In this paper, aiming to enhance the accuracy and dependability of fault identification in distribution network, a novel fault identification approach driven by power spectral density is proposed. First, the fault characteristics of current of lines are analyzed for a multi-branch distribution network, and according to the analysis result, it shows that the fault component characteristics of phase-current on a fault line are distinguished in comparison to those of a non-fault line, which can be utilized to fault lines and kinds identification. For the reduction of noise signal interference, the phase current is decomposed into several transient components by the improved local mean decomposition (ILMD) method. Then, the fault identification principle is established to achieve fault line and fault-phase identification via the high-frequency transient components’ entropy value of power spectral density. Simulations indicate the efficacy of the suggested strategy. The identification results are correct for various fault types and fault lines, and are less sensitive to the nonlinearity of high-impedance faults.