An improved online self-organizing dynamic fuzzy neural network for nonlinear dynamic system identification

In the area of neural fuzzy control, how to generate fuzzy rules for structural learning is a key issue. In this paper, an improved online self-organizing dynamic fuzzy neural network for nonlinear dynamic system identification. The system is a five-layered network, which features coalescence between Takagi-Sugeno-kang fuzzy architecture and dissymmetrical Gaussian functions as membership functions. The partitioning made by the dissymmetrical Gaussian functions introduces the dissymmetry to the left and right widths of the input space to increase the flexibility of the design, thus resulting in a parsimonious fuzzy neural network with higher performance under online learning. We apply two criteria for rule generation, namely system error and ε-completeness, reflecting both the performance and sample coverage of an existing rule base. During the parameters estimation phase, we adjust the Gaussian centers according to the adjustment of the widths. Parameters in the premise and the consequents are adjusted online based on the ε-completeness of the fuzzy rules and Kalman Filter (KF) approach, respectively. The error reduction ratio (ERR) method is used as the pruning strategy. Simulation studies demonstrate the efficacy and superiority of the proposed algorithm in terms of the approximation accuracy and the generalization performance.