Electrical arc faults, a major cause of electrical fires, are difficult to detect reliably due to the scarcity of labeled fault data and severe class imbalance between normal and arcing states. To address these challenges, this paper proposes a time-series data augmentation framework based on a Denoising Diffusion Probabilistic Model (DDPM), which enhances the U-Net with attention and residual blocks to capture complex temporal and high-frequency features, enabling the generation of high-quality synthetic arc-fault samples by reversing the noise diffusion process. By generating arc-fault samples that are difficult to collect and underrepresented in the positive class, the proposed method helps rebalance the dataset. Experimental results demonstrate that the data produced by the diffusion model closely align with the original distribution and lead to improved classification accuracy in CNN-based arc classifier.