Using source data to aid and build variational state-space autoencoders with sparse target data for process...

Area of Science:

• Industrial Process Monitoring
• Machine Learning
• Chemical Engineering

Background:

• Industrial processes often face challenges in monitoring due to rapidly changing production grades and scarce data in specific operating regions.
• Nonlinearity, process dynamics, and uncertainty complicate the development of reliable process monitoring models.

Purpose of the Study:

• To propose a novel source-aided variational state-space autoencoder (SA-VSSAE) for enhanced industrial process monitoring.
• To address the challenge of sparse target data by enabling information sharing from source grades.
• To develop a one-step procedure for information sharing and modeling, avoiding information loss.

Main Methods:

• Integration of variational state-space autoencoder (VSSAE) with Gaussian mixture models.
• Utilizing neural networks within VSSAE to extract dynamic and nonlinear features from process variables.
• Incorporating process uncertainty into a probabilistic framework for feature description.
• Development of monitoring indices for fault detection based on probability density estimates of residual and latent variables.

Main Results:

• The proposed SA-VSSAE method demonstrates enhanced reliability in monitoring processes with sparse target data.
• The one-step information sharing and modeling procedure effectively prevents information loss.
• The VSSAE component successfully captures dynamic and nonlinear process features.
• Validation through a numerical example and an industrial polyvinyl chloride drying process confirms the method's advantages.

Conclusions:

• SA-VSSAE offers a superior approach for monitoring industrial processes with limited data by leveraging information from related source grades.
• The probabilistic VSSAE framework provides a robust way to handle process uncertainty and detect faults.
• The proposed method outperforms traditional state-space models and two-step information sharing techniques.