A parallel adaptive neural network control system, with application to real-time control of a servomechanism with asymmetrical loading

Servomechanisms with nonlinear dynamics appear in many applications [1]-[4]. In previous work in [4], we had considered the application of a nonlinear coiitrol strategy based on neural networks to address the position control problein in such servomechanisms. The anuerse i ioii lmear coniroller usang neural iteiworks described there was shown to be capable of providing excellent closed-loop control for several classes of servomechanisms with fairly severe nonlinearities. The work in [4] also included experimental results for real-time control of a pilot-scale nonlinear position control system. However, the technique requires an accurate approximation to be provided by the neural network realisation of the nonlinear open-loop system. This typically requires a large number of hidden nodes for real-life syateiiis (refer to real-time experimental results in [4]), with the attendaiit disaclvantage that large training data frames and long training times are required. A further drawback of the controller of [4] is that it does not provide a built-in capability to adapt to changes in the system to be controlled.