Automated algorithm to detect changes in geostationary satellite’s configuration and cross- tagging

Using broadband and color photometry, analysts can evaluate satellite operational status and affirm its identity. The process of ingesting photometry data and deriving satellite physical characteristics can be undertaken by analysts in batch mode, meaning using an entire batch of data at the conclusion of each observation night, or by automated algorithms in a batch or on-line mode of operation. In an on-line mode, the assessment is generated with each new data point. Tools used for detecting change to satellite’s status or identity, whether performed with a human in the loop or automated algorithms, are generally not built to detect with minimum latency and traceable confidence intervals. To alleviate those deficiencies, we investigate the use of Hidden Markov Models (HMM), in a Bayesian Network framework, to infer the hidden state (changed or unchanged) of a three-axis stabilized geostationary satellite using broadband and color photometry. Unlike frequentist statistics which exploit only the stationary statistics of the observables in the database, HMM also makes use of the temporal pattern of the observables as well. The algorithm also operates in “learning” mode to gradually optimize the HMM. Our technique is designed to operate with or without color data. The version that ingests both panchromatic and color data can accommodate gaps in color photometry data. That attribute is important because while color indices, e.g. Johnson R and B, enhance the quality of the belief (probability) of a hidden state, in real world situations, flux data is collected sporadically in an untasked collect, and color data may be absent. Fluxes are measured with experimental error whose effect on the algorithm will be studied. Photometry data in the AFRL’s Geo Color Photometry Catalog (GCPC) and Geo Observations with Latitudinal Diversity Simultaneously (GOLDS) data sets are used to simulate configuration changes and identity cross-tags. The algorithm is tested against simulated sequences of observed magnitudes, mimicking the cadence of untasked ground sensors, configuration changes and cross tags of in-cluster satellites. We would like to show that the on-line algorithm can detect change; sometimes right after the first post-change data point is analyzed. We would like to show the unsupervised “learning” capability that allows the HMM to evolve with time without user’s assistance. For example, the users are not required to “label” the true state of the data points.