Novel semi-supervised sparse stacked autoencoder integrated with local linear embedding for industrial soft sensing

Area of Science:

• Industrial Process Control
• Machine Learning
• Data Science

Background:

• Industrial soft sensor modeling is crucial for predicting key variables in complex processes.
• Intricate industrial processes generate data with temporal dependencies, high dimensionality, and local structures, challenging traditional soft sensing.
• Existing methods often focus on global fitting errors, neglecting spatio-temporal data characteristics.

Purpose of the Study:

• To propose a novel soft sensor modeling technique that addresses the challenges of complex industrial data.
• To improve the prediction accuracy of industrial soft sensors by incorporating spatio-temporal data features and semi-supervised learning.
• To develop a method that leverages both labeled and unlabeled data for enhanced soft sensor performance.

Main Methods:

• A novel Semi-Supervised Sparse Stacked Autoencoder integrated with the Local Linear Embedding algorithm (SS-SAE-LLE) was developed.
• The Local Linear Embedding algorithm was employed to capture spatio-temporal data characteristics.
• A semi-supervised learning framework was utilized, incorporating supervised tuning with labeled data.

Main Results:

• The SS-SAE-LLE model demonstrated higher prediction accuracy compared to other benchmark models.
• Experiments conducted on PTA solvent and SRU system datasets validated the model's effectiveness.
• The method successfully accounted for spatio-temporal data characteristics, leading to improved soft sensing performance.

Conclusions:

• The proposed SS-SAE-LLE method offers a significant advancement in industrial soft sensor modeling.
• The integration of Local Linear Embedding and semi-supervised learning enhances the ability to model complex industrial processes.
• SS-SAE-LLE shows strong applicability and potential for real-world industrial soft sensor applications.