In order to solve the deficiencies that the number of intrinsic mode function component and its frequency band obtained by the decomposition of the complete ensemble empirical mode decomposition (CEEMD) are not fixed, which makes it difficult to accurately extract fault arc characteristics thus leading to low accuracy of fault identification, the T-test and variance contribution rate are introduced to form an improved CEEMD method, and a series fault arc identification method based on improved CEEMD and random forest (RF) is proposed. First, the current signals under different loads are collected by the series arc fault test platform. Second, the improved CEEMD is used to analyze the signal and extract the fault feature quantity. Then, the TreeBagger function is used to reduce the feature dimension to form the feature vector sample set. Finally, a fault arc diagnosis model is constructed combined with RF to classify and identify the sample set. Experimental results show that the improved CEEMD can effectively extract fault features of different load currents, and the identification accuracy of the proposed fault arc recognition method reaches 97.50%. Ablation experiments of the influence of different feature extraction methods and different classification models on the diagnostic results is carried out, the feasibility of the proposed method is further proved.