Prediction of Thruster Performance in Hall Thrusters Using Neural Network with Auto Encoder

Abstract

A thrust prediction system using a neural network for controlling Hall thrusters automatically is under development. This network has described the time variation of discharge current within 1% error, but calculation of the current was very cumbersome (2100 s) and overfitting occurred. In order to reduce the calculation cost and to prevent the neural network from overfitting, we have adopted a stacked auto encoder and optimized the network model using a genetic algorithm. The calculation time was reduced from 2100 s to 100 s without overfitting. The present system cannot yet describe hysteresis of discharge current; this will be addressed in future work.