Parameters estimation in marine powertrain using neural
networks

Τζουμέζη, Βασιλική; Tzoumezi, Vasiliki

dc.contributor.author	Τζουμέζη, Βασιλική	el
dc.contributor.author	Tzoumezi, Vasiliki	en
dc.date.accessioned	2020-11-18T08:46:47Z
dc.date.available	2020-11-18T08:46:47Z
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/51932
dc.identifier.uri	http://dx.doi.org/10.26240/heal.ntua.19630
dc.rights	Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα	*
dc.rights.uri	http://creativecommons.org/licenses/by-nc-nd/3.0/gr/	*
dc.subject	Νεωρωνικά δίκτυα	el
dc.subject	Ναυτικοί κινητήρες ντίζελ	el
dc.subject	Ναυτική μηχανολογία	el
dc.subject	Neural networks	en
dc.subject	Marine powertrain	en
dc.subject	Virtual sensors	en
dc.subject	Εκτίμηση παραμέτρων μηχανής πρόωσης	el
dc.subject	Diesel engine parameters	en
dc.subject	Μηχανική μάθηση	el
dc.subject	Machine learning	en
dc.title	Parameters estimation in marine powertrain using neural networks	en
heal.type	bachelorThesis
heal.classification	Marine Engineering	en
heal.classification	Ναυτική Μηχανολογία	el
heal.language	en
heal.access	free
heal.recordProvider	ntua	el
heal.publicationDate	2020-11-05
heal.abstract	There is a wide range of engine performance parameters affecting the combustion process, while some of them have been considered of utmost importance for emission modeling and engine control issues. The Fuel Consumption, MAP, λ and NOx are part of the most indicative engine variables and have stimulated interest in this work. In fact, Artificial Neural Network (ANN) models for predicting the aforementioned quantities could replace the real sensors, which sometimes might be unable to perform direct measurements or to provide with trustworthy values, and on top of that could be quite expensive. These models are known as the neural network-based virtual sensors which could be implemented on-board and, through strong generalization performance, provide results with high accuracy. For this work, the Feed-Forward Neural Network (FFNN) and Time-Delay Neural Network (TDNN) architectures are investigated for the predictive models, and their results are accordingly compared. The networks are trained, validated and tested using experimental data collected from various trials on the laboratory test-bed. In addition, a fully parametric study have been conducted concerning the models inputs selection. The latter has been based on: the theoretical background of the engine function, the relationship between various engine parameters and the available quantities measured by real sensors. Of course, with regard to the acquired experience through modeling, both the inputs and the calculation mechanisms of the models were revised until achieving the most efficient performance. After that, the models were tested on data sets within the same range of the training set (i.e. the whole envelope of test-bed) and on completely unknown data, with different pattern and scaling. Both types of deep neural network models performed well, with no appreciable errors. For the whole modeling process the Python language was used and mainly the Keras library (interface of Tensorflow 2.0), except for the Data Preparation process which was completed in the MATLAB environment.	en
heal.advisorName	Παπαλάμπρου, Γεώργιος	el
heal.advisorName	Papalambrou, George	en
heal.committeeMemberName	Κυρτάτος, Νικόλαος	el
heal.committeeMemberName	Kyrtatos, Nikolaos	en
heal.committeeMemberName	Γρηγορόπουλος, Γρηγόρης	el
heal.committeeMemberName	Grigoropoulos, Gregory	en
heal.academicPublisher	Εθνικό Μετσόβιο Πολυτεχνείο. Σχολή Ναυπηγών Μηχανολόγων Μηχανικών. Τομέας Ναυτικής Μηχανολογίας	el
heal.academicPublisherID	ntua
heal.numberOfPages	120 σ.	el
heal.fullTextAvailability	false