Adaptive PID Control of a Nonlinear Servomechanism Using Recurrent Neural Networks

Recent progress in the theory of neural networks played a major role in the development of new tools and techniques for modelling, identification and control of complex nonlinear dynamic systems. Intelligent control, with a special focus on neuro-control has been used successfully to solve difficult real control problems which are nonlinear, noisy and relatively complex. This is due to the fact that neural networks have an inherent ability to learn from input-output data and approximate an arbitrarily nonlinear function well. The inclusion of semi-linear sigmoid activation functions offers nonlinear mapping ability for solving highly nonlinear control problems (Omatu et al., 1995). A large number of identification and control structures have been proposed on the basis of neural networks in recent years (Jain & Medsker, 1999). Most of the developed neural networks use a feed-forward structure along with the back-propagation training algorithm. Recently, more research interest is given to recurrent networks with special application to dynamic systems. A Recurrent Neural Network (RNN) exhibits internal memory due to its feedback structure, which gives the network the possibility of retaining information to be used later. By their inherent characteristic of memorizing past information, for long or shortterm periods, RNNs are good candidates for nonlinear system identification and control (Narendra & Pathasarathy, 1990). Although control theory has made great advances in the last few decades, which has led to many sophisticated control schemes, PID control remains the most popular type of control being used in industry today. This popularity is partly due to the fact that PID controllers have simple structures that are easily implemented. On-line self-tuning PID controller offer an advantage for plants that have uncertain dynamics, time varying parameters, and nonlinearities. Recently a lot of attentions have been focused on neural based PID controller, and many efforts have been done to investigate different aspects of deploying neural networks in the area of adaptive PID control (Puskorius & Feldkamp, 1993), (Saikalis, 2001) The concept of adaptive PID control was introduced to compensate the drawbacks of the fixed-gains PID controller. For example, if the operating point of a process is changed due to disturbances, there is a need to adjust the controller parameters manually in order to keep