LFArc-PFE: A Series Arc Fault Detection Method Based on Low-Frequency Current Data and Perturbation Feature Extraction

Arc faults deliver a significant threat to daily life and property caused by the catastrophic damage to the electrical system. However, developing effective detection models for arc faults is challenging due to the difficulties in obtaining arc fault data in real-world scenarios. Moreover, models trained in a specific scenarios often struggle to adapt to different situations. This paper proposes a series arc fault detection method based on low-frequency current data and perturbation feature extraction (LFArc-PFE). To address the problem of low adaptability of traditional features and information loss posed by the feature selection methods, a Shapelets-based perturbation feature extraction method and a DTW-Hierarchical-Comentrepy (DTWHC) feature selection method are proposed to improve the detection accuracy and adaptability, which are used to comprehensively characterize the current variation during arc faults and to further refine the perturbation feature, respectively. In this paper, a hybrid CNN-LSTM deep neural network is developed to effectively extract key information from the current data and achieve accurate diagnosis of arc faults. A fine-tuning-based transfer learning approach is employed to enhance the adaptability of models across various domestic scenarios. Furthermore, the proposed LFArc-PFE method was evaluated on a hardware platform, and the experimental results demonstrated high accuracy, confirming the method’s reliability and highlighting its substantial potential for engineering applications.