Integrating YOLO object detection and numerical simulation for the closed-loop optimization of squeeze casting processes

Abstract

To address the difficulties in extracting minute defect features and the lag in traditional open-loop process parameter optimization during high-pressure forming of 4032 wrought aluminum alloy scroll disks, this paper proposes an adaptive closed-loop control architecture that integrates a multi-scale frequency-domain enhanced YOLO algorithm and a physical field Bayesian surrogate model. At the visual inspection front end, a two-dimensional discrete wavelet frequency reduction and preservation operator and a cross-layer adaptive attention module are introduced to improve the recognition accuracy of low-contrast defects on reflective surfaces. At the process decision-making level, a surrogate network based on the solidification phase transition gradient and the inverse mapping of visual spatial coordinates is constructed to achieve online closed-loop compensation of servo injection execution parameters. Industrial dual-station test results show that the improved visual inspection model achieves an mAP@0.5:0.95 index of 85.7% with an edge inference delay of 28.4 ms. Under closed-loop optimization, the mold thermal balance disturbance amplitude of the forming equipment is reduced by 87.0%, and the overall casting scrap rate converges from 8.45% in the open-loop state to 0.82%. This study verifies the application value of cross-modal feature perception and physical mechanism mapping in the automated quality control of complex metal precision components.