Autoencoder-based nonlinear Bayesian locally weighted regression for soft sensor development.

Area of Science:

• Process control and automation
• Machine learning applications
• Soft sensor development

Background:

• Locally weighted learning (LWL) is widely used for nonlinear soft sensors in industry.
• Accurate localized models are crucial for high-performing LWL-based soft sensors.
• Existing methods face challenges with process nonlinearities and uncertainties.

Purpose of the Study:

• To propose a novel nonlinear local model training algorithm, nonlinear Bayesian weighted regression (NBWR).
• To enhance the performance of LWL-based soft sensors by improving localized model accuracy.
• To address limitations in handling process nonlinearities, uncertainties, overfitting, and numerical issues.

Main Methods:

• Feature extraction using autoencoders to capture nonlinear process data characteristics.
• Local dataset selection in feature space based on query samples.
• Development of a fully Bayesian regression model with differentiated sample weights.
• Integration of NBWR within the locally weighted learning framework for soft sensor construction.

Main Results:

• The NBWR algorithm demonstrated superior performance in developing soft sensors.
• The method effectively handles process nonlinearities and uncertainties.
• Improved consistency of correlation and mitigation of overfitting and numerical issues were observed.
• Experimental evaluation on a real-world industrial process confirmed superior performance compared to benchmarking methods.

Conclusions:

• The proposed nonlinear Bayesian weighted regression (NBWR) method significantly advances LWL-based soft sensor technology.
• NBWR offers a robust solution for accurate localized modeling in complex industrial processes.
• The approach provides enhanced performance and reliability for soft sensing applications.