DCSN: Focusing on hard samples mining in small-sample fault diagnosis of marine engine

• Published in: Measurement : journal of the International Measurement Confederation 2024-08, Vol.235, p.114929, Article 114929
• Main Authors: Zhong, Baihong, Zhao, Minghang, Wang, Lin, Fu, Song, Zhong, Shisheng
• Format: Article
• Language: English
• Subjects: Concentric loss; Deep learning; Fault diagnosis; Hard sample; Metric learning; Siamese networks
• ISSN: 0263-2241; 1873-412X
• DOI: 10.1016/j.measurement.2024.114929

•This article constructs a deep concentric Siamese network (DCSN) for mining hard samples in small-sample fault diagnosis of marine engines.•DCSN transforms the multi-classification problem under small-sample conditions to a binary-classification problem, which allows for deep models to learn more inter-class discriminative information.•Under the supervision of concentric loss, DCSN imposes strong forces on hard sample pairs, while completely neglects easy sample pairs that can already be correctly classified.•DCSN promotes intra-class aggregation and inter-class separability via shrinking inner boundary.•The efficacy of the constructed DCSN has been validated in a publicly available ship main engine fault dataset. Fault samples of marine engine are extremely scarce, and there are unavoidably some hard samples with small inter-class differences, which pose a serious challenge to fault diagnosis of marine engines. This paper proposes a deep metric learning method, namely deep concentric Siamese network (DCSN), to apply strong forces to hard samples towards their corresponding correct distribution areas under small-sample conditions. First, DCSN is committed to learn discriminative information from limited fault samples through a carefully designed metric learning strategy. Then, DCSN distinguishes hard samples using inner and outer boundaries, and applies strong forces to them, making the deep model more focusing on the correct classification of hard samples. Third, DCSN shrinks the distribution area of intra-class samples, which improves intra-class compactness and inter-class separability. Finally, the experimental results on the marine engine fault dataset show that the proposed DCSN yields higher diagnostic performance compared to the considered competitive methods.