Real-time monitoring of Escherichia coli fermentation via hyperspectral imaging and lightweight edge artificial intelligence

Real-time monitoring of key metabolites in Escherichia coli (E. coli) fermentation is essential for ensuring the quality and consistency of fermentation products. Hyperspectral imaging (HSI) technology provides a promising approach for pixel-level fermentation monitoring. However, the large data volume and high computational demands pose significant challenges for real-time in-situ monitoring. In this paper, we propose a hyperspectral-based pipeline for real-time, pixel-level E. coli fermentation monitoring. First, a hardware-friendly feature spectral-band distillation method and a lightweight convolutional neural network (CNN) model are employed to enhance real-time performance and analysis capabilities. Experimental results show that the proposed method achieves root mean square error of prediction (RMSEP) of 1.592, 0.327, 1.537 (g/L) and relative prediction deviation (RPD) of 11.324, 11.983, and 9.372 for the target metabolites valine, ammonia, and acetic acid, respectively. Furthermore, we deployed the algorithms on the neural network processing unit (NPU) chip of RK3588 and the results show that our method processes one frame of hyperspectral image with an average inference time of approximately 0.76 s. Compared with traditional methods, the proposed pipeline enables real-time, pixel-level spatial concentration mapping of fermentation samples with improved accuracy and computational efficiency. This work facilitates efficient on-chip deployment for practical online fermentation monitoring applications and lays the groundwork for the development of miniaturized, low-cost hyperspectral monitoring systems.