Abstract: To address the issues of weak fault features being easily submerged by noise and the low detection rate of traditional methods in single-phase grounding fault detection of secondary circuits in substation relay protection equipment, this paper proposes a grounding fault detection method oriented to weak fault feature enhancement. An improved D-S evidence theory based on evidence distance weighting is employed to adaptively fuse multi-source heterogeneous information, suppress interference from conflicting evidence, and construct a highly reliable decision criterion. An adaptive variational mode decomposition (VMD) method with a dynamically adjusted penalty factor based on local waveform curvature is proposed to decompose the zero-sequence current signal with high precision, combined with Z-score standardization to enhance feature comparability. A gate-coordinated optimized improved long short-term memory (LSTM) network is constructed as the classifier, which strengthens the extraction and memory capabilities of weak fault temporal features through nonlinear coupling and cooperative modulation of the input gate, forget gate, and output gate. Experimental results show that the proposed method achieves a fault recognition rate of over 95% and a missed detection rate below 1% on the test set, and maintains a recognition rate of 88.2% even under a 10dB strong noise environment, verifying its effectiveness and robustness for weak fault detection.