Optimization of ANFIS with Applications in Machine Defect Severity Classification

High accuracy and high generalization capability are two conflicting objectives in the design of adaptive neuro-fuzzy inference system (ANFIS). Motivated by previous studies on handling similar conflicting situations in model selection and autoregressive order estimation, this paper investigates information criteria for the optimization of ANFIS model with applications in machine defect severity classification. The studied criteria include the Akaike Information Criterion (AIC), the corrected AIC (AIC_c), and the Generalized Information Criterion (GIC). By introducing a novel model complexity function and replacing the variances in the original criteria with weighted mean square error, the criteria extended for ANFIS are defined. Based on these criteria, the optimized ANFIS model is chosen to be the one which leads to the minimized criterion value. The performance of these criteria is experimentally studied using bearing defect severity classification as an example.