Multiple feature fusion transformer for modeling penicillin fermentation process with unequal sampling intervals.

Area of Science:

• Biotechnology
• Chemical Engineering
• Data Science

Background:

• Quality prediction in biological fermentation batch processes is crucial but challenging due to dynamic nonlinearity, variable sampling, and multiple data features.
• Existing methods struggle with the complexities of batch process data, including uneven durations and sampling intervals.

Purpose of the Study:

• To propose a novel Multiple-Feature Fusion Transformer (MFFT) model for accurate time series quality prediction in biological fermentation batch processes.
• To address the challenges posed by dynamic nonlinearity, unequal sampling, and diverse data features in batch process monitoring.

Main Methods:

• Developed the Multiple-Feature Fusion Transformer (MFFT) model utilizing a sequence-to-sequence architecture.
• Incorporated a transformer parallel operation model to handle variable sequence input and unequal sampling intervals.
• Integrated image data by using a pretrained ResNet50 as a mycelium status classifier and developed a multiple-feature encoding structure.

Main Results:

• The MFFT model demonstrated superior performance in time series prediction tasks for penicillin fermentation.
• The proposed model effectively fused image information and sampling time data for enhanced feature representation.
• MFFT significantly outperformed existing methods in predicting batch process quality.

Conclusions:

• The MFFT model offers a robust solution for time series quality prediction in complex biological fermentation batch processes.
• The integration of image data and a flexible sequence handling mechanism makes MFFT a powerful tool for industrial biotechnology.
• This study highlights the potential of advanced deep learning architectures for improving process monitoring and control in fermentation.