Abstract: Fast and accurate fault detection is of great significance for guaranteeing the operational security of power distribution grids. Considering the poor fault tolerance of matrix algorithms and the high complexity of intelligent optimization algorithms, a data-driven method was put forward to detect faults in low-voltage distribution networks using random matrix theory. The method adopts an edge-cloud collaborative architecture. When a fault occurred in the low voltage distribution network, firstly, edge devices conducted a distributed diagnostic analysis of the local measurements time series based on time-lag correlations and determine the type of data to upload. Secondly, the regional master station conducted a centralized analysis modeling uploaded data with a high-dimensional random matrix and mining the temporal and spatial characteristics of the fault. The accuracy and efficiency of the proposed method were finally validated by conducting case studies on the IEEE European low voltage test feeder. The results of case studies show that our method has following advantages: on the one hand, the method is highly versatile and no physical model and topology information but measurement data are needed; on the other hand, the method has good robustness to bad data and missing data.