MKDNet: A fault detection method based on multi-scale entropy mode denoising and collaborative knowledge distillation for distribution networks

To address the issues of low accuracy, poor universality, and susceptibility to noise interference in singlephase grounding fault detection of distribution networks, a fault detection method based on multi-scale mode denoising and collaborative knowledge distillation networks is proposed. Firstly, zero-sequence current is selected as the basis for differentiating between faulty and non-faulty lines. To alleviate the influence of noise, a multi-scale mode denoising algorithm is designed. To enhance the feature representation capabilities of the zero-sequence current, the gram transform is utilized to convert the zero-sequence current to the gram angular difference field (GADF). Secondly, a teacher-student network is established within a focal-global collaborative knowledge distillation framework to address the issue of weak fault currents in high-impedance grounding faults. Moreover, Wavelet transform cross-stage feature fusion (WTC2f) is devised in the teacher network to further enhance the model's ability to capture fault information. Finally, to validate the superiority of the proposed method, actual operation data from a substation are tested and compared with nine classical algorithms. The result demonstrates that the method achieves high precision, with a line selection accuracy of 95.9%, while also maintaining a relatively low parameter count of only 1.7M.