This paper investigates the architectural effect on adaptive neuro-fuzzy inference system (ANFIS) for machine condition assessment. The study was motivated by ANFIS’s limitation in adapting its architecture to map the modeled input output relationship. Based on the grid input space partition method, two elements in defining an ANFIS architecture were studied: the type of the membership function (MF) and the MF number assigned to ANFIS inputs. A new modeling accuracy index was introduced to address the limitation of the traditional root mean square error (RMSE) in describing the effect of the MF type. The analysis showed that wide core membership functions enabled a smaller RMSE than narrow core membership functions for machine defect severity classification. It is further shown that selecting appropriate MF number is critical to ensuring accuracy of ANFIS, considering the overfitting problem. These results were experimentally investigated on a bearing test bed, where defect severity classification and dynamic load estimation were evaluated. The experiments agreed well with the theoretical analysis.
