Machine learning hardware accelerator for single-event-latchup (SEL) and micro-SEL detection in commercial-off-the-shelf (COTS) systems

The space industry is undergoing the evolution of ‘New Space’, where commercial-off-the-shelf (COTS) integrated circuits (ICs) and System-on-Chip (SoC) are increasingly employed in space missions, particularly in small Low Earth Orbit (LEO) satellites. However, most COTS ICs and SoCs are intolerant...

Full description

Saved in:
Bibliographic Details
Main Author: Liao, Chenqing
Other Authors: Chang Joseph
Format: Thesis-Master by Coursework
Language:English
Published: Nanyang Technological University 2024
Subjects:
Online Access:https://hdl.handle.net/10356/178793
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:The space industry is undergoing the evolution of ‘New Space’, where commercial-off-the-shelf (COTS) integrated circuits (ICs) and System-on-Chip (SoC) are increasingly employed in space missions, particularly in small Low Earth Orbit (LEO) satellites. However, most COTS ICs and SoCs are intolerant to radiation, e.g., cosmic rays and solar rays in space. Among the various radiation effects, the single-event-latchup (SEL) is of the most concern – it is catastrophic to semiconductor devices when it occurs. In addition, the micro-SEL, a low-current SEL phenomenon, may also occur, where the induced current is lower than an SEL – therefore more difficult to detect, potentially compromising device reliability and lifespan. This dissertation involves a machine-learning-based (ML-based) and deep-learning-based (DL-based) method for micro-SEL detection, alongside the development of a hardware accelerator implemented on a Field-Programmable Gate Array (FPGA) to accelerate the detection process. Through laser-induced micro-SEL experimentation, several current profiles of micro-SELs are acquired and modeled using MATLAB, generating a comprehensive dataset for ML-based/DL-based models for micro-SEL detection. Subsequently, six algorithms (namely, Random Forest (RF), XGBoost, LightGBM, Support Vector Machines (SVMs), Deep Neural Networks (DNNs), and Long Short-Term Memory (LSTM) Neural Networks) are utilized. The LSTM achieved a superior accuracy of ~98%. Logarithmic and linear quantization strategies are thereafter proposed to process the weights in the LSTM layer and fully connected (FC) layer, respectively. Subsequently, the quantized model is reconstructed and validated. Finally, the hardware resource utilization of our developed FPGA-based prototype of the accelerator is evaluated. The accelerator demonstrated high accuracy and reasonable hardware resource utilization. Experimental results show that the accelerator can correctly detect the advent of SEL and micro-SEL and achieve 90% accuracy with small amount of hardware resource. Consequently, the hardware accelerator effectively protects the COTS system against SEL/micro-SEL, rendering it appropriate for small satellite (space application) deployments.