<p>Arc faults in low-voltage distribution circuits are major causes of electrical fires. The current amplitude of a series arc fault is small, and its waveform distortion is not obvious, which means that protection devices are unable to detect this type of fault effectively. The arc characteristics extracted by the traditional time-domain method are not sensitive to the partial distortion of the current waveform, which reduces the fault recognition rate. To solve this problem, the partial waveform characteristics of current are used as the basis of fault diagnosis. The improved shapelet method is used to extract the partial waveform subsequence with the best discriminative ability among various types of waveforms. According to the similarity between the current sample and all shapelets, the distance feature vector is constructed and sent to the fuzzy K-nearest neighbor (FKNN) classifier for arc fault recognition. The experimental results show that the arc detection accuracy of the Shapelets-FKNN method proposed in this paper reaches 99.65%, and the ability of the partial current to express arc features is better than that of the full current and other arc features. The FKNN classifier is suitable for sample sets with unclear category boundaries and has good robustness to noisy data, and the recognition effect is better than that of KNN. In addition, compared with other arc fault detection methods in the literature, the fault detection accuracy of Shapelets-FKNN is greater, the calculation speed is higher, and it meets the accuracy and real-time requirements of the system, providing a new method for series arc fault detection.</p>