A Dynamics-Learning Multirate Estimation Approach for the Feeding Condition Perception of Complex Industry Processes

Area of Science:

• Chemical Engineering
• Process Control
• Machine Learning

Background:

• Quality-related indices (QRIs) are crucial intermediate indicators in industrial processes.
• Estimating QRIs accurately is challenging due to multirate data and process delays.
• Existing methods struggle with the complexity of inter-unit process data dynamics.

Purpose of the Study:

• To develop a dynamics-learning multirate estimation approach for QRIs.
• To address the challenges of multirate data and process delays in quality estimation.
• To improve the accuracy and efficiency of quality-related index perception.

Main Methods:

• A two-stage estimation problem utilizing production data from adjacent unit processes.
• Dynamics-learning bidirectional long short-term memory (BiLSTM) with tailored inputs for forward/backward layers.
• Integration of a cycle control gate within BiLSTM to capture QRIs dynamics.
• Combination with a Bayesian estimation model to manage process delays.

Main Results:

• The proposed approach effectively estimates quality-related indices (QRIs) under multirate conditions.
• Demonstrated high-rate estimation capability, outperforming existing methods.
• Ablation and comparative experiments validated the feasibility and effectiveness of the dynamics-learning BiLSTM model.

Conclusions:

• The dynamics-learning multirate estimation approach offers a robust solution for real-time QRIs perception.
• The integration of BiLSTM with a cycle control gate and Bayesian model enhances estimation accuracy and speed.
• This method provides significant advancements in process monitoring and quality control.