Incremental Multi-Step Learning MLP Model for Online Soft Sensor Modeling.

Area of Science:

• Chemical Engineering
• Data Science
• Process Control

Background:

• Industrial production faces challenges with time-varying conditions and continuous time-series data.
• Traditional soft sensor models struggle with dynamic changes, leading to suboptimal performance.
• Online analytical systems are costly, have maintenance issues, and suffer from measurement delays, hindering real-time control.

Purpose of the Study:

• To develop an adaptive and accurate soft-sensing model for industrial processes.
• To address the limitations of traditional soft sensors in dynamic and time-varying environments.
• To enable real-time monitoring and control through improved predictive capabilities.

Main Methods:

• Introduction of a multivariate multi-step predictive multilayer perceptron regression soft-sensing model (MVMS-MLP).
• Integration of incremental learning strategies for enhanced adaptability and accuracy.
• Development of a pre-trained MVMS-MLP model incorporating temporal data handling and MLP regression, followed by incremental model construction.

Main Results:

• The incremental MVMS-MLP model demonstrates enhanced adaptability to dynamic changes in industrial processes.
• The model achieves improved accuracy in multivariate predictions compared to traditional methods.
• Effectiveness validated through benchmark problems and real-world industrial case studies.

Conclusions:

• The proposed incremental MVMS-MLP offers a robust solution for real-time soft sensing in complex industrial settings.
• Incremental learning significantly improves the performance and adaptability of multivariate soft sensors.
• The method provides a viable alternative to costly and delayed online analytical systems for process control.