FPGA-Enabled Deep Learning for Real-Time DC Series Arc Fault Detection

A direct current (DC) series arc fault is an unintended and hazardous electrical discharge in a DC circuit. It poses a safety risk as it can potentially cause fire and damage electrical components. The urgency of detecting DC series arc faults lies in their potential to escalate into hazardous situations swiftly. Rapid identification of abnormal patterns associated with DC series arc faults is essential to trigger intervention measures immediately. In recent years, the focus of studies for the detection algorithms has been shifting from conventional rule-based methods toward machine learning (ML)-based approaches, because ML-based algorithms have the advantage of adapting to different circuit dynamics. Although MLbased approaches have yielded positive results in fault identification, implementing such computationally intensive algorithms in a time-critical application remains a challenge. This paper addresses this challenge through a hardware solution, proposing an embedded real-time detection system with an FPGA-based hardware accelerator to enhance efficiency. The proposed solution implements optimized ML-based algorithms, including fully connected neural networks and convolutional neural networks (CNN), on an Avnet Ultra96-V2 FPGA using the STANN HLS library. The CNN implementation results in the highest accuracy; it achieves a good balance between low detection latency, high numerical accuracy after automated precision reduction, and moderate power consumption. Comparisons with the state-of-the-art demonstrate STANN's high efficiency. Additionally, the performance of the FPGA solution is evaluated by comparing it to a GPU implementation on an Nvidia Jetson Xavier NX board. The FPGA version shows superior performance, with a minimum runtime of 52 <tex-math notation="LaTeX">μ</tex-math>s and total energy consumption of 146 <tex-math notation="LaTeX">μ</tex-math>J per inference, while the Jetson Xavier NX achieves a minimum runtime of 170 <tex-math notation="LaTeX">μ</tex-math>s and consumes 343 <tex-math notation="LaTeX">μ</tex-math>J per inference.