Abstract

Object detection is an important operation for satellites to be able to perform, and in outer-space missions it is crucial that machine learning models perform inference on satellite images accurately and reliably. Where size, weight, power, and other constraints exist, meeting this goal for accurate and reliable object detection is challenging. Additionally, soft errors caused by radiation further disrupt and degrade the operation of object detection in satellites. This thesis studies the performance of a deep learning model on an embedded device, the AMD Versal, in the presence of soft errors. The well-known and high-performing YOLO convolutional neural network was used as the deep learning model in this thesis, and various sizes of YOLO were trained on a collection of satellite images of airplanes known as the RarePlanes dataset. The target device is AMD's XCVC1902 Versal Adaptive SoC part. The Versal part was chosen as the deployment target because it is capable of high-performance computing at low power, and because it has a space grade version, the XQRVC1902. The trained models were deployed on the Versal using two flows: (1) AMD's Vitis AI framework and Deep Learning Processing Unit (DPU) IP and (2) manual hardware design. The main contribution of this thesis is an experimental analysis of YOLO running on the Versal in a radiation environment. First, the performance of YOLO on the Versal was measured and reported. Next, a test setup was developed to simulate and discover the impact of soft errors on YOLO object detection through injecting random faults into the programmable logic (PL) CRAM of the Versal device. Various system-level and silent data corruption (SDC) effects were observed from fault injection. The effect of PL faults varied depending on the DPU configuration and model size, with up to 9% resulting in either a system hang or SDC. A proton radiation test was performed on the DPU running YOLO to validate the fault injection results and examine how well fault injection predicts device behavior in radiation. Finally, a non-DPU YOLO hardware design was created for the Versal and found to be affected by 2.55% of random PL CRAM faults, with TMR reducing this percentage by 4-5x.

Degree

College and Department

Ira A. Fulton College of Engineering; Electrical and Computer Engineering

Rights

https://lib.byu.edu/about/copyright/