TY - THES

T1 - Influence of Different Hyperparameter Settings and Data Preprocessing Methods on the Classification of Nonmetallic Inclusions with Machine Learning Algorithms

AU - Musi, Robert

N1 - no embargo

PY - 2023

Y1 - 2023

N2 - The automated scanning electron microscopy with energy dispersive X-ray spectroscopy (SEM/EDS) is a state-of-the-art method for the analysis of non-metallic inclusions (NMI) and well established for research and industry applications. Even though NMIs can be evaluated thoroughly with this method, the crucial disadvantage is its time effort, taking up to several hours per sample. Machine learning algorithms offer an interesting alternative for inclusion characterization, as data can be processed very fast. The present work deals with the time and energy efficient classification of non-metallic inclusions, carried out through the training of different machine learning algorithms on automated SEM/EDS generated image data. Features extracted from backscattered electron images, which were generated through the automated SEM/EDS analysis of seven different steels, served the purpose of training and evaluating the machine learning models. Various algorithms from the python libraries scikit-learn and PyTorch were used. The highest accuracy score of 73,1 % could be achieved with the Random Forest classifier trained on gray values of image pixels. Neural networks were not as suited for inclusion characterization. The used features, type of algorithm, and the NMI dimensions significantly influenced the classification performance. For further studies, parameters such as labeling criteria for NMI within the data preprocessing as well as contrast and brightness settings during automated SEM/EDS measurement need to be adapted in order to enhance accuracy scores.

AB - The automated scanning electron microscopy with energy dispersive X-ray spectroscopy (SEM/EDS) is a state-of-the-art method for the analysis of non-metallic inclusions (NMI) and well established for research and industry applications. Even though NMIs can be evaluated thoroughly with this method, the crucial disadvantage is its time effort, taking up to several hours per sample. Machine learning algorithms offer an interesting alternative for inclusion characterization, as data can be processed very fast. The present work deals with the time and energy efficient classification of non-metallic inclusions, carried out through the training of different machine learning algorithms on automated SEM/EDS generated image data. Features extracted from backscattered electron images, which were generated through the automated SEM/EDS analysis of seven different steels, served the purpose of training and evaluating the machine learning models. Various algorithms from the python libraries scikit-learn and PyTorch were used. The highest accuracy score of 73,1 % could be achieved with the Random Forest classifier trained on gray values of image pixels. Neural networks were not as suited for inclusion characterization. The used features, type of algorithm, and the NMI dimensions significantly influenced the classification performance. For further studies, parameters such as labeling criteria for NMI within the data preprocessing as well as contrast and brightness settings during automated SEM/EDS measurement need to be adapted in order to enhance accuracy scores.

KW - Nonmetallic Inclusions

KW - Automated SEM/EDS Analysis

KW - Inclusion Characterization

KW - Machine Learning

KW - Neural Networks

KW - Nichtmetallische Einschlüsse

KW - Automatisierte REM/EDX Analyse

KW - Einschlusscharakterisierung

KW - Maschinelles Lernen

KW - Neuronale Netze

U2 - 10.34901/mul.pub.2023.160

DO - 10.34901/mul.pub.2023.160

M3 - Master's Thesis

ER -