Sabogal, Sebastian
(2021)
Strategies for Selective and Adaptive Resilience in Reconfigurable Space Systems and Apps.
Doctoral Dissertation, University of Pittsburgh.
(Unpublished)
Abstract
Due to ongoing advancements in sensor technology and innovations in spacecraft autonomy enabled by compute-intensive deep-learning (DL) methods, modern spacecraft increasingly require more onboard processing capabilities that address the computational demands required for future space missions. Spacecraft designers are challenged to create dependable, high-performance space computers capable of converting onboard an immense volume of raw sensor data into actionable information that can be used to formulate critical decisions autonomously. Furthermore, this space-computing challenge is further exacerbated with stringent constraints in size, weight, power, and cost (SWaP-C) and dependability requirements due to radiation effects in the harsh environment.
The proliferation of small satellites (SmallSats) has enabled a paradigm for low-SWaP-C missions that frequently employ commercial-off-the-shelf (COTS) devices, including FPGAs and hybrid system-on-chips (SoCs), to improve onboard processing capabilities. These commercial devices have numerous architectural advantages that provide superior performance, energy efficiency, and affordability compared to radiation-hardened (rad-hard) alternatives but are highly susceptible to radiation-induced single-event effects (SEEs) that can impact mission dependability. To improve dependability, hardware-redundancy techniques are frequently employed for SEE mitigation; however, these methods incur a significant overhead that can be impractical for resource-constrained systems and can limit system performance. To create space computers capable of onboard DL, it is essential to create efficient methods for SEE mitigation and dependability evaluation.
In this dissertation research, we propose both selective and adaptive strategies for efficient SEE mitigation in reconfigurable space systems and applications. We devise, evaluate, and demonstrate these approaches in reconfigurable architectures to maximize performability subject to mission availability constraints. The first is HARFT, an environmentally adaptive, gracefully degradable system architecture that maximizes system performability in reconfigurable systems. The second is RECON, a selectively and adaptively resilient semantic-segmentation accelerator that maximizes inference performability in reconfigurable systems. Finally, we propose a methodology for evaluating the performance and dependability characteristics of FPGA-accelerated DL models, which includes a hierarchical fault-injection approach to accelerate the dependability evaluation. This methodology allows spacecraft designers to create a design tradespace and select the optimal DL solution. This dissertation research demonstrates the efficacy of these methods to enable dependable, high-performance onboard processing for next-generation missions.
Share
| Citation/Export: |
|
| Social Networking: |
|
Details
| Item Type: |
University of Pittsburgh ETD
|
| Status: |
Unpublished |
| Creators/Authors: |
|
| ETD Committee: |
|
| Date: |
13 June 2021 |
| Date Type: |
Publication |
| Defense Date: |
2 April 2021 |
| Approval Date: |
13 June 2021 |
| Submission Date: |
15 March 2021 |
| Access Restriction: |
1 year -- Restrict access to University of Pittsburgh for a period of 1 year. |
| Number of Pages: |
159 |
| Institution: |
University of Pittsburgh |
| Schools and Programs: |
Swanson School of Engineering > Electrical and Computer Engineering |
| Degree: |
PhD - Doctor of Philosophy |
| Thesis Type: |
Doctoral Dissertation |
| Refereed: |
Yes |
| Uncontrolled Keywords: |
FPGAs, SoCs, Single-Event Effects, Deep Learning, Semantic Segmentation, Adaptive Computing, Reconfigurable Computing, Space Computing |
| Date Deposited: |
13 Jun 2021 18:45 |
| Last Modified: |
13 Jun 2022 05:15 |
| URI: |
http://d-scholarship.pitt.edu/id/eprint/40370 |
Metrics
Monthly Views for the past 3 years
Plum Analytics
Actions (login required)
 |
View Item |
![[feed]](/images/feed-icon-32x32.png){kind=icon}