Due to the non-stationary and non-linear characteristics of arc fault currents, detecting series arc faults in low-voltage systems poses significant challenges. Discrete Wavelet Transform (DWT) is an algorithm with notable advantages in processing transient signals. Combining DWT with Machine Learning (ML) has significantly improved the accuracy of series arc fault detection, attracting widespread attention in the academic community. However, wavelet transforms involve numerous parameters to be determined, and there is limited research on how to select appropriate wavelet parameters. This paper proposes a Particle Swarm Optimization (PSO)-driven wavelet decomposition scheme and feature selection method for series arc faults (PSO-WDFS). The method first classifies the load based on the fundamental waveform of the signal, then uses PSO to automatically select suitable wavelet parameters and feature subsets from over 40,000 combinations for each category, and finally constructs an artificial neural network with the selected features to achieve accurate arc fault detection. Experimental results on the test dataset show that this method effectively discriminates between arc faults and normal currents, achieving an accuracy of over 97% in datasets with mixed load types.