Anlamsal-rehberli ağ, sinterleme verilerindeki eksiklikleri gideriyor

Özet (abstract)

The sintering process is a critical stage in iron and steel production, where downstream optimization relies on sensor-collected process data. However, missing data frequently occur in industrial environments, and accurate imputation is challenging due to complex spatio-temporal dependencies and diverse missing patterns. To address this problem, this paper proposes a semantic-guided multi-scale spatio-temporal imputation network, named SGMSIN. Specifically, a bidirectional architecture with a temporal decay mechanism is designed to exploit contextual information before and after missing points while dynamically adjusting feature reliability according to observation intervals. A long-term and short-term feature extraction module is further developed to capture both global trends and local fluctuations. Moreover, a scale-wise spatio-temporal interaction module is introduced to model and fuse multi-scale dynamic features. To handle spatial dependency variations caused by sparse observations, a density-aware graph convolutional network adaptively adjusts the spatial structure. Finally, a coarse-to-fine training strategy is employed to jointly recover global trends and refine local details. Experiments on real-world sintering datasets demonstrate the superiority of SGMSIN under various missing rates and modes.

Yazarların özeti; kaynağından alınmıştır. Control Engineering Practice, 2026 · DOI ↗