Arc faults are significant causes of electrical fires. Recent studies on detecting arc faults typically adopt supervised machine learning, relying on experimentally obtained data with labeled normal and arc fault samples, which have achieved high accuracy on the experimental dataset. However, as household electrical environments become increasingly complex, these approaches face several challenges, including limited data coverage, model overfitting, and the high cost of acquiring high-quality labeled data. This study introduces an unsupervised learning approach for arc-fault detection in which an autoencoder is trained to capture the intrinsic representations of normal current waveforms. Arc-fault current data are collected and annotated following IEC 62606. During inference, the autoencoder reconstructs only the normal components of the input signal, thereby suppressing nominal patterns and enhancing the distinctive features associated with fault conditions. Results demonstrate that this approach can effectively enhance the fault signal.