Fault-Tolerant Runtime Reconfiguration Techniques Using Machine Learning for Space-Grade FPGA Systems

Abstract

Mitigation against radiation-induced Single Event Upset (SEU) that can impair system integrity and mission success is required because of the increasing use of SRAM-based FPGAs in high-performance satellite and deep-space applications. In the current paper, I suggest a new fault-tolerant architecture that combines specialised Machine Learning (ML) classifiers with Dynamic Partial Reconfiguration (DPR) in order to offer autonomous real-time error-detection and recovery to space-grade FPGA systems. Our method contrasts with Triple Modular Redundancy (TMR), which would incur prohibitive area, power overhead, or with static scrubbing, which would lacks the ability to predict and classify transient faults because it is contextual and has a high latency. Using a hybrid feature-extraction layer, the system uses vernier bit-flips to distinguish vernacular component failures, and vital component failures, in the FPGA fabric. At the first indication of a localized anomaly, the intelligent controller will initiate localized runtime adjustment of functional tiles defined within the system, which does not affect the mission continuity and allows high system availability with no full system reconfiguration or interrupting parallel activities. The experimental data, obtained through the extensive fault-injection campaigns in a radiation-hardened SoC system, proves that the ML-based structure is able to detect the faults with an almost 98 percent accuracy and the recovery time is lower by 35 percent than the traditional blind scrubbing method. Moreover, the suggested architecture has a 20 percent higher power efficiency, as it does not require any unproductive hardware modules that are common in hardware based solution of voting schemes. Offering a sustainable, capable, and resource-efficient approach to the future generation of reconfigurable, autonomous, and space-based computing capability, this methodology can provide the high-relibility, scalable, and efficient way to operate the new generation of autonomous capabilities in the extreme orbital environment with reduced resource consumption.