Hierarchical deep convolution neural networks based on transfer learning for transformer rectifier unit fault diagnosis

Abstract Convolutional neural networks (CNNs) are able to extract features automatically. Fault identification and location of transformer rectifier units (TRUs) which are widely used as an avionic secondary power supply are significant for system reliability. Based on the analyzation of TRUs, this paper discusses the design process of the developed discrete time-series convolution neural network (DTCNN) and develops a hierarchical method for TRUs fault diagnosis along with a transfer learning-based fault diagnosis method instead of training new models for different TRUs. The DTCNN construction is determined firstly. Then, the performance of HDCNN is validated. On the basis, the conditions of a suitable source dataset for TRU fault diagnosis and the transfer layers from the pre-trained HDCNN are discussed. The comparison with other algorithms under different noise conditions shown that the transfer learning is an effective way to construct the diagnosis network for similar equipment and often can lead to better performance.