Integrating Autoencoder and Heteroscedastic Noise Neural Networks for the Batch Process Soft-Sensor Design.

Area of Science:

• Chemical Engineering
• Data Science
• Process Control

Background:

• Real-time in-process viscosity measurement is crucial for product quality but challenging.
• Complex mixing and dynamic conditions hinder traditional measurement techniques.

Purpose of the Study:

• To develop an innovative soft sensor for real-time viscosity prediction.
• To address the limitations of in-process viscosity measurement in dynamic industrial settings.

Main Methods:

• Utilized a deep learning autoencoder for feature extraction from high-dimensional industrial data.
• Employed a heteroscedastic noise neural network for simultaneous viscosity prediction and uncertainty estimation.
• Benchmarked against Gaussian process and Bayesian neural network models.

Main Results:

• The soft sensor achieved high accuracy and reliability in predicting product viscosity across industrial batches.
• Demonstrated superior performance compared to established probabilistic machine learning techniques.
• Successfully compressed high-dimensional industrial data for effective feature extraction.

Conclusions:

• The proposed soft sensor offers a viable solution for accurate and reliable real-time viscosity monitoring.
• Highlights the potential of advanced artificial neural networks in industrial process control and quality assurance.
• Provides a robust tool for managing product quality in dynamic manufacturing environments.