Neural Network NARMA Control of a Gyroscopic Inverted Pendulum

The objective herein is to demonstrate the feasibility of a real-time digital control of an inverted pendulum for modeling and control, with emphasis on nonlinear auto regressive moving average based neural network (NARMA). The plant of interest is a novel Gyroscopic Inverted Pendulum (GIP) device that is nonlinear and open-loop unstable. The GIP balances a pendulum on its free knife-edge base using a flywheel driven by DC motor fixated on the top. In this application, an indirect data-based technique is taken, where a model of the plant is identified on the basis of input-output data and then used in the model-based design of a NARMA controller. The plant under digital PID control with I-adaptation provides initial stability at the beginning of a single layer NARMA neural network training. NARMA models of increasing complexity are used successively to generate input-output data for the training of multilayered NARMA models. In using a NARMA neural network the control laws are nonlinear and online adaptation of the model is possible to capture un-modeled or time-varying dynamics. Such an environment provides for experimentation, data collection, system identification and real time control strategy implementation.