A Hybrid Approach to Longitudinal Crack Detection with Computer Vision, Image Processing, and Machine Learning

In the continuous casting process, longitudinal cracks are typical abnormalities that can affect smooth production and slab quality. It is essential to detect these cracks using effective and reliable methods. Since longitudinal cracks are accompanied by air gaps that disrupt the heat transfer between the molten steel and the mold, this paper proposes using thermocouples installed on the mold copper plate to capture the abnormal temperature signals generated during the formation of longitudinal cracks. These signals are visualized in a 2D plane that integrates both time and space. The abnormal longitudinal crack regions are then searched by leveraging image detection algorithms from computer vision libraries. The search results are further identified using machine learning models, leading to the development of a hybrid detection model that combines computer vision, image detection, and machine learning. The results indicate that the model effectively identifies the temperature signals associated with longitudinal cracks and converts them into bounding boxes on thermographic images. This approach reduces the false positive rate while ensuring that all crack samples are identified, providing a new data-driven perspective for abnormality detection in the continuous casting process.