Machine Learning for Complex Spacecraft Anomaly Prediction

GMSO-2 Machine Learning

Emergent plays a key role as a subcontractor to KBR on NASA GSFC’s Ground Systems and Mission Operations (GSMO) 2 contract in the areas of software and system engineering, system administration and project configuration management at the infrastructure, ground and flight mission, and advanced R&D organizational levels. Emergent also works with KBR to infuse innovative software solutions into the NASA Space Science Mission Operations (SSMO) Projects.

One such project is implementing machine learning algorithms for complex spacecraft anomaly prediction. Led by Eduardo Valente, project manager and principal software engineer, and Brett Carver, machine learning expert and senior software engineer, we are using machine learning to duplicate and extend the capability of the existing generation of the Fermi spacecraft “Antsy Meter” value.  The Antsy Meter value represents the spacecraft’s power system’s stress level, initially developed by KBR’s mission operations subject matter expert in anomaly prediction and resolution. Due to a mechanism failure in one of the solar panels, the spacecraft’s instruments operate at lower power levels while continuing to collect data. The health of the power system is calculated as an Antsy Meter value, which ultimately indicates if the spacecraft battery system is maintaining sufficient charge for spacecraft activities during orbit nights and days. It is calculated by hand once per day using eight different telemetry derived variables.

Emergent’s team exceeded expectations by designing a machine learning algorithm that analyzes the telemetry data from the Fermi spacecraft, in this instance one year’s worth of data. A significant milestone of the project was achieved when it was determined that fewer telemetry variables are needed to predict the Antsy Meter value.  “We reduced the number of variables from eight to four with a 98.5% confidence level,” said Brett Carver. Due to the large amount of data required to train the model, expanding the analyses over larger sets of data could be invaluable to the progression of the algorithm.

The machine learning algorithm shows great promise for future applications in spacecraft ground and flight systems. “This short-term project can serve as a springboard into the next set of tasks, should the customer choose to continue the project. Machine learning is more than just a set of algorithms for some type of decision making or pattern recognition. Successfully applying machine learning requires understanding the data for its virtues and flaws. That is to apply a rigorous process of analysis that brings into focus those elements that are important and separates those that are peripheral or even flawed. The Antsy meter turned to machine learning served as an example of that process, engagement of domain area expertise, and SSMO’s telemetry holdings. We hope to extend this approach to other satellites and telemetry sources. And in doing so create a variety of “meters” that will enhance the understanding of the spacecraft operations and mission management.” said Eduardo Valente.