Space Environment Anomaly Detection using AI/ML at the Edge

Professor Michael Ford of AFIT presented at the Vanderbilt Seminar Series.

Radiation effects on artificial‑intelligence inference constitute a critical challenge for space‑borne edge computing, where limited resources must sustain reliable operation under harsh particle environments. Understanding how high‑energy neutrons and, ultimately, heavy ions degrade the performance of convolutional neural networks on system‑on‑chip (SoC) heterogeneous platforms is essential for designing resilient autonomous sensors and spacecraft payloads. The resource investigates these issues through controlled neutron irradiation of a Xilinx Versal MPSoC running state‑of‑the‑art CNN models.

The study quantifies soft‑error cross‑sections for a ResNet‑50 network implemented on an UltraScale+ evaluation board (ZC104) and exposed to a collimated fast‑neutron beam at Ohio State University. Measurements capture size‑weight‑power (SWaP) metrics, accuracy, precision, and recall while varying error‑mitigation techniques such as memory scrubbing, ECC, and conditional triple‑modular redundancy (TMR). An escalation ladder approach evaluates the trade‑offs of dynamic re‑execution, selective re‑programming of CRAM, and on‑the‑fly instantiation of TMR only when an anomaly is detected, thereby reducing the SWaP burden compared with continuous TMR. The experimental framework also records the impact of most‑significant‑bit flips on inference degradation, informing future heavy‑ion testing and the exploration of deeper auto‑encoder CNNs.

Armed with this knowledge, learners can articulate how space radiation perturbs AI inference performance and assess competing mitigation strategies for radiation‑induced faults. They acquire the ability to devise conditional TMR architectures that activate voting only during suspected anomalies, optimizing resource consumption while preserving detection fidelity. Moreover, they can evaluate SWaP trade‑offs in radiation‑hardened AI systems, enabling informed design decisions for edge‑deployed anomaly‑detection payloads in spacecraft and other constrained platforms.

Previous
Previous

Turning Memory into Computing for AI and Beyond: Emerging Devices and Architectures

Next
Next

The Lunar Labs Initiative & Laser Interferometer Lunar Antenna