A Transformer-based framework with generative spectral augmentation for online monitoring of hyaluronic acid...

Accurate and real-time monitoring of the molecular weight and concentration of hyaluronic acid (HA) during fermentation is critical for ensuring product quality control. Near-infrared (NIR) spectroscopy offers rapid, cost-effective, and non-destructive on-site analysis capabilities. However, the effectiveness of NIR relies on sufficient and reliable spectral data and reference values, which are difficult to obtain given the complexity and variability of the fermentation process. To address this challenge, this study proposes a novel online monitoring framework integrating data augmentation and deep learning. By employing Wasserstein GAN with gradient penalty (WGAN-GP) to generate high-fidelity synthetic spectra, the calibration dataset was significantly expanded, thereby enhancing model robustness. Concurrently, an improved Transformer model was developed to effectively capture the complex spectral features inherent in the fermentation process. Experimental results demonstrate that the proposed method surpasses traditional PLSR, as well as commonly used 1D-CNN and LSTM models, in prediction accuracy and generalization on independent test sets. This research highlights the promising potential of integrating advanced data augmentation with deep learning, offering a powerful and broadly applicable intelligent quality control solution for precise and real-time monitoring of complex biological processes based on NIR spectroscopy.