Multi-branch AC arc fault detection based on ICEEMDAN and LightGBM algorithm

The operation characteristics of household appliances would interfere with the arc fault detection. Especially in the case of the multi-branch circuit, the arc fault information is mixed with the operation characteristics of various loads, which makes the traditional arc fault detection algorithm invalid. In this paper, arc fault signals are acquired from the designed multi-branch experimental platform composed of masking loads according to IEC 62606. By adding the adaptive noise, ICEEMDAN algorithm is found to be suitable to extract more sufficient arc fault information. By selecting the IMF and constructing its variance, the detection variables are constructed to distinguish the arc fault from the normal state. Compared with wavelet transform, the ability to extract arc fault information is improved by 8.47 times on average. In order to simplify the algorithm structure, the Boruta and LightGBM algorithm are used to reduce the feature dimension and improve the algorithm efficiency. Based on the well-trained LightGBM model, the arc fault detection algorithm is finally proposed, which could detect multi-branch arc faults accurately and fast. It is proved to have the higher detection accuracy of 97.06% and the lower detection time of no more than 300 ms on the Raspberry pi platform.