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

Area of Science:

• Chemical Engineering
• Data Science
• Machine Learning

Background:

• Industrial processes generate complex data with temporal dependencies and high dimensionality, challenging traditional soft sensing.
• Existing soft sensor models struggle with intricate industrial data characteristics, limiting prediction accuracy.

Purpose of the Study:

• To propose a novel Semi-Supervised Sparse Stacked Autoencoder integrated with Local Linear Embedding (SS-SAE-LLE) for enhanced industrial soft sensor modeling.
• To address the limitations of traditional autoencoders in capturing spatio-temporal data features and improve prediction accuracy.

Main Methods:

• The SS-SAE-LLE algorithm combines a semi-supervised stacked autoencoder with the Local Linear Embedding algorithm.
• It leverages hierarchical feature extraction, spatio-temporal data characteristics, and supervised tuning using labeled data.
• The model is trained within a semi-supervised learning framework.

Main Results:

• Experiments on PTA solvent and SRU system datasets demonstrate the effectiveness of SS-SAE-LLE.
• The proposed method achieved higher prediction accuracy compared to existing models.
• SS-SAE-LLE effectively handles the spatio-temporal characteristics of industrial process data.

Conclusions:

• SS-SAE-LLE offers a robust solution for industrial soft sensor modeling, outperforming traditional methods.
• The integration of Local Linear Embedding enhances the model's ability to capture complex data structures.
• The findings highlight the applicability and improved performance of SS-SAE-LLE in real-world industrial settings.