Deep Learning for Hardware-Based Real-Time Fault Detection and Localization of All Electric Ship MVDC Power System

Abstract

The tendency toward electrification of marine vessels has led the evolution of the all electric ship (AES). The harsh operating environment of the AES makes the shipboard power system (SPS) vulnerable, so a powerful monitoring system for fault detection and localization (FDL) is essential for safe navigation. We propose a machine learning based FDL method for monitoring the system condition with the problem of imbalanced training dataset. The generative adversarial network (GAN) comprising of deep convolutional neural networks was employed to synthesize numerous valid samples. Feature extraction and selection technologies were applied to time-series signals to reduce features for monitor training. Finally, the random forest (RF) model was trained using the augmented training dataset, combining real data with generated ones by GAN, to verify the capability of the GAN-RF based FDL method. Both real training and testing data were collected from the SPS model established in PSCAD/EMTDC. The results demonstrated that the monitor could distinguish different conditions in real-time with the help of hardware implementation on the FPGA and a 99% classification accuracy was achieved with excellent anti-noise capability.