Integral operator for chattering reduction in the training of a neural network in the control of a flexible robot

This thesis work aims to solve the problem of trajectory tracking and chattering attenuation at the end effector of the link of a flexible manipulator robot. Using the Generalized Proportional Integral (GPI) control technique in conjunction with a controller based on Feedforward Neural Network (FNN) trained online, to perform trajectory tracking and reduce chattering at the end effector as a result of flexible dynamics. The controller proposed in this research is divided into a linear control part and a data-driven controller to control the nonlinear dynamics of the system. The linear controller is based on a GPI controller, which uses integral reconstructors so that only the angular position of the manipulator is required to be measured to control the rigid dynamics of the system. The Neural Network based controller is known for its ability to approximate the dynamics of a system. In this case, two FNN trained online are used to approximate the dynamics of the flexible link. Likewise, an integral to the operator of the FNN adaptation algorithm is implemented to minimize the chattering of the end effector. The experimental platform to be used has a single actuator, which moves a single flexible link. The use of a single actuator complicates the control of the angular position of the end effector since the flexible dynamics cause i-th order vibration modes and the actuator has limited bandwidth to attenuate chattering. The advantage of using a single actuator to solve the control problem is that a low material cost is maintained. Increasing another actuator implies increasing its electronic components for control as well as the power stage. The system has a single input, which is the control signal that drives the actuator. On the other hand, the system has two outputs, an angular position signal of the link and an acceleration signal of the end effector of the link. The angular position signal is used to control the position of the link and perform trajectory tracking. While the acceleration signal of the end effector is used to control and compensate for unwanted accelerations caused by external oscillations and disturbances. Throughout this work, the changes made to the experimental platform are presented, as well as different NN to model the system. The contribution of this thesis work is presented, which is the implementation of an integral in the operator of the NN weight adaptation algorithm. The experimental results of a Feedforward Neural Network Module (FNNM) controller and a FNNM controller with the operator's integral are compared. Likewise, to obtain the position of the end effector in space, the results of the motion capture of the flexible manipulator robot of a link with each of the controllers are presented. Finally, the results obtained from the implementation of both controllers are compared and analyzed.

https://orcid.org/0000-0002-3192-1238