Adaptive interplanetary orbit determination
MetadataShow full item record
This work documents the development of a real-time interplanetary orbit determination monitoring algorithm for detecting and identifying changes in the spacecraft dynamic and measurement environments. The algorithm may either be utilized in a stand-alone fashion as a spacecraft monitor and hypothesis tester by navigators or may serve as a component in an autonomous adaptive orbit determination architecture. In either application, the monitoring algorithm serves to identify the orbit determination filter parameters to be modified by an offline process to restore the operational model accuracy when the spacecraft environment changes unexpectedly. The monitoring algorithm utilizes a hierarchical mixture-of-experts to regulate a multilevel bank organization of extended Kalman filters. Banks of filters operate on the hierarchy top-level and are composed of filters with configurations representative of a specific environment change called a macromode. Fine differences, or micromodes, within the macromodes are represented by individual filter con- figurations. Regulation is provided by two levels of single-layer neural networks called gating networks. A single top-level gating network regulates the weighting among macromodes and each bank uses a gating network to regulate member filters internally. Experiments are conducted on the Mars Pathfinder cruise trajectory environment using range and Doppler data from the Deep Space Network. The experiments investigate the ability of the hierarchical mixture-of-experts to identify three environment macromodes: (1) unmodeled impulsive maneuvers, (2) changes in the solar radiation pressure dynamics, and (3) changes in the measurement noise strength. Two methods of initializing the gating networks are examined in each experiment. One method gives the neurons associated with all filters equivalent synaptic weight. The other method places greater weight on the operational filter initially believed to model the spacecraft environment. The results will show that the equal synaptic weight initialization method is superior to the one favoring the operational filter and that processing range and Doppler data together is superior to processing Doppler data alone. When processing range and Doppler with an equally initialized hierarchy, all three macromodes are definitively identified by the top-level gating network weights. Additionally, in the case of multiple successive macromode changes, the hierarchy is generally able to recover from one macromode and identify a change to another macromode.