Series arc fault can occur in domestic electrical networks and lead to fire accidents. Many conventional detection algorithms based on time or frequency analysis have been published. Recently, machine learning technics have been adapted to the arc fault detection task and give promising results. However, the use of machine learning is frequently limited to the classification part. Manual feature extraction part which requires time and effort is always needed before obtaining the optimized features. In this paper, we used stacked autoencoders to replace the feature extraction part. The method presented can distinguish normal state and series arc fault with good accuracies.