Alleviating Data Scarcity In Industrial Machine Learning Applications

Θεοδωρόπουλος, Σπύρος-Χριστόφορος

dc.contributor.author	Θεοδωρόπουλος, Σπύρος-Χριστόφορος
dc.date.accessioned	2025-09-22T09:40:47Z
dc.date.available	2025-09-22T09:40:47Z
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/62510
dc.identifier.uri	http://dx.doi.org/10.26240/heal.ntua.30206
dc.rights	Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα	*
dc.rights.uri	http://creativecommons.org/licenses/by-nc-nd/3.0/gr/	*
dc.subject	Έλλειψη Δεδομένων	el
dc.subject	Υπερδειγματοληψία	el
dc.subject	Βιομηχανικός Οπτικός Έλεγχος Ποιότητας	el
dc.subject	Βαθιά Μάθηση	el
dc.subject	Αναγνώριση Καινοφανών Εισόδων	el
dc.subject	Data Scarcity	en
dc.subject	Oversampling	el
dc.subject	Deep Learning	el
dc.subject	Open-set Recognition	el
dc.subject	Industrial Visual Quality Inspection	el
dc.title	Alleviating Data Scarcity In Industrial Machine Learning Applications	en
dc.contributor.department	Τεχνολογίας Πληροφορικής και Υπολογιστών	el
heal.type	doctoralThesis
heal.classification	Τεχνητή Νοημοσύνη	el
heal.language	en
heal.access	free
heal.recordProvider	ntua	el
heal.publicationDate	2025-06-05
heal.abstract	Visual defect recognition and its manufacturing applications have been an upcoming topic in recent AI research as an integral part of the manufacturing process that is becoming increasingly automated with the advent of Industry 4.0 and Industry 5.0. While being a very beneficial solution to this problem, AI-driven Computer Vision Algorithms and Deep Neural Networks face several issues that may impede their adoption in practical real-life settings such as a manufacturing shop floor. For instance, defect datasets are often severely imbalanced and can be additionally burdened with separating classes of high visual similarity. Another issue arising during an AI classifier's continuous operation is the frequent lack of robustness to novel defects appearing for the first time. The aim of this thesis is to deal with such challenges by providing augmentations to AI solutions, either on the data or the model level, addressing real-life and benchmark scenarios from the domain of manufacturing. The initial focus is Imbalanced Learning. Although various methods of data augmentation have been proposed to mitigate class imbalances, they often fail to cope with tinier minority classes or have fidelity issues with smaller defects while, at the same time, needing significant computational resources to train. Also, augmentation based on vector-based oversampling struggles to produce high-fidelity inputs and is hard to apply on custom CNN architectures, which often perform better for this type of problem. Our work presents an image-level oversampling method based on an instance-based image generator that can be applied to any CNN directly during the training process without increasing the order of training time required. It is based on identifying a small number of the most uncertain base samples close to the estimated class boundaries and using them as seeds for augmentation. The resulting images are of high visual quality preserving small class differences, and they also improve the classifier boundary leading to higher recall scores than other state-of-the-art approaches. Aside from class imbalance, lack of real-world data as well as the strict safety constrains that need to be imposed to manufacturing AI deployments dictate the need for handling novel inputs. Such unanticipated inputs can pose a significant risk to cyber-physical applications as a resulting out-of-context decision could compromise the integrity of the production process. While recent Machine Learning methods can theoretically tackle this problem from different angles (e.g., open-set recognition, semi-supervised learning, intelligent data augmentation), applying them to a real-life setting with a small, imbalanced dataset and high inter-class similarity can be challenging. This work confronts such a use case aiming at the automation of the visual quality inspection of shaver shell brand prints from the electronics industry, which is characterized by data scarcity and the existence of small local defects. To that end, we introduce a novel data augmentation approach based on the latent space manipulation of StyleGAN, where defect data is intentionally synthesized to simulate novel inputs that can help form a boundary of the model's knowledge. Our approach shows promising results compared to well-established open-set recognition and semi-supervised methods applied to the same problem, while its consistent performance across classifier embeddings indicates lower coupling to the final classifier. The above mentioned method still requires enough data to train a GAN, which might not always be possible or cost-effective. Collecting more and more defect data is also often not a solution as defects occur rarely in production and the ramp-up time of the AI-driven quality inspector becomes significantly slower. To cope with smaller datasets we apply an innovative approach based on Neurosymbolic AI. Specifically, we use a Logic Tensor Network that expresses the outputs of an unsupervised out-of-distribution detector as symbolic rules and uses them to drive the training of a neural network classifier. The resulting algorithm shows improved results in comparison to other related methods, especially in terms of defect recall, meaning that few defects remain undetected even if completely novel. More specifically, it achieves similar or better recall scores than semi-supervised and unsupervised methods when handling novel defects, but significantly outperforms them in defects that were seen during training. Similarly, when compared to supervised methods, it maintains high performance on known defects but significantly improves on novel ones. These best-of-both-worlds results are illustrated through higher F1-scores in the majority of the test datasets of manufacturing products.	en
heal.advisorName	Τσανάκας, Παναγιώτης
heal.committeeMemberName	Κυριαζής, Δημοσθένης
heal.committeeMemberName	Αμδίτης, Άγγελος
heal.committeeMemberName	Σταφυλοπάτης, Ανδρέας-Γεώργιος
heal.committeeMemberName	Βουλόδημος, Αθανάσιος
heal.committeeMemberName	Μαρινάκης, Ευάγγελος
heal.committeeMemberName	Ξύδης, Σωτήριος
heal.academicPublisher	Σχολή Ηλεκτρολόγων Μηχανικών και Μηχανικών Υπολογιστών	el
heal.academicPublisherID	ntua
heal.numberOfPages	144
heal.fullTextAvailability	false