Abstract: To address the low accuracy of fault arc identification caused by changing working conditions and changing loads in DC systems,this paper proposes a DC series fault arc diagnosis method based on the complete ensemble empirical modal decomposition with adaptive noise-Hilbert transform（ CEEMDAN-HT）envelope spectrum and stacking automatic encoder（SAE）. To begin with,a DC series arc fault experimental platform with mixed load is built to collect the current signals under multiple operating conditions and establish the arc fault database. In addition,CEEMDAN-HT is used to decompose the original signal to obtain multiple intrinsic mode functions, and HT transform is performed to analyze the envelope spectrum and form the first and last high-dimensional feature samples,and finally the samples are input into SAE model to learn the features and realize the DC fault arc identification under variable load. The experimental results show that the method can make good use of CEEMDAN-HT to extract fault arc features from the original signal and the unsupervised learning ability of the SAE,and can accurately identify fault arcs and perform load classification without manually setting thresholds,with an accuracy rate of 98.9% on average.