Neural Network Modeling Of Hysteresis Characteristics Of Flexible Joint For Lightweight Industrial Robot

The lightweight industrial robot is increasingly applied in the 3C industry,welding,medical,parts assembly,and unmanned retail because of its safety,simplicity,and adaptability.Flexible joints ensure the security of lightweight industrial robots operating in unknown environments.However,the flexible joint actuators will result in hysteresis characteristics during the reciprocating motion of the robot,which will affect the dynamics performance,positioning accuracy and smoothness of the motion of the robot.For increasing the accuracy of the robot transmission and model-based compensated control,high-precision modeling of the hysteresis characteristics of flexible joint actuators is a crucial prerequisite.Traditional modeling approaches for hysteresis characteristics of flexible joint actuators need data from dual encoders on the motor and load sides in addition to load-side torque sensors.However,torque sensors are absent from low-cost lightweight industrial robots with flexible joints.Due to the lack of sensors,the nonlinear relationship between torque and torsion angle cannot be used to describe the hysteresis properties of the flexible joint.The research done in this paper:1.A method proposed to indirectly describe the hysteresis characteristics of a low-cost flexible joint actuator.The hysteresis characteristics of the joint actuator are described using the motor end torsion angle and the embodied load torque variation motor drive current.The difference between the set value and the measured value of the output angle of the motor end is used to obtain the motor end torsion angle.2.In this paper,the non-linear relationship between motor drive current and motor end torsion angle is used to indirectly describe the hysteresis characteristics exhibited by low-cost flexible joint actuator.The following two modeling approaches are proposed for the strong nonlinear hysteresis characteristics exhibited by the low-cost flexible joint actuator.(1)Forward and reverse trajectory features fused GRU neural network for hysteresis characteristics modeling of jointThe relationship between drive current and torsion angle is employed to describe the hysteresis characteristics of flexible joints,the forward and reverse trajectory features-based GRU neural network hysteresis model is proposed.The hysteresis characteristics in low-cost flexible joints are analyzed in the forward and reverse trajectory features.The Kalman filter-based current increments are employed to extract the features of the forward and reverse trajectory to describe the multi-valued characteristics exhibited by the forward and reverse trajectory features in the current-torsion angle hysteresis characteristics.A lightweight industrial robot experimental platform is built to validate the hysteresis data obtained under the excitation of sinusoidal input signals with different frequency decay.Compared with the GRU neural network hysteresis model,the experimental results show that the proposed hysteresis model has better prediction capability and high model accuracy.(2)Extended feedback correcting-based GRU neural network hysteresis modelBased on the hysteresis characteristics between motor drive current and motor torsion angle,the extended feedback correcting-based GRU neural network hysteresis model is proposed.The feedback correction structure in the predictive control is extended by using multiple historical prediction errors to form a compensation quantity that is fed back to the model to correct the new output value.A Q-learning algorithm is constructed to optimize the compensation coefficients for the problem of mismatch between the neural network weights and the convergence rate of the compensation coefficients.The lightweight industrial robot experimental platform is built to validate the hysteresis data collected under the excitation of sinusoidal superimposed signals with different frequencies,and the extended feedback correcting-based GRU neural network hysteresis model has higher model accuracy compared with the hysteresis model before the extended feedback correction.Among the two hysteresis models proposed in this paper,the extended feedback correcting-based GRU neural network hysteresis model has better generalization ability and higher model accuracy,and can effectively describe the hysteresis characteristics exhibited in low-cost flexible joints.Two kinds of hysteresis models are proposed for different operation speed and precision.