A novel approach based on an improved stacked denoised autoencoder for detecting high impedance faults in distribution networks is proposed. This aims to tackle challenges such as the selection of suitable thresholds, susceptibility to high noise levels, and low recognition accuracy. This method enhances reliability and noise resistance through two key avenues: feature extraction and network model refinement. First, by integrating time-frequency data analysis techniques, the method captures distinctive physical characteristics distinguishing high impedance grounding faults from normal operational states. This forms the foundation for constructing a fault sample feature library. Secondly, fault features across time, frequency, and time-frequency domains are filtered using Pearson correlation coefficients to create a streamlined multi-domain feature fusion sample library, reducing redundancy and enhancing computational efficiency within the network model. Leveraging the strong high-dimensional feature classification capabilities of an enhanced extreme learning machine, the stacked denoising autoencoder model is refined to boost the robustness and accuracy of the high impedance grounding fault classifier. Finally, a Matlab/Simulink simulation model of a 10 kV distribution network is used for illustrative analysis. The results show that the method still has 95.57% accuracy of high impedance fault detection in the condition of –1 dB strong noise, which has high engineering practical value.