A novel adaptive soft sensing framework for label delay in industrial data streams.

Area of Science:

• Chemical Engineering
• Data Science
• Machine Learning

Background:

• Industrial data streams often have delayed quality variables, hindering adaptive soft sensor performance.
• Timely updates are crucial for accurate soft sensor models in dynamic environments.

Purpose of the Study:

• To propose a novel framework, Adaptive Soft Sensor for Label Delay (ASSLD), to address challenges of label delay in industrial data streams.
• To enhance the online adaptability and accuracy of soft sensors in the presence of data acquisition delays.

Main Methods:

• Developed an adaptive multilevel regression model that integrates outputs from different network depths.
• Implemented an online diverse database for efficient reuse of historical labeled data through sample selection and weighting.
• Utilized unlabeled samples within the delay period to adapt to recent data.

Main Results:

• The proposed ASSLD framework demonstrated effectiveness in handling label delay.
• Achieved accuracy improvements exceeding 12% compared to existing baseline methods.
• Validated on sulfur recovery unit and polyester datasets.

Conclusions:

• ASSLD provides a robust solution for soft sensor applications with delayed quality variables.
• The integration of multilevel regression and diverse data management enhances model adaptability and predictive accuracy.
• This framework offers a significant advancement for real-time quality monitoring in industrial processes.