Ship Ballast Water System Fault Diagnosis Method Based on Multi-Feature Fusion Graph Convolution

Abstract

To tackle the issues of limited fault data, inadequate information availability, and subpar fault diagnosis within the realm of ship ballast water system condition monitoring, this paper presents a novel fault diagnosis methodology known as the Probabilistic Similarity and Linear Similarity-based Graph Convolutional Neural Network (PCGCN) model. PCGCN initially converts the ship’s ballast water system dataset into two distinct graph structures: a probabilistic topology graph and a correlation topology graph. It delves into data similarity by employing T-SNE for probabilistic similarity and Pearson’s correlation coefficient for linear similarity to establish the inter-sample neighbor relationships. Subsequently, an early fusion of these two graph structures is conducted to extract more profound multi-scale feature information. Following this step, the graph convolutional neural network (GCN) is introduced to amalgamate the feature information from neighboring nodes in addition to its inherent features. This is aimed at enhancing the available information for the classification task and addressing the issues of limited fault data and inadequate label information. In conclusion, we employ a simulated ship fault dataset for testing experiments, and the PCGCN model demonstrates superior classification accuracy, reaching 97.49%, outperforming traditional diagnostic methods. These experimental outcomes underscore the applicability of the model introduced in this study to the realm of ship ballast water system fault diagnosis, even under challenging conditions characterized by limited sample sizes and insufficient labeling information.