SOYMILK ISOFLAVONE CONVERSION PREDICTION BY ADAPTIVE NEURO-FUZZY INFERENCE SYSTEM

In this study, a model for prediction of soymilk isoflavone glycoside conversion was constructed using adaptive neuro-fuzzy inference system (ANFIS) techniques. We chose aeration rate, cultivation duration, and the amount of isoflavone glycoside as the three inputs and the yield of isoflavone aglycone as the single output to develop the prediction model. The average root mean square error (RMSE) of the output over 50 training epochs for genistin and daidzin conversion processes were 3.43 x 10(-5) and 4.59 x 10(-5), respectively, which demonstrates that the established models were significantly well-trained. The values of RMSE and MAE for genistin and daidzin conversion processes were (0.46, 0.83) and (0.36, 0.63) during testing, which suggests that the yield values predicted by the ANFIS model closely matched the actual values. The results implied that ANFIS is a powerful tool for predicting isoflavone conversion during fermentation processes. Compared with the one-factor-at-a-time approach, ANFIS exhibited superior performance for scale-up of soybean fermentation.