overall load level, the peak value reflects extreme load conditions, and the variance rate of change reflects the trend of load fluctuation. The calculated average value, peak value, and variance rate of change are input into a pre-trained structural health assessment model. This model is a machine learning model based on gradient boosting decision trees, which uses a large amount of data containing the above statistical features and corresponding known structural state labels during the training phase. The structural health index output by the structural health assessment model is obtained. The structural health index is a value between zero and one, for example, 0.92. The lower the value, the higher the degree of stiffness decay. If the structural health index falls below a preset warning threshold, for example, a threshold of 0.80, a stiffness degradation warning message is generated, containing the specific arm number and recommended inspection items. The warning message format is "Warning: The structural health index of the left front arm is 0.75, which is below the threshold of 0.80. It is recommended to check the torque of the hinge bolts at the arm root and whether there are cracks in the composite skin." The stiffness degradation warning message is recorded in the system's structured maintenance log database. Optionally, the stiffness degradation warning message can be displayed via a pop-up window on the ground station software interface and simultaneously sent to maintenance personnel via email. The training data for the structural health assessment model comes from the correspondence between structural status and flight data accumulated by a fleet of the same model of UAVs during long-term operation and confirmed by manual inspection. The process from retrieving historical data from the onboard storage unit to generating warning messages is fully automated, requiring no manual intervention in data analysis. In some embodiments, when calculating the variance change rate, the standard deviation of the strain signal sequence is calculated for each sub-time period, then the rate of change of the standard deviation of adjacent sub-time periods is calculated, and then all rates of change are averaged. This method can reduce the influence of dimensions. It is understandable that the arm stiffness attenuation early warning step based on structural health enables long-term, quantitative trend monitoring of the arm's structural status, providing data support for predictive maintenance.

[0099] See Figure 4This is a comparison chart of the deformation compensation effect at the end of a drone's folding arm, visually demonstrating the trend of deformation at the end of the drone's folding arm over time with and without active compensation control. The deformation without compensation generally shows an upward trend, ranging from approximately 13.5mm to 18.3mm, reaching a peak of 18.3mm at 60 seconds, reflecting the cumulative deformation of the structure under continuous load. The deformation with compensation is consistently suppressed within the range of 10.5mm to 15.7mm, reaching only 11.3mm at 60 seconds, significantly lower than the uncompensated state. Over time, the deformation with compensation remains lower than the uncompensated state, and the suppression rate gradually increases to over 38%, verifying the effectiveness of the active compensation control strategy. The deformation curve with compensation fluctuates less, indicating that the system can stably resist load changes and structural deformation, avoiding the risk of continuous deformation accumulation in the uncompensated state. Controlling the end-effector deformation to around 11mm can effectively maintain the drone's attitude stability and operational accuracy, reducing the risk of flight accidents caused by structural deformation.

[0100] In one embodiment of the present invention, the optimal arm deployment angle decision-making step for different mission modes is executed before the start of the flight mission. The mission mode code for the upcoming flight mission is obtained. This mission mode code is specified by the flight planning software and transmitted to the arm control system. The mission mode code corresponds to one of the following: cruise mode, high-speed maneuvering mode, hovering observation mode, or wind-resistant attitude stabilization mode. The historical best flight database corresponding to the mission mode code is retrieved. This historical best flight database is a relational database stored in onboard memory or on a ground server. It stores the correlation records between different arm deployment angles, flight states, and final structural deformation data under the same mission mode. An example record includes the following fields: mission mode = "HOVER_OBS", arm deployment angle = 85 degrees, average wind speed = 5.0 m / s, load = 1.0 kg, and strain signal peak value = 72.5 microstrain.

[0101] Based on the current estimated mission environment parameters, a matching query is performed in the historical best flight database. The estimated mission environment parameters include at least the predicted wind speed and predicted payload. For example, the current estimated mission environment parameters are a predicted wind speed of 5.5 m / s and a predicted payload of 0.8 kg. The specific process of the matching query includes reading the estimated mission environment parameters from the flight mission planning system, and calculating the similarity distance between the environmental parameters of each historical mission record and the current estimated mission environment parameters in the historical best flight database. The formula for calculating the similarity distance is:

[0102]

[0103] in: The similarity distance between the environmental parameters of the j-th historical task record and the estimated parameters of the current task environment is dimensionless. The projected wind speed represents the current mission. This represents the average wind speed recorded in the j-th historical task record; Represents the expected payload for the current mission; Represents the load recorded in the j-th historical mission record; and These are the weighting coefficients for wind speed and load parameters, used to adjust the importance of different parameters in distance calculation. A similarity threshold is set, for example, a similarity threshold of 2.0. All historical task records with similarity distances less than the similarity threshold are selected to form a candidate record set. For each record in the candidate record set, it is sorted according to the size of its final structural deformation data. The final structural deformation data is quantified by the average or peak value of the strain signal in the corresponding historical task, for example, sorted in ascending order using the peak value of the strain signal. From the sorting results, several records with the smallest final structural deformation data are selected as the optimal matching records, used to extract the initial recommended deployment angle. For example, the top 3 records are selected, with arm deployment angles of 84 degrees, 85 degrees, and 86 degrees, respectively. In some embodiments, the average (85 degrees) or median of these angles can be calculated as the initial recommended deployment angle. Before the UAV takes off, the initial recommended deployment angle is sent as a preset command to the drive motor of the folding arm structure via the communication bus. The drive motor controls the arm to deploy to the initial recommended deployment angle according to the angle command.

[0104] Optionally, the historical best flight database can be automatically updated with data after each successful flight. The actual environmental parameters, deployment angle, and calculated structural deformation data of the current flight are stored as a new record in the database, enabling self-learning and optimization. When selecting the records with the smallest final structural deformation data from the sorting results, if the candidate record set is empty or the number of records is insufficient, the default safe deployment angle for that mission mode is used. It can be understood that the process of matching and querying the historical best flight database based on the current mission environment's estimated parameters is essentially an instance-based reasoning process, using historical best flight experience to provide a data-driven, personalized pre-configuration of the arm deployment attitude for the current mission. In some embodiments, the mission environment's estimated parameters may also include the expected flight altitude or ambient temperature; in this case, the similarity distance calculation formula needs to be updated accordingly. The initial recommended deployment angle is not fixed during flight; the main flight control system or real-time active compensation control signals can still dynamically fine-tune it.

[0105] See Figure 5This is a similarity matching analysis chart for UAV mission environments, showing the similarity distance between historical and current missions under different wind speeds. The blue curve represents the similarity distance at different wind speeds, peaking at 2.0 at 5.0 m / s. The red dot indicates the current mission's wind speed is 5.5 m / s, corresponding to a similarity distance of 1.6. The similarity distance initially increases and then decreases with wind speed, reaching the upper threshold at 5.0 m / s. The current mission's similarity distance (1.6 < 2.0) at 5.5 m / s satisfies the candidate matching condition. The current mission's similarity distance at 5.5 m / s is 1.6, below the threshold of 2.0, and can be included in the candidate mission set for extracting the optimal arm deployment angle. The similarity distance at 5.0 m / s is exactly equal to the threshold, representing the matching boundary; deviating from this wind speed, the similarity distance decreases, and environmental similarity increases. This data provides a quantitative basis for "selecting the record with the least deformation from candidate missions," ensuring that the optimal deployment angle decision is based on historical experience under similar environments.

[0106] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for operating a folding arm structure for a drone, characterized in that, The working method includes: When the UAV is in the unfolded operation state, real-time strain signals are obtained by stress sensors deployed at the hinge of the folding arm structure. The real-time strain signals include time history data of the bending moment and torsional load borne by the arm during flight. Displacement sensors installed at the drive motor of the folding arm structure synchronously collect the instantaneous value sequence of the arm deployment angle, and the instantaneous value sequence of the arm deployment angle is aligned with the real-time strain signal on the timeline; The real-time strain signal and the instantaneous value sequence of the boom deployment angle are input into a pre-trained load-deformation correlation model, which is used to analyze the dynamic response of the boom structure under different load modes. Based on the analysis results of the pre-trained load-deformation correlation model, the theoretical deformation distribution cloud map of the folding arm structure at multiple preset feature sections is calculated. The theoretical deformation distribution cloud map reflects the deformation trend and weak points of the arm structure in space. Based on the theoretical deformation distribution cloud map, an active compensation control signal is generated for the drive motor of the folding arm structure. The active compensation control signal is used to fine-tune the unfolding angle of the arm to suppress the structural deformation predicted by the theoretical deformation distribution cloud map.

2. The working method of the folding arm structure of a UAV according to claim 1, characterized in that, The real-time strain signal and the instantaneous value sequence of the boom deployment angle are input into a pre-trained load-deformation correlation model, including: Modal decomposition is performed on the real-time strain signal to separate the trend component representing the low-frequency principal stress and the detail component representing the high-frequency vibration stress. The instantaneous value sequence of the arm deployment angle is calculated by difference to obtain the variation curves of the arm deployment angular velocity and angular acceleration; The separated trend component, detail component and the change curves of the arm deployment angular velocity and angular acceleration are aligned and fused according to the timestamps to form a structural load state matrix. The structural load state matrix is ​​input into a convolutional neural network model trained using historical flight data, i.e., the pre-trained load-deformation correlation model. The forward propagation calculation of the convolutional neural network model outputs the overall stiffness evaluation vector of the boom structure corresponding to the current structural load state matrix.

3. The working method of the folding arm structure of a UAV according to claim 1, characterized in that, Based on the analysis results of the pre-trained load-deformation correlation model, the theoretical deformation distribution cloud map of the folding arm structure at multiple preset characteristic sections is calculated, including: Receive the overall stiffness evaluation vector of the boom structure output by the pre-trained load-deformation correlation model; Based on the UAV 3D structure database, a 3D solid model and material physical property parameters of the folding arm structure are extracted. The material physical property parameters include at least the elastic modulus and Poisson's ratio. On the three-dimensional solid model, multiple preset feature sections are selected along the axis of the boom. The location of the preset feature sections is selected based on the abrupt change in structural geometry or the critical node of the load transfer path. Using the overall stiffness evaluation vector of the boom structure as the boundary condition, and in combination with the material physical property parameters, a quasi-static mechanical solution is performed on the simplified finite element model of the boom containing multiple preset feature sections. Displacement field data on each preset feature section is extracted from the quasi-static mechanical solution, and the displacement field data of all sections are mapped back to three-dimensional spatial coordinates to generate the theoretical deformation distribution cloud map.

4. The working method of the folding arm structure of a UAV according to claim 3, characterized in that, Using the overall stiffness evaluation vector of the boom structure as boundary conditions, and combining the material physical property parameters, a quasi-static mechanical solution is performed on the simplified finite element model of the boom containing multiple preset feature sections, including: The overall stiffness evaluation vector of the arm structure is decomposed into force components and moment components corresponding to the local coordinate system of the arm. The force components and torque components are used as nodal loads and applied to the corresponding fixed nodes at the root of the boom in the simplified finite element model of the boom. The initial displacement and rotation of all nodes in the simplified finite element model of the robotic arm, except for the root fixed node, are set to zero; The static solver based on the finite element method is invoked, and the displacement and stress response of all nodes of the simplified finite element model of the boom under the nodal load is calculated using the material physical property parameters. From the displacement response output by the statics solver, the displacement data of the grid nodes located on the preset characteristic cross section are selected to form the preliminary displacement field of each cross section.

5. The working method of the folding arm structure of a UAV according to claim 1, characterized in that, Based on the theoretical deformation distribution cloud map, an active compensation control signal is generated for the drive motor of the folding arm structure, including: Extract the predicted deformation displacement vector of the feature points at the end of the boom from the theoretical deformation distribution cloud map; The predicted deformation displacement vector is compared with the theoretical end position vector in the zero deformation state to obtain the position deviation vector of the computer arm end. The position deviation vector is converted into the theoretical compensation angle increment required for each joint of the folding arm through inverse kinematics calculation; The current angle feedback value of the drive motor of the folding arm structure is queried, and the theoretical compensation angle increment is superimposed on the target angle command of the drive motor to form an angle adjustment command. Based on the difference between the angle adjustment command and the current angle feedback value, the active compensation control signal is generated by a proportional-integral-derivative control algorithm. The active compensation control signal directly acts on the torque output of the drive motor.

6. The working method of the folding arm structure of a UAV according to claim 5, characterized in that, The working method also includes a dynamic compensation gain adjustment step based on flight status feedback: Continuously receive real-time flight status data from the flight control unit of the UAV, wherein the real-time flight status data includes at least flight speed, flight altitude and attitude angle; Calculate the current airspeed vector and angle of attack from the real-time flight status data based on the flight speed and attitude angle; Establish a lookup table indexed by airspeed and angle of attack, which stores the compensation gain coefficients corresponding to different flight states; Based on the current airspeed vector and airflow angle of attack, a set of real-time compensation gain coefficients are obtained by querying the lookup table. The real-time compensation gain coefficient is applied to the gain parameters of the proportional, integral, and derivative terms of the proportional-integral-derivative control algorithm to dynamically adjust the strength and response speed of the active compensation control signal.

7. The working method of the folding arm structure of a UAV according to claim 1, characterized in that, The working method also includes a boom stiffness attenuation early warning step based on structural health: After each UAV flight mission is completed, historical data of all real-time strain signals from the mission are retrieved from the onboard storage unit. Statistical features are extracted from the historical data to calculate the average value, peak value, and variance rate of change of the strain signal throughout the entire flight mission cycle. The calculated average value, peak value, and variance rate of change of the strain signal are input into a pre-trained structural health assessment model, which is used to identify small attenuation trends in the stiffness of the boom structure. Obtain the structural health index output by the structural health assessment model. The structural health index is a value between zero and one, and the lower the value, the higher the stiffness decay. If the structural health index is lower than the preset warning threshold, a stiffness attenuation warning message containing the specific arm number and suggested inspection items will be generated and recorded in the maintenance log.

8. The working method of the folding arm structure of a UAV according to claim 7, characterized in that, Statistical feature extraction is performed on the historical data to calculate the average value, peak value, and variance rate of change of the strain signal throughout the entire flight mission cycle, including: From the historical data of the real-time strain signal, extract the signal data of two independent channels representing bending moment and torsional load respectively; For the signal data of each channel, the pulse interference segment caused by known maneuvering motion is removed to obtain a smoothed strain signal sequence; The arithmetic mean of the smoothed strain signal sequence over the entire mission duration is calculated to obtain the average value of the strain signal. Search for the point with the maximum absolute value in the smoothed strain signal sequence, record its value and the time of occurrence, and the value is the peak value of the strain signal. The task duration is divided into several equal-length sub-time periods. The variance of the strain signal sequence in each sub-time period is calculated, and the change in variance between adjacent sub-time periods is calculated. The average value of all changes is then taken to obtain the variance change rate.

9. The working method of the folding arm structure of a UAV according to claim 1, characterized in that, The working method also includes a decision-making step for the optimal arm deployment angle for different task modes: Obtain the mission mode code for the upcoming flight mission, wherein the mission mode code corresponds to one of the following: cruise mode, high-speed maneuvering mode, hovering observation mode, or wind-resistant attitude stabilization mode; Retrieve the historical best flight database corresponding to the mission mode code. The historical best flight database stores the correlation records between different arm deployment angles, flight states and final structural deformation data under the same mission mode. Based on the current mission environment prediction parameters, a matching query is performed in the historical best flight database. The mission environment prediction parameters include at least the predicted wind speed and the predicted payload. From the results of the matching query, the record with the smallest structural deformation data under the estimated parameters of similar task environment is selected, and the corresponding arm deployment angle in the record is extracted as the initial recommended deployment angle. Before the drone takes off, the initial recommended deployment angle is sent as a preset command to the drive motor of the folding arm structure to control the arm to deploy to the initial recommended deployment angle. The matching query in the historical best flight database based on the estimated parameters of the current mission environment specifically includes: Read the predicted mission environment parameters from the flight mission planning system; In the historical best flight database, the similarity distance between the environmental parameters of each historical mission record and the estimated environmental parameters of the current mission is calculated. Set a similarity threshold, and select all historical task records whose similarity distance is less than the similarity threshold to form a candidate record set; For each record in the candidate record set, sort them according to the size of the final structural deformation data of the record, wherein the final structural deformation data is quantified by the average or peak value of the strain signal in the corresponding historical task; From the sorting results, select several records with the smallest final structural deformation data as the optimal matching records, which are used to extract the initial recommended unfolding angle.

10. A folding arm structure for a drone, characterized in that, include: processor; Memory, which stores computer programs; When the processor executes the computer program, it implements a method for operating the folding arm structure of a drone according to any one of claims 1 to 9.

Images