Combining convolutional neural networks and on-line Raman spectroscopy for monitoring the Cornu Caprae Hircus hydrolysis process.

Abstract

Cornu Caprae Hircus (goat horn, GH) is one of the frequently used medicinal animal horns in traditional Chinese medicine (TCM). Hydrolysis is one of the key steps for GH pretreatment in pharmaceutical manufacturing. However, the physicochemical complexity of the hydrolysis samples imposes a challenge for hydrolysis process analysis and monitoring. In this study, convolutional neural networks (CNNs), one of the most popular deep learning methods, were used to develop quantitative calibration models based on on-line Raman spectroscopy for monitoring the GH hydrolysis process. Partial least squares (PLS) calibration models were also developed for model performance comparison. For CNN modeling, raw Raman spectra were used as inputs and hyperparameters in the CNN structure were optimized. Results show for four of the seven analytes, the optimized CNN models using raw spectra as inputs outperform the optimized PLS models developed with preprocessed spectra. Therefore, compared with the commonly used PLS algorithm, CNN modeling is also a practicable regression method and can be employed for the analytical purpose of this study. Models with better performance are expected to be obtained by improving the CNN model structure and using more effective hyperparameter optimization approaches in further studies. To the best of our knowledge, this is the first reported case study of combining CNNs and on-line Raman spectroscopy for a regression task.