Early fault diagnosis of transformer winding based on leakage magnetic field and DSAN learning method

Abstract

Aiming at the problem of lack of training samples and low accuracy in transformer early winding fault diagnosis, this paper proposes a transformer early faults diagnosis method based on transfer learning and leakage magnetic field characteristic quantity. The method uses the leakage magnetic field waveform on the measuring point of the simulated transformer winding to draw the Lissajous figure to calculate the characteristic quantity. The characteristic quantity of the simulation model is used to train the convolutional neural network (CNN) faults classification model. The CNN fault classification model is transferred to the actual transformer fault detection through the improved deep subdomain adaptive network (DSAN), so as to realize the fault diagnosis of the actual transformer by the classification model trained by the simulation data. The test examples of the actual transformer early fault experimental platform and the leakage magnetic field measurement platform are established, and the feasibility of the transfer learning method based on the leakage magnetic field feature quantity proposed in this paper is verified.