Page 991 - The Mechatronics Handbook

FIGURE 33.18 LHM position.

FIGURE 33.19 LHM velocity.

inside a neural network [50]. This section deals with the control of an electrohydraulic axis using a neural
controller that has a widely spread structure, namely, multilayer perceptron (MLP).

Neural Control Techniques
Learning Based on Mimic
Inspired from biological systems, learning by mimic is applied to control systems. A supervised neural
network can mimic the behavior of another system. A first method to develop a neural controller is to
replicate a human controller. The neural controller tries to behave like the human operator. Neural
training means learning the correspondence between the information received by the human operator
and the control input (Fig. 33.20).
Inverse Learning
The purpose of inverse control is to control a system by using its inverse dynamic. In this case, the neural
network receives the output of the system as input and has the input of the system as output. The system
works in open loop and has to be in the region where the controller will operate. Inverse learning
(Fig. 33.21) is an indirect approach to minimize the network output error instead of the overall system
error.
Specialized Inverse Learning
According to Psaltis, who proposed in 1988 a specialized inverse learning, the neural network should be
trained online in order to minimize the control error e y = r - y (see Fig. 33.22).

©2002 CRC Press LLC