Event Detection and Characterization During Autonomous Interplanetary Navigation

A prototype autonomous/adaptive interplanetary navigation system employing neural networks and genetic algorithms is introduced. This system consists of a near real-time autonomous monitoring component and an offline adaptive modeling component. Unexpected dynamic or measurement events trigger 1) alerts to navigators and 2) recommendations for modifying the appropriate model parameters to improve filtering. The autonomous component analyzes tracking measurement residuals to first detect and then characterize the unexpected event. Once an environment change has been detected and characterized, the adaptive modeling component then modifies the necessary model parameters to bring the tracking filter back into optimal operation. The autonomous monitor employs a hierarchical mixture-of-experts model where the experts are extended Kalman filters organized into banks regulated by two levels of single-layer neural networks. The autonomous monitor is the focus of this presentation and demonstrates the ability to detect successfully the occurrence and to differentiate between the characteristics of unexpected discrete velocity changes and continuous dynamic changes. These environment changes are represented by an unmodeled impulsive maneuver and by solar radiation pressure surface mismodeling, respectively. The robust decision-making capability of this approach is further demonstrated by successfully characterizing three successive environment changes. All experiments were performed on recorded Mars Pathfinder two-way Doppler data from the period of 4 February 1997 to 17 April 1997.