Abstract: Aiming at the problem that it is difficult to extract the features of early weak arc fault for large current level of photovoltaic (PV) system, this paper proposes a method based on Chirplet sparse representation to extract the time-frequency information of early weak arc fault. First, a DC arc fault experimental platform with resistance, electronic load and inverter load is built. The extraction effect of Chirplet sparse representation on the time-frequency information of early weak arc fault for different system topologies with large current level is studied. Through multi-strategy improved Harris Hawks algorithm optimize Chirplet time-frequency dictionary, the further interference of inverter noise on the time-frequency information of weak arc fault is eliminated. Then the optimal detection feature construction based on Chirplet sparse representation is realized. The universality of the features from the large current DC bus for the early weak arc fault detection when the arc fault occurs on small current branch and aluminum electrode material are verified by the experimental data. Finally, an arc fault detection algorithm is designed based on unsupervised classifier K-means. The detection results show that the detection accuracy of the proposed Chirplet sparse representation feature is 100 %, and the detection time is 0.31 s on average. Compared with the existing methods, our proposed method achieves an average improvement in accuracy by 48.22%, reduces the average detection time by 1.9s, and has been successfully implemented on the Raspberry Pi platform.