A Master of Science thesis in Mechatronics Engineering submitted by Amir Hossein Jafari entitled, "Adaptive Neuro-Observer for Plastic Drive Systems," December 2011. Available are both soft and hard copies of the thesis.
This work presents the design, implementation, testing, and validation of a Neural Network based observer for a real hardware elastic drive system. Diagonal recurrent neural network architecture and two types of controllers are proposed. Critical states of the elastic drive system are estimated online and offline. In this design, the proposed observer is of the hybrid type, composed of a conventional linear observer, augmented by a neural network (NN). All software implementation has been performed in Matlab/Simulink environment. The dynamic modeling of the two-mass model platform has been derived using Newtonian methodologies. This modeling includes the mechanical structure, actuator, sensors dynamics, and friction modeling. The motor drive control system consists of two cascaded loops. A fast inner current control is used to control the motor torque. This hardware controller is cascaded to the outer speed control loop and gains are optimized for maximum system performance. In order to get a better performance, the outer loop employs a modified version of a proportional integral controller. The system is driven by two different types of controllers and state variables are estimated first by a linear observer, then by the NN observer. The first observer is a traditional Luenberger observer that has been implemented online on the real system as well as in the simulation. The second observer is a neural network based observer that consists of a linear observer cascaded with a diagonal recurrent neural network that learns over time and estimates the nonlinear parameters of the system. Compare to the linear observer, the proposed neural network based observer shows a significant enhancement in state estimation. The integration of the neural network with a linear observer has shown better performance that the linear one due to the unmodeled dynamics that the linear observer cannot cope with. The training of the neural network has proven to be critical to find the friction forces that are invisible in the linear design.