A Deep Probabilistic Transfer Learning Framework for Soft Sensor Modeling With Missing Data.

Area of Science:

• Process Systems Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Data-driven soft sensors are crucial in the process industry but often suffer from insufficient labeled data.
• Transferring knowledge from related operating conditions is key to enhancing soft sensor performance on target applications.

Purpose of the Study:

• To introduce a novel deep transfer learning framework for soft sensor modeling.
• To address the challenges of limited labeled data and knowledge transfer in data-driven soft sensors.
• To develop a method capable of handling missing data in target operating conditions.

Main Methods:

• Proposes a Deep Probabilistic Transfer Regression (DPTR) framework.
• Utilizes a deep generative regression model within a stochastic gradient variational Bayes framework to learn Gaussian latent features.
• Implements a probabilistic latent space transfer strategy to minimize discrepancies between source and target data.
• Extends DPTR to effectively handle missing process data using generative and reconstruction capabilities.

Main Results:

• Demonstrates successful knowledge transfer from source to target operating conditions.
• Significantly enhances soft sensor performance by leveraging related data.
• Effectively manages and imputes missing values in target process data.
• Validates the DPTR framework's effectiveness on an industrial multiphase flow process.

Conclusions:

• The DPTR framework offers a robust solution for data-driven soft sensor modeling, particularly when labeled data is scarce.
• Deep transfer learning, combined with probabilistic latent space transfer, is a powerful approach for improving soft sensor accuracy.
• The proposed method's ability to handle missing data makes it highly practical for real-world industrial applications.