Closed-loop control using a backpropagation algorithm: a practicable approach for energy loss minimisation in electrical drives

In general, optimal controls are computed off line and subsequently applied in real time but this approach is impracticable due to lack of robustness with respect to the plant modelling errors and unknown external disturbances. Closed loop versions of these optimal controls could circumvent this problem but are only available in the analytical form for very simple cases, not including minimisation of frictional energy loss in motion control systems, which is the aim of the research programme. The approach suggested by Matieni and Dodds (2009), however, overcomes this obstacle by training an artificial neural network (ANN) to reproduce the optimal control values computed off-line from given states and reference inputs, thereby yielding a closed loop solution.

The purpose of this paper is to present the results of an initial simulation experiment to assess the capability of a Multilayered Perceptron (MLP), in the backpropagation mode, to perform a direct state feedback function, which, to the authors‘ knowledge, is new. A known linear state feedback controller for a double integrator plant is used for this purpose. The control law is used to train the MLP. Then a simulation of the closed loop system formed using this MLP is compared with a simulation of the known linear state feedback control system. The results show that the closed loop step response with the MLP closely follows that of the conventional system