• 목록

H-beam rolling is a high-dimensional, nonlinear, multi-pass forming process that must satisfy several quality criteria at once, chiefly cross-sectional shape accuracy and material soundness. Finite element method (FEM) analysis is accurate but far too costly for iterative multi-objective optimization, at about eight hours per four pass condition. Moreover, the decisive quality is the spatial distribution of the cross-sectional shape — image-form information — yet existing surrogate-based optimization has pursued productivity-centered scalar objectives, not such image-form quality. This paper proposes a surrogate-assisted multi-objective Pareto optimization framework for the final four-pass stage. A pre-trained convolutional neural network (CNN) image surrogate evaluates the final shape and equivalent-strain distribution of each candidate, and NSGA II searches a constrained feasible space to optimize three image-derived objectives spanning shape and material quality. In an industrial case study, it obtained 30 Pareto-optimal solutions with a hypervolume of 0.6743 and quantified the trade-off structure — a strong shape-accuracy versus uniformity conflict in which the strain objective acts as a stabilizing regularizer. The full search completes in about 10.63 minutes, over