Indirect field oriented control of an induction motor implemented with an artificial neural network

Presented in this dissertation is an artificial neural network that is designed to perform indirect field orientation control for an induction motor speed control system. A multi-layer feedforward network with the hypertangent sigmoid transfer function is chosen as the controller that generates current command signals. The neural network inputs are the induction motor speed, synchronous frame q-axis current and a delayed sample of this current, the stationary frame q and d axis currents are the current control command outputs, which are also fed back to the neural network input after one sample delay. The neural network has two neurons at the output layer, ten neurons in the hidden layer, and five input neurons.; It is shown that due to the recurrent structure of the neural network and the excessive output error of the neural network, off-line training is not sufficient for stable operation of the system. An improved training algorithm employing experimental data is proposed to remedy this problem.; The synchronous frame space vector current regulator method is chosen as the current control approach employed for experimental verification of the controller. In this standard current regulator method the zero voltage sequence is applied equally in the beginning and at the end of each sampling interval which results in low ripple current, but it does not guarantee the least current error during the sampling interval. A modified space vector current regulator is proposed in this dissertation that optimizes the zero voltage time interval which reduces the transient current error in the motor stator and improves the output torque tracking capability.; The experimental setup to test the neural network includes a Texas Instruments TMS320C30 floating-point digital signal processing board operating at 33MHz. Interface circuits are designed to measure the current and speed feedback signals and to control the switching of the DC/AC converter.; The neural network calculations are performed on the digital signal processing board. Programming techniques are presented to increase the speed of the computation. Experimental results indicate successful design and implementation of the neural network. A 1ms sampling interval is employed with a total of 330{dollar}mu s{dollar} required for all neural network, current regulator and system calculations. This leaves 67% of the digital signal processor computation capacity available for other purposes as might be required in a commercial drive system. It is also verified that as expected from the simulation results, two stages of training are required for stable system operation employing the neural network. The main limitation of the controller is shown to be the low-speed region of operation {dollar}(<200rpm){dollar}, where the neural network output error causes large output speed error.