Online-Dynamic-Clustering-Based Soft Sensor for Industrial Semi-Supervised Data Streams.

Area of Science:

• Process Systems Engineering
• Data Science
• Machine Learning

Background:

• Industrial processes generate high-speed, sequential data streams.
• Traditional soft sensors built offline fail with changing process states (concept drift).
• Real-world data streams exhibit nonlinearity, time-variability, and label scarcity.

Purpose of the Study:

• To develop a soft sensor model for real-time quality variable prediction in industrial semi-supervised data streams.
• To address challenges of nonlinearity, time-variability, and label scarcity in process data.
• To improve the accuracy and reliability of soft sensors in dynamic industrial environments.

Main Methods:

• Online dynamic clustering for automatic cluster generation, sample deletion, and process state identification.
• Selective ensemble learning and just-in-time learning (JITL) with adaptive switching for handling gradual and abrupt changes.
• Semi-supervised learning to leverage unlabeled data and generate pseudo-labeled samples.

Main Results:

• The proposed Online Dynamic Clustering-based Soft Sensor (ODCSS) effectively identifies process states online.
• The adaptive switching strategy successfully mitigates performance degradation due to concept drift.
• Semi-supervised learning enhances the training dataset by utilizing unlabeled data.
• Case studies demonstrate the superiority of ODCSS over conventional soft sensors in semi-supervised data streams.

Conclusions:

• The ODCSS approach provides a robust solution for real-time quality prediction in industrial data streams.
• The method effectively handles nonlinearity, time-variability, and label scarcity.
• ODCSS offers superior performance compared to traditional soft sensors in dynamic, semi-supervised environments.