MDJA-Trans: a multivariate data joint analysis transformer architecture with dynamic gating for load-aware low-voltage AC Arc fault detection

To address the challenges of multivariate signal coupling, noise interference, and load adaptability in arc fault detection within low-voltage distribution systems, this paper introduces a novel arc fault detection architecture, termed Multivariate Data Joint Analysis with Transformer (MDJA-Trans) architecture. The architecture employs Noise-Suppressing Time Convolutional Learning (NS-TCL) to dynamically mitigate noise and extract multiscale arc features through frequency-domain masking. Load-Aware Cross-Variate Attention Fusion (CVAF) facilitates dynamic alignment of voltage and current features, while Dynamic Gating Decision (DGD) integrates electrical features with load attribute embeddings for robust arc fault classification. This approach enhances physical interpretability and load generalization. Extensive experiments demonstrate superior noise immunity and cross-load adaptability, achieving an arc fault detection accuracy and F1-Score of 98.55%. The proposed solution will provide technical support for improving household safety and reducing electrical fire risks.