Enhancing Fermentation Process Monitoring through Data-Driven Modeling and Synthetic Time Series Generation.

Area of Science:

• Biotechnology
• Chemical Engineering
• Data Science

Background:

• Soft sensors utilizing deep learning regression models show promise for real-time fermentation quality prediction.
• Sparse, outlier-prone experimental datasets hinder model performance.
• A fully distributed solution space is needed for effective model training.

Purpose of the Study:

• To enhance the robustness and predictive capability of soft sensor models.
• To address data scarcity and quality issues in fermentation datasets.
• To improve deep learning model performance through synthetic data generation.

Main Methods:

• Variational autoencoders (VAEs) were used to generate synthetic datasets.
• Synthetic datasets were combined with original experimental data.
• Neural network regression models were trained on original and augmented datasets for soft sensor development.
• Performance was evaluated using R² scores on intensified ethanol fermentation data.

Main Results:

• Soft sensor predictive capability improved by 34% using augmented datasets.
• Data variability was reduced by 82% with the proposed method.
• Enhanced models demonstrated superior performance compared to those trained solely on experimental data.

Conclusions:

• Synthetic data generation via VAEs effectively improves soft sensor reliability and robustness.
• The method offers significant time and cost savings for deep learning in fermentation processes.
• This approach is adaptable to various fermentation monitoring applications, advancing soft sensor technology.