Abstract: To meet the stringent requirements for high surface quality and flatness in the finishing process of grain-oriented silicon steel, and to address the issue of surface damage caused by traditional contact profilometers, this paper proposed a non-contact online flatness detection method based on 3D line laser scanning. In accordance with silicon steel flatness control standards, the system achieves real-time reconstruction of the full-width 3D topography by employing encoder-synchronized triggering, Gaussian noise reduction, point cloud tilt correction, and dual-sensor image fusion. Furthermore, a convolutional neural network (CNN) is utilized to realize the automatic classification and localization of typical defects, such as edge waves and buckles. The system has been successfully applied to the flatness control line of a silicon steel plant. It achieves an identification accuracy of over 95% for typical flatness defects, effectively avoiding the damage to the insulating coating and surface quality associated with contact measurement. By generating intelligent shearing strategies based on the defect distribution of the entire coil, the system guided precise shearing in downstream processes, reducing the flatness concession rate from 17.32% to 8.67%. This significantly improves the yield rate and the level of intelligent production for silicon steel.

Key words: grain-oriented silicon steel, 3D laser scanning, flatness defect, online detection, strip quality control