RegGAN: A Virtual Sample Generative Network for Developing Soft Sensors with Small Data.

Area of Science:

• Chemical Engineering
• Process Monitoring
• Artificial Intelligence

Background:

• Critical quality variables in chemical production often lack online measurement capabilities.
• Soft sensors are essential for real-time monitoring but require extensive labeled data for development.
• Acquiring sufficient labeled data for soft sensors is challenging due to time and cost constraints.

Purpose of the Study:

• To propose a novel regression generative adversarial network (GAN) for generating virtual samples.
• To address the challenge of insufficient labeled data in developing high-performance soft sensors.
• To improve the predictive performance of soft sensors in chemical production systems.

Main Methods:

• Developed a regression GAN that learns the data distribution and mapping between auxiliary and target variables.
• Incorporated an importance-weighted autoencoder to enhance the stability of the generative model training.
• Utilized a similarity measurement algorithm to select relevant virtual samples for integration into the training set.

Main Results:

• The proposed GAN effectively generates virtual samples that closely resemble actual data.
• Virtual samples generated by the proposed method showed higher proximity to real samples than other approaches.
• Integrating virtual samples into a long short-term memory (LSTM)-based soft sensor reduced root-mean-square error by 21.03%.

Conclusions:

• The proposed generative method successfully creates valuable virtual samples to augment limited real-world data.
• The integration of these virtual samples significantly enhances the predictive accuracy of soft sensors.
• This approach offers a viable solution for improving real-time process monitoring in chemical production.