Use of Deep Learning for Sticker Detection During Continuous Casting

Breakout is the most expensive and dangerous issue of continuous casting, which causes the loss of production time and significant yield penalties. The common cause of breakout is sticker, that is a part of strand shell, which adheres to a mold surface. Stickers can be detected by a temperature pattern in a mold heat-map. SMS group GmbH (Germany) develops HD mold, a cyber-physical system for sticker detection by monitoring and analysis of the temperature data from the fiber optical sensors during casting. Currently, HD mold exploits an analytical sticker detection algorithm that gives a large number of false alarms. This leads to a significant loss of time and quality overheads. We design a special Convolutional Neural Network (CNN), whicht recognizes a sticker pattern and can be employed as a full-fledged substitute or an assistant of the current algorithm. The experiments show that being an assistant, CNN reduces the number of false alarms of the current algorithm by 47%.