Robot self-adaptive force control and admittance control method and system for wallboard assembly

Abstract

The invention discloses a robot self-adaptive force control and admittance control method and system for wallboard assembly, and the method comprises the steps: firstly, measuring the pose of the tail end of a robot in real time in combination with a high-precision photogrammetry system, carrying out the track correction according to a measurement result and an expected pose, achieving the positioning of the robot in a position control direction, and carrying out the positioning of the robot in a force control direction; the method comprises the following steps: establishing a mapping relation between robot end force and contact force of a contact surface through a heterovariance neural network mapping model, and jointly training an inverse model to map expected contact force back to expected end force, thereby controlling the contact force on the contact surface by adjusting the end force, and constructing a self-adaptive force control and admittance control model; a virtual energy tank is connected with a self-adaptive force control and admittance control model, so that updating of parameters of the control model is guided, and the passivity of a system is ensured; meanwhile, the parameters of the virtual energy tank are dynamically adjusted according to the system state, passivity can be kept, and the assembly precision and performance of the robot can be improved to the maximum extent.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent manufacturing, and particularly relates to a robot adaptive force control and admittance control method and system for panel assembly. Background Art

[0002] The panel assembly is the main load-bearing part of the aircraft fuselage structure, usually assembled from multiple parts such as skins, frames, and stringers. As an important part of the aircraft aerodynamic shape and fuselage structure, the assembly quality of the panel assembly directly affects the aerodynamic performance and service life of the aircraft. In the traditional assembly method, the assembly and connection of the panel assembly mainly rely on workers to manually position the parts and manually connect them with pneumatic tools. The assembly quality of this method highly depends on the experience and operation proficiency of the workers. However, the traditional method has the problems of long assembly cycle and low efficiency, and it is difficult to meet the requirements of high-efficiency and high-quality mass production.

[0003] With the development of robot and robotic equipment technology, robots have the advantages of flexible movement, large working space, strong parallel coordination operation ability, etc., and can integrate a variety of sensors to adapt to complex processing environments. Therefore, the intelligent manufacturing technology with robots as the core has gradually become a new trend in the high-quality manufacturing of aircraft panel assemblies. During the assembly process of the panel assembly, the frame, as a key component of the fuselage load-bearing skeleton, has a complex multi-joint surface structure. Its assembly with the fuselage skin requires controlling the contact force of each mating joint surface to ensure full contact and avoid a decrease in structural strength caused by excessive contact force. Therefore, it is necessary to monitor the contact force of each joint surface in real time. However, industrial robots usually only install force sensors at the end. If the numerical values of the robot end sensors can be pre-collected, compared with the contact forces of each joint surface, and the mapping relationship between them is established, the contact forces of each joint surface can be controlled by adjusting the end force of the robot, thereby ensuring the assembly quality.

[0004] Industrial robots usually adopt a position control mode for driving. Therefore, if the desired trajectory can be dynamically updated using the tracking error of the robot end force, the assembly force can be indirectly adjusted by adjusting the pose of the robot. During the assembly process, environmental factors are often uncertain. Admittance control is an effective safe assembly method, but its performance depends on the interactive environment and admittance parameters. Adaptive admittance control can adjust the admittance parameters online according to task requirements to adapt to unknown environmental changes. In addition, the force controller also plays an important role in achieving force error convergence. Combining force control with adaptive admittance control and designing an adaptive force control and admittance control system can effectively ensure the safety of the assembly process and the convergence of force. However, the force controller or adjusting the stiffness value of the adaptive admittance system may inject energy into the system, resulting in the system not meeting passivity, or even causing the system to diverge.

[0005] Based on this, a robot adaptive force control and admittance control method and system for panel assembly are proposed. Summary of the Invention

[0006] In view of the above technical problems, the present invention provides a robot adaptive force control and admittance control method and system for panel assembly.

[0007] The technical solution adopted by the present invention to solve its technical problems is as follows: A robot adaptive force control and admittance control method for panel assembly, the method comprising the following steps: S100: Based on an external digital measurement device, the pose of the robot end is measured in real time, the end force and contact force data are synchronously collected, and a data set is constructed; S200: An inhomogeneous variance neural network mapping model is established. After being trained by the data set and a preset first loss function, the mean value and sampling variance of the contact force on each contact surface are predicted in combination with the end force of the robot in the force control direction; S300: A neural network inverse mapping model is established. After being trained by the data set, a preset second loss function, and the predicted mean value and sampling variance of the contact force on each contact surface, the mapping from the desired contact force to the desired end force is realized; S400: Define the desired contact force range constraint and uniformity index for each contact surface, and optimize the desired end force through gradient calculation; S500: Define a basic adaptive force control and admittance control model, update the admittance parameters according to the control target of end force convergence, and perform stability analysis; S600: Design a virtual energy tank with adaptive power limit and energy injection and freezing functions, and connect it to the basic adaptive force control and admittance control model. Through the interconnected structure, correct the adaptive stiffness adjustment law and force controller of the admittance parameters to obtain a corrected adaptive force control and admittance control model; S700: According to the corrected adaptive force control and admittance control model and its parameter values, obtain the corrected trajectory expected value in the force control direction, and combine the actual pose of the robot end in the force control direction to obtain the trajectory correction amount in the force control direction; S800: Measure the actual pose of the robot end in the position control direction in real time, subtract it from the desired pose of the robot in the position control direction to obtain the trajectory correction amount in the position control direction, and send the trajectory correction amount in the position control direction and the trajectory correction amount in the force control direction to the industrial robot to achieve high-precision compliant assembly.

[0008] Preferably, S100 includes: S110: Measure the pose of the robot end in real time with the aid of an external digital measurement device , according to the process requirements, it is divided into the pose in the position control direction and the pose in the force control direction , where the goal of the position control direction is to achieve precise positioning, and the goal of the force control direction is to achieve the convergence of the end force error; the dimension of the force control direction and the dimension of the position control direction depend on the process requirements. When only position control is considered, ; when force control is considered in all directions, ; S120: Data collection. At the robot pose , perform times of measurement of the end force and the contact force to obtain the end force sample of the robot in the force control direction and the contact force samples of each contact surface , calculate and record the corresponding sample statistics: ; Among them, represents the rd pose, is the total number of poses to obtain the pose, represents the th measurement of the end force and the contact force, represents the total number of measurements of the end force and the contact force, is the mean value of the robot end force calculated by times of sampling at the robot pose in the force control direction, , is the mean value of the corresponding pressure sensor, and its sampling variance is: .

[0009] Preferably, S200 includes: S210: Construct a heteroscedastic neural network mapping model. Divide the data set into a training set, a test set, and a validation set. During the training process, use the mean sample of the robot end force in the training set after being normalized as the input to train the heteroscedastic neural network, and establish the mapping relationship between the robot end force in the force control direction and the mean value and sample sampling variance of the contact force of each contact surface of the bulkhead. This mapping relationship is expressed as: ; Among them, is the number of samples in the training set, and are the neural network functions used to predict the mean value and variance logarithm of the contact force of each contact surface respectively; The output obtains the variance through exponential transformation to ensure the non-negativity of the variance; and are the parameter sets of the corresponding networks respectively; and are the mean contact forces of each contact surface predicted by the neural network and the sample sampling variance when the input of the neural network is the mean end force respectively; S220: Design the loss function of the neural network parameters, combining the maximum likelihood estimation and the parameter set of the neural network , and the specific loss function design is as follows: ; Use the optimization algorithm to update the parameters of the neural network: ; where is the learning rate; S230: Use the heteroscedastic neural network mapping model to infer the samples in the validation set to obtain the predicted sample sampling variance ; Calibrate the predicted variance using the actual sample sampling variance corresponding to the contact force in the validation set. For this purpose, define the following calibration loss function: ; where is the sample variance calibration factor, is the number of samples in the validation set; The process of using the optimization algorithm to update the sample variance calibration factor is as follows: ; where is the learning rate, and the calibrated predicted sampling variance is ; S240: According to the trained neural network model and the sample variance calibration factor, combine the end force of the robot in the force control direction to predict the mean contact force of each contact surface and the calibrated predicted sampling variance .

[0010] Preferably, S300 includes: S310: Construct a neural network inverse mapping model, and use the data set to train the neural network inverse mapping model by combining the predicted mean contact forces and sampling variances on each contact surface. Define the joint training objective as minimizing the prediction error of the neural network inverse mapping model and the consistency error of the heteroscedastic neural network mapping model: ; where is the average contact force of the th sample in the training set The end-effector force predicted by the inverse mapping model, that is, ; while and are respectively the average contact force and the sampling variance predicted by the heteroscedastic neural network mapping model; Use an optimization algorithm to update the parameters of the neural network inverse mapping model: ; where is the learning rate; S320: Use the trained neural network inverse mapping model to achieve real-time inference from the desired average contact force to the desired end-effector force , that is: ; where is the parameter set of the neural network inverse mapping model, is the neural network inverse mapping model.

[0011] Preferably, S400 includes: S410: Define the objective function of the desired contact force range for each contact surface. The objective function needs to ensure sufficient contact of each contact surface, and at the same time avoid a decrease in structural strength caused by excessive contact force. Therefore, the contact force value of each contact surface during assembly should meet the upper and lower limit requirements of the process; Assume is the contact force of the th predicted contact surface, where , is the total number of contact surfaces; the actual contact force of each contact surface should meet the following conditions: ; where, and are respectively the desired lower bound value of the contact force and the desired upper bound value of the contact force; Considering the uncertainty factors in the sampling data of the sensor, when considering uncertainty, the conditions that the actual contact force of each contact surface should meet are modified to: ; where is the sampling standard deviation of the contact force of the th contact surface predicted by the neural network; Fully considering uncertainty and contact force uniformity, the allowable value of the contact force upper bound is set to , and the allowable value of the contact force lower bound is set to , where is the standard deviation of the contact force with the largest contact area among the contact surfaces, is the maximum value function; in addition, the upper bound value and the lower bound value of the desired contact force are respectively converted into the upper bound value and the lower bound value of the desired end force in the force control direction through the neural network inverse mapping model; Since the contact force is indirectly adjusted by adjusting the end force, the objective function of the contact force range constraint is designed as: ; where, is the rectified linear unit function, and its specific expression is , , and are respectively , and the -th component of S420: Define the contact uniformity objective function. To make the contact force distribution more uniform and minimize the difference between contact forces, that is, minimize the difference between the contact force and their average value , map the contact uniformity to the end through the neural network inverse model, and the objective function is defined as: ; where, represents the expected average contact force of each contact surface as , and the expected end force regarding contact uniformity in the corresponding force control direction is inferred by the neural network inverse model; therefore, the gradient of with respect to can be obtained: S430: Calculate the gradient of the desired contact force range. According to the objective function of S410, let and , and the gradient of with respect to can be obtained: ; where, is the indicator function, and its specific expression is: ; S440: Design the desired end force. Combining the gradients calculated in S420 and S430, the ideal end force variation law is designed as: ; Among them, is a weight coefficient, is a constant, is a projection operator, which is used to ensure that once it enters the predetermined range it will always remain within the predetermined range, and its specific expression is: ; Regarding the end-effector force obtained according to the ideal end-effector force variation law as the desired end-effector force , that is . In addition, since industrial robots usually can only receive position commands, it is necessary to adjust the position of the robot according to the end-effector force error, so as to achieve the control of the end-effector force.

[0012] Preferably, S500 includes: S510: Define the end-effector force tracking error in the force control direction, specifically: ; S520: Design a force controller, and use the following PI force controller to adjust the end-effector force tracking error of the robot: ; Among them, and are positive definite diagonal matrices, is the adaptive factor of the force controller coefficient. The role of this adaptive factor is to increase the integral coefficient to accelerate the convergence of the end-effector force error when the force tracking error is large; while when the force tracking error is small, reduce the integral coefficient to avoid over-correction; the adaptive factor has the following specific form: ; Among them, is a preset threshold, is a constant; S530: Design the activation function of the PI force controller. When the frame and the skin come into contact, turn on the force controller. In addition, to ensure the safety of the frame and the robot, when the contact force exceeds the set value, turn off the force controller. Therefore, the activation function matrix of the force controller is defined as: , where the th diagonal element of the activation function matrix of the force controller has the following specific form: ; Among them, , is a constant is a positive constant and satisfies , represents a diagonal matrix. When is greater than the threshold , the force controller is activated. When , the force controller starts to be smoothly turned off until when the force controller is completely turned off to avoid excessive contact force; S540: Design the activation function of the adaptive admittance controller, defined as: , where the specific form of ; Among them, ; is the width of the transition section of the adaptive admittance control activation function; S550: Design the basic adaptive force control and admittance control model, specifically: ; Among them, , , are positive definite diagonal matrices, representing the inertia matrix, damping matrix, and stiffness matrix respectively. , , and are the adjustments of the desired position, velocity, and acceleration in the force control direction respectively. is the planned desired assembly point in the force control direction. is the updated desired assembly point in the force control direction; S560: Update the parameters of the adaptive admittance control. Design the following adaptation law to update the stiffness matrix , and its specific expression is: ; Among them, is the basic stiffness update gradient, and its specific expression is ; and are the learning rates. and are the and th elements of the vectors and and are the th diagonal elements of the stiffness lower bound matrix and the stiffness upper bound matrix respectively. Let , , the function is used to penalize the stiffness value outside a predetermined range ; S570: Conduct a stability analysis of the adaptive force control and admittance control strategies for the base, considering the following energy function: ; For taking the differential gives: ; Substitute the adaptive force control and admittance dynamic equations of the base into the equation of to obtain: ; Since is a positive definite matrix, so is always non-positive. However, since it cannot be guaranteed in advance that and are non-positive, therefore, the passivity of the adaptive force control and admittance control system of the base with respect to cannot be guaranteed.

[0013] Preferably, S600 includes: S610: Design the energy tank dynamics with an online energy flow injection and freezing mechanism, and define the energy stored in the energy tank as , where is the state of the energy tank, is the time; regard the energy difference of as the initial energy margin available for energy exchange pre-stored in the energy tank, where represents the initial time, is the initial energy value stored in the energy tank, is the lower bound value of the energy in the energy tank, and their relationship satisfies ; To take into account system performance while ensuring safety, combined with the goal of the force control direction, that is, the convergence of the end-effector force error of the robot, the following dynamics of the energy tank with an online energy flow injection and freezing mechanism are proposed to dynamically adjust the energy in the energy tank to ensure that the system can operate efficiently and safely during actual operation: ; where, is a constant used to store part of the energy consumed by damping; is the corrected velocity update amount, is the corrected planned velocity vector of the th element. At this time, and can be obtained by integration and differentiation respectively; the sum of the power of the d - dimensional force controller and stiffness update in the force control direction is expressed as , where , and are adaptive power scaling factors; and are the -th and -th elements of vectors and respectively; is the basic stiffness update gradient in S560; is the function of the online energy injection and freezing mechanism in the -th force control direction, and the coefficient functions to ensure that the energy value stored in the energy tank does not exceed the predetermined energy upper bound , so ; where the function is used to smoothly interrupt the energy exchange between the adaptive force control and admittance control system and the virtual energy tank when the energy reaches the preset lower bound of the energy tank, so as to ensure that the value of the energy tank is not lower than this lower bound value. Therefore is defined as follows: ; where is the width of the smooth transition section of the energy value; is the function of the online energy injection and freezing mechanism in the -th force control direction, which functions to ensure the system performance by online injecting energy when the end - effector force of the robot remains within the safe range and the energy stored in the energy tank is too low, resulting in limited system performance; in addition, when the contact force is too large, i.e., , the energy exchange with the energy tank is frozen to avoid the adaptive force control and admittance system extracting energy from the energy tank, thus ensuring the safety of the system; to achieve a smooth transition from energy injection to energy freezing, a transition section from energy injection to energy freezing is set when the contact force is large within the safe range, which functions to start a smooth transition from the energy injection function to the energy freezing function when , and completely freeze the energy exchange when . Its specific expression is: ; where is a constant, and the function ; constant; the adaptive proportionality coefficient of energy injection and freezing in the th dimension in the force control direction has the following specific expression ; S620: Design an adaptive power flow. According to the mission requirements and system status, adjust and update the maximum power allowed to be injected into the actual system in real time; for this purpose, define the following adaptive power scaling coefficient , where has the following specific form: ; where the power of the force controller is , and the power of the stiffness update gradient corrected by the energy tank is . According to the target in the force control direction, design the following update law for : ; where is the learning rate; is the th end force component in the force control direction; is the forgetting factor, which is used to prevent the maximum power value allowed to be injected into the actual system from being too large; S630: Design a corrected force control and adaptive admittance control model, and its form is designed as: ; where is the -dimensional identity matrix, is the function matrix of the online energy injection and freezing mechanism in the force control direction, is the adaptive power scaling coefficient matrix. At this time, the update law of the corrected stiffness matrix is designed as: ; where is the update law of the th diagonal element of the stiffness matrix, and are the th diagonal element of the stiffness lower bound matrix and the th diagonal element of the stiffness upper bound matrix respectively; to ensure the stability and performance of the system, determine the damping value of the system according to the critical damping condition: ; S640: Stability and performance analysis. Define the total energy function of the interconnected system composed of the robot system and the energy tank as follows: ; where is the energy of the robot system, and the total power sum of the force controller and the stiffness update in the force control direction The expression of is rewritten as where is the original stiffness update gradient matrix with power limitation, and its expression is ; Taking the derivative of yields: ; To analyze the stability of the interconnected system composed of the robot system and the energy tank, define as the set of all force control directions, and define as the set of in the force control direction; Analyze the stability and performance of the system through and two cases, where represents the empty set. In addition, from the expression of in S610, the following inequality always holds: ; Case 1: ; Under this condition, for , there is always , so there is and . At this time, can be simplified to: ; From , it can be seen that in this case, the interconnected system is stable; Case 2: ; For , there is and for , . At this time, there is and , where and when , , is simplified to: ; Since the matrix is a diagonal positive semi - definite matrix, so when , there is . Therefore, when , the interconnected system is stable; When When is decomposed into , where is the corresponding Lyapunov function, is the corresponding Lyapunov function. At this time, there are and . Therefore, when is satisfied, the direction of the system with respect to is passive. For is stable. Since passivity can ensure the stable interaction between the robot system and any passive environment, the system is stable in this case. In summary, both system cases 1 and 2 are stable, so the system is always stable. S650: Analysis of the convergence of the force tracking error. The total system variables are defined as , where is the state of the environment when the environmental energy is . At this time, the total energy function of the system is defined as . Considering the optimal state and such that: ; where is the space of all possible values. Since the environment always dissipates energy, at this time, when and , there is . So at this time, there is . Therefore is asymptotically stable in the whole system. Considering at the equilibrium point , the modified adaptive force control and admittance control model becomes: ; It is easy to prove that is the unique steady-state solution of the system, so the end-effector force of the robot converges to the desired force.

[0014] Preferably, S700 includes: According to the admittance parameter values obtained in S500, substitute them into the modified adaptive force control and admittance control model to obtain the pose adjustment amount: ; where and are the adjustment amounts of the desired position and desired velocity at the previous sampling moment respectively, is the sampling period; Update the desired position of the force control direction: ; Thus, the trajectory correction amount in the force control direction is obtained: , where is the actual position of the end of the force-controlled robot measured by an external digital measurement device in the force control direction.

[0015] Preferably, S800 includes: S810: According to the actual position of the end of the position control direction robot measured by an external digital measurement device , compare it with the expected position in the position control direction , and calculate the trajectory correction amount in the position control direction according to the difference between the two: ; S820: Compile a robot control program. Since the industrial robot receives position control instructions, send the updated trajectory correction amount to the robot, thereby realizing the precise control of the robot.

[0016] A robot adaptive force control and admittance control system for panel assembly, including a data set acquisition module, a heteroscedastic neural network mapping model establishment module, a neural network inverse mapping model establishment module, an expected end force optimization module, an admittance parameter update module, an adaptive adjustment law of admittance parameters and a force controller correction module, a trajectory expected value correction module and a trajectory correction; The data set acquisition module is used to measure the pose of the end of the robot in real time based on an external digital measurement device, synchronously collect the end force and contact force data, and construct a data set; The heteroscedastic neural network mapping model establishment module is used to establish a heteroscedastic neural network mapping model. After training with the data set and a preset first loss function, combine the end force of the robot in the force control direction to predict the mean value and sampling variance of the contact force on each contact surface; The neural network inverse mapping model establishment module is used to establish a neural network inverse mapping model. After training the neural network inverse mapping model with the data set, a preset second loss function combined with the predicted mean value and sampling variance of the contact force on each contact surface, realize the mapping from the expected contact force to the expected end force; The expected end force optimization module is used to define the expected contact force range constraint and uniformity index on each contact surface, and optimize the expected end force through gradient calculation; The admittance parameter update module is used to define a basic adaptive force control and admittance control model, update the admittance parameters according to the control target of end force convergence, and perform stability analysis; The adaptive adjustment law of admittance parameters and the force controller correction module are used to design a virtual energy tank with adaptive power limitation and energy injection and freezing functions, and connect it to the basic adaptive force control and admittance control model. Through the interconnected structure, the adaptive stiffness adjustment law of admittance parameters and the force controller are corrected to obtain the corrected adaptive force control and admittance control model; The trajectory expected value correction module is used to obtain the corrected trajectory expected value in the force control direction according to the corrected adaptive force control and admittance control model and its parameter values, and combine the actual pose of the robot end in the force control direction to obtain the trajectory correction amount in the force control direction; The trajectory correction module is used to measure the actual pose of the robot end in the position control direction in real time, subtract it from the expected pose of the robot in the position control direction to obtain the trajectory correction amount in the position control direction, and send the trajectory correction amounts in the position control direction and the force control direction to the industrial robot to achieve high-precision compliant assembly.

[0017] The above-mentioned robot adaptive force control and admittance control method and system for panel assembly combine a high-precision photogrammetry system to measure the pose of the robot end in real time, and perform trajectory correction according to the measurement results and the expected pose, so as to achieve high-precision positioning of the robot in the position control direction. In the force control direction, a mapping relationship between the force at the robot end and the contact force on the contact surface is established through a heteroscedastic neural network mapping model, and an inverse model is jointly trained to map the expected contact force back to the expected end force, so as to adjust the end force to control the contact force on the contact surface and further construct an adaptive force control and admittance control model; In addition, by connecting the virtual energy tank with the adaptive force control and admittance control model, the update of the control model parameters can be guided to ensure the passivity of the system. At the same time, the parameters of the virtual energy tank are dynamically adjusted according to the system state, which can not only maintain the passivity of the system, but also maximize the assembly accuracy and performance of the robot on this basis. Description of the Drawings

[0018] Figure 1 It is a flowchart of the robot adaptive force control and admittance control method for panel assembly in an embodiment of the present invention; Figure 2 It is a control block diagram in an embodiment of the present invention. Detailed Embodiment

[0019] In order to enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0020] In one embodiment, as Figure 1 shown, the robot adaptive force control and admittance control method for panel assembly, the method includes the following steps: S100: Based on an external digital measurement device, the pose of the robot end is measured in real time, the end force and contact force data are synchronously collected, and a data set is constructed; that is, the end force and contact force are measured a preset number of times under the pose as the data set. S200: Establish a heteroscedastic neural network mapping model, train the heteroscedastic neural network mapping model using the data set and a preset first loss function, and use the trained neural network model to combine the end force of the robot in the force control direction to predict the mean value and sampling variance of the contact force on each contact surface. S300: Establish a neural network inverse mapping model, train the neural network inverse mapping model using the data set, a preset second loss function, combined with the predicted mean value and sampling variance of the contact force on each contact surface, and use the trained neural network inverse mapping model to realize the mapping from the desired contact force to the desired end force. S400: Define the constraint of the desired contact force range and the uniformity index on each contact surface, and optimize the desired end force through gradient calculation; that is, define the objective function of the desired contact force range and the contact uniformity objective function on each contact surface, calculate the gradient of the desired contact force range according to the objective function of the desired contact force range on each contact surface, and combine the contact uniformity objective function and the gradient of the desired contact force range to optimize the desired end force. S500: Define a basic adaptive force control and admittance control model, update the admittance parameters according to the control objective of end force convergence, and perform stability analysis. S600: Design a virtual energy tank with adaptive power limit and energy injection and freezing functions, connect it to the basic adaptive force control and admittance control model, and modify the adaptive stiffness adjustment law and force controller of the admittance parameters through the interconnected structure to obtain a modified adaptive force control and admittance control model. S700: According to the modified adaptive force control and admittance control model and its parameter values, obtain the modified trajectory expected value in the force control direction, and combine the actual pose of the robot end in the force control direction measured in real time to obtain the trajectory correction amount in the force control direction. S800: Use a high-precision instrument to measure the actual pose of the robot end in the position control direction in real time, subtract it from the desired pose of the robot in the position control direction to obtain the trajectory correction amount in the position control direction, and send the trajectory correction amount in the position control direction and the trajectory correction amount in the force control direction to the industrial robot to achieve high-precision compliant assembly.

[0021] Specifically, the present invention proposes a robot adaptive force control and admittance control method for panel assembly. However, the following problems will occur during the assembly process of the robot aircraft panel components: (1) The absolute positioning accuracy of the robot is low, making it difficult to meet the high-precision positioning requirements of panel component assembly. In addition, in mass production, there are no sensors on the contact surfaces of the frames, resulting in inaccurate measurement of the actual contact force. (2) The controller is an effective means to achieve contact force convergence, but it does not meet the passivity condition, easily leading to instability of the robot system and being difficult to dynamically adjust according to environmental changes and lacking compliance. (3) Admittance control has good compliance. Combining admittance control with a force controller to form an adaptive force control and admittance control system helps to improve the interaction ability of the robot in a complex environment. However, the performance of the admittance system is affected by environmental factors and admittance parameters. In the case of unknown environment, it is difficult to select appropriate admittance parameters to ensure the performance of the system. Although variable admittance control can achieve adaptive adjustment of admittance parameters, changing the stiffness of the admittance system may cause the system to violate passivity, thus triggering instability. (4) Although the energy tank can ensure the passivity of the adaptive force control and admittance control system, the safety and performance of the system still depend on the energy storage value in the energy tank and the maximum power allowed to be injected into the system, and these two parameters are difficult to pre-determine in practical applications. Based on the above problems, the following improvements are specifically proposed: (1) Real-time measure the end pose of the robot through a digital photogrammetry system, and design a closed-loop control algorithm according to the measurement results to correct the robot's trajectory, thereby achieving high-precision positioning. At the same time, based on the data of the robot end force and its corresponding contact force collected in advance, use a heteroscedastic neural network to establish a mapping relationship between the end force and the contact force, so that during the actual mass production process, the contact force can be predicted through the robot end force. In addition, jointly train the inverse mapping model of the neural network to convert the desired contact force into the desired end force. (2) Define a loss function and update the desired end force according to indicators such as the ideal contact force range of the contact surface, and design the activation functions of the force controller and the adaptive admittance control to ensure the tracking performance of the robot system in free space. The force controller is combined with the adaptive admittance control to construct a unified force control and adaptive admittance control system. (3) Interconnect the adaptive force control and admittance system with a virtual energy tank to ensure the passivity of the system. (4) To avoid too rapid parameter changes, set the power limit of the energy tank to ensure the safety of the system. At the same time, dynamically adjust the power limit of the energy tank according to the contact force error, and design an energy injection and freezing mechanism to further improve the system performance on the premise of ensuring system safety. Through the above steps, this method effectively solves the positioning accuracy and force control problems of the robot during the panel component assembly process, and ensures the operation of the system with high precision and high stability by dynamically adjusting and optimizing control parameters.

[0022] In one embodiment, S100 includes: S110: Since industrial robots usually have relatively low absolute positioning accuracy, in order to improve the motion accuracy of the robot, it is necessary to use an external digital measurement device (such as a digital photogrammetry system, a laser tracker, etc.) to measure the pose of the robot end in real time , and according to the process requirements, it is divided into the pose in the position control direction and the pose in the force control direction , where the goal of the position control direction is to achieve precise positioning, and the goal of the force control direction is to achieve the convergence of the end force error; the dimension of the force control direction and the dimension of the position control direction depend on the process requirements. When only position control is considered, ; when force control is considered in all directions, ; S120: Data collection. When the robot is in the pose , , perform measurements of the end force and contact force, obtain the end force samples of the robot in the force control direction and the contact force samples of each contact surface , calculate and record the corresponding sample statistics: ; where, represents the th pose, is the total number of poses obtained, represents the th measurement of the end force and contact force, represents the total number of measurements of the end force and contact force, is the mean value of the robot end force calculated by times of sampling in the force control direction when the robot is in the pose , is the mean value of the corresponding pressure sensor, and its sampling variance is: .

[0023] In one embodiment, S200 includes: S210: Construct a heteroscedastic neural network mapping model. Divide the data set into a training set, a test set, and a validation set. During the training process, use the mean samples of the robot end force in the training set after being standardized as the input to train the heteroscedastic neural network, and establish the mapping relationship between the robot end force in the force control direction and the mean value and sample sampling variance of the contact force of each contact surface of the bulkhead. This mapping relationship is expressed as: ; Among them, is the number of samples in the training set, and are neural network functions for predicting the mean and variance logarithm of the contact force of each contact surface respectively; The output of and are the parameter sets of the corresponding networks respectively; and are the mean contact forces of each contact surface of the spacer frame predicted by the neural network and the sample sampling variance when the input of the neural network is the mean end force respectively; S220: Design the loss function of the neural network parameters, combined with the maximum likelihood estimation and the parameter set of the neural network, and the specific loss function design is as follows: ; Use the optimization algorithm to update the parameters of the neural network: ; Among them, is the learning rate; S230: To ensure that the predicted variance of the neural network model can accurately reflect the uncertainty in the real data, it is necessary to calibrate the predicted variance; use the heteroscedastic neural network mapping model to infer the samples in the validation set to obtain the predicted sample sampling variance ; Use the actual sample sampling variance corresponding to the contact force in the validation set to calibrate the predicted variance; for this purpose, define the following calibration loss function: ; Among them, is the sample variance calibration factor, is the number of samples in the validation set; The process of using the optimization algorithm to update the sample variance calibration factor is as follows: ; Among them, is the learning rate, and the calibrated predicted sampling variance is ; S240: According to the trained neural network model and the sample variance calibration factor, combined with the end force of the robot in the force control direction, predict the mean contact force of each contact surface and the calibrated predicted sampling variance .

[0024] In one embodiment, S300 includes: S310: Construct a neural network inverse mapping model, and train the neural network inverse mapping model by using a data set combined with the mean contact force and sampling variance predicted on each contact surface. Define the joint training objective as minimizing the prediction error of the neural network inverse mapping model and the consistency error of the heteroscedastic neural network mapping model: ; Wherein, is the mean contact force of the th sample in the training set The end-effector force predicted by passing through the inverse mapping model, that is, ; while and are respectively The mean contact force and sampling variance predicted by passing through the heteroscedastic neural network mapping model; Use an optimization algorithm to update the parameters of the neural network inverse mapping model: ; Wherein, is the learning rate; S320: Use the trained neural network inverse mapping model to implement real-time inference from the expected mean contact force to the expected end-effector force , that is: ; Wherein, is the parameter set of the neural network inverse mapping model, is the neural network inverse mapping model.

[0025] In one embodiment, S400 includes: S410: Define the objective function of the expected contact force range for each contact surface. The objective function needs to ensure the full contact of each contact surface, and at the same time avoid the decrease of the structural strength caused by excessive contact force. Therefore, the contact force value of each contact surface should meet the upper and lower limit requirements of the process when the assembly is completed; Assume is the contact force of the th contact surface predicted, wherein, , is the total number of contact surfaces; the actual contact force of each contact surface should meet the following conditions: ; Wherein, and are respectively the expected lower bound value of the contact force and the expected upper bound value of the contact force; Considering the uncertainty factors in the sampled data of the sensor, when considering uncertainty, the conditions that the actual contact force of each of the above contact surfaces should satisfy are modified as follows: ; where, is the sampling standard deviation of the contact force of the th contact surface predicted by the neural network; Fully considering uncertainty and contact force uniformity, the allowable value of the upper bound of the contact force is set to , and the allowable value of the lower bound of the contact force is set to , where is the maximum contact force standard deviation of the contact surfaces, is the maximum value function; in addition, the desired upper bound value and the lower bound value of the contact force are respectively converted into the desired upper bound value and the lower bound value of the end-effector force in the force control direction through the neural network inverse mapping model;; where, is the rectified linear unit function, and its specific expression is , , and are respectively the th, th and th components of ; S420: Define the contact uniformity objective function. In order to make the contact force distribution more uniform and minimize the difference between contact forces, that is, minimize the difference between the contact force and their average value , map the contact uniformity to the end through the neural network inverse model, and the objective function is defined as: ; where, represents the expected end-effector force regarding contact uniformity in the force control direction inferred by the neural network inverse model when the expected average contact force of each contact surface is ; therefore, the gradient of with respect to can be obtained: ; S430: Calculate the gradient of the desired contact force range. According to the objective function in S410, let and , the gradient of with respect to can be obtained as: ; where is the indicator function, and its specific expression is: ; S440: Desired end - effector force design. Combining the gradient calculated in S420 and S430, the ideal end - effector force variation law is designed as: ; where is a weight coefficient, is a constant, is the projection operator, which is used to ensure that once it enters the predetermined range it will always remain within this predetermined range, and its specific expression is: ; It should be noted that industrial robots usually only have position control and can only indirectly change the end - effector force by adjusting the position. Therefore, the end - effector force obtained according to the ideal end - effector force variation law is regarded as the desired end - effector force , that is . Then, an adaptive force control and admittance model is designed. By adjusting the desired position in the force control direction, the actual end - effector force in the force control direction of the robot will finally converge to the desired end - effector force , so as to achieve the convergence of the contact force of the angle piece to the desired contact force.

[0026] In one embodiment, S500 includes: S510: Define the end - effector force tracking error in the force control direction of the robot, specifically: ; S520: Design a force controller. The following PI force controller is used to adjust the end - effector force tracking error of the robot: ; where and are positive - definite diagonal matrices, is the adaptive factor of the force controller coefficient. The role of this adaptive factor is to increase the integral coefficient to accelerate the convergence of the end - effector force error when the force tracking error is large; while when the force tracking error is small, reduce the integral coefficient to avoid over - correction; the specific form of the adaptive factor is as follows: ; Among them, is a preset threshold value, is a constant; S530: Design the activation function of the PI force controller. When the frame and the skin come into contact, turn on the force controller. In addition, to ensure the safety of the frame and the robot, when the contact force exceeds the set value, turn off the force controller. Therefore, the activation function matrix of the force controller is defined as: , where the -th diagonal element of the activation function matrix of the force controller has the specific form of: ; Among them, , is a constant, is a positive constant and satisfies , represents a diagonal matrix. When is greater than the threshold value , the force controller is activated. This design helps to avoid false triggering caused by factors such as sensor measurement noise and zero drift, so as to ensure that the tracking performance of the robot in free space is not affected; when , start to smoothly turn off the force controller until when the force controller is completely turned off to avoid excessive contact force; S540: Design the activation function of the adaptive admittance controller. The design of the activation function of the adaptive admittance controller aims to ensure the motion performance of the robot in free space, and at the same time be able to accurately execute adaptive admittance control when the frame and the skin are in contact to achieve compliance to ensure the safety of the frame and the skin; define the activation function of the adaptive admittance controller as: , where has the specific form of: ; Among them, ; is the width of the transition section of the activation function of the adaptive admittance control; S550: Design the basic adaptive force control and admittance control model, specifically: ; Among them, , , are positive definite diagonal matrices, representing the inertia matrix, the damping matrix, and the stiffness matrix respectively, , , and They are the adjustment amounts of the desired position, velocity, and acceleration in the force control direction, respectively, is the planned desired assembly point in the force control direction, is the updated desired assembly point in the force control direction; S560: Update the parameters of the adaptive admittance control, and design the following adaptation law for updating the stiffness matrix , and its specific expression is: ; where, is the basic stiffness update gradient, and its specific expression is ; and are the learning rates, and are the and th elements of the vectors and are the th diagonal element of the lower stiffness bound matrix and the th diagonal element of the upper stiffness bound matrix , let , and the function is used to penalize the stiffness values outside the predetermined range ; S570: Conduct a stability analysis on the basic adaptive force control and admittance control strategies, and consider the following energy function: ; Differentiating yields: ; Substituting the basic adaptive force control and admittance dynamic equations into the equation of gives: ; Since is a positive definite matrix, so is always non-positive. However, since it cannot be guaranteed in advance that and are non-positive, the passivity of the basic adaptive force control and admittance control systems with respect to cannot be guaranteed.

[0027] In one embodiment, S600 includes: S610: Design the energy tank dynamics with an online energy flow injection and freezing mechanism, and define the energy stored in the energy tank as , where, is the state of the energy tank, is the time; Regarding the energy difference as the initial energy margin that can be used for energy exchange pre-stored in the energy tank, where represents the initial time, is the initial energy value stored in the energy tank, is the lower bound value of the energy in the energy tank, and their relationship satisfies ; Since energy exchange will occur when the actual system is interconnected with the energy tank, the system performance will be affected by the energy storage value in the energy tank: If the initial energy margin is too small, it may not meet the system performance requirements; while if the initial energy margin is too large, it may lead to system insecurity; To balance system performance while ensuring safety, combined with the goal of force control direction, that is, the end-effector force error of the robot converges, the following dynamics of the energy tank with an online energy flow injection and freezing mechanism are proposed to dynamically adjust the energy in the energy tank to ensure that the system can operate efficiently and safely during actual operation: ; Among them, is a constant used to store part of the energy consumed by damping; is the updated velocity after correction, is the planned velocity vector after correction of the th element. At this time, and can be obtained by integration and differentiation from respectively; The sum of the power of the -dimensional force controller and stiffness update in the force control direction is expressed as , where and are adaptive power scaling coefficients; and are the and th elements of the vectors and is the basic stiffness update gradient in S560; is the function of the th online energy injection and freezing mechanism in the force control direction. The coefficient functions to ensure that the energy value stored in the energy tank does not exceed the predetermined energy upper bound , so is defined as follows: ; Among them, the function For when the energy reaches the preset lower bound of the energy tank smoothly interrupt the energy exchange between the adaptive force control and admittance control system and the virtual energy tank, so as to ensure that the value of the energy tank is not lower than this lower bound value. Therefore is defined as follows: ; wherein is the width of the smooth transition section of the energy value; is the function of the online energy injection and freezing mechanism in the th force control direction. Its function is that when the end force of the robot remains within the safe range if the energy stored in the energy tank is too low, resulting in the system performance being limited, energy is injected online to ensure the system performance; in addition, when the contact force is too large, that is the energy exchange with the energy tank is frozen to avoid the adaptive force control and admittance system extracting energy from the energy tank, thus ensuring the safety of the system; in order to achieve a smooth transition from energy injection to energy freezing, a transition section from energy injection to energy freezing is set when the contact force is large within the safe range whose function is that when starts a smooth transition from the energy injection function to the energy freezing function, and when the energy exchange is completely frozen. Its specific expression is: ; wherein is a constant, and the function ; constant; the adaptive proportionality coefficient of energy injection and freezing in the th dimension in the force control direction has the specific expression ; S620: Design an adaptive power flow. Although the energy tank can ensure the passivity of the interconnected system, if the power allowed to be injected into the system is too large, it may lead to system insecurity; on the contrary, if the injected power is too small, it may limit the actual performance and response speed of the system; therefore, according to the task requirements and system status, the maximum power allowed to be injected into the actual system is adjusted and updated in real time; for this purpose, the following adaptive power scaling coefficient is defined, where has the following specific form: ; wherein, the power of the force controller is and the power of the stiffness update gradient corrected by the energy tank is . According to the target of the force control direction, the following update law about is designed: ; Among them, is the learning rate; is the th end - force component in the force - control direction; is the forgetting factor, which is used to prevent the maximum power value allowed to be injected in the actual system from being too large; S630: Design the modified force - control and adaptive - admittance control model, and its form is designed as: ; Among them, is the - dimensional identity matrix, is the function matrix of the online energy injection and freezing mechanism in the force - control direction, is the adaptive power - scaling coefficient matrix. At this time, the update law of the modified stiffness matrix is designed as: ; Among them, is the update law of the th diagonal element of the stiffness matrix, and are the th diagonal elements of the lower - bound stiffness matrix and the upper - bound stiffness matrix respectively; To ensure the stability and performance of the system, determine the damping value of the system according to the critical - damping condition: ; S640: Stability and performance analysis. Define the total energy function of the interconnected system composed of the robot system and the energy tank as: ; Among them, is the energy of the robot system, and the total power sum of the force controller and stiffness update in the force - control direction is rewritten as , where is the original stiffness - update gradient matrix with power limitation, and its expression is ; Differentiating yields: ; To analyze the stability of the interconnected system composed of the robot system and the energy tank, define as the set of all force - control directions, and define as the set of in the force - control direction; Through and For two cases, analyze the stability and performance of the system, where represents the empty set. In addition, from the expression in S610 , the following inequality always holds: ; Case 1: ; Under this condition, for , there is always , so there is and . At this time, can be simplified to: ; From , it can be seen that in this case, the interconnected system is stable; Case 2: ; For there is and for , . At this time, there is and , where and when at this time , is simplified to: ; Since the matrix is a diagonal positive semi - definite matrix, so when , there is . Therefore, when , the interconnected system is stable; when , is decomposed into , where is the corresponding Lyapunov function, is the corresponding Lyapunov function. At this time, there is and . Therefore, when , the system is passive with respect to the direction of and is stable for . Since passivity can ensure the stable interaction between the robot system and any passive environment, the system is stable in this case; In summary, both system cases 1 and 2 are stable, so the system is always stable; S650: Analysis of the convergence of the force tracking error. The total system variables are defined as , where is the state of the environment when the environmental energy is . At this time, the total energy function of the system is defined as ; Consider the optimal state and such that: ; wherein, is the space of all possible values; Since the environment always dissipates energy, at this time when and , there is , so at this time there is , so is asymptotically stable in the whole system; Consider at the equilibrium point , the modified adaptive force control and admittance control model becomes: ; It is easy to prove that is the unique steady-state solution of the system, so the end-effector force of the robot converges to the desired force.

[0028] In one embodiment, S700 includes: Substitute the admittance parameter value obtained in S500 into the modified adaptive force control and admittance control model to obtain the pose adjustment amount: ; wherein, and are the adjustment amounts of the desired position and the desired velocity at the previous sampling moment respectively, is the sampling period; Update the desired position in the force control direction: ; Thereby obtaining the trajectory correction amount in the force control direction: , where is the actual position of the end-effector of the robot in the force control direction measured by the external digital measurement device.

[0029] In one embodiment, S800 includes: S810: Compare the actual position of the end-effector of the robot in the position control direction measured by the external digital measurement device with the desired position in the position control direction, and calculate the trajectory correction amount in the position control direction according to the difference between the two: ; S820: Compile the robot control program. Since the industrial robot accepts position control instructions, send the updated trajectory correction amount to the robot, thereby realizing the precise control of the robot.

[0030] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) An adaptive force control and admittance control model is innovatively proposed. Based on performance indicators, the expected force at the end of the robot is optimized, and the admittance parameters are adaptively adjusted according to the goal of force control. This model is interconnected with a virtual energy tank with adaptive power limitation and online energy injection and freezing mechanisms to ensure the safety and performance stability of the system. At the same time, by ensuring the accurate tracking of the force at the end of the robot to the expected force, the precise compliant assembly of the panel assembly is achieved. (2) A contact force prediction model for each contact surface of the bulkhead based on a heteroscedastic neural network is proposed. This model can real-time predict the contact force and its sampling variance of each contact surface of the bulkhead according to the force at the end of the robot, so as to adjust the contact force by adjusting the force at the end. This method effectively ensures the assembly quality of the bulkhead and avoids problems such as insufficient contact or excessive contact force. (3) By designing a force control activation function and an adaptive admittance activation function, the proposed scheme can be applied to both free motion spaces and scenarios with rich contacts. (4) A real-time high-precision control method for robot motion based on digital measurement is designed to improve the motion accuracy of the robot.

[0031] A robot adaptive force control and admittance control system for panel assembly, comprising a data set acquisition module, a heteroscedastic neural network mapping model establishment module, a neural network inverse mapping model establishment module, an expected end force optimization module, an admittance parameter update module, an adaptive adjustment law of admittance parameters and a force controller correction module, a trajectory expected value correction module and a trajectory correction; The data set acquisition module is used to measure the pose of the end of the robot in real time based on an external digital measurement device, synchronously collect the end force and contact force data, and construct a data set; The heteroscedastic neural network mapping model establishment module is used to establish a heteroscedastic neural network mapping model. After training through the data set and a preset first loss function, the mean value and sampling variance of the contact force on each contact surface are predicted in combination with the end force of the robot in the force control direction; The neural network inverse mapping model establishment module is used to establish a neural network inverse mapping model. After training the neural network inverse mapping model through the data set, a preset second loss function, and the predicted mean value and sampling variance of the contact force on each contact surface, the mapping from the expected contact force to the expected end force is realized; The expected end force optimization module is used to define the expected contact force range constraint and uniformity index of each contact surface, and optimize the expected end force through gradient calculation; The admittance parameter update module is used to define a basic adaptive force control and admittance control model, update the admittance parameters according to the control goal of end force convergence, and perform stability analysis; The adaptive adjustment law of admittance parameters and the force controller correction module are used to design a virtual energy tank with adaptive power limit and energy injection and freezing functions, and connect it to the basic adaptive force control and admittance control model. Through the interconnected structure, the adaptive stiffness adjustment law of the admittance parameters and the force controller are corrected to obtain the corrected adaptive force control and admittance control model; The trajectory expected value correction module is used to obtain the corrected trajectory expected value in the force control direction according to the corrected adaptive force control and admittance control model and its parameter values, and combine the actual pose of the robot end in the force control direction to obtain the trajectory correction amount in the force control direction; The trajectory correction module is used to measure the actual pose of the robot end in the position control direction in real time, subtract it from the expected pose of the robot in the position control direction to obtain the trajectory correction amount in the position control direction, and send the trajectory correction amount in the position control direction and the trajectory correction amount in the force control direction to the industrial robot to achieve high-precision compliant assembly.

[0032] For the specific limitations of the robot adaptive force control and admittance control system for panel assembly, reference can be made to the limitations of the robot adaptive force control and admittance control method for panel assembly in the above text, which will not be elaborated here. Each module in the above-mentioned robot adaptive force control and admittance control system for panel assembly can be implemented in whole or in part by software, hardware and their combination. The above-mentioned modules can be embedded in the processor of the computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to the above-mentioned modules.

[0033] The above has introduced in detail the robot adaptive force control and admittance control method and system provided by the present invention. Specific examples are used in this article to elaborate the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the core idea of the present invention. It should be pointed out that for those of ordinary skill in the art in this technical field, without departing from the principle of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A robot adaptive force control and admittance control method for panel assembly, characterized in that The method includes the following steps: S100: Based on an external digital measurement device, the pose of the robot end is measured in real time, and the end force and contact force data are synchronously collected to construct a data set; S200: Establish a heteroscedastic neural network mapping model. After training with the data set and a preset first loss function, the average value and sampling variance of the contact force on each contact surface are predicted in combination with the end force of the robot in the force control direction; S300: Establish a neural network inverse mapping model. After training the neural network inverse mapping model with the data set, a preset second loss function, and the predicted average value and sampling variance of the contact force on each contact surface, the mapping from the desired contact force to the desired end force is realized; S400: Define the desired contact force range constraint and uniformity index for each contact surface, and optimize the desired end force through gradient calculation; S500: Define a basic adaptive force control and admittance control model, update the admittance parameters according to the control objective of end force convergence, and perform stability analysis; S600: Design a virtual energy tank with adaptive power limit and energy injection and freezing functions, and connect it to the basic adaptive force control and admittance control model. Through the interconnected structure, correct the adaptive stiffness adjustment law of the admittance parameters and the force controller to obtain a corrected adaptive force control and admittance control model; S700: According to the corrected adaptive force control and admittance control model and its parameter values, obtain the corrected trajectory expectation value in the force control direction, and combine the actual pose of the robot end in the force control direction to obtain the trajectory correction amount in the force control direction; S800: Measure the actual pose of the robot end in the position control direction in real time, subtract it from the desired pose of the robot in the position control direction to obtain the trajectory correction amount in the position control direction, and send the trajectory correction amount in the position control direction and the trajectory correction amount in the force control direction to the industrial robot to achieve high-precision compliant assembly.

2. The method according to claim 1, wherein S100 includes: S110: Measure the pose of the robot end in real time with the help of an external digital measuring device , and according to process requirements, it is divided into poses in the position control direction and poses in the force control direction . Among them, the goal of the position control direction is to achieve precise positioning, and the goal of the force control direction is to achieve the convergence of the end force error; the dimension of the force control direction and the dimension of the position control direction depend on process requirements. When only position control is considered, ; when force control is considered in all directions, ; S120: Data collection. While the robot is in the pose , , perform measurements of the end-effector force and contact forces to obtain the end-effector force samples of the robot in the force control direction and the contact force samples of each contact surface , calculate and record the corresponding sample statistics: ； Among them, represents the th pose, is the total number of times to obtain the pose, represents the th measurement of the end force and the contact force, represents the total number of measurements of the end force and the contact force, is the mean value of the robot end force calculated by times of sampling in the force control direction under the robot pose , is the mean value of the corresponding pressure sensor, and its sampling variance is: 。 3. The method according to claim 2, wherein S200 includes: S210: Construct a heteroscedastic neural network mapping model, divide the data set into a training set, a test set, and a validation set. During the training process, use the mean sample of the end-effector force of the robot in the training set After standardization processing, use it as the input to train the heteroscedastic neural network, and establish a mapping relationship between the end-effector force of the robot in the force control direction and the mean contact force and sample sampling variance of each contact surface of the bulkhead. This mapping relationship is expressed as: ； Among them, is the number of samples in the training set, and are neural network functions for predicting the mean and logarithmic variance of the contact force of each contact surface, respectively; The output of is transformed by exponentiation to obtain the variance, so as to ensure the non-negativity of the variance; are the parameter sets of the corresponding networks, respectively; and are the mean contact force of each contact surface of the spacer predicted by the neural network and the sample sampling variance when the input of the neural network is the mean end force respectively; S220: Design the loss function for the neural network parameters, combining maximum likelihood estimation and the parameter set of the neural network , and the specific loss function design is as follows: ； Use an optimization algorithm to update the parameters of the neural network: ； Among them, is the learning rate; S230: Infer the samples in the validation set using the heteroscedastic neural network mapping model to obtain the predicted sample sampling variances ; Calibrate the predicted variances using the actual sample sampling variances corresponding to the contact forces in the validation set , for which the following calibration loss function is defined: ； Among them, is the sample variance calibration factor, is the number of samples in the validation set; The process of using an optimization algorithm to update the sample variance calibration factor is as follows: ； Among them, is the learning rate, and the calibrated predicted sampling variance is ; S240: Based on the trained neural network model and the sample variance calibration factor, combined with the end force of the robot in the force control direction , predict the mean contact force of each contact surface and the calibrated predicted sampling variance .

4. The method according to claim 3, wherein S300 includes: S310: Construct a neural network inverse mapping model, train the neural network inverse mapping model with the data set combined with the predicted average value and sampling variance of the contact force on each contact surface, and define the joint training objective as minimizing the prediction error of the neural network inverse mapping model and the consistency error of the heteroscedastic neural network mapping model: ； Among them, is the mean contact force of the th sample in the training set The end effector force predicted by the inverse mapping model, that is, ; while and are respectively the mean contact force and the sampling variance predicted by the heteroscedastic neural network mapping model; Use an optimization algorithm to update the parameters of the neural network inverse mapping model: ； wherein, is the learning rate; S320: Implement real-time inference from the desired average contact force to the desired end-effector force using the trained neural network inverse mapping model, i.e.: ； Among them, is the parameter set of the neural network inverse mapping model, is the neural network inverse mapping model.

5. The method according to claim 4, wherein S400 includes: S410: Define the desired contact force range objective function for each contact surface. The objective function needs to ensure the full contact of each contact surface and avoid the decrease of the structural strength caused by excessive contact force. Therefore, the contact force value of each contact surface should meet the upper and lower limit requirements of the process when the assembly is completed; Hypothesis is the contact force of the th predicted contact surface, where , is the total number of contact surfaces; the actual contact force of each contact surface should satisfy the following conditions: ； Among them, and are the lower bound value and the upper bound value of the desired contact force respectively; Considering the uncertainty factors in the sampling data of the sensor, when considering the uncertainty, the conditions that the actual contact force of each contact surface should meet are modified as: ； Among them, is the sampling standard deviation of the contact force of the th contact surface predicted by the neural network; Fully considering uncertainty and contact force uniformity, the allowable value of the upper bound of the contact force is set to , and the allowable value of the lower bound of the contact force is set to , where is the maximum standard deviation of the contact forces of contact surfaces, and is the maximum value function; in addition, the desired upper bound value and lower bound value of the contact force are respectively converted into the desired upper bound value and lower bound value of the end effector force in the force control direction through the neural network inverse mapping model; Since the contact force is indirectly adjusted by adjusting the end force, the objective function design of the contact force range constraint is: ； Among them, is the linear rectification function, and its specific expression is , , and are respectively , and the th component of S420: Define the contact uniformity objective function. To make the contact force distribution more uniform, minimize the difference between contact forces, that is, minimize the contact force and their average value The difference between them. Map the contact uniformity to the end using the neural network inverse model. The objective function is defined as: ； Among them, indicates that when the expected average contact force of each contact surface is , the expected end force regarding contact uniformity in the corresponding force control direction is inferred by the neural network inverse model; thus, the gradient regarding can be obtained: ​ ； S430: Calculating the gradient of the desired contact force range. According to the objective function of S410, let and , we can obtain with respect to : ； Among them, is an indicator function, and its specific expression is: ； S440: Desired end - force design. Combining the gradient calculated in S420 and S430, the ideal end - force variation law is designed as: ； Among them, is a weight coefficient, is a constant, is a projection operator used to ensure that once it enters the predetermined range it will always remain within the predetermined range, and its specific expression is: ； The end - effector force obtained according to the variation law of the ideal end - effector force is regarded as the desired end - effector force , that is , in addition, since industrial robots usually can only receive position commands, it is necessary to adjust the position of the robot according to the end - effector force error, so as to realize the control of the end - effector force.

6. The method according to claim 5, wherein S500 includes: S510: End-effector force tracking error that defines the force control direction , specifically: ； S520: Design a force controller. Use the following PI force controller to adjust the end - force tracking error of the robot: ； Among them, and are positive definite diagonal matrices, is an adaptive factor of the force controller coefficient. The function of this adaptive factor is to increase the integral coefficient to accelerate the convergence of the end force error when the force tracking error is large; while when the force tracking error is small, reduce the integral coefficient to avoid overcorrection; the adaptive factor has the following specific form: ； Among them, is a preset threshold value, is a constant; S530: Design the activation function of the PI force controller. When the frame contacts the skin, activate the force controller. In addition, to ensure the safety of the frame and the robot, when the contact force exceeds the set value, turn off the force controller. Therefore, the activation function matrix of the force controller is defined as: , where the th diagonal element of the activation function matrix of the force controller has the following specific form: ； Among them, , is a constant, is a positive constant and satisfies , represents a diagonal matrix. When is greater than the threshold , the force controller is activated. When , the force controller starts to smoothly turn off until when the force controller is completely turned off to avoid excessive contact force; S540: Design the activation function of the adaptive admittance controller, defined as: , where The specific form of is: ; Among them, ; is the width of the transition section of the adaptive admittance control activation function; S550: Design a basic adaptive force control and admittance control model, specifically: ; Among them, , , are positive definite diagonal matrices, representing the inertia matrix, damping matrix, and stiffness matrix respectively, , , and are the adjustment amounts of the desired position, velocity, and acceleration in the force control direction respectively, is the planned desired assembly point in the force control direction, is the updated desired assembly point in the force control direction; S560: Update the parameters of the adaptive admittance control and design the following adaptation law to update the stiffness matrix , and its specific expression is as follows: ； Among them, is the basic stiffness update gradient, and its specific expression is ; and are the learning rates, and are the and -th elements of the vectors and and are the -th and -th diagonal elements of the lower stiffness bound matrix and the upper stiffness bound matrix , let , is a function used to penalize stiffness values outside the predetermined range ; S570: Conduct a stability analysis on the basic adaptive force control and admittance control strategy. Consider the following energy function: ; For Differentiating gives: ; Substitute the basic adaptive force control and admittance dynamic equations into the equation of, and we can get: ; Since is a positive definite matrix, so is always non-positive. However, since it cannot be guaranteed in advance that and are non-positive, thus the passivity of the basic adaptive force control and admittance control systems with respect to cannot be guaranteed.

7. The method according to claim 6, wherein S600 includes: S610: Design the dynamics of the energy tank with an online energy flow injection and freezing mechanism, and define the energy stored in the energy tank as , where is the state of the energy tank, is the time; regard the energy difference of as the initial energy margin pre-stored in the energy tank available for energy exchange, where represents the initial time, is the initial energy value stored in the energy tank, is the lower bound value of the energy in the energy tank, and their relationship satisfies ; To balance system performance while ensuring safety, combining the goal in the force control direction, that is, the convergence of the robot end - force error, the following dynamics of the energy tank with an online energy - flow injection and freezing mechanism are proposed to dynamically adjust the energy in the energy tank and ensure that the system can operate efficiently and safely during actual operation: ； Among them, is a constant used to store a part of the energy consumed by damping; is the updated velocity correction, is the corrected planned velocity vector The -th element, at this time, and can be obtained by integration and differentiation respectively; the sum of the power of the -th dimensional force controller and stiffness update in the force control direction is expressed as where, and and are adaptive power scaling coefficients; and are the -th and -th elements of the vectors respectively; is the basic stiffness update gradient in S560; is the function of the -th online energy injection and freezing mechanism in the force control direction, and the coefficient functions to ensure that the energy value stored in the energy tank does not exceed the predetermined energy upper bound , so is defined as follows: ； Among them, the function is used to smoothly interrupt the energy exchange between the adaptive force control and admittance control system and the virtual energy tank when the energy reaches the preset lower bound of the energy tank , so as to ensure that the value of the energy tank is not lower than this lower bound value. Therefore is defined as follows: ； Among them, is the width of the smooth transition section of the energy value; is the function of the online energy injection and freezing mechanism in the th force control direction. Its function is to ensure the system performance by online injecting energy when the end force of the robot remains within the safe range ; in addition, when the contact force is too large, that is , the energy exchange with the energy tank is frozen to avoid the adaptive force control and admittance system extracting energy from the energy tank, thus ensuring the safety of the system; to achieve a smooth transition from energy injection to energy freezing, a transition section from energy injection to energy freezing is set when the contact force is large within the safe range , and its function is when starts a smooth transition from the energy injection function to the energy freezing function, and when , the energy exchange is completely frozen. Its specific expression is: ； Among them, is a constant, and the function ; constant; the adaptive proportionality coefficient of energy injection and freezing in the th dimension in the force control direction has the specific expression of ; S620: Design an adaptive power flow to adjust and update the maximum power allowed to be injected into the actual system in real time according to the mission requirements and system status; for this purpose, define the following adaptive power scaling factor , where has the following specific form: ； Among them, the power of the force controller is , and the power of the stiffness update gradient corrected by the energy tank is . According to the goal of the force control direction, the following update law about is designed: ； Among them, is the learning rate; is the th end force component in the force control direction; is the forgetting factor, which is used to prevent the maximum power value allowed to be injected in the actual system from being too large; S630: Design a modified force control and adaptive admittance control model, and its form is designed as: ； Among them, is an identity matrix of dimension is a function matrix of the online energy injection and freezing mechanism in the force control direction, is an adaptive power scaling coefficient matrix, and the update law of the corrected stiffness matrix at this time is designed as: ； Among them, is the update law of the th diagonal element of the stiffness matrix, and are respectively the th diagonal element of the lower bound stiffness matrix and the th diagonal element of the upper bound stiffness matrix; To ensure the stability and performance of the system, determine the damping value of the system according to the critical damping condition: ； S640: Stability and performance analysis. Define the total energy function of the interconnected system composed of the robot system and the energy tank as: ； Among them, is the energy of the robot system, the total power sum of the force controller and the stiffness update in the force control direction The expression of is rewritten as , where is the original stiffness update gradient matrix with power limitation, and its expression is ; Taking the derivative of ； To analyze the stability of the interconnected system composed of a robotic system and an energy tank, define as the set of all force control directions, and define as the set on the force control direction ; analyze the stability and performance of the system through and two cases, where represents the empty set. In addition, from the expression of in S610, the following inequality always holds: ； Case 1: ; Under this condition, for , there is always , so there is and . At this time can be simplified to: ； From , it can be seen that in this case, the interconnected system is stable; Case 2: ; For there is and for , , at this time there is and , where and when then , Simplify to: ； Since the matrix is a diagonal positive semi - definite matrix, when , we have . Thus, when , the interconnected system is stable; when , is decomposed into , where is the corresponding Lyapunov function, is the corresponding Lyapunov function. At this time, we have and . Therefore, when , the system is passive with respect to the direction of . For , it is stable. Since passivity can ensure the stable interaction between the robot system and any passive environment, the system is stable in this case; In summary, both system cases 1 and 2 are stable, so the system is always stable; S650: Analysis of the convergence of the force tracking error. The total system variables are defined as , where is the environmental energy, and is the state of the environment at this time. At this time, the total energy function of the system is defined as ; Considering the optimal state and such that: ； Among them, is the space of all possible values; since the environment always dissipates energy, at this time when and , there is , so at this time there is , so the whole system is asymptotically stable; considering at the equilibrium point the modified adaptive force control and admittance control model becomes: ； It is easy to prove is the unique steady-state solution of the system, so the end-effector force of the robot converges to the desired force.

8. The method according to claim 7, characterized in that S700 includes: According to the admittance parameter values obtained in S500, substitute them into the modified adaptive force control and admittance control model to obtain the pose adjustment amount: ； wherein, and are respectively the adjustment amounts of the expected position and the expected speed at the previous sampling moment, is the sampling period; Update the desired position in the force control direction: ； Thus, the trajectory correction amount in the force control direction is obtained: , where is the actual position of the robot end in the force control direction measured by the external digital measurement device.

9. The method according to claim 8, characterized in that, S800 includes: S810: Control the actual position of the end of the orientation robot according to the position measured by an external digital measuring device , and compare it with the desired position in the position control direction . Calculate the trajectory correction amount in the position control direction according to the difference between the two: ; S820: Write a robot control program. Since the industrial robot receives position control instructions, the updated trajectory correction amount is sent to the robot, thereby achieving precise control of the robot.

10. The robot adaptive force control and admittance control system for wall panel assembly, characterized in that, It includes a dataset acquisition module, a heteroscedastic neural network mapping model establishment module, a neural network inverse mapping model establishment module, a desired end - force optimization module, an admittance parameter update module, an adaptive adjustment law for admittance parameters and a force controller correction module, a trajectory expected value correction module, and a trajectory correction; The dataset acquisition module is used to measure the pose of the robot end in real - time based on an external digital measurement device, synchronously collect end - force and contact - force data, and construct a dataset; The heteroscedastic neural network mapping model establishment module is used to establish a heteroscedastic neural network mapping model. After training with the dataset and a preset first loss function, combine the end - force of the robot in the force control direction to predict the mean and sampling variance of the contact force on each contact surface; The neural network inverse mapping model establishment module is used to establish a neural network inverse mapping model. After training the neural network inverse mapping model with the dataset, a preset second loss function, and the predicted mean and sampling variance of the contact force on each contact surface, realize the mapping from the desired contact force to the desired end - force; The desired end - force optimization module is used to define the range constraint and uniformity index of the desired contact force on each contact surface, and optimize the desired end - force through gradient calculation; The admittance parameter update module is used to define a basic adaptive force control and admittance control model, update the admittance parameters according to the control goal of end - force convergence, and conduct a stability analysis; The adaptive adjustment law of admittance parameters and the force controller correction module are used to design a virtual energy tank with adaptive power limitation and energy injection and freezing functions, and connect it to the basic adaptive force control and admittance control model. Through the interconnected structure, the adaptive stiffness adjustment law of admittance parameters and the force controller are corrected to obtain the corrected adaptive force control and admittance control model; The trajectory expected value correction module is used to obtain the corrected trajectory expected value in the force control direction according to the corrected adaptive force control and admittance control model and its parameter values, and combine the actual pose of the robot end in the force control direction to obtain the trajectory correction amount in the force control direction; The trajectory correction module is used to measure the actual pose of the robot end in the position control direction in real time, subtract it from the expected pose of the robot in the position control direction to obtain the trajectory correction amount in the position control direction, and send the trajectory correction amount in the position control direction and the trajectory correction amount in the force control direction to the industrial robot to achieve high-precision compliant assembly.

Images