Fault diagnosis of marine diesel engines based on graph convolutional network under unbalanced datasets

  Objectives  The class imbalance problems which are prevalent in the condition monitoring data of marine diesel engines significantly deteriorate the performance of data-driven models for the automatic and accurate identification of the health condition of engines. In this paper, a graph convolutional network (GCN) model based on probability similarity between samples is proposed to solve the classification problem of unbalanced datasets.

  Methods  First, the Kullback-Leibler divergence is introduced to calculate the probability similarity between samples and mine the nonlinear relationship, and a probability topological graph is constructed to represent the similarity of samples. Graph learning is then introduced to learn and extract the correlations between adjacent samples in addition to their own features, providing more information for the imbalanced classification task. After multi-layer graph learning, the higher-level features of each node are extracted.

  Results  The two cases of the simulation model and bench test clearly show that the proposed method efficiently extracts more information based on aggregating neighboring samples'features, and improves the classification accuracy under imbalanced datasets.

  Conclusions  The precision and recall rate of the proposed GCN model are higher than those of other state-of-the-art algorithms, giving it certain practical application value.