In-line monitoring and endpoint determination of percolation process of herbal medicine using ultraviolet spectroscopy combined with convolutional neural network

Abstract

Abstract Objectives As a common step in the herbal medicine production process, percolation usually lacks effective process monitoring methods and is often conducted with fixed process parameters. In this study, an in-line ultraviolet (UV) spectroscopy was used for monitoring the Caulis Sinomenii percolation process. Methods The spectra and concentration data of 156 percolation samples from five batches were collected. Convolutional neural networks (CNNs) were used to develop quantitative calibration models. The mean squared error (MSE), mean absolute percentage error (MAPE) and mean absolute error (MAE) were compared to select the proper loss function for developing the CNN models. Meanwhile, partial least square regression (PLSR) was also used to develop calibration models for performance comparison. Key findings The CNN models with MAPE or MAE as the loss function could provide accurate predictions for all samples. However, CNN models adopting MSE as the loss function tended not to predict low-concentration samples accurately. The CNN models mostly achieved satisfactory results without any preprocessing techniques and surpassed PLSR models in all the performance metrics. Conclusions An in-line UV spectroscopy system combining the CNN algorithm was implemented to monitor the percolation process of Caulis Sinomenii. The system can accurately determine the endpoint of the percolation process.