Abstract: To address the challenge of detecting high-impedance grounding faults, where arc signals are extremely weak and deeply embedded in harmonic interference, this paper proposes a new fault identification paradigm for extremely low signal-to-noise ratio scenarios. Its core is a collaborative architecture integrating a one-dimensional convolutional neural network (1D-CNN) and a self-attention mechanism. The proposed model adaptively extracts multi-scale local features through the 1D-CNN and intelligently focuses on key distortion periods using the attention mechanism, thereby eliminating the dependence on prior knowledge. Full-condition simulations based on ATP-EMTP show that the method achieves a detection accuracy of 99.2% and an F1-Score of 98.9%, significantly outperforming baseline methods. Ablation experiments confirm that the attention mechanism contributes approximately 2.5% to the performance improvement. Visualization using Gradient-Weighted Class Activation Mapping (Grad-CAM) reveals a high consistency between the model"s attention focus and the physical distortion moments of the arc, providing a paradigm for the explainable application of deep learning in safety-critical systems. This work not only enhances fault detection performance but also offers a transferable theoretical framework for weak signal detection problems in the energy and power field.