Investigation of burrs formed on the material surface during drilling with image processing and optimization of cutting parameters

In recent years, efforts to improve quality and efficiency in manufacturing technologies have accelerated, leading to the adoption of automated defect detection and classification methods, particularly in machining processes. With the advancement of computer vision and artificial intelligence approaches, deep learning models have emerged as an alternative to conventional burr measurement and manual inspection methods, offering faster, more cost-effective, and more repeatable solutions. In this thesis study, the aim is to classify the burrs formed after drilling operations on AA6013 aluminum alloy using the YOLOv10n model on datasets created from microscope images. In addition to the original dataset, the dataset was augmented by applying horizontal flipping (HF), 180° rotation (R180°), and their combinations (CF) to the original dataset to improve the model's generalization ability. Furthermore, different decision threshold values (k.e. = 0.25 and 0.65) were applied to each dataset to evaluate the model's sensitivity and selectivity. During training on these datasets, good–medium–poor classes were predicted, and performance metrics such as accuracy, precision, recall, and F1 score were calculated. As a result of evaluating the established system, the highest performance was achieved with dataset 8, which had the lowest decision threshold and the largest number of data sets (Accuracy=0.985; F1=0.977). A significant decrease in success was observed in non-augmented or single-augmented Datasets from 1 to 4, while combined augmentations improved generalization ability by datasets from 5 to 7. Furthermore, it was determined that Recall decreased and false negatives (FN) increased as the threshold value increased. In the study, the model camera system obtained from training was integrated into real-time application, enabling the instant display of the spatter class on the screen. The results demonstrate that the proposed approach can be applied in industrial production lines and quality control processes. Accordingly, the system is recommended to planning and process engineers as a practical method for fast and reliable burr classification, particularly by using low/medium confidence thresholds and combined data augmentation strategies.