To more accurately and efficiently detect series arc faults under different loads and address the difficulty of setting a unified threshold for different loads when using a single feature, a multi-feature fault arc recognition method based on t-distributed stochastic neighbor embedding (t-SNE) and simulated annealing particle swarm optimization algorithm-optimized BP neural network (SAPSO-BP) is proposed. Firstly, considering the rich high-frequency components in arc fault currents, the traditional coefficient of variation (CV) feature is improved by extracting the CV of current frequency. The improved CV achieves an average recognition accuracy of 96% across different loads. Secondly, wavelet packet detail components and energy entropy, which are time-frequency domain features, are further extracted and fused with CV for the identification of arc faults. During the fusion process, various nonlinear dimensionality reduction algorithms are used, and clustering visualization comparisons are carried out. It is found that reducing the multidimensional features to a three-dimensional space using t-SNE dimensionality reduction provides the highest distinction for fault arcs. Finally, the reduced features are input into SAPSO-BP for training, and ablation experiments are designed to verify the recognition performance and robustness of the proposed method. The results show that the recognition performance of the fusion algorithm tSNE-SAPSO-BP is improved by 3.2%, 16.8%, 27.66%, and 33.5% respectively compared with the recognition accuracy of single features on different loads. t-SNE dimensionality reduction and clustering effectively deal with the nonlinear correlations between features, providing key feature information for the identification of fault arcs by the fusion machine learning method.