Machine learning approaches for identifying substandard ships in port state control inspections with imbalanced data

To effectively identify substandard ships in port state control (PSC) inspections, this study proposes a novel machine learning (ML) approach that integrates support vector machines (SVMs) and deep belief networks (DBNs). A comprehensive dataset sourced from the Paris Memorandum of Understanding (MoU) inspection database is used, covering the years 2018–2020. The proposed ML approach effectively addresses the imbalanced data problem in PSC inspections, where the probability of substandard ships is typically no more than 5 %. The results of prediction accuracy indicate that the ensembled SVM-DBN model outperforms the standalone SVM model on both the original and oversampling datasets in all scenarios, except for a slightly lower precision rate on the oversampling dataset. In contrast, the SVM-DBN model underperforms relative to the standalone SVM model on the under-sampling dataset. These results have practical implications for both shipowners and port authorities. Shipowners can use the results to proactively assess and improve ship conditions, thereby mitigating potential detention risks, and port authorities can use them to improve risk-based inspection strategies, thereby optimizing inspection resources and increasing the effectiveness of substandard ship identification. • Two machine learning techniques, support vector machines (SVMs) and deep belief networks (DBNs), are employed. • Imbalanced data was sourced from the Paris Memorandum of Understanding (MoU) inspection database. • The ensembled SVM-DBN model outperforms the individual SVM model in most cases. • Findings can offer valuable insights for identifying substandard ships.