Analysis And Modeling Of Nonlinear Hysteresis For Harmonic Reducer In Industrial Robots

Harmonic transmission,widely used in aerospace,mechanical equipment and robotics,is a kind of transmission technology developed with space technology.Harmonic drive has the advantages of small size,compact structure,light weight,large torque,high transmission ratio,close to zero clearance and good coaxial assembly,and it is the core component of industrial robots to realize motion.The harmonic reducer is mainly composed of three basic components: wave generator,flexspline and circular spline.The elastic deformation of the flexspline realizes the transmission of motion.The deformation of the flexspline,the nonlinear friction in the transmission,and the manufacturing and installation errors of the various components lead to the hysteresis of the harmonic drive,reflecting the hysteresis characteristic of the output torque and the angle between the input and output shaft.Hysteresis is an important factor affecting the repeatability of industrial robots during reciprocating motion.The analysis and modeling of the hysteresis characteristics of the harmonic drive is important for realizing high precision control of industrial robots.In view of the special non-smooth nonlinear hysteresis characteristics exhibited by harmonic drives in flexible joints of industrial robots,two modeling methods are proposed as follows shown:(1)A hybrid hysteresis model consisting of SDH model in series with neural networkThe SDH model is used as the pre-model of which hysteresis characteristics between the input and output signals is similar to the hysteresis characteristics of the harmonic drive;the dynamic RBF neural network which can describe the nonlinearity is used as the post-model to make the hybrid hysteresis model approach the hysteresis characteristic of the harmonic drive with high precision.The data obtained under different frequency input signals and different load are modeled.Compared with the classical RBF neural network model and SDH model,the experimental(modeling and experimenting)results verify that the constructed hybrid hysteresis model has high precision and strong adaptability.(2)Preprocessing-based dynamic RBF neural network hysteresis model1)the input signal preprocessing to make the processed signal have a hysteresis-like characteristic with the input signal;2)fully utilizing dynamic RBF neural network to achieve high-precision approximation from hysteresis-like to hysteresis characteristics of harmonic drive.The data obtained under different experimental conditions are modeled and verified under different input signals and loads,the verification accuracy of the neuralnetwork hysteresis hybrid model is much higher than that of the classical neural network model with the same input accuracy,which proves the effectiveness and adaptability of the constructed preprocessing-based dynamic RBF neural network hysteresis model.The mean square error(MSE)of the modeling and verification errors of the two models proposed is basically the same,but the constructed dynamic pre-processing based dynamic RBF neural network hysteresis model is simpler and easier to apply than the SDH hybrid hysteresis model.