SENGraph: A Self-Learning Evolutionary and Node-Aware Graph Network for Soft Sensing in Industrial Processes.

Area of Science:

• Industrial Process Monitoring
• Machine Learning for Chemical Engineering

Background:

• Traditional soft sensing models often overlook intricate relationships within industrial process data, limiting their effectiveness.
• Existing graph network approaches struggle to simultaneously model inter-variable structures and intra-variable temporal dependencies in dynamic industrial settings.
• A key challenge is learning from nodes with varying importance for accurate soft sensing.

Purpose of the Study:

• To develop an advanced graph network model, SENGraph, for enhanced industrial soft sensing.
• To address limitations in capturing complex industrial data relationships and node importance.
• To improve the accuracy and robustness of soft sensing in dynamic industrial processes.

Main Methods:

• Proposed a self-learning graph generation (SLG) module to create combined coarse- and fine-grained graphs.
• Implemented a self-evolutionary graph module (EGM) using genetic strategies for diversified node feature learning.
• Designed a node-aware module (NAM) to prioritize informative nodes and down-weight less significant ones.

Main Results:

• SENGraph successfully captures both global trends and local dynamics in industrial data.
• The evolutionary and node-aware modules enhance the model's ability to discern critical process variables.
• Demonstrated superior performance compared to state-of-the-art methods on four real-world industrial datasets.

Conclusions:

• SENGraph offers a significant advancement in industrial soft sensing by effectively modeling complex data structures.
• The proposed approach addresses critical challenges in graph-based soft sensing, improving discriminative power.
• The model's effectiveness is validated through extensive experiments on diverse industrial datasets.