Machine Learning and Instrument Autonomy Group

Principal Investigator

Jack Lightholder
Phone: (626) 710-3246
Email: jack.a.lightholder@jpl.nasa.gov


      Spacecraft mission operations requires the management of complex systems, with short response times, and limited margin for error. Machine learning technologies can provide focus of attention for operators, as well as identify relationships in data not immediately obvious to operators. These technologies require data to be prepared for ingestion, which is traditionally a significant component of a data scientist’s job. This effort requires reaching out to subject matter experts to learn everything from meaningful featurization to valid binning and interpolation strategies. Conducting proper, and comprehensive, data preparation across a spacecraft mission system would require the data scientist to talk with hundreds of subject matter experts and manage featurization of over 100,000 features. This approach is not tenable for small scale data science projects, and traditionally leads to a priori assumptions by the team as to which data to exclude from the analysis. Since subject matter expert knowledge is divided across spacecraft subsystem teams, contacting experts for each data set becomes challenging. This leads to high level assumptions about the data being applied to all features.

      The Operational Recommendations for Capturing History and Infusing Data Science (ORCHIDS) system provides a framework for parsing, featurization and metadata tracking across a spacecraft ground data system, allowing teams to store data in a format which is analysis ready. ORCHIDS tracks data from across the mission timeline. These include planning artifacts, commanding, onboard EVRs, spacecraft telemetry, science data products, vehicle parameter state, downlink route, ground data system logs and team schedules. These data sets, which are traditionally siloed, are then time harmonized, featurized and stored in their sparse matrix form. The resultant “Analysis Ready Data” artifacts are ingested into the Institution’s flight engineering data archive for eventual retrieval by data scientists. Accompanying this data set, are subject matter expert curated metadata which informs data science end users on the correct usage of the data. Data scientists can then query data in the raw featurized state, or apply binning/interpolation/standardization/normalization which follows the best-practices defined by the subject-matter experts. This flexibility in data access follows the traditional science data processing levels schema, a core component of science data access. This flexibility provides the end user with the correct amount of post-processing for their application.

      The ORCHIDS system provides significant reduction in time overhead for data scientists conducting data access and preparation of spacecraft mission system data. The ability to perform these data preparation strategies at the mission level reduces the need for each data scientist to write custom code to prepare data for analysis. This error prone process is time consuming and cannot capture comprehensive subject matter expert knowledge in large spacecraft mission systems. ORCHIDS is currently being demonstrated across in-flight mission systems, and being assessed for infusion into upcoming mission ground data systems.