Decentralized neural identifier and control for nonlinear systems based on extended Kalman filter

Area of Science:

• Robotics
• Control Systems
• Artificial Intelligence

Background:

• Nonlinear systems present significant challenges in identification and control.
• Recurrent high-order neural networks offer potential for modeling complex dynamics.
• Existing methods may struggle with nonstationary system dynamics.

Purpose of the Study:

• To propose a time-varying learning algorithm for recurrent high-order neural networks.
• To integrate a statistical framework, specifically the extended Kalman filter, for enhanced identification.
• To develop a robust method for dynamical modeling of nonstationary systems.

Main Methods:

• The learning algorithm is based on the extended Kalman filter, incorporating coupled variance in noise covariance matrices.
• A sliding window-based approach is used for dynamical modeling of nonstationary systems.
• The methodology was applied to a five degree-of-freedom (DOF) robot manipulator for trajectory tracking.

Main Results:

• The proposed algorithm effectively identifies interactions between plant states and neural network convergence.
• The sliding window method improved neural identification accuracy for nonstationary dynamics.
• Decentralized discrete-time block control and sliding mode techniques were successfully implemented for trajectory tracking.

Conclusions:

• The developed time-varying learning algorithm provides an accurate and efficient approach for identifying and controlling nonlinear systems.
• The integration of statistical frameworks and advanced control techniques enhances performance in complex dynamic environments.
• The method demonstrates significant potential for applications in robotics and other fields requiring precise control of nonlinear dynamics.