Fuzzy Kernel Stochastic Configuration Networks for Industrial Data Modeling

Abstract

Kernel stochastic configuration networks (KSCNs) belong to the randomized learner model with universal approximation property. However, the original KSCNs model lacks logical reasoning ability and the utilization of prior knowledge. This article proposes a novel neurofuzzy learner model based on KSCNs, termed fuzzy KSCNs (F-KSCNs), for enhancing model’s interpretability and modeling performance. First, the Takagi–Sugeno–Kang fuzzy inference system is integrated into the modeling process of KSCNs, where multiple submodels are constructed according to the associated fuzzy rules. Then, the parameters of F-KSCNs are determined by the kernel stochastic configuration algorithm to guarantee model’s universal approximation property. Finally, the output of F-KSCNs is obtained by comprehensively evaluating the significance of each submodel. Compared to the original KSCNs, the proposed F-KSCNs have stronger fuzzy inference ability and can deal with rule-based information. Moreover, computational efficiency is greatly improved under this fuzzy inference framework. Two industrial case studies are carried out for performance evaluation. The experimental results with comparisons with existing solutions clearly show the superiorities of our proposed method, including generalization performance and modeling efficiency.