Polish Maritime Research (Dec 2022)

A Convolutional Neural Network-Based Method of Inverter Fault Diagnosis in a Ship’s DC Electrical System

  • Yan Guohua,
  • Hu Yihuai,
  • Shi Qingguo

DOI
https://doi.org/10.2478/pomr-2022-0048
Journal volume & issue
Vol. 29, no. 4
pp. 105 – 114

Abstract

Read online

Multi-energy hybrid ships are compatible with multiple forms of new energy, and have become one of the most important directions for future developments in this field. A propulsion inverter is an important component of a hybrid DC electrical system, and its reliability has great significance in terms of safe navigation of the ship. A fault diagnosis method based on one-dimensional convolutional neural network (CNN) is proposed that considers the mutual influence between an inverter fault and a limited ship power grid. A tiled voltage reduction method is used for one-to-one correspondence between the inverter output voltage and switching combinations, followed by a combination of a global average pooling layer and a fully connected layer to reduce the model overfitting problem. Finally, fault diagnosis is verified by a Softmax layer with good anti-interference performance and accuracy.

Keywords