University of Twente Student Theses


A Deep Learning Approach for Assessing Micrographs of Fibre Reinforced Composite Laminates

Fernhout, M.M. (2023) A Deep Learning Approach for Assessing Micrographs of Fibre Reinforced Composite Laminates.

[img] PDF
Abstract:Fibre reinforced composites are becoming more common in a world that demands lightweight products. These materials play a crucial role in the aerospace and automotive industries in reducing fuel consumption. Additionally, they are essential in sustainable energy technologies, particularly for the production of improved wind turbines. Fibre reinforced composites consist of high-performance fibres, such as carbon or glass, embedded in a matrix material. The distribution of fibres in the matrix material and the presence of defects, such as air inclusions, greatly influence the mechanical performance of the material. Microscopic images of cross-sections of the material can be used to visualise these characteristics so that a human can identify any possible defects. However, automating this analysis would be highly beneficial and time-efficient. Currently, image analysis techniques are mainly based on manual pixel intensity thresholding, which is sensitive to illumination conditions during image generation and image quality. Machine learning offers a potential solution to automate the analysis of the micrographs. Therefore, the objective of this research is to explore the potential of using a machine learning model to recognise fibres and voids in micrographs of composites to determine the void and fibre volume fractions in the material. This thesis proposes a deep learning approach with a model based on a u-net architecture. The u-net model is trained, validated and tested with data sets composed of images of 256~x~256 pixels cut from microscopy images of carbon fibre composites. This research also includes the generation of these data sets with corresponding ground truth masks of the images. Furthermore, data augmentation is used to increase diversity in these training data sets, which was found to improve the prediction results of the trained models on the test data. The trained models demonstrate that this deep learning approach is capable of accurately recognising voids and fibres without the need for calibration, unlike traditional thresholding techniques. Thus, it is shown that this deep learning approach is a promising method for identifying fibres and voids in microscopy images, despite having a limited data set and not having ideal ground truth masks.
Item Type:Essay (Master)
Faculty:ET: Engineering Technology
Subject:51 materials science, 52 mechanical engineering, 54 computer science
Programme:Mechanical Engineering MSc (60439)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page