Abstract: In order to identify fault causes of distribution networks, currently used artificial patrol inspection not only consumes a lot of manpower and material but also prolongs the power outage time. For this reason, a data driven based fault cause identifying for distribution network was proposed. Firstly, by means of analyzing a lot of spot-recorded fault waveform data the mechanism of different fault causes and the wave characteristics were obtained, and a fault feature extraction method based on empirical mode decomposition (abbr. EMD) and principal component analysis (abbr. PCA) was proposed. Secondly, through EMD the time domain waveform of the fault was decomposed according to different time scales to obtain intrinsic mode function (abbr. IMF) components possessing local features of the signal. Thirdly, by use of PCA the dimensionality reduction of multi-IMF components were conducted and the principal characteristic components in IMF series were extracted to compose them into eigenvectors. Finally, a fault cause classification model based on long-short term memory (abbr. LSTM) network was put forward to extract dynamic time-scale feature and to realize the classification of fault causes. The experiment results, which utilizes practical field data, show that the proposed fault cause classification model possesses a higher accuracy.