Gated Stacked Target-Related Autoencoder: A Novel Deep Feature Extraction and Layerwise Ensemble Method for...

Area of Science:

• Chemical Engineering
• Data Science
• Artificial Intelligence

Background:

• Data-driven soft sensors are crucial for estimating unmeasurable quality variables in industrial processes.
• Extracting effective features from complex process data remains a significant challenge in soft sensing.
• Deep learning (DL) offers powerful nonlinear modeling and feature extraction capabilities for process monitoring and quality prediction.

Purpose of the Study:

• To propose a novel deep learning approach, the gated stacked target-related autoencoder (GSTAE), to enhance soft sensor modeling performance.
• To address limitations of conventional stacked autoencoders (SAE) by integrating target-related information during pretraining and optimizing feature utilization.
• To improve the accuracy and reliability of quality variable estimation in industrial settings.

Main Methods:

• Introduced a novel gated stacked target-related autoencoder (GSTAE) for soft sensor model construction.
• Incorporated target value prediction errors into the loss function during layerwise pretraining to guide feature learning.
• Utilized gated neurons to control information flow from different hidden layers, quantifying contributions of various abstraction levels.

Main Results:

• The proposed GSTAE method demonstrated improved modeling performance compared to conventional approaches.
• Target-related information effectively guided the feature learning process, leading to more relevant representations.
• Gated neurons successfully leveraged multi-level features and quantified their impact on the final prediction.
• The approach was validated in two real-world industrial case studies.

Conclusions:

• The GSTAE provides an effective method for constructing high-performance soft sensors by leveraging deep learning.
• Integrating target information and utilizing gated neurons enhances feature extraction and modeling accuracy in industrial processes.
• The validated effectiveness in industrial cases highlights the practical applicability of the proposed GSTAE approach.