Soft sensor modeling method for Pichia pastoris fermentation process based on substructure domain transfer learning.

Area of Science:

• Biotechnology
• Chemical Engineering
• Data Science

Background:

• Traditional transfer learning methods for soft sensor models often suffer from data loss and domain mismatch, reducing prediction accuracy.
• Existing approaches struggle to perfectly align source and target domains, hindering effective model transfer.

Purpose of the Study:

• To develop an advanced soft sensor modeling method that overcomes limitations of traditional domain-level transfer.
• To enhance the accuracy and robustness of soft sensor models by utilizing a substructure domain transfer framework.

Main Methods:

• Clustering source and target domains into substructure domains using Gaussian mixture models.
• Adaptively weighting substructure domains based on inter-domain distances.
• Employing optimal subspace domain adaptation for spatial alignment of data.
• Utilizing least squares support vector machine for final model prediction.

Main Results:

• The proposed method significantly reduces prediction errors for Pichia pastoris concentration (48.7% RMSE reduction) and inulinase concentration (54.9% RMSE reduction).
• Simulation results demonstrate the effectiveness of the substructure domain transfer approach in a Pichia pastoris fermentation example.

Conclusions:

• The developed soft sensor model accurately predicts Pichia pastoris and inulinase concentrations online across different working conditions.
• This novel method offers superior prediction accuracy compared to traditional soft sensor modeling techniques.