The series arc fault is one of the main causes of the electrical fire. Aiming at the problem that the series arc fault is difficult to detect accurately under unknown working conditions, a series arc fault detection method based on real-time training and updating kernel based extreme learning machine prediction model was proposed. First, the series arc fault experiments under different power supply harmonics, the carrier frequency and operating frequency of the frequency converter and current level conditions were carried out by using the three-phase motor with frequency converter load circuit. Second, the current signal was denoised by using singular value decomposition filtering and improved single exponential smoothing filtering in turn. Third, the kernel based extreme learning machine prediction model was trained and updated by using the first two cycle current signals, and the predicted residual of the next cycle current signal was calculated. Then, a matrix was constructed by using the absolute value of the predicted residual, the residual matrix was reduced to one-dimensional vector by combining the non-negative matrix factorization, and the maximum value of the one-dimensional vector was used as the fault feature. The series arc fault was detected by using a fixed threshold. Finally, the series arc fault detection and anti-noise performance of the proposed method were tested under unknown working conditions. The results indicated that the proposed method can effectively detect the series arc fault under four kinds of unknown working conditions, which are unknown power supply harmonics, carrier frequency and operating frequency of the frequency converter, and current level, respectively. It showed that method has a strong anti-noise ability.