Error minimising gradients for improving cerebellar model articulation controller performance

Abstract

In motion control applications where the desired trajectory velocity exceeds an actuator’s maximum velocity limitations, large position errors will occur between the desired and actual trajectory responses. In these situations standard control approaches cannot predict the output saturation of the actuator and thus the associated error summation cannot be minimised.An adaptive feedforward control solution such as the Cerebellar Model Articulation Controller (CMAC) is able to provide an inherent level of prediction for these situations, moving the system output in the direction of the excessive desired velocity before actuator saturation occurs. However the pre-empting level of a CMAC is not adaptive, and thus the optimal point in time to start moving the system output in the direction of the excessive desired velocity remains unsolved. While the CMAC can adaptively minimise an actuator’s position error, the minimisation of the summation of error over time created by the divergence of the desired and actual trajectory responses requires an additional adaptive level of control.This thesis presents an improved method of training CMACs to minimise the summation of error over time created when the desired trajectory velocity exceeds the actuator’s maximum velocity limitations. This improved method called the Error Minimising Gradient Controller (EMGC) is able to adaptively modify a CMAC’s training signal so that the CMAC will start to move the output of the system in the direction of the excessive desired velocity with an optimised pre-empting level.The EMGC was originally created to minimise the loss of linguistic information conveyed through an actuated series of concatenated hand sign gestures reproducing deafblind sign language. The EMGC concept however is able to be implemented on any system where the error summation associated with excessive desired velocities needs to be minimised, with the EMGC producing an improved output approximation over using a CMAC alone.In this thesis, the EMGC was tested and benchmarked against a feedforward / feedback combined controller using a CMAC and PID controller. The EMGC was tested on an air-muscle actuator for a variety of situations comprising of a position discontinuity in a continuous desired trajectory. Tested situations included various discontinuity magnitudes together with varying approach and departure gradient profiles.Testing demonstrated that the addition of an EMGC can reduce a situation’s error summation magnitude if the base CMAC controller has not already provided a prior enough pre-empting output in the direction of the situation. The addition of an EMGC to a CMAC produces an improved approximation of reproduced motion trajectories, not only minimising position error for a single sampling instance, but also over time for periodic signals.