Identification and prediction of dynamic systems using an interactively recurrent self-evolving fuzzy neural network

Area of Science:

• Computational intelligence
• Artificial neural networks
• Fuzzy logic systems

Background:

• Dynamic systems require accurate prediction and identification models.
• Existing recurrent fuzzy neural networks (FNNs) have limitations in mapping ability and online learning.
• Novel approaches are needed to improve the performance of FNNs for dynamic system modeling.

Purpose of the Study:

• To propose a novel interactively recurrent self-evolving fuzzy neural network (IRSFNN) for dynamic system prediction and identification.
• To enhance the mapping ability of fuzzy neural networks using functional link neural networks (FLNNs) in the consequent part.
• To develop an online learning mechanism for simultaneous structure and parameter optimization.

Main Methods:

• The interactively recurrent self-evolving fuzzy neural network (IRSFNN) architecture is introduced, featuring external loops and internal feedback.
• A functional link neural network (FLNN) is integrated into the consequent part of fuzzy rules to improve nonlinear mapping capabilities.
• Simultaneous online learning of structure and parameters is achieved using an on-line clustering algorithm for rule generation, a variable-dimensional Kalman filter for consequent parameter updates, and gradient descent for premise and recurrent parameter learning.

Main Results:

• The IRSFNN demonstrated superior performance in predicting and identifying dynamic plants compared to other well-known recurrent FNNs.
• The integration of FLNN in the consequent part significantly enhanced the model's mapping ability.
• The online learning approach enabled efficient and simultaneous optimization of both structure and parameters.

Conclusions:

• The proposed IRSFNN offers an effective and enhanced approach for the prediction and identification of dynamic systems.
• The novel architecture and online learning methodology provide a robust framework for complex system modeling.
• IRSFNN represents a significant advancement in the field of recurrent fuzzy neural networks.