Abstract: In order to address the challenges posed by weak and variable high-impedance fault signals and limited data availability in practical distribution networks, a novel method for detecting high-impedance faults is proposed. Initially, a multi-head variational autoencoder model based on squeeze-excitation networks is employed to augment the small sample dataset. Subsequently, the data are filtered, and the temporal and frequency domain features are extracted, respectively. Considering the weak characteristics of high impedance fault features and the limitations of the proliferation model in generating comprehensive and effective fault features, a categorical boosting algorithm based on the gradient harmonized mechanism (GHM-CatBoost) is introduced. The GHM-CatBoost algorithm incorporates a gradient harmonized mechanism loss function to address the imbalance in attention between easily distinguishable and challenging samples, thereby mitigating the issue of overfitting. The research findings suggest that the data proliferation model can produce fault samples with a blend of simulation data diversity and measured data randomness, thereby enhancing the richness of the dataset. Furthermore, the fault recognition accuracy achieved by the proposed GHM-CatBoost model is notably high at 97.21%, outperforming its counterpart classifier model. Moreover, the efficacy of the proposed approach is validated through rigorous testing and comparative analysis.