Online Self-Learning Fuzzy Recurrent Stochastic Configuration Networks for Modeling Nonstationary Dynamics

Abstract

Fuzzy recurrent stochastic configuration networks (FRSCNs) are a class of randomized neurofuzzy models that have shown success in modeling nonlinear dynamic systems. However, the data generated from the real-world industry process often exhibits nonstationary characteristics, which could lead to the resulting model with poor generalization performance. This article presents an online self-learning FRSCN (OSL-FRSCN) for problem-resolving. OSL-FRSCNs integrate the self-organizing learning strategies into FRSCNs, maintaining the model's strong fuzzy inference capabilities while enhancing their continuous learning abilities for nonstationary data streams. The network parameters are updated online using the projection algorithm based on the newly arriving data streams. Moreover, the network structure can be dynamically adjusted in the light of the fuzzy recurrent stochastic configuration (FRSC) algorithm and an improved sensitivity analysis. Comprehensive comparisons over two nonlinear system identification tasks and two industrial applications are carried out. Numerical results clearly demonstrate that the proposed OSL-FRSCNs outperform other neurofuzzy and nonfuzzy models with sound generalization, verifying their effectiveness in modeling nonlinear systems with nonstationary dynamics.