Six-degree-of-freedom robotic arm for aerial work, adaptive anti-disturbance control method and system thereof

Abstract

The present invention relates to the field of unmanned aerial vehicle control technology, and in particular to a six-degree-of-freedom robotic arm for aerial work, and its adaptive anti-disturbance control method and system. The adaptive anti-disturbance control method comprises: using an arm body designed with carbon fiber topology optimization and a joint shell provided with a guide groove to manufacture a robotic arm structure; constructing a dynamic disturbance observer based on the DDPG algorithm, outputting a dynamic compensation value through the dynamic disturbance observer, and combining the disturbance compensation algorithm to calculate the wind disturbance value and integrate it into the control law of the unmanned aerial vehicle; adjusting the impedance parameters online based on the TD3 algorithm; generating a real-time obstacle avoidance path based on an optimized multi-objective PPO path planning algorithm, and executing the path planning task of the corresponding real-time obstacle avoidance path. The present invention has the advantages of significantly enhanced wind disturbance resistance, significantly optimized load mutation stability, significantly reduced mission energy consumption, significantly improved adaptability to confined spaces, and light weight.

Description

Technical Field

[0001] The present invention relates to the field of unmanned aerial vehicle (UAV) control technology, and in particular to a six-degree-of-freedom robotic arm for aerial operations, and an adaptive anti-disturbance control method and system thereof. Background Art

[0002] As the application of unmanned aerial vehicles (UAVs) in power equipment inspection, wind turbine blade maintenance, and overhead line maintenance continues to grow, the manipulator arms of traditional flying robots still face significant limitations in several areas. For example, during high-altitude wind turbine blade maintenance operations, the impact of strong winds on the manipulator arms presents a major challenge. Traditional manipulator arms are easily disturbed by strong winds, causing their end-piece position to shift, which in turn affects the manipulator's operational accuracy, work quality, and efficiency. This is primarily due to the structural design of traditional manipulator arms' inadequacies in handling strong winds. The joint housings lack proper wind diversion, resulting in significant wind resistance. Furthermore, the heavy arm mass creates inertial shock, which can easily cause the end-piece position to shift. Another example is when carrying tools of varying weights at high altitude during overhead line maintenance, the load can suddenly change. Traditional manipulator arms, lacking effective mechanisms to identify and respond to these load changes, are prone to significant vibration, which not only affects operational stability but can also damage the arm itself. At the same time, in scenarios involving long, continuous operations, traditional path planning methods focus solely on path length while ignoring joint energy consumption. This results in frequent joint movements of conventional robotic arms. Failure to consider joint energy consumption leads to energy waste and increased operating costs. Furthermore, in complex working environments, existing obstacle avoidance algorithms, which rely on offline calculations, struggle to cope with dynamic obstacles. For example, at a bridge maintenance site, dynamic obstacles such as people and vehicles may appear. Traditional obstacle avoidance algorithms are unable to respond in a timely manner, easily causing collisions between the robotic arm and obstacles, impacting work safety and efficiency.

[0003] To this end, the present application provides a six-degree-of-freedom robotic arm for aerial work, and its adaptive anti-disturbance control method and system. Summary of the Invention

[0004] Based on this, it is necessary to provide a six-degree-of-freedom robotic arm for aerial work, and its adaptive anti-disturbance control method and system to address the above technical problems.

[0005] According to a first aspect of the present invention, an adaptive anti-disturbance control method for a six-degree-of-freedom robotic arm for aerial work is provided, comprising: manufacturing a robotic arm structure using an arm body designed with carbon fiber topology optimization and a joint shell provided with a guide groove; constructing a dynamic disturbance observer based on the DDPG algorithm, outputting a dynamic compensation amount through the dynamic disturbance observer, and combining the disturbance compensation algorithm to calculate the wind disturbance value and integrate it into the control law of the unmanned aerial vehicle for real-time dynamic anti-disturbance compensation; adjusting the impedance parameters online based on the TD3 algorithm to cope with sudden load changes; generating a real-time obstacle avoidance path based on an optimized multi-objective PPO path planning algorithm, and executing the path planning task of the corresponding real-time obstacle avoidance path to cope with dynamic obstacles.

[0006] Optionally, the arm body designed by carbon fiber topology optimization is an internal truss structure generated after multi-working condition topology optimization of the arm body based on finite element analysis software, and the internal truss structure is made of carbon fiber composite materials, wherein the load conditions of multi-working condition topology optimization include a wind pressure of 15m / s and an end load of 10kg, the constraints are maximum stress ≤300MPa and first-order modal frequency ≥50Hz, and the objective function includes minimizing the mass of the arm body and maximizing the stiffness of the arm body.

[0007] Optionally, the guide groove is arranged on the surface of the joint shell and has a V-shaped structure, defining geometric parameters including groove depth, inclination and spacing, and the drag coefficient under a wind pressure condition of 15m / s is simulated and verified by CFD software to optimize aerodynamic performance.

[0008] Optionally, the method constructs a dynamic disturbance observer based on the DDPG algorithm, outputs a dynamic compensation through the dynamic disturbance observer, and combines the disturbance compensation algorithm to calculate the wind disturbance value and integrate it into the control law of the unmanned aerial vehicle, including: collecting real-time wind pressure data, joint state data and historical error data to construct a state space data set; constructing and training the DDPG algorithm, and constructing a dynamic disturbance observer based on the trained DDPG algorithm; wherein the DDPG algorithm consists of a first Actor network for outputting a dynamic compensation and a first Critic network for evaluating the value of the dynamic compensation, the first Actor network and the first Critic network The c network consists of a three-layer fully connected structure, with 256, 128, and 64 output nodes in each layer, respectively. During the DDPG algorithm training process, a state-space dataset is input and the dynamic compensation value in the corresponding state space is output. The loss value in the real-time state space is calculated based on the mean square tracking error loss function, and the DDPG algorithm parameters are updated online with real-time wind pressure data to obtain the online updated DDPG algorithm. Based on the online updated DDPG algorithm, the dynamic compensation value in the real-time state space is output. Combined with the disturbance compensation algorithm, the wind disturbance value is calculated and integrated into the control law of the unmanned aerial vehicle for real-time dynamic anti-disturbance compensation.

[0009] Optionally, the real-time wind pressure data is collected by multiple pressure sensors arranged on the surface of the robotic arm structure, the joint state data includes the joint angle, joint speed and current loop torque fed back by the joint encoder, and the historical error data is the end posture tracking error data within a defined historical control period.

[0010] Optionally, the formula of the disturbance compensation algorithm is: Where, is the wind disturbance value, is the dynamic compensation amount, is the state space dataset, and is obtained after calibration by the particle swarm optimization algorithm, where The value is 12. The value is 0.5, To control the error.

[0011] Optionally, the online adjustment of impedance parameters based on the TD3 algorithm includes: collecting real-time end contact force, joint torque deviation data and trajectory error data, and constructing an observation state data set; wherein the end contact force includes three force components and three torque components collected by the six-dimensional force sensor, the three force components are Fx, Fy, and Fz, the three torque components are Mx, My, and Mz, the joint torque deviation data is the difference between the actual torque of each joint and the expected value, and the trajectory error data includes the end position deviation value and the end velocity deviation value; constructing and training the TD3 algorithm; wherein the TD3 algorithm consists of a second Actor network for outputting stiffness and damping and a In order to prevent over-estimation, the second Critic network is composed of a four-layer fully connected structure. The output nodes of each layer of the fully connected structure are 512, 256, 128 and 64 respectively. The second Critic network adopts a double-Q network structure. When training the TD3 algorithm, the observation state data set is input and combined with the defined OU noise to obtain the trained TD3 algorithm. Based on the trained TD3 algorithm, the stiffness and damping under the real-time observation state are output and used as the impedance parameters under the real-time observation state. Based on the impedance parameters under the real-time observation state, the corresponding impedance parameters are adjusted through PID control to cope with sudden load changes.

[0012] Optionally, the optimized multi-objective PPO path planning algorithm generates a real-time obstacle avoidance path and performs a path planning task for the corresponding real-time obstacle avoidance path, including: collecting real-time environmental perception data, the real-time environmental perception data including real-time point cloud data collected by a binocular vision camera and IMU attitude data collected by an IMU attitude sensor, downsampling the real-time point cloud data to obtain a 128×128 pixel depth map, and inputting it into a three-layer convolutional neural network for feature extraction, outputting dynamic obstacle features, and state encoding the dynamic obstacle features and IMU attitude data to obtain a state encoding result; optimizing the multi-objective PPO path planning algorithm through a defined dynamic reward function and a policy update mechanism, wherein the formula of the dynamic reward function is: ; Where R is the total reward, 、 、 is the weight coefficient, which is used to balance the importance of each reward item. The path completion degree indicates the degree of completion of the current path, and its value range is between 0 and 1. Energy consumption represents the energy consumed by the robotic arm structure during the operation task. It is calculated as the sum of the squares of the joint torques. The specific formula is: , For the The torque of each joint, is the total number of joints, The collision risk is expressed as the inverse of the distance between the robotic arm structure and the nearest obstacle, and the normalization is performed; a policy network structure is constructed and trained, wherein the policy network structure consists of an LSTM layer and a fully connected layer for outputting the probability distribution of each joint motion. During the policy network structure training process, the state encoding result is input, and the probability distribution of each joint motion under the corresponding environmental state is output. The total reward under the corresponding environmental state is calculated in combination with the dynamic reward function. The policy network structure parameters are updated based on the total reward under the corresponding environmental state to obtain the optimized multi-objective PPO path planning algorithm; dynamic obstacles in the environment are detected in real time by a binocular vision camera, and according to the corresponding state encoding result and the defined local path update time, the local path replanning mechanism is triggered to replan the local path. Based on the optimized multi-objective PPO path planning algorithm, the probability distribution of each joint motion under the corresponding local path is output, and the optimal joint motion is selected, and the optimal joint motion is used as the real-time obstacle avoidance path; in response to the real-time obstacle avoidance path, the path planning task of the corresponding real-time obstacle avoidance path is executed.

[0013] According to a second aspect of the present invention, an adaptive anti-disturbance control system for a six-degree-of-freedom robotic arm for aerial work is provided, comprising: a structure manufacturing module for manufacturing a robotic arm structure using an arm body designed with carbon fiber topology optimization and a joint shell provided with a guide groove; a real-time dynamic anti-disturbance compensation module for constructing a dynamic disturbance observer based on the DDPG algorithm, outputting a dynamic compensation amount through the dynamic disturbance observer, and calculating the wind disturbance value in combination with the disturbance compensation algorithm and integrating it into the control law of the unmanned aerial vehicle for real-time dynamic anti-disturbance compensation; an impedance parameter adjustment module for online adjusting the impedance parameters based on the TD3 algorithm to cope with sudden load changes; an obstacle avoidance path planning module for generating a real-time obstacle avoidance path based on an optimized multi-objective PPO path planning algorithm, and executing the path planning task of the corresponding real-time obstacle avoidance path to cope with dynamic obstacles.

[0014] According to a third aspect of the present invention, there is provided a six-degree-of-freedom robotic arm for aerial work, including a robotic arm structure, which is manufactured using the adaptive anti-disturbance control method.

[0015] The advantages and beneficial effects of the present invention are: 1. The ability to resist wind disturbance is significantly enhanced: through the structural design and optimization of the guide groove, the wind resistance coefficient can be effectively reduced. Combined with the lightweight characteristics of the carbon fiber arm, the inertial impact of wind load on the mechanical arm structure can be greatly reduced, and the terminal positioning accuracy can be improved. At the same time, combined with the dynamic disturbance observer, by estimating the wind disturbance value in real time and performing feedforward compensation, the terminal positioning error under the wind pressure condition of 15m / s can be effectively reduced, and the compensation response delay time can be effectively reduced; 2. The load mutation stability is significantly optimized: the six-dimensional force sensor is used to fuse the joint encoder data to achieve millisecond-level synchronous feedback of the terminal contact force and the joint torque, thereby improving the sensitivity of identifying load changes. At the same time, combined with the dynamic impedance control technology, it can be adjusted according to the load mass Adjust the impedance parameters in real time, thereby effectively reducing the stabilization time and vibration amplitude when the load is suddenly added; 3. Breakthrough improvement in energy efficiency: By optimizing the multi-objective PPO path planning algorithm, introducing energy consumption reward items and energy consumption weight coefficients in the dynamic reward function, the task energy consumption can be significantly reduced, thereby improving energy utilization efficiency. At the same time, combined with the arm body designed with carbon fiber topology optimization, the ineffective energy consumption of the joint motor during the acceleration stage is further reduced by reducing the arm body inertia; 4. Significantly improved adaptability to narrow spaces: Through the synergy of online real-time obstacle avoidance and compact joint design, not only can new paths be generated and responded to quickly, but the axial dimension of the joint of the robotic arm structure can also be reduced, so that the minimum passable gap of the robotic arm structure is ≤200mm, which meets the needs of narrow scenarios such as bridge maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present invention. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:

[0017] Figure 1 This is a flow chart of an embodiment of an adaptive anti-disturbance control method for a six-degree-of-freedom robotic arm for aerial work according to the present application.

[0018] Figure 2 FIG. 4 is a comparison curve diagram of an embodiment of a load mutation PID control response according to the present application.

[0019] Figure 3 The figure is a comparison curve diagram of an embodiment of the energy consumption of flight operations according to the present application.

[0020] Figure 4 Schematic diagram of the structure of an embodiment of an adaptive anti-disturbance control system for a six-degree-of-freedom robotic arm for aerial work according to the present application.

[0021] Figure 5 Schematic diagram of the structure of a robotic arm according to an embodiment of the present application.

[0022] Figure 6 It is a structural schematic diagram of an embodiment of the arm body according to the present application.

[0023] Among them, 100, robotic arm structure; 200, guide trough; 300, joint; 400, joint shell; 500, arm body. DETAILED DESCRIPTION

[0024] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art. It should be noted that, unless there is a conflict, the embodiments of the present disclosure and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0025] Example 1

[0026] Reference Attachment Figure 1 , the present application provides an adaptive anti-disturbance control method for a six-degree-of-freedom robotic arm for aerial work, the method comprising the following steps.

[0027] S1: The robotic arm structure is made using an arm body designed with carbon fiber topology optimization and a joint shell with guide grooves.

[0028] Specifically, traditional robotic arms mostly use aluminum alloy or steel arms, which have the following defects: (1) high inertia moment: traditional materials have high density (steel: 7.8g / cm³), which limits the dynamic response speed; (2) high drag coefficient: aerodynamic design is not optimized, the drag coefficient is ≥0.8, and the additional power consumption increases by 30% under strong winds. To this end, this application uses an arm body designed with carbon fiber topology optimization and a joint shell with guide grooves to manufacture a robotic arm structure. Among them, the internal truss structure is generated through finite element topology optimization, which reduces its mass by 40% and increases its stiffness by 25%. Combined with the guide groove design of the joint shell, and the geometric parameters including groove depth, inclination and spacing are defined, CFD software simulation and verification show that the drag coefficient under 15m / s wind pressure conditions is reduced from 0.8 to 0.35, and the joint drive power consumption under the same operation task is reduced by 28%.

[0029] Furthermore, the arm body designed by carbon fiber topology optimization is an internal truss structure generated after multi-working condition topology optimization of the arm body based on finite element analysis software, and the internal truss structure is made of carbon fiber composite materials. Among them, the load conditions of multi-working condition topology optimization include a wind pressure of 15m / s and an end load of 10kg. The constraints are that the maximum stress is ≤300MPa and the first-order modal frequency is ≥50Hz. The objective function includes minimizing the mass of the arm body and maximizing the stiffness of the arm body.

[0030] Furthermore, the guide groove is arranged on the surface of the joint shell and has a V-shaped structure. The geometric parameters including groove depth, inclination and spacing are defined. The drag coefficient under the wind pressure condition of 15m / s is simulated and verified by CFD software to optimize the aerodynamic performance; preferably, the groove depth is 3mm, the inclination is 15°, and the spacing is 10mm.

[0031] S2: A dynamic disturbance observer is constructed based on the DDPG algorithm. The dynamic compensation value is output by the dynamic disturbance observer. Combined with the disturbance compensation algorithm, the wind disturbance value is calculated and integrated into the control law of the unmanned aerial vehicle for real-time dynamic anti-disturbance compensation.

[0032] Specifically, when working at high altitude, traditional manipulators mostly use fixed feedforward compensation or equivalent degree of freedom models based on preset constant wind loads (such as CN117325143A). Their defects are: (1) dependence on offline wind field models: unable to perceive dynamic wind speed changes (such as gusts, turbulence, etc.) in real time, resulting in compensation lag; (2) sensitivity to high-frequency disturbances: fixed gain parameters are prone to oscillation under high-frequency random disturbances. Under a wind speed of 10m / s, the measured terminal positioning error is ≥±1.2mm. To this end, this application adopts a dynamic disturbance observer based on reinforcement learning, uses the DDPG algorithm to output real-time dynamic compensation, and combines it with a disturbance compensation algorithm to optimize the wind disturbance value in real time to achieve feedforward compensation. It can dynamically adapt to wind speed changes and reduce the compensation delay to within 10ms. At the same time, under a wind speed of 15m / s, the terminal positioning error is ≤±0.3mm, and the anti-disturbance accuracy is improved by 60%.

[0033] Furthermore, a dynamic disturbance observer is constructed based on the DDPG algorithm. The dynamic compensation is output through the dynamic disturbance observer. Combined with the disturbance compensation algorithm, the wind disturbance value is calculated and integrated into the control law of the unmanned aerial vehicle. The method includes: collecting real-time wind pressure data, joint state data and historical error data to construct a state space data set; constructing and training the DDPG algorithm, and constructing a dynamic disturbance observer based on the trained DDPG algorithm; wherein the DDPG algorithm consists of a first Actor network for outputting dynamic compensation and a first Critic network for evaluating the value of the dynamic compensation. The first Actor network and the first Critic network The networks consist of a three-layer fully connected structure, with 256, 128, and 64 output nodes in each layer, respectively. During the DDPG algorithm training process, a state-space dataset is input and the dynamic compensation value in the corresponding state space is output. The loss value in the real-time state space is calculated based on the mean square tracking error loss function, and the DDPG algorithm parameters are updated online in combination with real-time wind pressure data to obtain an online updated DDPG algorithm. Based on the online updated DDPG algorithm, the dynamic compensation value in the real-time state space is output. Combined with the disturbance compensation algorithm, the wind disturbance value is calculated and integrated into the control law of the unmanned aerial vehicle for real-time dynamic anti-disturbance compensation.

[0034] Furthermore, real-time wind pressure data is collected by multiple pressure sensors arranged on the surface of the robotic arm structure. For example, real-time wind pressure data can be collected by 16 pressure sensors (such as the pressure sensor model MPX5700AP) arranged on the surface of the robotic arm structure; the joint state data includes the joint angle, joint speed and current loop torque fed back by the joint encoder; the historical error data is the end posture tracking error data within the defined historical control cycle, for example, the end posture tracking error in the past 5 control cycles.

[0035] Furthermore, in the process of constructing and training the DDPG algorithm, this application adopts the initial DDPG algorithm constructed by the first Actor network and the first Critic network, wherein the first Actor network is used to output the dynamic compensation amount The first critic network is used to evaluate the value of dynamic compensation, specifically the state-action value. The first actor network and the first critic network have the same structure, consisting of three fully connected layers. Each fully connected layer has 256, 128, and 64 output nodes, respectively. Each layer's output undergoes a nonlinear transformation using the ReLU activation function to enhance the model's expressiveness. During the initial training phase of the DDPG algorithm, 10 hours of wind pressure data, joint state data, and historical error data within a wind speed range of 5-20 m / s are collected to construct a state-space dataset. This dataset is then fed into the initial DDPG algorithm for training. The loss value in the real-time state space is calculated using the mean squared tracking error loss function. The DDPG algorithm parameters are then updated online using the real-time wind pressure data (e.g., once every hour), resulting in an online updated DDPG algorithm.

[0036] Furthermore, the formula of the disturbance compensation algorithm is: Where, is the wind disturbance value, is the dynamic compensation amount, is the state space dataset, and is obtained after calibration by the particle swarm optimization algorithm, where The value is 12. The value is 0.5, To control the error.

[0037] Furthermore, based on the online updated DDPG algorithm, a real-time state-space dataset is input and the dynamic compensation value in the real-time state space is output. Combined with the disturbance compensation algorithm, the wind disturbance value is calculated and integrated into the control law of the unmanned aerial vehicle to implement feedforward compensation for real-time dynamic anti-disturbance compensation.

[0038] S3: Online adjustment of impedance parameters based on the TD3 algorithm to cope with sudden load changes.

[0039] Specifically, the existing technology mostly adopts the impedance control framework to achieve force / position hybrid control by setting fixed stiffness parameters (such as CN110653805A). Specifically, in the task constraint space, the impedance model is defined based on the fixed inertia, damping and stiffness matrix, and the impedance parameter library under different loads is calibrated offline. When the load is switched, the preset parameters are called. Its defects are: (1) the parameter fixation leads to response lag when the load suddenly changes (stabilization time> 2s); (2) the change of environmental stiffness (such as the difference in hardness and softness of the contact surface) is not considered, which easily leads to overshoot or undercompensation. To this end, the present application adopts the adaptive impedance control technology driven by the TD3 algorithm, which inputs the real-time observation state data set, outputs stiffness and damping in real time, and uses the real-time output stiffness and damping as the impedance parameters under the real-time observation state. The corresponding impedance parameters are adjusted through PID control and execution to deal with the sudden load change, which can shorten the stabilization time of the 5kg load from 1.2s of the traditional PID control to 0.5s (see Appendix). Figure 2 ), vibration amplitude attenuation 80%, and contact force tracking error ≤±3N, in order to adapt to the load mutation scenarios in different working environments.

[0040] Furthermore, the impedance parameters are adjusted online based on the TD3 algorithm, including: collecting real-time end contact force, joint torque deviation data and trajectory error data, and constructing an observation state data set; wherein, the end contact force includes three force components and three torque components collected by the six-dimensional force sensor, the three force components are Fx, Fy, and Fz, the three torque components are Mx, My, and Mz, the joint torque deviation data is the difference between the actual torque of each joint and the expected value, and the trajectory error data includes the end position deviation value and the end velocity deviation value; constructing and training the TD3 algorithm; wherein, the TD3 algorithm consists of a second Actor network for outputting stiffness and damping and a second Actor network for outputting stiffness and damping. To prevent over-estimation, the second Critic network is composed of a four-layer fully connected structure. The output nodes of each layer of the fully connected structure are 512, 256, 128 and 64 respectively. The second Critic network adopts a double-Q network structure. When training the TD3 algorithm, the observation state data set is input and combined with the defined OU noise to obtain the trained TD3 algorithm. Based on the trained TD3 algorithm, the stiffness and damping under the real-time observation state are output and used as the impedance parameters under the real-time observation state. Based on the impedance parameters under the real-time observation state, the corresponding impedance parameters are adjusted through PID control to cope with sudden load changes.

[0041] Furthermore, real-time end contact force, joint torque deviation data and trajectory error data are collected. The end contact force can be collected by a six-dimensional force sensor (such as the six-dimensional force sensor of model ATI Gamma) to collect three force components and three torque components. The three force components are Fx, Fy, and Fz, and the three torque components are Mx, My, and Mz. The joint torque deviation data is the actual torque of each joint. and expected value The trajectory error data includes the end position deviation value and the end velocity deviation value.

[0042] Furthermore, in the process of constructing and training the TD3 algorithm, the present application adopts the second Actor network and the second Critic network to construct the initial TD3 algorithm, wherein the second Actor network is used to output the stiffness and damping , , The second Actor network is composed of a four-layer fully connected structure. The output nodes of each layer of the fully connected structure are 512, 256, 128 and 64 respectively. The second Critic network adopts a double Q network structure to prevent over-estimation. At the same time, during the training process, exploration noise is introduced, specifically OU noise. Preferably, θ (the rate of mean reversion) in the OU noise is set to 0.15, and σ (volatility, that is, the degree of disturbance) is set to 0.2 to accelerate the model convergence process.

[0043] Furthermore, based on the trained TD3 algorithm, the stiffness and damping under real-time observation conditions are output and used as impedance parameters under real-time observation conditions (e.g., the impedance parameters are updated every 10 ms). Based on the impedance parameters under real-time observation conditions, PID control is used to adjust the corresponding impedance parameters to cope with sudden load changes.

[0044] S4: Generate a real-time obstacle avoidance path based on the optimized multi-objective PPO path planning algorithm and execute the path planning task of the corresponding real-time obstacle avoidance path to deal with dynamic obstacles.

[0045] Specifically, the existing technologies mostly use obstacle avoidance algorithms based on geometric rules, such as the improved FM algorithm proposed in CN110653805A (specifically using a static cost function) to plan collision-free paths. Its defects are: (1) Lack of dynamic obstacle avoidance capability: the planning relies on a static environment model and cannot respond to dynamic obstacles (such as drifting debris) in real time; (2) Unoptimized energy consumption: the path length priority strategy ignores changes in joint torque, resulting in a 30% increase in ineffective energy consumption. To this end, the multi-objective PPO path planning algorithm involved in this application has made targeted adjustments to the dynamic reward function design and strategy update mechanism. By introducing energy consumption reward items and energy consumption weight coefficients into the dynamic reward function, the optimized sum of square joint torques is reduced by 42%, and the overall energy consumption is reduced by 35% under the same task (see Appendix). Figure 3 ), it can also shorten the response time and effectively improve the speed of obstacle avoidance path planning.

[0046] Furthermore, based on the optimized multi-objective PPO path planning algorithm, a real-time obstacle avoidance path is generated, and the path planning task of the corresponding real-time obstacle avoidance path is performed, including: collecting real-time environmental perception data, the real-time environmental perception data includes real-time point cloud data collected by a binocular vision camera and IMU attitude data collected by an IMU attitude sensor, downsampling the real-time point cloud data to obtain a 128×128 pixel depth map, and inputting it into a three-layer convolutional neural network for feature extraction, outputting dynamic obstacle features, and state encoding the dynamic obstacle features and IMU attitude data to obtain a state encoding result; optimizing the multi-objective PPO path planning algorithm through the defined dynamic reward function and policy update mechanism, wherein the formula of the dynamic reward function is: ; Where R is the total reward, 、 、 is the weight coefficient, which is used to balance the importance of each reward item. The path completion degree indicates the degree of completion of the current path, and its value range is between 0 and 1. Energy consumption represents the energy consumed by the robotic arm structure during the operation task. It is calculated as the sum of the squares of the joint torques. The specific formula is: , For the The torque of each joint, is the total number of joints, The collision risk is expressed as the inverse of the distance between the robotic arm structure and the nearest obstacle, and the normalization is performed; a policy network structure is constructed and trained, wherein the policy network structure consists of an LSTM layer and a fully connected layer for outputting the probability distribution of each joint motion. During the policy network structure training process, the state encoding result is input, and the probability distribution of each joint motion under the corresponding environmental state is output. The total reward under the corresponding environmental state is calculated in combination with the dynamic reward function. The policy network structure parameters are updated based on the total reward under the corresponding environmental state to obtain the optimized multi-objective PPO path planning algorithm; dynamic obstacles in the environment are detected in real time by a binocular vision camera, and according to the corresponding state encoding result and the defined local path update time, the local path replanning mechanism is triggered to replan the local path. Based on the optimized multi-objective PPO path planning algorithm, the probability distribution of each joint motion under the corresponding local path is output, and the optimal joint motion is selected, and the optimal joint motion is used as the real-time obstacle avoidance path; in response to the real-time obstacle avoidance path, the path planning task of the corresponding real-time obstacle avoidance path is executed.

[0047] Furthermore, environmental perception and state encoding are key steps in flying robot path planning. Their goal is to extract useful environmental information from sensor data and encode it into a machine-understandable state representation. In this embodiment, real-time environmental perception data includes real-time point cloud data collected by a binocular vision camera (such as the Intel RealSense D435) and IMU attitude data collected by an IMU attitude sensor. The real-time point cloud data contains three-dimensional spatial information about objects in the environment. However, raw point cloud data is typically very large, and direct processing consumes significant computing resources. Therefore, the point cloud data must first be downsampled to convert it into a 128×128 pixel depth map. A depth map is a two-dimensional image in which each pixel value represents the distance between the corresponding object and the binocular vision camera. Downsampling significantly reduces the data size while preserving critical environmental information. The downsampled depth map is then fed into a three-layer convolutional neural network (CNN) for feature extraction. A CNN is a deep learning model specifically designed for processing image data. Its core concept is to extract useful features from images through convolution operations. In this embodiment, the three-layer structure of the CNN performs different feature extraction tasks. For example, the first convolutional layer uses 32 3×3 convolution kernels to convolve the input depth map and extract local features. The convolution kernels slide across the image, calculating the feature values ​​of a local area at a time. Through this first convolutional layer, the edge and contour information in the depth map can be captured. The second convolutional layer uses 64 3×3 convolution kernels to further convolve the feature map output by the first layer. The purpose of this layer is to extract higher-level features, such as object shape and texture information. The third convolutional layer uses 128 3×3 convolution kernels to perform a final convolution operation on the feature map output by the second layer. The features extracted by this layer are used for subsequent path planning decisions. After each convolution operation, a nonlinear transformation is performed using the ReLU activation function to enhance the model's expressive power. In addition, each convolution layer is followed by a maximum pooling operation to further reduce the size of the feature map while retaining the most important feature information. Through the convolution operation of these three layers of CNN, the final feature map can effectively represent key information in the environment, such as the location and shape of obstacles and the distance relationship between the robot and the obstacles; these features will be encoded into a machine-understandable state representation and serve as input for subsequent path planning strategies.

[0048] Furthermore, in the path planning of flying robots, strategy optimization is a key step to ensure that flying robots can complete tasks efficiently and safely in complex environments. This application adopts a multi-objective PPO path planning algorithm to optimize the path planning strategy. Specifically, the multi-objective PPO path planning algorithm makes targeted adjustments in the design of dynamic reward functions and the strategy update mechanism. Among them, the design of the dynamic reward function directly affects the behavior of the flying robot in path planning. In this embodiment, the dynamic reward function consists of three reward items, namely, path completion degree and strategy update mechanism. , energy consumption , collision risk , the specific formula is: ; Where R is the total reward, 、 、 is the weight coefficient, which is used to balance the importance of each reward item. Path completion degree indicates the degree of completion of the current path. The value range is between 0 and 1. The higher the path completion degree, the greater the reward. The specific calculation formula is: = Length of the path completed by the flying robot / total path length; Energy consumption represents the energy consumed by the robotic arm structure during the operation task. It is calculated as the sum of the squares of the joint torques. The lower the energy consumption, the greater the reward. The specific formula is: , For the The torque of each joint, is the total number of joints; The collision risk represents the inverse of the distance between the robot structure and the nearest obstacle, and is normalized. The lower the collision risk, the greater the reward. The specific formula is: =1 / the distance between the robotic arm structure and the nearest obstacle.

[0049] Furthermore, this application adopts a strategy network structure composed of an LSTM layer (long short-term memory network) and a fully connected layer to output the probability distribution of each joint movement, wherein LSTM can capture the long-term dependency in time series data and is suitable for path planning in a dynamic environment; the fully connected layer is used to output the probability distribution of each joint movement; the specific structure of the strategy network structure includes: (1) LSTM layer: the input is the state encoding result , the output is the hidden state , the calculation formula is: , where is the hidden state of the previous moment; (2) Fully connected layer: the output of the LSTM layer The probability distribution of each joint movement is mapped to the fully connected layer; in a dynamic environment, the flying robot needs to detect dynamic obstacles in real time and quickly update the local path when an obstacle is detected. The specific implementation process is: the dynamic obstacles in the environment are detected in real time through the binocular vision camera, and according to the corresponding state encoding results, based on the defined local path update time (such as 50ms), the local path replanning mechanism is triggered to replan the local path, and based on the optimized multi-objective PPO path planning algorithm, the probability distribution of each joint movement under the corresponding local path is output, the optimal joint movement is selected, and the optimal joint movement is used as the real-time obstacle avoidance path. In response to the real-time obstacle avoidance path, the flying robot performs the path planning task of the corresponding real-time obstacle avoidance path.

[0050] Furthermore, in order to trigger the local path update within the defined local path update time (e.g., 50 ms), this embodiment can also optimize the replanning mechanism, for example, (1) parallel computing: parallelizing the dynamic obstacle detection and path replanning tasks, and utilizing a multi-core CPU or GPU to accelerate the computing; (2) path caching: pre-calculating multiple possible local paths and storing them in the path cache; when a dynamic obstacle is detected, directly selecting the appropriate path from the cache, thus reducing the computing time; (3) incremental updating: during the local path update, only the portion of the path affected by the dynamic obstacle is updated, rather than replanning the entire path, further reducing the computational effort.

[0051] Example 2

[0052] This embodiment provides an adaptive anti-disturbance control system for a six-degree-of-freedom manipulator for aerial work based on the above embodiment 1. Figure 4 , an adaptive anti-disturbance control method for a six-degree-of-freedom robotic arm for aerial work applied in Example 1, the system includes a structure manufacturing module, a real-time dynamic anti-disturbance compensation module, an impedance parameter adjustment module and an obstacle avoidance path planning module.

[0053] Furthermore, a structural manufacturing module is used to manufacture a robotic arm structure using an arm body with a carbon fiber topology optimized design and a joint shell with a guide groove.

[0054] Furthermore, a real-time dynamic anti-disturbance compensation module is used to build a dynamic disturbance observer based on the DDPG algorithm. The dynamic compensation amount is output by the dynamic disturbance observer and combined with the disturbance compensation algorithm to calculate the wind disturbance value and integrate it into the control law of the unmanned aerial vehicle for real-time dynamic anti-disturbance compensation.

[0055] Furthermore, the impedance parameter adjustment module is used to adjust the impedance parameters online based on the TD3 algorithm to cope with sudden load changes.

[0056] Furthermore, the obstacle avoidance path planning module is used to generate a real-time obstacle avoidance path based on the optimized multi-objective PPO path planning algorithm, and perform the path planning task of the corresponding real-time obstacle avoidance path to deal with dynamic obstacles.

[0057] Example 3

[0058] This embodiment provides a six-degree-of-freedom robot arm for aerial work based on the above embodiment 1. Figure 5-6 The structure of the robotic arm is manufactured using the adaptive anti-disturbance control method of the six-degree-of-freedom robotic arm for aerial work in Example 1.

[0059] Furthermore, the structure of the robotic arm is composed of an arm body designed with carbon fiber topology optimization and a joint shell provided with a guide groove.

[0060] For further information, see the attached Figure 6 The arm body designed by carbon fiber topology optimization is an internal truss structure generated after multi-condition topology optimization of the arm body based on finite element analysis software, and the internal truss structure is made of carbon fiber composite materials. Among them, the load conditions of multi-condition topology optimization include 15m / s wind pressure and 10kg end load. The constraints are maximum stress ≤300MPa and first-order modal frequency ≥50Hz. The objective function includes minimizing the mass of the arm body and maximizing the stiffness of the arm body.

[0061] For further information, see the attached Figure 5 The guide groove is set on the surface of the joint shell and has a V-shaped structure. The geometric parameters including groove depth, inclination and spacing are defined. The drag coefficient under the wind pressure condition of 15m / s is simulated and verified by CFD software to optimize the aerodynamic performance.

[0062] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0063] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0064] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0065] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0066] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. An adaptive anti-disturbance control method for a six-degree-of-freedom manipulator for aerial work, characterized in that: include: The robotic arm structure is made of an arm body designed with carbon fiber topology optimization and a joint shell with guide grooves; A dynamic disturbance observer is constructed based on the DDPG algorithm. The dynamic compensation value is output by the dynamic disturbance observer and combined with the disturbance compensation algorithm to calculate the wind disturbance value and integrate it into the control law of the unmanned aerial vehicle for real-time dynamic anti-disturbance compensation. Online adjustment of impedance parameters based on TD3 algorithm to cope with sudden load changes; Based on the optimized multi-objective PPO path planning algorithm, a real-time obstacle avoidance path is generated and the path planning task of the corresponding real-time obstacle avoidance path is performed to deal with dynamic obstacles; The guide groove is set on the surface of the joint shell and has a V-shaped structure. The geometric parameters including groove depth, inclination and spacing are defined. The drag coefficient under the wind pressure condition of 15m / s is simulated and verified by CFD software to optimize aerodynamic performance. The optimized multi-objective PPO path planning algorithm generates a real-time obstacle avoidance path and performs the path planning task of the corresponding real-time obstacle avoidance path, including: Collecting real-time environmental perception data, including real-time point cloud data collected by a binocular vision camera and IMU attitude data collected by an IMU attitude sensor, downsampling the real-time point cloud data to obtain a 128×128 pixel depth map, and inputting it into a three-layer convolutional neural network for feature extraction. The system outputs dynamic obstacle features, and performs state encoding on the dynamic obstacle features and IMU attitude data to obtain a state encoding result. The multi-objective PPO path planning algorithm is optimized by defining a dynamic reward function and a policy update mechanism, where the formula of the dynamic reward function is: Where R is the total reward, 、 、 is the weight coefficient, which is used to balance the importance of each reward item. The path completion degree indicates the degree of completion of the current path, and its value range is between 0 and 1. Energy consumption represents the energy consumed by the robotic arm structure during the operation task. It is calculated as the sum of the squares of the joint torques. The specific formula is: , is the torque of the i-th joint, n is the total number of joints, The collision risk is the inverse of the distance between the robot structure and the nearest obstacle, and is normalized. Construct and train a policy network structure, which consists of an LSTM layer and a fully connected layer for outputting the probability distribution of each joint motion. During the policy network structure training process, the state encoding result is input, and the probability distribution of each joint motion under the corresponding environmental state is output. The total reward under the corresponding environmental state is calculated in combination with the dynamic reward function. The policy network structure parameters are updated based on the total reward under the corresponding environmental state, and an optimized multi-objective PPO path planning algorithm is obtained. The system detects dynamic obstacles in the environment in real time through a binocular vision camera, triggers the local path replanning mechanism based on the corresponding state coding results and the defined local path update time, replans the local path, and outputs the probability distribution of each joint motion under the corresponding local path based on the optimized multi-objective PPO path planning algorithm. The optimal joint motion is selected and used as the real-time obstacle avoidance path. In response to the real-time obstacle avoidance path, a path planning task corresponding to the real-time obstacle avoidance path is performed.

2. The adaptive anti-disturbance control method for a six-degree-of-freedom manipulator for aerial work according to claim 1, characterized in that: The arm body designed by carbon fiber topology optimization is an internal truss structure generated after multi-working condition topology optimization of the arm body using finite element analysis software, and the internal truss structure is made of carbon fiber composite materials. The load conditions of the multi-working condition topology optimization include a wind pressure of 15m / s and an end load of 10kg, the constraints are a maximum stress ≤300MPa and a first-order modal frequency ≥50Hz, and the objective function includes minimizing the mass of the arm body and maximizing the stiffness of the arm body.

3. The adaptive anti-disturbance control method for a six-degree-of-freedom manipulator for aerial work according to claim 1, characterized in that: The dynamic disturbance observer is constructed based on the DDPG algorithm. The dynamic compensation value is output by the dynamic disturbance observer. In combination with the disturbance compensation algorithm, the wind disturbance value is calculated and integrated into the control law of the unmanned aerial vehicle, including: Collect real-time wind pressure data, joint status data and historical error data to construct a state space data set; A DDPG algorithm is constructed and trained, and a dynamic disturbance observer is constructed based on the trained DDPG algorithm. The DDPG algorithm consists of a first actor network for outputting dynamic compensation and a first critic network for evaluating the value of the dynamic compensation. Both the first actor network and the first critic network consist of a three-layer fully connected structure, with 256, 128, and 64 output nodes in each layer, respectively. During the DDPG algorithm training process, a state-space dataset is input, and the dynamic compensation in the corresponding state space is output. The loss value in the real-time state space is calculated based on the mean square tracking error loss function, and the DDPG algorithm parameters are updated online in combination with real-time wind pressure data, resulting in an online updated DDPG algorithm. Based on the online updated DDPG algorithm, the dynamic compensation value in the real-time state space is output. Combined with the disturbance compensation algorithm, the wind disturbance value is calculated and integrated into the control law of the unmanned aerial vehicle for real-time dynamic anti-disturbance compensation.

4. The adaptive anti-disturbance control method for a six-degree-of-freedom manipulator for aerial work according to claim 3, characterized in that: The real-time wind pressure data is collected by multiple pressure sensors arranged on the surface of the robotic arm structure. The joint state data includes the joint angle, joint speed and current loop torque fed back by the joint encoder. The historical error data is the end position tracking error data within the defined historical control cycle.

5. The adaptive anti-disturbance control method for a six-degree-of-freedom manipulator for aerial work according to claim 3, characterized in that: The formula of the disturbance compensation algorithm is: Where, is the wind disturbance value, is the dynamic compensation amount, is the state space dataset, and is obtained after calibration by the particle swarm optimization algorithm, where The value is 12. The value is 0.5, To control the error.

6. The adaptive anti-disturbance control method for a six-degree-of-freedom manipulator for aerial work according to claim 1, characterized in that: The online adjustment of impedance parameters based on the TD3 algorithm includes: Collect real-time end contact force, joint torque deviation data, and trajectory error data to construct an observation state dataset. The end contact force includes three force components and three torque components collected by the six-dimensional force sensor. The three force components are Fx, Fy, and Fz, and the three torque components are Mx, My, and Mz. The joint torque deviation data is the difference between the actual torque of each joint and the expected value. The trajectory error data includes the end position deviation value and the end velocity deviation value. Construct and train the TD3 algorithm. The TD3 algorithm consists of a second Actor network for outputting stiffness and damping, and a second Critic network for preventing overestimation. The second Actor network consists of a four-layer fully connected structure, with each layer having 512, 256, 128, and 64 output nodes, respectively. The second Critic network uses a dual-Q network structure. During TD3 algorithm training, an observation state dataset is input and combined with the defined OU noise to obtain the trained TD3 algorithm. Based on the trained TD3 algorithm, the stiffness and damping under real-time observation state are output and used as the impedance parameters under real-time observation state; Based on the impedance parameters under real-time observation, the corresponding impedance parameters are adjusted through PID control to cope with sudden load changes.

7. The adaptive anti-disturbance control system of the six-degree-of-freedom manipulator for aerial work is characterized by: include: The structural manufacturing module is used to manufacture the robotic arm structure using an arm body designed with carbon fiber topology optimization and a joint shell with guide grooves; The real-time dynamic disturbance rejection compensation module is used to build a dynamic disturbance observer based on the DDPG algorithm. The dynamic disturbance observer outputs the dynamic compensation value, which is combined with the disturbance compensation algorithm to calculate the wind disturbance value and integrate it into the control law of the unmanned aerial vehicle for real-time dynamic disturbance rejection compensation. Impedance parameter adjustment module, used to adjust impedance parameters online based on the TD3 algorithm to cope with sudden load changes; The obstacle avoidance path planning module is used to generate a real-time obstacle avoidance path based on the optimized multi-objective PPO path planning algorithm and execute the path planning task of the corresponding real-time obstacle avoidance path to deal with dynamic obstacles; The guide groove is set on the surface of the joint shell and has a V-shaped structure. The geometric parameters including groove depth, inclination and spacing are defined. The drag coefficient under the wind pressure condition of 15m / s is simulated and verified by CFD software to optimize aerodynamic performance. The optimized multi-objective PPO path planning algorithm generates a real-time obstacle avoidance path and performs the path planning task of the corresponding real-time obstacle avoidance path, including: Collecting real-time environmental perception data, including real-time point cloud data collected by a binocular vision camera and IMU attitude data collected by an IMU attitude sensor, downsampling the real-time point cloud data to obtain a 128×128 pixel depth map, and inputting it into a three-layer convolutional neural network for feature extraction. The system outputs dynamic obstacle features, and performs state encoding on the dynamic obstacle features and IMU attitude data to obtain a state encoding result. The multi-objective PPO path planning algorithm is optimized by defining a dynamic reward function and a policy update mechanism, where the formula of the dynamic reward function is: Where R is the total reward, 、 、 is the weight coefficient, which is used to balance the importance of each reward item. The path completion degree indicates the degree of completion of the current path, and its value range is between 0 and 1. Energy consumption represents the energy consumed by the robotic arm structure during the operation task. It is calculated as the sum of the squares of the joint torques. The specific formula is: , is the torque of the i-th joint, n is the total number of joints, The collision risk is the inverse of the distance between the robot structure and the nearest obstacle, and is normalized. Construct and train a policy network structure, which consists of an LSTM layer and a fully connected layer for outputting the probability distribution of each joint motion. During the policy network structure training process, the state encoding result is input, and the probability distribution of each joint motion under the corresponding environmental state is output. The total reward under the corresponding environmental state is calculated in combination with the dynamic reward function. The policy network structure parameters are updated based on the total reward under the corresponding environmental state, and an optimized multi-objective PPO path planning algorithm is obtained. The system detects dynamic obstacles in the environment in real time through a binocular vision camera, triggers the local path replanning mechanism based on the corresponding state coding results and the defined local path update time, replans the local path, and outputs the probability distribution of each joint motion under the corresponding local path based on the optimized multi-objective PPO path planning algorithm. The optimal joint motion is selected and used as the real-time obstacle avoidance path. In response to the real-time obstacle avoidance path, a path planning task corresponding to the real-time obstacle avoidance path is performed.

8. Six-degree-of-freedom robotic arm for aerial work, characterized by: It comprises a mechanical arm structure and is manufactured using the adaptive anti-disturbance control method according to any one of claims 1 to 6.

Images