Maritime Accident Prediction with Machine Learning

Σπαθιάς, Γεώργιος; Spathias, Georgios

dc.contributor.author	Σπαθιάς, Γεώργιος	el
dc.contributor.author	Spathias, Georgios	en
dc.date.accessioned	2023-01-26T09:13:50Z
dc.date.available	2023-01-26T09:13:50Z
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/56923
dc.identifier.uri	http://dx.doi.org/10.26240/heal.ntua.24621
dc.rights	Default License
dc.subject	Machine Learning	en
dc.subject	Supervised Learning	el
dc.subject	Classification algorithms	el
dc.subject	Maritime accidents prediction	el
dc.subject	Python	el
dc.subject	Μηχανική Μάθηση	el
dc.subject	Πρόβλεψη ναυτικών ατυχημάτων	el
dc.subject	Ανάλυση δεδομένων	el
dc.subject	Μοντέλα πρόβλεψης	el
dc.subject	Τεχνητή Νοημοσύνη	el
dc.title	Maritime Accident Prediction with Machine Learning	en
dc.title	Ανάπτυξη Μοντέλου Πρόβλεψης Ναυτικών Ατυχημάτων με τη χρήση Αλγορίθμων Μηχανικής Μάθησης	el
dc.contributor.department	Division of Ship Design and Maritime Transport	el
heal.type	bachelorThesis
heal.classification	Naval - Research engineering	en
heal.language	el
heal.language	en
heal.access	campus
heal.recordProvider	ntua	el
heal.publicationDate	2022
heal.abstract	The present thesis aims to predict information about maritime accidents by using machine learning models. Maritime accidents affect the smooth operation of the maritime industry. Their consequences are important because they lead to human deaths, injuries and environmental disasters. Hence, it is important to reduce the dangers and try to minimize the number of accidents. Accident records from two different databases are collected, cleaned and modified accordingly in order to fit machine learning models. The data offer information about maritime accidents and their outcome. The type of the accidents, their severity and all their consequences are the output features, while, information about the vessels and the circumstances during the accident represent the input features. Data modification and their preparation for the models is one of the most important procedures in this thesis. Databases quality and the existence of data imbalance is one of the main factors that affect the final results. The existence of 2 different databases allows their comparison and helps to understand which one of them is ideal and why. Machine learning is a subfield of Artificial Intelligence. This thesis implements 10 different machine learning models that are commonly used for supervised learning (MLP, Logistic Regression, Gaussian NB, Random Forest, Decision Trees, AdaBoost, Gradient Boosting, Hist Gradient Boosting, XGBoost, LGBM). These models are being used because they are ideal for classification (the output features have categorical classes) and they are also used in similar studies. The models are based on neural networks, statistics, decision trees and boosting. A wide variety of models is being used in order to compare their results and understand which one is ideal for similar projects. The main goal of this thesis is to predict the type of an accident. Given that this feature is encoded to have 4 classes, their weighted f1-score is 52%. This result is the optimal out of all the studied cases. It occurs from the first database, which has more features that the second one, but it contains less accident records. Also, the result occurs from random forest classifier which along with the boosting-based algorithms (Gradient Boosting, Hist Gradient Boosting, XGBoost, LGBM) produce the best results.	en
heal.advisorName	Βεντίκος, Νικόλαος	el
heal.advisorName	Ventikos, Nikolaos P.
heal.committeeMemberName	Ginnis, Alexandros
heal.committeeMemberName	Anyfantis, Konstantinos
heal.academicPublisher	Εθνικό Μετσόβιο Πολυτεχνείο. Σχολή Ναυπηγών Μηχανολόγων Μηχανικών	el
heal.academicPublisherID	ntua
heal.numberOfPages	170 σ.	el
heal.fullTextAvailability	false