A Dynamically Adaptive Soft Sensor for End‐Point Carbon Content and Temperature Prediction in Multimodal Basic Oxygen Furnace Steelmaking using a Meta‐Learning Stacked Autoencoder

Accurate prediction of endpoint carbon and temperature is critical for basic oxygen furnace (BOF) steelmaking. Traditional soft sensor models often assume uniform or single‐mode data distributions, making them ineffective under frequently changing conditions and in a multimodal industry. Sensor aging further introduces data drift, increasing modeling complexity. This article proposes a meta ‐learning‐guided dynamic adaptive soft sensor method using Stacked Autoencoders (SAE). First, historical data are clustered using a Wasserstein distance‐weighted Dirichlet Process Gaussian mixture model to identify operating modes automatically. Then, a meta ‐learned SAE is trained across all modes to enhance generalization. For new samples, their posterior mode is inferred, and a most similar historical sample set is selected via mutual‐information‐weighted Jensen–Shannon divergence. This set is used to fine‐tune the model in real time, enabling adaptation to data drift. Experiments on real‐world BOF steelmaking data validate the method's effectiveness, demonstrating improved robustness, and prediction accuracy under complex, multimodal conditions.