A Graph-Based Time-Frequency Two-Stream Network for Multistep Prediction of Key Performance Indicators in Industrial...

Area of Science:

• Industrial Process Control
• Machine Learning
• Soft Sensor Modeling

Background:

• Deep learning soft sensor models primarily focus on real-time, current-step predictions.
• Industrial applications increasingly require advance multistep predictions for key performance indicators.
• Existing methods struggle with complex process variable coupling and long-term dependency learning.

Purpose of the Study:

• To develop an advanced soft sensor model for accurate multistep prediction.
• To address limitations in modeling complex variable relationships and long-term dependencies.
• To enhance industrial process monitoring and control through predictive capabilities.

Main Methods:

• Proposed a graph-based time-frequency two-stream network for multistep prediction.
• Introduced a multigraph attention layer to model dynamical coupling between process variables.
• Utilized a time-frequency two-stream network with multi-GAT for extracting time and frequency domain features.
• Implemented a feature fusion module based on minimum redundancy and maximum correlation.

Main Results:

• The proposed model significantly outperforms state-of-the-art methods on real-world industrial datasets.
• Demonstrated substantial improvements in prediction accuracy (RMSE, MAE, MAPE) for the three-step prediction task.
• Achieved 12.40% RMSE, 22.49% MAE, and 21.98% MAPE improvement on the waste incineration dataset.

Conclusions:

• The graph-based time-frequency network effectively handles complex variable couplings and long-term dependencies.
• The developed model offers superior performance for industrial multistep prediction tasks.
• This approach provides a valuable tool for enhancing predictive maintenance and operational efficiency in industrial settings.