High impedance Grounding fault (HIF) often occurs in distribution network, it usually accompanies with arc, and tend to cause security risks. But due to its weak fault characteristics, it is difficult to be detected. This paper improves the DenseNet and proposes a lightweight one-dimensional multi-scale DenseNet to realize HIF detection. Specifically, multi-scale features of sequence signals are extracted by multi-scale convolution layer, features are transferred by dual channel fully connected module, and the performance of the model is improved by using the ELU activation function. At the same time, the simulation model is improved to simulate the phenomenon of random fluctuation and zero rest offset of fault waveform in the actual situation. The measured data and simulation data verify that the algorithm can identify the high resistance grounding faults in different network structures and different grounding media, and verify the anti-noise ability of the model.