The aerospace sector has been aiming for years to produce structures that are resistant but at the same time as much as possible lightweight. This request, has led the scientific study to intensify research concerning the phenomenon of buckling. One of the methods currently studied to increase the value of the critical load associated with this phenomenon consists in the use of curvilinear fibers, through which, it is possible to locally change the stiffness of the structure. The present thesis, developed in collaboration with Delft University of Technology, wants to fit into this context, going to propose a framework for the analysis and optimization variable stiffness (VS) cylindrical shells. The developed methodology consists of a synergic work between the finite element (FE) method and artificial intelligence (AI) techniques. The former can produce very accurate results of the structure's behavior while the latter, are useful tools for reducing computational costs in an optimization problem. In the first phase of the thesis a new mathematical formulation for the description of the fiber path is proposed, which is able to extend the design space of classical composites and whose variables represent the design parameters during the optimization. In the second phase, a artificial neural network (ANN) is built, trained on simulations produced with FE analysis in order to accurately approximate the relationship between the input parameters of the optimization and the related structural outputs. In this way, it is possible to use the ANN to predict the response of the VS cylindrical shells, specifically buckling load and pre-buckling stiffness, so as to to be able to drastically reduce computational times during the optimization. In the last phase, a code for the Particle Swarm Optimization (PSO) able to consider the manufacturing constraint on the minimum radius of curvature is developed. The higher accuracy offered by ANN with respect to other global approximation techniques and the time saving, resulting from the developed methodology are both highlited.
Il settore aerospaziale mira ormai da anni a produrre strutture resistenti ma allo stesso tempo il più possibile leggere. Tale richiesta, ha portato lo studio scientifico ad intensificare la ricerca riguardante il fenomeno del buckling. Uno dei metodi attualmente studiati per aumentare il valore del carico critico associato a tale fenomeno consiste nell'utilizzo di fibre curvilinee, attraverso le quali, è possibile modificare localmente la rigidezza della struttura. La presente tesi, sviluppata in collaborazione con Delft University of Technology, vuole inserirsi in questo contesto, andando a proporre una metodologia per l'analisi e l'ottimizzazione di gusci cilindrici a rigidezza variabile. La metodologia sviluppata consiste in un lavoro sinergico fra il metodo ad elementi finiti e le tecniche di intelligenza artificiale (IA). Il primo permette di produrre risultati molto accurati del comportamento della struttura mentre l' IA, è uno strumento utile per poter ridurre i tempi computazionali in un problema di ottimizzazione. Nella prima fase della tesi viene presentata una nuova formulazione matematica per la descrizione della traiettoria delle fibre, la quale permette di estendere lo spazio di progetto dei compositi classici e le cui variabili rappresentano i parametri di progetto durante l'ottimizzazione. Nella seconda fase, viene costruita una rete neurale artificiale, allenata sui risultati prodotti dalle analisi a elementi finiti al fine di approssimare con accuratezza la relazione che intercorre fra i parametri di input dell'ottimizzazione e le relative risposte strutturali. In questo modo, è possibile utilizzare la rete per prevedere la risposta dei gusci cilindrici a rigidezza variabile, in particolare il carico di buckling e la rigidezza pre-instabilità, così da poter ridurre drasticamente i tempi computazionali durante l'ottimizzazione. Nell'ultima fase, viene sviluppato un codice per l'ottimizzazione tramite sciame di particelle in grado di considerare anche il vincolo progettuale sul minimo raggio di curvatura delle fibre. La maggior accuratezza offerta dalla rete rispetto alle altre tecniche di approssimazione globale e il risparmio di tempo, derivante dalla metodologia qui proposta, vengono entrambi evidenziati.
Artificial intelligence techniques for the optimization of variable stiffness cylindrical shells
PITTON, STEFANO FRANCESCO
2017/2018
Abstract
The aerospace sector has been aiming for years to produce structures that are resistant but at the same time as much as possible lightweight. This request, has led the scientific study to intensify research concerning the phenomenon of buckling. One of the methods currently studied to increase the value of the critical load associated with this phenomenon consists in the use of curvilinear fibers, through which, it is possible to locally change the stiffness of the structure. The present thesis, developed in collaboration with Delft University of Technology, wants to fit into this context, going to propose a framework for the analysis and optimization variable stiffness (VS) cylindrical shells. The developed methodology consists of a synergic work between the finite element (FE) method and artificial intelligence (AI) techniques. The former can produce very accurate results of the structure's behavior while the latter, are useful tools for reducing computational costs in an optimization problem. In the first phase of the thesis a new mathematical formulation for the description of the fiber path is proposed, which is able to extend the design space of classical composites and whose variables represent the design parameters during the optimization. In the second phase, a artificial neural network (ANN) is built, trained on simulations produced with FE analysis in order to accurately approximate the relationship between the input parameters of the optimization and the related structural outputs. In this way, it is possible to use the ANN to predict the response of the VS cylindrical shells, specifically buckling load and pre-buckling stiffness, so as to to be able to drastically reduce computational times during the optimization. In the last phase, a code for the Particle Swarm Optimization (PSO) able to consider the manufacturing constraint on the minimum radius of curvature is developed. The higher accuracy offered by ANN with respect to other global approximation techniques and the time saving, resulting from the developed methodology are both highlited.File | Dimensione | Formato | |
---|---|---|---|
2019_4_Pitton.pdf
solo utenti autorizzati dal 05/04/2020
Descrizione: Stefano Francesco Pitton Thesis
Dimensione
9.63 MB
Formato
Adobe PDF
|
9.63 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/146761