ProMix-DGNet: A Process-Aware Spatiotemporal Network for Sintering System Prediction.

Area of Science:

• Industrial Process Control
• Machine Learning for Manufacturing
• Spatiotemporal Data Analysis

Background:

• Multistep-ahead prediction of critical states in iron ore sintering is crucial for production stability, energy efficiency, and emission reduction.
• Existing spatiotemporal graph neural networks (STGNNs) struggle with challenges like large time delays, strong coupling, and condition drifts inherent in sintering processes.

Purpose of the Study:

• To develop an advanced deep learning model capable of robustly predicting critical states in iron ore sintering processes.
• To address the limitations of current STGNNs in handling complex industrial environments characterized by significant time delays and process coupling.

Main Methods:

• Proposes Process-aware Mixed Dynamic Graph Network (ProMix-DGNet) integrating a Decoupled Two-Stream Topology Learning mechanism.
• Employs an Adaptive Static Graph fused with a Radial Basis Function (RBF)-driven Dynamic Graph Constructor for robust spatial modeling.
• Introduces a Process-View Global Mixer to capture long-range process coupling and a future control-informed module with BiLSTM for decoding.

Main Results:

• ProMix-DGNet demonstrated superior performance over mainstream baselines on two real-world industrial datasets (Sinter-A and Sinter-B).
• The model achieved significant improvements in accuracy and robustness, as evidenced by metrics like Mean Absolute Error (MAE) and Root Mean Square Error (RMSE).

Conclusions:

• ProMix-DGNet offers a reliable framework for intelligent monitoring and closed-loop optimization of complex sintering processes.
• The model's ability to handle large time delays and process coupling provides a significant advancement in industrial AI applications.