Dar ve trafiği yoğun su yollarında riskli gemi karşılaşmalarının kümeleme tabanlı yapay öğrenme ve sıralı derin öğrenme ile tahminlenmesi

This thesis provides a machine learning framework to predict risky encounters between ships before ships establish visual contact in narrow and congested waterways. There are three parts in this thesis. The first one is clustering for exploration of encounters, the second is ensemble machine learning application to predict risky encounters without distance as a decision factor, and the third one is sequential deep learning based prediction of risky encounters. The Strait of Istanbul (SOI) serves as a case study. Ship–ship interaction database is constructed using historical Automatic Identification System (AIS) messages. Interactions are analyzed via clustering to explore risky encounters using degree of ship domain violation. Findings confirm that ship length and ship speed can serve as reliable indicators to understand the patterns in risky encounters. Results suggest that ship domain violation exists within a nuanced grey zone, requiring cautious consideration instead of rigid categorization. On this basis, clustering based ensemble machine learning framework is developed to predict close encounters and overcome class imbalance. The model successfully predicts each 4 out of 5 risky encounters without the knowledge of distance between two ships. To demonstrate applicability of risky encounter prediction in real time, a practical sequential prediction framework is introduced as the final step. Long Short-Term Memory (LSTM) networks are used to predict risky encounters based on sequential navigation data. The methodology presents an improved encounter model to identify ship-to-ship interactions and classify them as risky, gray-zone, or non-risky encounters based on ship domain violations. The developed approach can be integrated to narrow and congested waterways as a decision support tool to improve maritime safety, and can be a guide to autonomous vessels for safe navigation.