Multimodal computer vision and machine learning for predictive monitoring of planar-flow casting of amorphous steel foils

Methodology/Principal Findings

We present a multimodal computer vision and machine learning framework for predictive monitoring of industrial PFC using synchronised visible and infrared imagery acquired during full-scale production. More than 300,000 RGB and thermal image frames collected from approximately 200 production runs were processed to extract process-informed geometric, thermal and texture descriptors, including ridge density, ridge width, reflection characteristics and thermal-gradient features. Conventional image-processing baselines were compared with gradient-boosted decision trees (XGBoost) and a multimodal deep-learning fusion model combining RGB, infrared and process metadata. Both learning-based approaches substantially improved prediction of the nozzle-to-wheel gap compared with conventional computer-vision heuristics, achieving mean absolute errors of approximately 0.01 mm and R ² values of approximately 0.97. In contrast, prediction of the optical-flow-based melt-flow proxy remained substantially more challenging, indicating that this variable is influenced by factors not fully represented in the available imaging modalities. Feature-importance, calibration and ablation analyses demonstrated that ridge-density and thermal-gradient descriptors contributed most strongly to predictive performance, while complementary analyses showed that unstable production exhibited broader ridge-width distributions and larger edge-to-centre thermal gradients than stable operation.