Real-Time AC Arc Detection Technique Based on Artificial Intelligence with Frequency Feature Extraction

This paper proposes a real-time ac arc detection method that utilizes frequency-domain feature analysis and Random Forest algorithms to extract key frequency components sensitive to arc faults. These selected components are then used as inputs to a lightweight deep learning model. The proposed approach significantly reduces input dimensionality and computational complexity, while achieving approximately a 50% reduction in detection time without compromising detection accuracy. The deep learning model is designed based on 1D Convolutional layers, Inverted Residual (IR) Blocks, and Squeeze-and-Excitation (SE) structures, and employs non-linear activation functions such as LeakyReLU and h-swish to enhance representation capability. Experimental validation was conducted under various load conditions and circuit configurations in compliance with IEC 62606 standards, and the proposed model maintained high detection accuracy even in complex electrical environments. This study demonstrates the feasibility of implementing a real-time arc detection system on an embedded platform using Raspberry Pi 5.