Adaptive soft sensor using stacking approximate kernel based BLS for batch processes.

Area of Science:

• Chemical Engineering
• Machine Learning
• Data Science

Background:

• Batch processes exhibit complex nonlinear and time-varying dynamics.
• Traditional Broad Learning Systems (BLS) face challenges with noisy data and performance uncertainty due to random mappings.
• Existing methods struggle to adapt effectively to gradual changes in industrial batch processes.

Purpose of the Study:

• To introduce an adaptive stacking approximate kernel based broad learning system (AKBLS) for robust batch process modeling.
• To enhance the nonlinear fitting, generalization, and adaptive capabilities of broad learning systems.
• To improve prediction accuracy and computational efficiency for industrial online applications.

Main Methods:

• Developed an Approximate Kernel based Broad Learning System (AKBLS) algorithm to reduce uncertainty and improve prediction accuracy by projecting feature nodes into kernel space.
• Implemented an Adaptive Stacking framework using ensemble learning with a meta-learner to integrate multiple AKBLS models.
• Utilized a moving window method for adaptive ability to handle gradual changes in batch processes.
• Optimized kernel matrix computation by searching for approximate kernels.

Main Results:

• The proposed AKBLS algorithm demonstrated reduced uncertainty and improved prediction accuracy.
• The Adaptive Stacking framework enhanced generalization capabilities.
• The moving window method provided adaptive ability to changing process dynamics.
• Extensive experiments on public datasets and penicillin simulations validated superior performance over common algorithms.

Conclusions:

• The adaptive stacking approximate kernel based broad learning system offers a powerful solution for modeling complex batch processes.
• The model exhibits strong nonlinear fitting, generalization, and adaptive abilities, outperforming existing methods.
• This approach is suitable for industrial online applications requiring accurate and adaptive predictions.