The purpose of this thesis is the classification of geometric defects on airfoils using Machine Learning techniques. The motivation for the work lies in the analogy between geometric defects on an airfoil and deformations of standard sinus-nasal anatomy that produce pathology. The study intends to use a simple and tractable problem in two dimensions to gain experience with respect to the clinical problem. The classification algorithm was trained by creating a dataset of more than 4000 airfoils containing five classes of geometric anomalies. A function in Python was extended to generate the defect airfoils. The mesh was created using Construct2D software. Fluid dynamics simulations were conducted with OpenFOAM software, using the Spalart-Allmaras turbulence model, keeping the angle of attack and undisturbed flow velocity fixed. The dataset was created by implementing a pipeline that automates generation, meshing, and numerical simulation for each of the airfoils using the resources of CINECA's Galileo supercomputer. The dataset was then mapped maintaining only a few features for each airfoil. A Machine Learning algorithm, specifically a neural network, was created and trained to solve a multi-label classification problem on five defects, thus having as objective the recognition of the presence or absence of each of the five defects on each airfoil but also having to be able to distinguish between defects and geometric properties. The results show that the neural network captures the information needed to link fluid dynamic quantities and geometric defects.
Lo scopo di questa tesi è la classificazione di difetti geometrici su profili alari utilizzando tecniche di Machine Learning. La motivazione del lavoro risiede nell’analogia fra difetti geometrici di un profilo alare e deformazioni della normale anatomia sino-nasale che producono patologie. Lo studio intende utilizzare un problema semplice e trattabile in due dimensioni per acquisire esperienza rispetto al problema clinico. L'algoritmo di classificazione è stato allenato creando un dataset di più di 4000 profili contenenti cinque classi di anomalie geometriche. Per generare i profili difettosi è stata estesa una funzione in Python. La mesh è stata creata con il software Construct2D. Le simulazioni fluidodinamiche sono state condotte con il software OpenFOAM, usando il modello di turbolenza di Spalart-Allmaras, mantenendo fissati l'angolo di attacco e la velocità del flusso indisturbato. Il dataset è stato creato implementando una pipeline che automatizza generazione, meshing e simulazione numerica per ciascuno dei profili alari utilizzando le risorse del supercomputer Galileo di CINECA. Il dataset è stato poi mappato mantenendo solo alcune features per ciascun profilo. Un algoritmo di Machine Learning, in particolare una rete neurale, è stato creato e allenato per risolvere un problema di classificazione multi-label su cinque difetti, avendo quindi come obiettivo il riconoscimento della presenza o meno di ciascuno dei cinque difetti su ogni profilo ma dovendo essere anche in grado di distinguere tra difetti e proprietà geometriche. I risultati mostrano che la rete neurale coglie le informazioni necessarie a legare grandezze fluidodinamiche e difetti geometrici.
Machine learning-based classification of geometrically defective airfoils
Casella, Jacopo
2019/2020
Abstract
The purpose of this thesis is the classification of geometric defects on airfoils using Machine Learning techniques. The motivation for the work lies in the analogy between geometric defects on an airfoil and deformations of standard sinus-nasal anatomy that produce pathology. The study intends to use a simple and tractable problem in two dimensions to gain experience with respect to the clinical problem. The classification algorithm was trained by creating a dataset of more than 4000 airfoils containing five classes of geometric anomalies. A function in Python was extended to generate the defect airfoils. The mesh was created using Construct2D software. Fluid dynamics simulations were conducted with OpenFOAM software, using the Spalart-Allmaras turbulence model, keeping the angle of attack and undisturbed flow velocity fixed. The dataset was created by implementing a pipeline that automates generation, meshing, and numerical simulation for each of the airfoils using the resources of CINECA's Galileo supercomputer. The dataset was then mapped maintaining only a few features for each airfoil. A Machine Learning algorithm, specifically a neural network, was created and trained to solve a multi-label classification problem on five defects, thus having as objective the recognition of the presence or absence of each of the five defects on each airfoil but also having to be able to distinguish between defects and geometric properties. The results show that the neural network captures the information needed to link fluid dynamic quantities and geometric defects.File | Dimensione | Formato | |
---|---|---|---|
TESI_Jacopo_Casella.pdf
accessibile in internet solo dagli utenti autorizzati
Dimensione
1.46 MB
Formato
Adobe PDF
|
1.46 MB | Adobe PDF | Visualizza/Apri |
TESI_DEFINITIVA.pdf
accessibile in internet solo dagli utenti autorizzati
Dimensione
3.68 MB
Formato
Adobe PDF
|
3.68 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/10589/174213