A novel quality prediction model based on dual-layer graph supervised embedding with multi-granularity attention...

Area of Science:

• Industrial Process Monitoring
• Machine Learning for Manufacturing
• Data-driven Quality Control

Background:

• Traditional soft sensing struggles with complex industrial data, including nonlinear interactions and dynamic changes.
• Existing methods often fail to adequately represent multivariate couplings, limiting prediction accuracy.
• Adapting to evolving industrial settings remains a significant challenge for current quality prediction models.

Purpose of the Study:

• To develop an advanced manufacturing quality prediction model addressing limitations of traditional soft sensing.
• To enhance the representation of complex relationships within industrial process data.
• To improve the adaptability and accuracy of quality prediction in dynamic industrial environments.

Main Methods:

• Proposed a Dual-layer Graph Supervised Embedding with Multi-Granularity Attention Enhancement Mechanisms (DGS-MA) model.
• Constructed a dual-layer graph: feature similarity graph for local associations and a supervised Node2vec graph for global topology.
• Implemented a multi-granular graph attention mechanism with dual-pathway and cross-layer attention for feature fusion and explicit supervision constraints.

Main Results:

• The DGS-MA model demonstrated significant improvements in quality prediction accuracy.
• Effectively captured both local static associations and global quality-driven topology through dual-view representation.
• The attention enhancement mechanism successfully fused neighborhood information from original features and supervised embeddings.
• Explicit supervision constraints enhanced both prediction accuracy and model interpretability.

Conclusions:

• The DGS-MA model offers a robust solution for data-driven quality prediction in complex industrial settings.
• The integration of dual-layer graphs and multi-granular attention significantly outperforms traditional methods.
• This approach provides a more accurate and interpretable framework for industrial process quality monitoring.