Session: CT-08-01 The L. Eugene Hulbert Memorial Session on New and Emerging Methods of Analysis and Applications

Paper Number: 106773

106773 - Crack Size Measurements on Fracture Surface Images Using Deep Neural Networks for Semantic Segmentation 

For the safe evaluation of nuclear safety components, a precise and reliable analysis of the fracture surface after the test procedure is required. As specified in ASTM E1820 nine equally spaced crack size measurements are required for the determination of the crack size a. The manual classification of fracture mechanisms and the crack size measurements are a time-consuming process that can depend strongly on the subjective perception of the technician and the given working conditions. The aim of the automated analysis and measurement of fracture surfaces using artificial neural networks is to ensure high quality, comparability and, above all, cost-effectiveness of the evaluation of the fracture surfaces.

In a first step, a data set was setup, which represents relevant materials (e.g. 22 NiMoCr3-7 and its weld metal) and specimen geometries (e.g. SE(B) and C(T) specimens) or test types. For this purpose, digital camera images of fracture surfaces commonly used in practice were provided by various test labs and bundled with an image data set of the Fraunhofer-Institute for Mechanics of Materials IWM. Subsequently, suitable network architectures for semantic segmentation (pixel-fine classification) of fracture surfaces were identified and models of this type were successfully trained. On a representative test data set a high agreement (mean intersection over union) between manual annotation (ground truth) and prediction of the neural network could be found with good generalizability of the model. Currently, the implementation of semi-supervised architectures is being worked on with the goal of further increasing precision and generalizability of the models. The automated measurement of the fracture surfaces based on the model predictions proves to be accurate, so that the determined crack lengths can be used to determine fracture mechanics parameters. The paper will present and discuss the development process, the strengths and weaknesses of the model approach, and further work in this context.

Presenting Author: Johannes Rosenberger RWTH Aachen University Steel Institute IEHK

Presenting Author Biography: Johannes Rosenberger is a research assistant at the Steel Institute IEHK which is part of the RWTH Aachen University. Additionally, he is a permanent guest at Fraunhofer Institute for Mechanics of Materials IWM in Freiburg. His PhD studies focus on the usage of machine learning and especially computer vision for various applications in fracture mechanics, such as automated crack size measurements and fracture mechanism classification. He studied Mechanical Engineering at Karlsruhe Institute of Technology and Aalto University in Helsinki with a heavy focus on material science and lightweight construction.

Authors:

Johannes Rosenberger RWTH Aachen University Steel Institute IEHK
Johannes Tlatlik Fraunhofer Institute for Mechanics of Materials IWM
Sebastian Münstermann RWTH Aachen University Steel Institute IEHK