This paper introduces a novel approach to fault line selection, addressing the limitations of traditional methods, especially when dealing with weak zero-sequence current fault characteristics that are both complex and variable. The method proposed utilizes a dilated convolutional neural network (CNN) for enhanced feature extraction and self-classification capabilities. Initially, the zero-sequence current sequence is decomposed into natural mode functions of different frequencies using variational mode decomposition, which improves the stationarity and distinctiveness of fault signal characteristics. Subsequently, the dilated CNN is employed as the line selection network. The effectiveness of the proposed method is demonstrated through example analysis on a 10 kV distribution network constructed in MATLAB/Simulink, showing superior classification accuracy and rapidity compared to conventional CNNs.