A comparative investigation of data-driven approaches based on one-class classifiers for condition monitoring of marine machinery system. (1st April 2020)

• Record Type: Journal Article
• Authors: Tan, Yanghui; Tian, Hui; Jiang, Ruizheng; Lin, Yejin; Zhang, Jundong
• Is Part Of: Ocean engineering. Volume 201(2020)
• Publication Date: 2020-04-01
• Subjects: One-class classifier -- Condition monitoring -- Marine machinery -- Condition based maintenance; Ocean engineering -- Periodicals; Ocean engineering; Periodicals; 620.4162
• DOI: 10.1016/j.oceaneng.2020.107174
• Languages: English
• ISSNs: 0029-8018

Abstract

Abstract: The safety and reliability of ship navigation depend heavily on the performance of the marine machinery system, which can be maintained at a high level by condition based maintenance. Intelligent condition monitoring is the key to achieve condition based maintenance. However, it is not easy to realize intelligent condition monitoring in a real ship, because there are not enough labeled fault samples to train an accurate detection model. In fact, the potential value of a large number of normal samples collected by the shipboard monitoring system has not been mined out. Since one-class classifiers only need one-class samples to train the detection models, they are very suitable for such occasions. In this paper, we investigate the performance of several representative one-class classifiers, such as OCSVM (One class support vector machine), SVDD (Support vector data description), GKNN (Global k-nearest neighbors), LOF (Local outlier factor), IForest (Isolation forest) and ABOD (Angle-based outlier detection), in condition monitoring of the marine machinery system. The dataset of a marine gas turbine propulsion system is used for a case study, which is a simulation dataset verified by the real ship data. Compared with previous literatures, our new work is to investigate the performance of a variety of key one-class classifiers, not only considering the common evaluation indexes for machine learning, but also considering the sample combinations of the training dataset, distribution of misclassified samples, and the tolerance to contaminated data. The experimental results show that these algorithms have good performance in the dataset in some common evaluation indexes. However, they show some obvious performance differences in some novel evaluation indexes we proposed. This can provide decision support for the application of one-class classifiers in the condition monitoring of many other marine machinery systems. Highlights: A variety of representative one-class classifiers are evaluated for anomaly detection of marine machinery system. Performance stability, distribution of misclassified samples and tolerance to contaminated data are considered. The best algorithm is selected through the comparison and some special characteristics are proposed. It can provide decision support for application of one-class classifiers on condition monitoring of marine machinery system.