An uncertainty inversion technique using two-way neural network for parameter identification of robot arms

Abstract

Due to structural complexity of robot arms, constraints of experiments, especially uncertainty of design parameters, numerical models for dynamics analysis of robot arms can produce erroneous results, which can seriously affect the performance of the designed robot arms. Reliable parameter uncertainty identification for robot arms becomes important. The current methods for uncertainty analysis have double-layered processes, in which the inner layer is for uncertainty propagation and the outer layer is an iterative optimization process. Such a nested double-layered approach limits computational efficiency. This work proposes a novel inverse method for parameter uncertainty identification using a two-way neural network. First, an element (FE) model of a robot arm is established and validated experimentally. Sensitivity analysis is then conducted using the FE model to determine a set of major parameters to be identified. A two-way neural network is next established, and the explicit formulae of direct weight inversion (DWI) use to inverse these parameters of the robot arm. Finally, the inverse result is validated by experiments. Our study show that the present inverse method can greatly improve the computational efficiency. It provides a new avenue to tackle complex inverse problems in engineering and sciences.