Leakage protection is one of the critical measures to ensure mine safety. This study aims to achieve real-time monitoring of leakage faults in mine power circuits. To this end, a MATLAB simulation model replicating the mine production environment is developed, incorporating the instantaneous symmetrical component method for transient fault analysis. By comparing the simulated fault waveforms with normal waveforms, set transformation is applied to expand the dataset focused on transient waveform images. Subsequently, based on classic deep learning models such as VGGNet and ResNet, various structured deep learning models are constructed for the classification and recognition of transient waveform images. Experimental results demonstrate that: 1) The proposed deep learning framework effectively distinguishes leakage fault patterns with high reliability; 2) Evaluation metrics on unseen test data-accuracy (T_Acc?) , precision (T_Pre?) , recall (T_Rec?) , and T_F?1? score, reached 0.983 0, 0.986 7, 0.983 3, and 0.983 3 respectively, validating the model's robust classification capability.