Increasing autonomy of precision spacecraft using neural network adaptive control

Abstract

In recent years, there has been a significant interest in the use of adaptive methods for controlling structures in high precision aerospace applications. This is because adaptive methods offer the potential to autonomously adjust to system characteristics different from those modeled or seen in qualification testing. This is especially true of spacecraft, which are generally tested in a 1-g environment. Despite extensive research, it remains extremely difficult to predict on-orbit 0-g behavior. In addition, system dynamics often tend to be time varying. This can take the form of slow changes due to degradation of materials and aging of the spacecraft or sudden failures such as the loss of a sensor or actuator. These events become increasingly likely as spacecraft become more and more complex. By decreasing modeling and testing requirements, lowering operations and maintenance activities that require human intervention, and increasing reliability, adaptive methods have the potential to significantly reduce cost and increase performance of these systems. One class of adaptive control methods are those which utilize artificial neural networks. The use of neural networks has become increasingly mature in a number of areas such as image processing and speech recognition. However, despite a number of publications on the subject, very few instances exist where neural networks have actually been used in control and in particular, structural control applications. The United States Air Force Research Laboratory (AFRL) is currently engaged in advancing adaptive neural control technologies for application to precision space systems. This paper gives an overview of several past and current ground and space based adaptive neural control experiments.