Development of artificial intelligence-supported quality control system in industrial production lines

Image processing technology is used in many industries, especially in the automotive sector, and its technological infrastructure continues to develop day by day. In the manufacturing sector, image processing systems are frequently utilized to ensure product quality and optimize production processes. Image processing technology is used to inspect the surfaces of manufactured parts and detect defects such as roughness, cracks, holes, or other imperfections. Camera systems scan product surfaces with high-resolution images, and algorithms detect any flaws that do not meet predetermined criteria. The dimensions and geometry of manufactured parts are measured using image processing technology. Camera systems check whether parts conform to specified dimensions and tolerances, issuing warnings in case of any deviation. In addition, image processing technology is widely used on assembly lines in various production and control stages, such as verifying whether parts are correctly assembled or ensuring label accuracy. This thesis aims to develop suitable methods for detecting and classifying defects on industrial parts by leveraging machine learning techniques. Initially, a vision-based quality control system was developed to identify dimensional defects in parts according to predefined tolerances. Subsequently, a dataset comprising 2,650 images was created to address issues such as scratches and burrs on parts. This dataset was divided into three main classes: defect- free, scratched, and burr-containing. 80% of the dataset was used for training, while 20% was reserved for testing. DWT, CWT and UWT methods were applied to the images. The features extracted from the images were fed into various types of classifiers to determine the most effective machine learning method. After the feature extraction process, classification methods such as artificial neural networks, support vector machines, k-nearest neighbors and decision trees were tested and evaluated. Subsequently, a deep learning approach was employed to obtain more distinctive features using the basic textural properties extracted from part images. As a result of the study, dimensional inspections of industrial parts were performed with deviations ranging between 0.003 and 0.05. The machine learning method implemented using a combination of UWT and DT achieved a 98.2% success rate in detecting and classifying visual defects in industrial parts. In the deep learning process, images obtained using Gabor filters were trained with the YOLOv5 model, enabling the high- accuracy classification of scratches and burrs. The methods developed in this thesis, including both dimensional control and AI-based defect detection, provide a practical and applicable solution for quality control of parts at the output stage of industrial production lines.