This paper proposes a scheme whereby the entire
conventional mathematical model that relies on the
machine parameters is replaced by a universal function
approximator, either a neural network model or a fuzzy
logic model. The control models proposed do not use any
machine parameters. The control operates entirely in the
two-phase domain. Actual current control is done by
transforming the control output, the flux and the torque
current back to the actual three phase currents.
I1 THE NEURAL NETWORK CONTROL MODEL
The neural network control model basically takes the
speed feedback compare to the command speed to
compute the cost function-the speed error. This speed
error is then used to adjust the neural network weights in
order to achieve the optimum performance. Conventional
neural network model normally uses a multilayer neural
network trained by the backpropagation algorithm
[4][5][6]. However, such models are often complicated
and slow in convergence. This paper proposed a very
simple neural network structure with a competitive
training algorithm. The network structure consists of only
an input layer and an output layer as shown in fig I .