Title: Autonomous On-board Processing for Sensor Systems: Initial Fault Tolerance and Autonomy Results
Author: Matthew French
Organization: USC / ISI
Co-Authors: John Paul Walters, Mark Bucciero

The A-OPSS project fuses high performance reconfigurable processors with emerging fault-tolerance and autonomous processing techniques in order to realize high performance, autonomous sensor systems. By enabling fault-tolerant use of high performance on-board processing the utility of sensor systems will be greatly enhanced, achieving 10-100x decrease in processing time, which directly translates into more science experiments conducted per day and a more thorough, timely analysis of captured data. This research also addresses the ability to quickly react and adapt processing or mission objectives in real-time, by combining autonomous agents with reconfigurable computing techniques. Using A-OPSS, satellites, airborne or ground sensors will be able to perform high performance, fault tolerant computation, and develop situational awareness about their operating environment and tune or adapt the application algorithm such that they return the most useful and significant data to human and automated decision support systems.

This paper focuses on the second yearís efforts which revolve around validating the radiation hardening by software techniques that were developed in the projectís first year, and developing an autonomous framework for remote hyperspectral imaging applications. To validate the fault tolerance, a software fault injector, the Memory Sentinel and Injection System (MSIS) was developed, which can inject faults into 99% of the sensitive bits in the Xilinx embedded PowerPC 405. The fault tolerance tech-niques were then tested using both MSIS and radiation testing. Our results show that 94% of all faults were recoverable without reset, while only adding a 2% overhead, freeing up substantial computational resources for high performance scientific applications. We then provide an overview of hyperspectral applications mapped to the NASA GSFC SpaceCube 1.0 using A-OPSS techniques and show how it can use the additional computational resources to perform autonomous look-ahead and decision making operations.