At present, most AC arc fault protective devices adopt arc frequency characteristics as fault indicators, which tend to cause failure detection in multi-branch systems. To address this issue, based on analyzing the load and connection factors that mask and weaken arc noise features, this paper proposes an enhanced arc detection method based on local characteristic-scale decomposition (LCD). Firstly, in accordance with the GB/T 31143—2014 standard, an experimental platform of AC arc faults in multi-branch systems was established. The bus-level current data were collected under conditions of different arc-generating branches and load types. After processing via wavelet packet transform (WPT), the existence of arc fault feature attenuation was found. The attenuation degree is closely related to the specific load characteristics and arc-generation configuration. Then, the propagation law of arc noise features in multi-branch systems was investigated through an arc simulation model, revealing that the produced filtering effects of different arc-generating branches and load types are the core reason for fault feature weakening. Subsequently, a multi-dimensional index-based evaluation strategy for arc feature saliency was designed. The proposed LCD method could construct decomposition basis functions based on the temporal scale characteristics and local scale properties of arc signals themselves. Its fine-scale priority ensures that significant arc features are all concentrated in the same intrinsic scale component (ISC), which effectively solves the issue of dynamic changes in traditional time-frequency features due to arc filtering effects. Finally, an arc fault detection algorithm was constructed for AC multi-branch systems by combining with a deep neural network (DNN). The experimental results confirm that the dynamically selected multi-dimensional evaluation metrics could evaluate the effectiveness of arc fault features comprehensively. The proposed method is also proved to have stronger feature extraction capability than existing methods, which would improve the detection accuracy by 17.75%.