Abstract: The aim of this study was to effectively solve the problem of membrane fouling when integrated techniques of the enzymolsis extraction-ultrafiltration purification were used in fibrous rhizome herbs and keep the integrated techniques using smoothly.The prediction model of critical flux and pressure was established based on the ultrafiltration data of enzymatic hydrolysate of Hedysari Radix by BP neural network.The performance and applicability of the model were evaluated.Then sensitivity analysis of input variables were performed using connection weights method to assess the relative contribution of input variables.Results indicated that the model had better performance and applicability.Mean absolute error,mean error rate and R² of critical flux and pressure were 1.5228 L/(m²·h) and 0.0032 MPa,3.46% and 2.50%,0.96 and 0.95,respectively.Mean absolute error and mean error rate of critical flux and pressure for Astragali Radix were 1.4360 L/(m²·h) and 0.0034 MPa,3.93% and 2.80%,respectively.The relative contribution of input variables to critical flux and pressure presented the same order of viscosity＞concentration＞pH＞temperature.

Key words: Hedysari Radix, Astragali Radix, enzymolsis extraction, ultrafiltration, critical flux, critical pressure, BP neural network