A Soft Sensor Modeling Method Based on Local Migration Modeling Framework.

Area of Science:

• Biotechnology
• Machine Learning
• Process Engineering

Background:

• Pichia pastoris fermentation presents challenges in soft sensor modeling due to multi-stage characteristics and data heterogeneity.
• Existing methods struggle with low model fitting accuracy and prediction capability in dynamic fermentation environments.

Purpose of the Study:

• To develop a novel soft sensor modeling method using deep transfer learning (DTL) for Pichia pastoris fermentation.
• To address data distribution heterogeneity and improve prediction accuracy in multi-condition fermentation processes.

Main Methods:

• A local transfer modeling framework partitioned fermentation data into sub-source domains using K-means clustering.
• Deep Neural Networks (DNNs) were optimized with an improved firefly algorithm for prediction models.
• Correlation analysis identified the most relevant sub-source domain for deep transfer fine-tuning to create the target model.

Main Results:

• The proposed DTL method effectively extracts local feature information from fermentation data.
• Enhanced prediction accuracy and improved model generalization performance were achieved.
• The method demonstrated viability for soft sensor modeling in diverse Pichia pastoris fermentation scenarios.

Conclusions:

• The novel soft sensor modeling approach significantly improves upon existing methods for Pichia pastoris fermentation.
• DTL offers a robust solution for handling data heterogeneity in bioprocess monitoring.
• This study provides a valuable tool for optimizing fermentation processes through accurate soft sensor development.