Adaptive control optimization method for robot

Abstract

The application discloses a kind of adaptive control optimization methods of manipulator, it is related to intelligent control technical field, utilize manipulator ontology perception network to obtain the real-time time series signal stream of driving pressure and current, real-time time series signal stream is converted into continuous time sequence characteristic matrix, and equivalent rigid body dynamics model is constructed by physical structure data, and state variable and error variable are defined, continuous time sequence characteristic matrix is input into echo state network, and reserve pool state vector is output by reserve pool structure mapping, and reserve pool state vector is defined as characteristic base function vector, Bellman error is calculated based on error variable, and reinforcement learning evaluation network is constructed using characteristic base function vector, weight estimation value and disturbance estimation parameter are calculated, aggregate uncertainty function is obtained, output control moment is calculated based on aggregate uncertainty function, and actual output moment of actuator is calculated by output control moment, to complete the control of manipulator.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to an adaptive control optimization method for a robotic arm. Background Technology

[0002] In the field of industrial automation, robotic arms, as core equipment for performing complex tasks, are widely used in chip assembly, precision machining, and material handling. As working environments become increasingly unstructured and highly interactive, optimizing the performance of robotic arm control systems has become crucial for improving production efficiency and product quality. Most existing robotic arm control systems rely on preset control strategies. However, in actual operation, limited by the physical characteristics of the robotic arm itself and changing external environmental factors, preset static control strategies often struggle to adapt to complex dynamic conditions, resulting in suboptimal operational performance.

[0003] Currently, Chinese patent application number 202511073167.5 discloses an optimization method and system for a robotic arm control system. This method involves acquiring historical operation datasets, merging and extracting associated motion features, and using a pre-trained model to generate strategy adjustment parameters to update the control command sequence. However, the existing technology has the following shortcomings: it is difficult to accurately characterize the complex physical coupling relationships within the flexible actuator; due to the lack of a high-dimensional feature mapping mechanism with time-delay memory characteristics, the system struggles to fully exploit the deep dynamic features in the original sensor signals when facing strong interactive interference; and the existing technology involves offline or quasi-online strategy adjustment, which becomes problematic when the actuator experiences efficiency loss. In the event of sudden faults such as loss or bias, the lack of instantaneous and continuous online compensation mechanisms leads to significant chattering of the control torque at the moment of the fault, severely affecting operational accuracy and equipment safety. Existing technologies lack the ability to approximate the lumped uncertainty of the system in real time. Without continuous guidance from an online evaluation network, the actuator cannot adjust the control gain in real time according to environmental changes, and cannot achieve true strategy self-evolution. Under complex interactive disturbances, existing control commands cannot guarantee that the tracking error reaches a steady state within a finite time. Conventional sliding mode control frameworks are prone to inducing control singularity problems when pursuing convergence speed, which limits the application of robots in high-requirement scenarios such as precision machining. Summary of the Invention

[0004] The technical problem solved by this invention is that in the case of unstructured dynamic environment and actuator sudden failure, the existing technology suffers from insufficient extraction of sensing signal features, failure compensation mechanism relies on discrete classification leading to response lag, and traditional neural networks lack temporal memory ability, making it difficult to achieve real-time and accurate correction of control gain under strong interactive interference. This results in problems such as severe control torque chattering, tracking error not converging within a finite time, and control singularity, which cannot meet the stringent requirements of robustness and steady-state accuracy for complex flexible operations.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: an adaptive control optimization method for a robotic arm, comprising the following steps:

[0006] Step S1: Use the robot's body perception network to obtain the real-time time series signal stream of driving pressure and current, convert the real-time time series signal stream into a continuous time series feature matrix, construct an equivalent rigid body dynamic model through physical structure data, and define state variables and error variables.

[0007] Step S2: Input the continuous temporal feature matrix into the echo state network, output the reserve pool state vector through the reserve pool structure mapping, and define the reserve pool state vector as the feature basis function vector;

[0008] Step S3: Calculate the Bellman error based on the error variable, construct a reinforcement learning evaluation network using the feature basis function vector, calculate the weight estimate and perturbation estimate parameters of the reinforcement learning evaluation network, obtain the execution network weights from the weight estimate, and obtain the lumped uncertainty function based on the execution network weights and perturbation estimate parameters.

[0009] Step S4: Calculate the output control torque based on the lumped uncertainty function and sliding mode function values, obtain the actual output torque of the actuator through the output control torque calculation, and complete the control of the robot arm through the actual output torque of the actuator.

[0010] Preferably, step S1 includes the following sub-steps:

[0011] Step S101: Arrange sensor groups within the parallel integrated over-constraint structure of each joint of the manipulator, construct a body perception network using the physical coupling relationship between the software actuators, and collect multi-dimensional physical quantities generated by the manipulator's operation and interaction with the outside through the body perception network to generate a real-time time series signal stream. The multi-dimensional physical quantities include the liquid pressure signal inside the software actuator and the drive current signal of the drive motor.

[0012] Preferably, step S1 further includes the following sub-steps:

[0013] Step S102, the preset length is A sliding time window is used to slide along the time axis on the real-time time series signal stream with a preset sampling step size. At each sampling moment, multi-dimensional signal data within the current sliding time window range is extracted. The multi-dimensional signal data includes multi-channel hydraulic pressure waveform signals and driving current sequences. Wavelet scattering network is used to extract the time-domain waveform feature vector of the multi-dimensional signal data, and the features of each channel are fused through matrix dimension transformation to construct a continuous time series feature matrix.

[0014] Preferably, step S1 further includes the following sub-steps:

[0015] Step S103: Collect the physical structure data of the robot arm, including joint geometry, component mass, moment of inertia, center of mass position, joint velocity vector, and joint position vector;

[0016] The inertia matrix is ​​calculated based on the component mass, moment of inertia, and center of mass position.

[0017] The gravity vector is calculated based on the component mass, joint position vector, and gravitational acceleration constant.

[0018] The Coriolis force matrix is ​​calculated based on the partial derivative of the inertia matrix with respect to the joint position vector and the joint velocity vector.

[0019] Based on physical structure data, an equivalent rigid body dynamic model of the manipulator is constructed using the rigid body equivalence assumption. The mathematical expression of the equivalent rigid body dynamic model is as follows:

[0020] ;

[0021] in, Let the joint position vector be... For joint velocity vectors, The joint acceleration vector, The inertia matrix, The Coriolis force matrix, The gravity vector For lumped disturbance torque, This refers to the actual output torque of the actuator.

[0022] The mathematical expression for the actual output torque of the actuator is:

[0023] ;

[0024] in, For ideal control torque, For an unknown time-varying fault matrix, Set the actuator bias fault vector;

[0025] Define state variables and error variables, specifically including:

[0026] The state variables include a first state variable and a second state variable, and the error variables include a first error variable and a second error variable;

[0027] Define the joint position vector as the first state variable and the joint velocity vector as the second state variable;

[0028] Obtain the desired trajectory vector and the first derivative of the desired trajectory vector as set for the robot's operation task. Subtract the desired trajectory vector from the first state variable to obtain the first error variable. Subtract the first derivative from the second state variable to obtain the second error variable.

[0029] Preferably, step S2 includes the following sub-steps:

[0030] Step S201: Input the continuous temporal feature matrix into the echo state network, and use the reservoir structure of the echo state network to perform high-dimensional mapping on the input continuous temporal feature matrix to output the reservoir state vector.

[0031] Step S202: Define the state vector of the reserve pool as a feature basis function vector, which includes the Actor basis function vector and the Critic basis function vector.

[0032] Preferably, step S3 includes the following sub-steps:

[0033] Step S301: Based on the first error variable, the second error variable, and the ideal control torque, calculate the Bellman error. The mathematical expression for the Bellman error is:

[0034] ;

[0035] in, For Bellman error, This is the transpose of the first error variable. As the first error variable, For the transpose of the ideal control torque, It is the first positive definite constant matrix. It is the second positive definite constant matrix;

[0036] A reinforcement learning evaluation network is established based on the Critic basis function vector, and the weight estimates of the reinforcement learning evaluation network are calculated.

[0037] Construct a gradient term based on the Critic basis function vector and the second error variable;

[0038] The reinforcement learning evaluation network is iterated based on the gradient term, Bellman error, and weight estimates to obtain the weight change rate of the reinforcement learning evaluation network. The mathematical expression for the weight change rate is:

[0039] ;

[0040] in, The rate of change of weight. For gradient terms, These are the weight estimates. For Bellman error, The preset learning rate, This is the transpose of the gradient term.

[0041] Preferably, step S3 further includes the following sub-steps:

[0042] Step S302: Construct a sliding mode function based on the first error variable and the second error variable. The mathematical expression of the sliding mode function is:

[0043] ;

[0044] in, The value of the sliding mode function. The preset convergence exponent parameter, The preset gain adjustment parameters, As the first error variable, This is the second error variable.

[0045] Preferably, step S3 further includes the following sub-steps:

[0046] Step S303: For the dynamic disturbance caused by the sudden failure of the actuator, an online update law for the disturbance estimation parameters is designed using the nonlinear damping term generated by the state vector of the reservoir, and the rate of change of the disturbance estimation parameters is calculated. The mathematical expression for the rate of change of the disturbance estimation parameters is:

[0047] ;

[0048] in, For the rate of change of the disturbance estimation parameters, It is the first positive constant. It is the second positive constant. For the disturbance estimation parameters, As the second error variable, This is point-to-point multiplication. For switching functions, This represents the sliding mode function value.

[0049] Preferably, step S3 further includes the following sub-steps:

[0050] Step S304: Map the weight estimates to the execution network to obtain the execution network weights. Based on the execution network weights and the perturbation estimation parameters, perform vectorized recombination to construct a lumped uncertainty function. The mathematical expression of the lumped uncertainty function is:

[0051] ;

[0052] in, For lumped uncertainty functions, As the first state variable, For the second state variable, To perform the transpose of network weights, Let Actor be the basis function vector. For the disturbance estimation parameters, This is point-to-point multiplication. For switching functions, This represents the sliding mode function value.

[0053] Preferably, step S4 further includes the following sub-steps:

[0054] Step S401: Calculate the output control torque based on the lumped uncertainty function and the sliding mode function value. The mathematical expression for the output control torque is:

[0055] ;

[0056] in, To output control torque, For lumped uncertainty functions, The first derivative of the desired trajectory vector. The preset positive definite reaching law gain parameters, For the sliding mode approaching law term;

[0057] The output control torque is assigned to the ideal control torque, and the actual output torque of the actuator is calculated. The control of the robot is completed by the actual output torque of the actuator.

[0058] The beneficial effects of this invention are as follows: Addressing the problems of insufficient sensor signal feature extraction and the lack of temporal memory in traditional neural networks when unstructured dynamic environments and actuator abrupt failures coexist, leading to lag in control gain correction and torque chattering, this invention utilizes wavelet scattering networks to extract continuous temporal feature matrices with translation invariance. The echo state network's reservoir state vector is defined as the feature basis function vector for reinforcement learning. The reservoir's unique high-dimensional dynamic memory characteristics accurately characterize the nonlinear physical coupling and external interactive micro-vibrations of the soft actuator. Furthermore, combining a Bellman error-based evaluation network with a nonlinear damping term generated from the reservoir state vector, an online update law for disturbance estimation parameters is designed to construct a lumped uncertainty function. Finally, in conjunction with a non-singular terminal sliding mode control framework, this invention eliminates control law singularities while achieving rapid and smooth suppression of actuator failures and severe environmental disturbances, ensuring the steady-state accuracy and robustness of the robot under complex working conditions. Attached Figure Description

[0059] Figure 1 The flowchart illustrates the steps of an adaptive control optimization method for a robotic arm, as provided in one embodiment of the present invention.

[0060] Figure 2 The diagram below shows a control flowchart of an adaptive control optimization method for a robotic arm, as provided in one embodiment of the present invention. Detailed Implementation

[0061] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0062] Example, refer to Figure 1 An adaptive control optimization method for a robotic arm is provided, comprising the following steps:

[0063] Step S1: Use the robot's body perception network to obtain the real-time time series signal stream of driving pressure and current, transform the real-time time series signal stream into a continuous time series feature matrix, construct an equivalent rigid body dynamic model through physical structure data, and define state variables and error variables.

[0064] Step S2: Input the continuous temporal feature matrix into the echo state network, map the reservoir state vector through the reservoir structure, and define the reservoir state vector as the feature basis function vector.

[0065] Step S3: Calculate the Bellman error based on the error variable, construct the reinforcement learning evaluation network using the feature basis function vector, calculate the weight estimate and perturbation estimate parameters of the reinforcement learning evaluation network, obtain the execution network weights from the weight estimate, and obtain the lumped uncertainty function based on the execution network weights and perturbation estimate parameters.

[0066] Step S4: Calculate the output control torque based on the lumped uncertainty function and sliding mode function value. Obtain the actual output torque of the actuator through the output control torque calculation. Complete the control of the robot arm through the actual output torque of the actuator.

[0067] To address the issues of lag in control gain correction and torque chattering in existing technologies when faced with unstructured dynamic environments and sudden actuator failures, due to insufficient extraction of sensing signal features and the lack of temporal memory in traditional neural networks, this paper proposes a new approach. This approach utilizes wavelet scattering networks to extract translation-invariant continuous temporal feature matrices and defines the state vector of the echo state network's reservoir as the feature basis function vector for reinforcement learning. Leveraging the reservoir's unique high-dimensional dynamic memory, the nonlinear physical coupling and external interactive micro-vibrations of the soft actuator are accurately characterized. Furthermore, by combining a Bellman error-based evaluation network with a nonlinear damping term generated from the reservoir state vector, an online update law for disturbance estimation parameters is designed to construct a lumped uncertainty function. Finally, this is coupled with a non-singular terminal sliding mode control framework. This approach eliminates control law singularities while achieving rapid and smooth suppression of actuator failures and severe environmental disturbances, ensuring the steady-state accuracy and robustness of the robot under complex working conditions.

[0068] In a specific embodiment, step S1 includes the following sub-steps:

[0069] Step S101: Arrange sensor groups in the parallel integrated over-constraint structure of each joint of the robot, construct a body perception network by utilizing the physical coupling relationship between the software actuators, collect multi-dimensional physical quantities generated by the operation of the robot and its interaction with the outside through the body perception network, and generate a real-time time series signal stream. The multi-dimensional physical quantities include the liquid pressure signal inside the software actuator and the drive current signal of the drive motor.

[0070] It should be noted that ontology-aware networks refer to distributed heterogeneous sensing systems built upon physical connections. The specific construction method is as follows:

[0071] The hardware layout involves embedding micro-MEMS hydraulic sensors in the inner wall of each hydraulic chamber of the soft actuator to collect internal fluid pressure, and connecting Hall current sensors in series at the motor drive end connected to the soft actuator to collect drive current.

[0072] The physical coupling mechanism utilizes the fluid, solid, and electrical coupling characteristics of the soft actuator; that is, changes in the motor current cause changes in the hydraulic pump output pressure, which in turn causes deformation of the soft material. The body perception network captures the dynamic response of the physical coupling relationship in the time domain by synchronously acquiring heterogeneous signals, and indirectly senses the force state and contact vibration of the robot in the unstructured environment.

[0073] Step S102, the preset length is The sliding time window slides along the time axis on the real-time time series signal stream with a preset sampling step size. At each sampling moment, multi-dimensional signal data within the current sliding time window range is extracted. The multi-dimensional signal data includes multi-channel hydraulic pressure waveform signals and driving current sequences. The time-domain waveform feature vector of the multi-dimensional signal data is extracted using a wavelet scattering network, and the features of each channel are fused through matrix dimension transformation to construct a continuous time series feature matrix.

[0074] Multi-channel hydraulic pressure waveform signals characterize the internal state and external interactive vibration of the robot. The time-domain waveform feature vector has displacement invariance and deformation stability characteristics, and the continuous time-series feature matrix characterizes the dynamic characteristics and collision interaction characteristics of the robot.

[0075] It should be noted that the preset sliding time window length is 100ms, and the sampling frequency is set to 1kHz. The preset sliding time window length can cover the complete waveform cycle of a transient collision between the robot and the external environment.

[0076] The wavelet scattering network configuration adopts a two-layer wavelet scattering transform network. Because the complex Morlet wavelet has good time-frequency localization characteristics, the complex Morlet wavelet is selected as the mother wavelet for the wavelet basis function. The hyperparameters are set to 8 wavelets per octave, and the maximum scale parameter is 3, which corresponds to the time support domain covering the entire preset sliding time window length.

[0077] Zero-order scattering, first-order scattering, and second-order scattering calculations are performed on the intercepted multidimensional signal to generate a time-domain waveform feature vector with translation invariance and deformation stability.

[0078] The specific operation of matrix dimension transformation is as follows:

[0079] Assuming the number of hydraulic signal channels is The number of current signal channels is The feature dimension of each channel extracted by wavelet scattering is A channel cascading strategy is adopted to integrate the hydraulic feature matrix. With current characteristic matrix Concatenate along the channel dimension to construct a dimension of... The continuous temporal feature matrix is ​​then reshaped into a one-dimensional vector stream or a two-dimensional sequence suitable for the input of the echo-state network.

[0080] Step S103: Collect the physical structure data of the robot arm. The physical structure data includes joint geometry, component mass, moment of inertia, center of mass position, joint velocity vector, and joint position vector.

[0081] The inertia matrix is ​​calculated based on the component mass, moment of inertia, and center of mass position.

[0082] The gravity vector is calculated based on the component mass, joint position vector, and gravitational acceleration constant.

[0083] The Coriolis force matrix is ​​calculated based on the partial derivative of the inertia matrix with respect to the joint position vector and the joint velocity vector.

[0084] Based on physical structure data, an equivalent rigid body dynamic model of the manipulator is constructed using the rigid body equivalence assumption. The mathematical expression of the equivalent rigid body dynamic model is as follows:

[0085] ;

[0086] in, Let the joint position vector be... For joint velocity vectors, The joint acceleration vector, The inertia matrix, The Coriolis force matrix, The gravity vector For the lumped disturbance torque that includes external disturbances, This represents the actual output torque of the actuator.

[0087] The mathematical expression for the actual output torque of the actuator is:

[0088] ;

[0089] in, The ideal control torque to be designed, To characterize the unknown time-varying fault matrix of actuator efficiency loss, For unknown actuator bias fault vectors.

[0090] The actual output torque of the actuator is the physical torque that actually acts on the joint of the robotic arm, reflecting the actual driving capability that the actuator can ultimately output after being distorted by fault interference.

[0091] Ideal control torque is the command torque calculated by the control algorithm. It represents the theoretical torque value required to achieve accurate trajectory tracking under the assumption that the actuator is functioning perfectly.

[0092] The unknown time-varying fault matrix consists of multiplicative fault terms used to characterize the operating efficiency of the actuator.

[0093] The actuator bias fault vector is an additive fault term used to characterize the zero-point drift disturbance of the actuator.

[0094] Define state variables and error variables, specifically including:

[0095] The state variables include the first state variable and the second state variable, and the error variables include the first error variable and the second error variable.

[0096] Define joint position vector First state variable Joint velocity vector For the second state variable .

[0097] Obtain the desired trajectory vector of the robot's task. and the first derivative of the expected trajectory vector The first error variable is obtained by subtracting the expected trajectory vector from the first state variable, and the second error variable is obtained by subtracting the first derivative from the second state variable.

[0098] Actuator fault data, model parameter disturbance data, and external interactive disturbance data are combined into a lumped uncertainty function. This serves as the input basis for subsequent adaptive compensation control using continuous time-series feature matrices.

[0099] In this embodiment, firstly, miniature pressure sensors and Hall current sensors are respectively arranged in the hydraulic chambers of the soft actuators and the motor drive ends of each joint of the manipulator. The physical coupling characteristics of the electrical, hydraulic and deformation of the soft actuator are used to construct a body perception network, and multi-dimensional physical quantities reflecting internal deformation and external interaction are collected synchronously to generate real-time signal streams. Then, a sliding time window with a length of 100ms is set to intercept the multi-dimensional signals. The time-domain waveform features with translation invariance are extracted using a wavelet scattering network with preset scale parameters. A continuous time-series feature matrix implicitly representing the nonlinear state of the soft actuator is constructed through channel cascading. Based on the collected physical structure data, an equivalent rigid body dynamic equation including the inertia matrix, Coriolis force matrix and gravity vector is constructed using the rigid body equivalence assumption. At the same time, an actuator fault model structure including an unknown efficiency loss matrix and a bias fault vector is established. The unmodeled dynamic deviation of the soft actuator relative to the rigid body model, external disturbances and parameter perturbations are combined and defined as a lumped uncertainty function. Finally, the first error variable and the second error variable are calculated in combination with the expected trajectory to provide a dynamic model basis and state input for subsequent adaptive compensation control.

[0100] To address the problems of inaccurate nonlinear dynamic modeling and severe interference from native signal noise caused by the infinite degrees of freedom of soft actuators, this invention constructs a body perception network by utilizing the physical coupling relationship between soft actuators. It introduces a wavelet scattering network with displacement invariance and deformation stability to extract time-domain waveform feature vectors, solving the problem of the difficulty in robustly representing flexible deformation characteristics. Furthermore, it constructs nominal dynamic equations as a benchmark using the rigid body equivalence assumption. It unifies and defines the unmodeled dynamics of the soft actuators relative to the rigid body model, unknown actuator fault data, and external interaction interference data into a lumped uncertainty function. This allows the system to provide a standardized input basis for subsequent adaptive control without establishing a complex continuum analytical model. It achieves the goal of accurately solving the model mismatch and fault tolerance problems of soft manipulators under complex interactions by approximating the lumped uncertainty function while maintaining the computational efficiency of the rigid body model.

[0101] In a specific embodiment, step S2 includes the following sub-steps:

[0102] Step S201: Input the continuous temporal feature matrix into the echo state network, and use the reservoir structure of the echo state network to perform high-dimensional mapping on the input continuous temporal feature matrix, and output a high-dimensional feature vector containing the dynamic time-varying information of the robot arm.

[0103] Step S202: Define the high-dimensional feature vector as the feature basis function vector, which includes the Actor basis function vector and the Critic basis function vector.

[0104] To address the problem that traditional static neural networks lack temporal memory capabilities, making it difficult to accurately represent the dynamic time-varying information and hysteresis effects of soft manipulators in unstructured environments, this invention innovatively utilizes the unique reservoir structure of echo state networks to process continuous temporal feature matrices. Through high-dimensional mapping, low-dimensional temporal signals are transformed into high-dimensional features rich in historical states. This breaks through the limitation of traditional reinforcement learning requiring separate feature layer design, directly defining the output reservoir state vector as the feature basis function vector. This architecture design, which replaces the static Gaussian kernel with a dynamic reservoir, cleverly endows the subsequent evaluation and execution networks with the ability to dynamically remember the system's historical states. Without adding complex recursive structures, the echo characteristics of the reservoir can be used to effectively solve the nonlinear time delay problem caused by soft deformation, significantly improving the real-time approximation accuracy and convergence speed of the control algorithm in complex dynamic environments.

[0105] In a specific embodiment, step S3 includes the following sub-steps:

[0106] Step S301: Based on the first error variable, the second error variable, and the ideal control torque, construct the Bellman error. The mathematical expression for the Bellman error is:

[0107] ;

[0108] in, For Bellman error, This is the transpose of the first error variable. As the first error variable, For the transpose of the ideal control torque, It is the first positive definite constant matrix. It is the second positive definite constant matrix.

[0109] The first positive definite constant matrix is ​​the state weight matrix, which determines the system's emphasis on tracking accuracy. The larger the element value in the first positive definite constant matrix, the heavier the system's penalty for position error, prompting the control algorithm to prioritize reducing tracking error.

[0110] The second positive definite constant matrix is ​​the control weight matrix, which determines the degree of restriction on the control energy consumption of the system. The larger the element value in the second positive definite constant matrix, the heavier the penalty for large control torques, prompting the control algorithm to minimize energy output while ensuring accuracy, thus achieving compliant control.

[0111] A reinforcement learning evaluation network is established based on the Critic basis function vector, and the weight estimates of the reinforcement learning evaluation network are calculated.

[0112] The gradient term is constructed based on the Critic basis function vector and the second error variable.

[0113] The reinforcement learning evaluation network is iterated based on the gradient term, Bellman error, and weight estimates to obtain the weight change rate of the reinforcement learning evaluation network. The mathematical expression for the weight change rate is:

[0114] ;

[0115] in, The rate of change of weight. For gradient terms, These are the weight estimates. For Bellman error, The preset learning rate, This is the transpose of the gradient term.

[0116] Step S302: Construct a sliding mode function based on the first error variable and the second error variable. The mathematical expression of the sliding mode function is:

[0117] ;

[0118] in, The value of the sliding mode function. The preset convergence exponent parameter, The preset gain adjustment parameters, As the first error variable, This is the second error variable.

[0119] Step S303: For the dynamic disturbance caused by the sudden failure of the actuator, an online update law for the disturbance estimation parameters is designed using the nonlinear damping term generated by the state vector of the reservoir, and the rate of change of the disturbance estimation parameters is calculated. The mathematical expression for the rate of change of the disturbance estimation parameters is:

[0120] ;

[0121] in, For the rate of change of the disturbance estimation parameters, It is the first positive constant. It is the second positive constant. For the disturbance estimation parameters, As the second error variable, This is point-to-point multiplication. This is a switching function.

[0122] The first positive constant, used as the adaptive learning rate, is a preset positive number used to adjust the sensitivity of the perturbation estimate to the sliding mode function value. The larger the value, the more violently the perturbation estimation parameter reacts to the error and the faster the convergence speed, but too large a value may cause oscillations.

[0123] The second positive constant, as the robust leakage term coefficient, introduces a negative feedback damping term to prevent the disturbance estimation parameter from drifting to infinity over time in the absence of continuous excitation or in the presence of measurement noise, thereby ensuring the boundedness of the disturbance estimation parameter and the robust stability of the system.

[0124] The disturbance estimation parameter is used to estimate the upper bound of the norm of the system's lumped uncertainty in real time. By introducing the disturbance estimation parameter into the sliding mode switching term, it is ensured that the control gain can cover the energy intensity of external disturbances and actuator failures in real time, thereby ensuring the robust stability of the system under non-ideal operating conditions.

[0125] The magnitude of the actual lumped disturbance moment is always limited to the range defined by the disturbance estimation parameters.

[0126] When Bellman error When a transient fluctuation in the system state is detected, the disturbance estimation step size is automatically increased through the update law to achieve rapid online compensation for actuator failures and severe environmental disturbances.

[0127] Step S304: Map the weight estimates to the execution network to obtain the execution network weights. Based on the execution network weights and the perturbation estimation parameters, perform vectorized recombination to construct the lumped uncertainty function. The mathematical expression for the lumped uncertainty function is:

[0128] ;

[0129] in, For lumped uncertainty functions, As the first state variable, For the second state variable, To perform the transpose of network weights, Let Actor be the basis function vector. For the disturbance estimation parameters, This is point-to-point multiplication. For switching functions, This represents the sliding mode function value.

[0130] in, It is a vector composed of the first error variable and the second error variable:

[0131] ;

[0132] Lumped uncertainty function The combined negative impacts of model parameter disturbances, actuator failures, and external interactive disturbances on the system are mapped into a single continuous numerical compensation function, which is then output as the control gain correction value.

[0133] To address the challenge of existing adaptive control systems in simultaneously optimizing long-term energy consumption and suppressing the instantaneous impact of actuator failures, this invention constructs a reinforcement learning evaluation network based on Bellman error to iterate weight estimates in real time. It also creatively utilizes a nonlinear damping term generated from the reservoir state vector to design an online update law for the disturbance estimation parameters, solving the problem of gain adjustment lag in traditional adaptive laws at fault moments. By vectorizing the weight estimates and disturbance estimation parameters to obtain a lumped uncertainty function, the control gain correction value is calculated from this function. This correction value automatically increases the compensation force when Bellman error detects instantaneous fluctuations in the system state, thus achieving rapid online compensation and robust control for actuator failures and severe environmental disturbances.

[0134] In a specific embodiment, step S4 includes the following sub-steps:

[0135] Step S401: Calculate the output control torque based on the lumped uncertainty function and sliding mode function values. The mathematical expression for the output control torque is:

[0136] ;

[0137] in, To output control torque, For lumped uncertainty functions, The first derivative of the desired trajectory vector. The preset positive definite reaching law gain parameters, For the sliding mode approaching law term, This is a nonlinear feedback term.

[0138] The first derivative of the desired trajectory vector serves as a feedforward control term, representing the desired speed information set in the robot's task, and is used to guide the robot to follow the preset motion trajectory.

[0139] The preset positive definite reaching law gain parameter is used as an adjustment parameter. It is a preset positive number used to adjust the speed at which the system state approaches the sliding surface in sliding mode control. The larger the positive definite reaching law gain parameter, the faster the approach speed.

[0140] The sliding mode reaching law term serves as a robust feedback term. It constructs a reaching law using the sliding mode function value, forcing the system's state trajectory to converge and remain on the sliding surface within a finite time, thus ensuring that the tracking error converges to zero.

[0141] The nonlinear feedback term is a dynamic elimination term derived from the derivative of the nonsingular terminal sliding surface. It is used to counteract the nonlinear dynamics of the system itself and prevent control singularities.

[0142] The output control torque is assigned to the ideal control torque, and the actual output torque of the actuator is calculated. The control of the robot is completed by the actual output torque of the actuator.

[0143] To address the issues of control singularities and insufficient robustness under abrupt actuator failures in traditional sliding mode control, this invention constructs a non-singular sliding mode function with a convergence exponent parameter. This eliminates control singularities while ensuring error convergence within a finite time. The execution network weights are calculated online using an adaptive law, and reinforcement learning compensation terms are generated using Actor basis function vector mapping to offset uncertainties in system model parameters. Furthermore, a fault-tolerant compensation term is constructed using nonlinear damping terms generated from perturbation estimation parameters and the reserve pool state vector. Based on the strategy of synthesizing control torque using reinforcement learning compensation terms and fault-tolerant compensation terms, a dual adaptive decoupling compensation for internal parameter perturbations and external actuator failures is achieved, significantly improving the trajectory tracking accuracy and fault tolerance of the robot under extreme conditions.

[0144] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0145] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A method for adaptive control optimization of a robot, characterized by, Includes the following steps: Step S1: Use the robot's body perception network to obtain the real-time time series signal stream of driving pressure and current, convert the real-time time series signal stream into a continuous time series feature matrix, construct an equivalent rigid body dynamic model through physical structure data, and define state variables and error variables. Step S2: Input the continuous temporal feature matrix into the echo state network, output the reserve pool state vector through the reserve pool structure mapping, and define the reserve pool state vector as the feature basis function vector; Step S3: Calculate the Bellman error based on the error variable, construct a reinforcement learning evaluation network using the feature basis function vector, calculate the weight estimate and perturbation estimate parameters of the reinforcement learning evaluation network, obtain the execution network weights from the weight estimate, and obtain the lumped uncertainty function based on the execution network weights and perturbation estimate parameters. Step S4: Calculate the output control torque based on the lumped uncertainty function and sliding mode function value, obtain the actual output torque of the actuator through the output control torque calculation, and complete the control of the robot arm through the actual output torque of the actuator; Step S3 further includes the following sub-steps: Step S304: Map the weight estimates of the reinforcement learning evaluation network to the execution network to obtain the execution network weights. Based on the execution network weights and perturbation estimation parameters, perform vectorized recombination to construct a lumped uncertainty function. The mathematical expression of the lumped uncertainty function is: ； wherein, is a lumped uncertainty function, is a first state variable, is a second state variable, is a transpose of an execution network weight, is an actor base function vector, is a disturbance estimation parameter, is a point-to-point multiplication, is a switching function, is a sliding mode function value; Step S4 further includes the following sub-steps: Step S401: Calculate the output control torque based on the lumped uncertainty function and the sliding mode function value. The mathematical expression for the output control torque is: ； in, To output control torque, For lumped uncertainty functions, The first derivative of the desired trajectory vector. The preset positive definite reaching law gain parameters, For the sliding mode approaching law term, The preset convergence exponent parameter, The preset gain adjustment parameters, It is the second error variable; The output control torque is assigned to the ideal control torque, and the actual output torque of the actuator is calculated. The control of the robot is completed by the actual output torque of the actuator.

2. The adaptive control optimization method for a robotic arm as described in claim 1, characterized in that, Step S1 includes the following sub-steps: Step S101: Arrange sensor groups within the parallel integrated over-constraint structure of each joint of the manipulator, construct a body perception network using the physical coupling relationship between the software actuators, and collect multi-dimensional physical quantities generated by the manipulator's operation and interaction with the outside through the body perception network to generate a real-time time series signal stream. The multi-dimensional physical quantities include the liquid pressure signal inside the software actuator and the drive current signal of the drive motor.

3. The adaptive control optimization method for a robotic arm as described in claim 2, characterized in that, Step S1 further includes the following sub-steps: Step S102, the preset length is A sliding time window is used to slide along the time axis on the real-time time series signal stream with a preset sampling step size. At each sampling moment, multi-dimensional signal data within the current sliding time window range is extracted. The multi-dimensional signal data includes multi-channel hydraulic pressure waveform signals and driving current sequences. Wavelet scattering network is used to extract the time-domain waveform feature vector of the multi-dimensional signal data, and the features of each channel are fused through matrix dimension transformation to construct a continuous time series feature matrix.

4. The adaptive control optimization method for a robotic arm as described in claim 3, characterized in that, Step S1 further includes the following sub-steps: Step S103: Collect the physical structure data of the robot arm, including joint geometry, component mass, moment of inertia, center of mass position, joint velocity vector, and joint position vector; The inertia matrix is ​​calculated based on the component mass, moment of inertia, and center of mass position. The gravity vector is calculated based on the component mass, joint position vector, and gravitational acceleration constant. The Coriolis force matrix is ​​calculated based on the partial derivative of the inertia matrix with respect to the joint position vector and the joint velocity vector. Based on physical structure data, an equivalent rigid body dynamic model of the manipulator is constructed using the rigid body equivalence assumption. The mathematical expression of the equivalent rigid body dynamic model is as follows: ； in, Let the joint position vector be... For joint velocity vectors, The joint acceleration vector, The inertia matrix, The Coriolis force matrix, The gravity vector For lumped disturbance torque, This refers to the actual output torque of the actuator. The mathematical expression for the actual output torque of the actuator is: ； in, For ideal control torque, For an unknown time-varying fault matrix, Set the actuator bias fault vector; Define state variables and error variables, specifically including: The state variables include a first state variable and a second state variable, and the error variables include a first error variable and a second error variable; Define the joint position vector as the first state variable and the joint velocity vector as the second state variable; Obtain the desired trajectory vector and the first derivative of the desired trajectory vector as set for the robot's operation task. Subtract the desired trajectory vector from the first state variable to obtain the first error variable. Subtract the first derivative from the second state variable to obtain the second error variable.

5. The adaptive control optimization method for a robotic arm as described in claim 4, characterized in that, Step S2 includes the following sub-steps: Step S201: Input the continuous temporal feature matrix into the echo state network, and use the reservoir structure of the echo state network to perform high-dimensional mapping on the input continuous temporal feature matrix to output the reservoir state vector. Step S202: Define the state vector of the reserve pool as a feature basis function vector, which includes the Actor basis function vector and the Critic basis function vector.

6. The adaptive control optimization method for a robotic arm as described in claim 5, characterized in that, Step S3 includes the following sub-steps: Step S301: Based on the first error variable, the second error variable, and the ideal control torque, calculate the Bellman error. The mathematical expression for the Bellman error is: ； in, For Bellman error, This is the transpose of the first error variable. As the first error variable, For the transpose of the ideal control torque, It is the first positive definite constant matrix. It is the second positive definite constant matrix; A reinforcement learning evaluation network is established based on the Critic basis function vector, and the weight estimates of the reinforcement learning evaluation network are calculated. Construct a gradient term based on the Critic basis function vector and the second error variable; The reinforcement learning evaluation network is iterated based on the gradient term, Bellman error, and weight estimates to obtain the weight change rate of the reinforcement learning evaluation network. The mathematical expression for the weight change rate is: ； in, The rate of change of weight. For gradient terms, These are the weight estimates. For Bellman error, The preset learning rate, This is the transpose of the gradient term.

7. The adaptive control optimization method for a robotic arm as described in claim 6, characterized in that, Step S3 further includes the following sub-steps: Step S302: Construct a sliding mode function based on the first error variable and the second error variable. The mathematical expression of the sliding mode function is: ； in, The value of the sliding mode function. The preset convergence exponent parameter, The preset gain adjustment parameters, As the first error variable, This is the second error variable.

8. The adaptive control optimization method for a robotic arm as described in claim 7, characterized in that, Step S3 further includes the following sub-steps: Step S303: For the dynamic disturbance caused by the sudden failure of the actuator, an online update law for the disturbance estimation parameters is designed using the nonlinear damping term generated by the state vector of the reservoir, and the rate of change of the disturbance estimation parameters is calculated. The mathematical expression for the rate of change of the disturbance estimation parameters is: ； in, For the rate of change of the disturbance estimation parameters, It is the first positive constant. It is the second positive constant. For the disturbance estimation parameters, As the second error variable, This is point-to-point multiplication. For switching functions, This represents the sliding mode function value.

Images