Advanced science.  Applied technology.


Machine Learning for Space Data Compression, 10-R8733

Principal Investigators
Daniel Davila
Joshua Anderson
Inclusive Dates 
01/01/17 to 01/01/18


The objective of this research was to investigate the adaptation of machine learning algorithms to the task of space science data compression. This capability is pivotal for reducing the strain on the telemetry links of many science missions that is imposed by a combination of the high (and ever increasing) data volume and low telemetry bandwidths of modern spacecraft instrumentation systems.


Compression was achieved by using a machine learning algorithm to intelligently select the portions of the signal that represented “interesting” data. This approach was motivated by observation that, within most scientific instrumentation data, only a small fraction of the data is non-noise. In this case, the regions of interest corresponded to specie peaks from mass spectrometer data acquisitions. Data from the instrument was used to train a neural network with the ability to accurately segment the data for relevant peaks. The rest of the signal was rejected as noise. The resultant compressed signal was the combination of the extracted peaks and some necessary metadata for establishing where in the original time-of-flight spectrum these peaks occurred.


This effort resulted in an algorithm that performed well on simulated data, correctly extracting 95 percent of the relevant peaks and reducing the signal by a factor of 10:1. Initial work was also performed for determining the feasibility of embedding a neural network onto a space flight field programmable gate array. The successful completion of this project positions SwRI favorably for pursuing immediate leads:

  • The application of machine learning for intelligent data handling for ground and space Internet of things (IoT)
  • The application of machine learning tools to space instrumentation data
  • The deployment of machine learning to hardware-constrained space mission hardware

This project investigated the use of machine learning for intelligent science data selection onboard a spaceflight instrument. Artifacts and results from this project also have direct applicability in other industries that utilize embedded intelligence at the edge, including automotive, energy, and aerospace applications.