Marine Fault Diagnosis With Self-Supervised Learning Incorporating Out-of-Distribution Detection

In marine machinery systems, the scarcity of labeled samples and the inability of the training set to cover all mechanical conditions often hinder the effectiveness of data-driven fault diagnosis models. In this paper, a self-supervised contrastive learning framework incorporating an out-of-distribution detection classifier (SCLODC) is designed for the fault diagnosis of marine machinery systems. First, this paper develops a combination one-dimensional (1D) data augmentation strategy to aid in extracting valuable features from unannotated conditional signals through self-supervised contrastive learning. Second, an out-of-distribution detection classifier in this framework based on two output distributions, which are the class prediction probability and the confidence level value, is proposed to not only recognize known health states but also to identify previously unseen fault conditions. In the SCLODC framework, only 5% of the total samples are annotated to determine the most effective feature representations and to train the classifier. This framework is applied in four case studies across fault diagnosis scenarios. Impacts of training sample sizes on classification performance are also examined. Experimental results indicate that the SCLODC framework performs efficiently in the extraction of the valuable feature information from abundant unlabeled samples, and functions well in obtaining more excellent diagnostic results than other state-of-the-art methods.