Arc fault detection method based on improved wavelet packet threshold noise reduction

Low-voltage series fault arcs are a significant source of electrical fire hazards, where noise interference complicates accurate detection. This paper improves upon traditional wavelet threshold denoising by optimizing the threshold estimation criteria and processing function. The denoised signals are then converted into 2D time-frequency datasets via a short-time Fourier transform (STFT) to provide rich inputs to an improved neural network (arc-DRSN), which incorporates both channel and spatial attention mechanisms. Experimental results show that arc-DRSN significantly outperforms traditional models in accuracy, robustness, and generalization, achieving a state recognition accuracy of 99.86% and a type recognition accuracy of 99.17%. These findings present a robust and efficient solution for arc fault detection.