Prediction of residual stress profiles in welding using machine learning

Ρισσάκη, Δήμητρα; Rissaki, Dimitra

dc.contributor.author	Ρισσάκη, Δήμητρα	el
dc.contributor.author	Rissaki, Dimitra	en
dc.date.accessioned	2021-12-23T11:11:45Z
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/54254
dc.identifier.uri	http://dx.doi.org/10.26240/heal.ntua.21952
dc.rights	Default License
dc.subject	Residual stresses	en
dc.subject	Welding	en
dc.subject	Machine learning	en
dc.title	Prediction of residual stress profiles in welding using machine learning	en
heal.type	bachelorThesis
heal.secondaryTitle	Πρόβλεψη κατατομών παραμενουσών τάσεων στη συγκόλληση με χρήση μηχανικής μάθησης	el
heal.classification	Engineering	en
heal.dateAvailable	2022-12-22T22:00:00Z
heal.language	el
heal.language	en
heal.access	embargo
heal.recordProvider	ntua	el
heal.publicationDate	2021-10
heal.abstract	Pressure vessels and other critical for safety components used in today’s industry require careful structural integrity assessment. A factor that should be determined in order to assess the integrity and lifetime of structures is residual stresses. Residual stresses are encountered in the welding region of these components, due to the uneven thermal distribution during welding. Measurements of residual stresses can be achieved by various techniques, but have their disadvantages, for instance cost, inaccuracy, and that in some of them the removal of material is required. The prediction of residual stresses can be achieved computationally, with Finite Element Simulation, but again there are disadvantages, for example inaccuracy due to assumptions for welding parameters and computational cost. Another method that has been proposed recently for the prediction of residual stresses is machine learning, which is cost-effective, and it appears promising in terms of accuracy of predictions. Machine learning has been used in literature for predicting results of Finite Element Simulations as well as predicting results of measurements. In the current study, the residual stresses induced by welding were predicted using an ensemble of Artificial Neural Networks (ANNs). The acquired data concern measurements of through-thickness residual stress profiles of austenitic stainless steel pipes in Weld Centre Line (WCL) and Heat Affected Zone (HAZ) regions. To assess the various architectures of Artificial Neural Networks an objective function was formed, which was based on four criteria, namely training and generalisation error, solution space consistency and feedforward architecture. One hidden layer was selected due to the simplicity of the problem, and fifteen different architectures were assessed, with number of nodes varying from one to fifteen. To eliminate any bias induced by the initial weight selection, each architecture was initialised to fifteen different random initial weights. The objective function was computed, and the five architectures with lowest median value of objective function were then selected to form the final model. The proposed methodology was then assessed by performance metrics, and it was compared with the performance of two other methodologies. Firstly, it was compared with a single architecture model, where only one architecture is selected based on the minimum median value of objective function. Secondly, the proposed methodology was compared with the scenario of using all the architectures that were considered for the ensemble, without any filtering by the objective function. The results obtained showed that the proposed methodology performs better than selecting a single architecture and slightly better than selecting all the architectures considered.	en
heal.advisorName	Μπενάρδος, Πανώριος	el
heal.committeeMemberName	Βοσνιάκος, Γιώργιος-Χριστόφορος	el
heal.committeeMemberName	Μανωλάκος, Δημήτριος	el
heal.academicPublisher	Εθνικό Μετσόβιο Πολυτεχνείο. Σχολή Μηχανολόγων Μηχανικών. Τομέας Τεχνολογίας των Κατεργασιών	el
heal.academicPublisherID	ntua
heal.numberOfPages	80 σ.	el
heal.fullTextAvailability	false