Crack detection in sheet metal parts with dynamic tests using deep learning algorithms

In recent years, quality control and precision measurement processes have begun to play a critical role in modern manufacturing industries, particularly in the automotive industry. Traditional inspection processes rely heavily on human oversight, resulting in significant limitations such as subjective judgments, inconsistencies, and insufficient response times in high-frequency testing environments. In this thesis, a deep learning-based method has been developed to detect cracks in sheet metal components operating under dynamic and repetitive loads at an early stage. A large and balanced dataset has been created for pre-crack, crack formation (microcrack–early stage), and advanced damaged stages. It includes YOLOv8, Faster R-ESA, Mask R-ESA, and classic ESA architectures trained on a comprehensive dataset of high-resolution images labeled using the Roboflow platform. The data was obtained from both steel defect datasets available on Kaggle and industrial tests. Model performance was improved using hyperparameter tuning techniques such as Genetic Algorithms, Artificial Bee Colony, and Ant Colony Optimization.Experimental validation was performed using a specially designed test setup that creates cyclic stress conditions via hydraulic pistons operating at 220 bar pressure. The findings demonstrate that deep learning algorithms are significantly more successful than traditional threshold-based image processing methods, particularly in detecting microscopic cracks prior to fracture. This study is unique in that it is the first in the literature to perform real-time crack detection under dynamic test conditions using a Raspberry Pi Camera three-based imaging system. Furthermore, a hybrid dataset combining industrial and laboratory data was created to adapt deep learning models to industrial reality. Furthermore, it not only contributes to defect recognition technologies but also presents a scalable and cost-effective solution for quality assurance in industrial forming processes. The findings aim to contribute to safer and more sustainable production cycles in the automotive spare parts sector.