To deal with the issue of insignificant series arc fault characteristics in disturbing loads, this paper proposes a voltage-type series arc fault detection technique that utilizes a convolutional bidirectional long- and short-term memory neural network (CNN-BiLSTM) combined with the Keplerian optimization algorithm (KOA) and Attention Mechanism (AM).  Moreover, through experimental verifications, the accuracy of detecting experimental loads during the formation of series arc faults can exceed 99%, which demonstrate the effectiveness of the proposed method.