Development of a complete marine data science pipeline: from data collection to model deployment

Rigas, Spyros; Ρήγας, Σπύρος

dc.contributor.author	Rigas, Spyros	en
dc.contributor.author	Ρήγας, Σπύρος	el
dc.date.accessioned	2023-03-29T07:46:48Z
dc.date.available	2023-03-29T07:46:48Z
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/57357
dc.identifier.uri	http://dx.doi.org/10.26240/heal.ntua.25055
dc.description	Εθνικό Μετσόβιο Πολυτεχνείο--Μεταπτυχιακή Εργασία. Διεπιστημονικό-Διατμηματικό Πρόγραμμα Μεταπτυχιακών Σπουδών (Δ.Π.Μ.Σ.) "Επιστήμη Δεδομένων και Μηχανική Μάθηση"	el
dc.rights	Default License
dc.subject	SR-CNN	en
dc.subject	MTAD-GAT	en
dc.subject	Ανίχνευση ανωμαλιών	el
dc.subject	Υπολογισμοί σε υπηρεσίες υπολογιστικού νέφους	el
dc.subject	Νευρωνικά δίκτυα γράφων	el
dc.subject	SR-CNN	en
dc.subject	Anomaly detection	en
dc.subject	MTAD-GAT	en
dc.subject	Cloud computing	en
dc.subject	Graph neural networks	en
dc.title	Development of a complete marine data science pipeline: from data collection to model deployment	en
heal.type	masterThesis
heal.classification	Data Science	en
heal.language	en
heal.access	free
heal.recordProvider	ntua	el
heal.publicationDate	2023-03-03
heal.abstract	The application of data science and machine learning techniques to the maritime industry has emerged as a promising field with significant potential for improving vessel performance and ensuring safe marine operations, due to the abundance of data generated by various sensors and systems. Despite this potential, the industry has been slow to embrace digitalization and the application of advanced analytics. This thesis represents a step towards the direction of automation in the maritime industry, presenting an approach for detecting anomalies in shipboard systems through the application of machine learning. More specifically, the deployment of univariate and multivariate anomaly detection algorithms is studied, with the aim of monitoring the health of vessel systems and identifying potential issues before they become critical. The Spectral Residual Convolutional Neural Network model is used for univariate anomaly detection, while an advanced version of Microsoft's Multivariate Time-Series Anomaly Detection via Graph Attention Network model is used for multivariate anomaly detection. As part of the study, three separate dataset types are developed to train and evaluate the performance of the anomaly detection models: a family of datasets with artificially induced anomalies that is used for the pre-training and evaluation of the models before deployment, a family of synthetic time-series datasets which is utilized to evaluate the expressivity of the trained models using unseen types of anomalies and a family of datasets corresponding to the operational sensor data feed. The work also includes a comprehensive analysis of the data engineering process followed for the construction of a complete pipeline, which involves the ingestion of data, all necessary computations, data storing and models serving. This is achieved by leveraging the advancements in cloud computing and in particular the features and services provided by Microsoft's Azure Cloud Ecosystem. The results achieved from training and evaluating the models in the artificially induced anomalies dataset demonstrate how promising the proposed approach in detecting anomalies in different types of data and settings is. Importantly, two instances of the deployed models are able to detect anomalous behaviour in two vessel systems using their sensors' live feed and consequent inspections by engineers on-board validate the models' predictions. This thesis provides valuable insights into the potential for machine learning to revolutionize the way marine operations are conducted, and offers a roadmap for future research in the field.	en
heal.advisorName	Kollias, Stefanos	en
heal.committeeMemberName	Stamou, Giorgos	en
heal.committeeMemberName	Voulodimos, Athanasios	en
heal.academicPublisher	Εθνικό Μετσόβιο Πολυτεχνείο. Σχολή Ηλεκτρολόγων Μηχανικών και Μηχανικών Υπολογιστών	el
heal.academicPublisherID	ntua
heal.numberOfPages	106 σ.	el
heal.fullTextAvailability	false