University of Thessaly (UTH)
Optimization of deep reinforcement learning and fuzzy logic hybrid methodologies with application to smart manufacturing
Abstract
dc:descriptionSmart manufacturing increasingly relies on optimizing production parameters to improve quality and reliability, with Deep Learning (DL) now the prominent approach for visual defect detection and segmentation. Its performance, however, remains constrained by illumination conditions, typically configured manually through trial and error rather than actively optimized—particularly problematic for reflective metallic surfaces common in industrial manufacturing. This dissertation develops Deep Reinforcement Learning (DRL) methodologies that treat illumination as a controllable quantity optimized directly for downstream inspection performance, validated on real industrial use cases. The dissertation develops this argument across three connected studies of illumination as a controllable acquisition variable. It first shows that illumination can be actuated reliably by a learned policy guided by a perceptually weighted image-quality reward, combining entropy and luminance, under real disturbances on a physically instrumented vehicle-documentation system. It then shows that acquisition-related artifacts, particularly those associated with surface specularity, persist across different segmentation architectures under fixed illumination, while also addressing the latency constraints required for deployment. Building on both, a task-driven reinforcement learning framework selects illumination directly against the performance of a vision-based segmentation model, guided by a segmentation-driven reward benchmarked against alternative reward formulations and image-quality-driven baselines, outperforming fixed and perceptually-driven alternatives in aggregate, with negligible added overhead, and validated on complete production-scale surfaces. Finally, this soft-sensor-coupled paradigm is extended to process calibration in microelectronics manufacturing through hybrid DL, RL, and Fuzzy Logic methodologies: a Q-learning agent selects pressure adjustments from a visual soft sensor’s state estimate, while a fuzzy-weighted action-selection strategy incorporates the complete class-probability distribution under ambiguous state estimates, improving control performance and stability. This extension provides evidence that the proposed paradigm can transfer beyond illumination to a distinct controllable process variable, supporting its broader applicability to autonomous process optimization in industrial manufacturing.
Degree
thesis:*- Grantor dc:publisher
- University of Thessaly (UTH)
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Leontaris, Lampros
- Λεοντάρης, Λάμπρος
Subjects
dc:subject × 16- Βαθειά μάθηση
- Βαθιά ενισχυτική μάθηση
- Οπτικός ποιοτικός έλεγχος
- Ασαφής λογική, Έλεγχος με
- Μηχανική όραση
- Deep learning
- Deep reinforcement learning (DRL)
- Optical quality control
- Fuzzy logic control
- Machine vision
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/62872