Data-Based Feedback Relearning Control for Uncertain Nonlinear Systems With Actuator Faults.

Area of Science:

• Control Systems Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Uncertain nonlinear systems pose challenges for data-based reinforcement learning (RL) due to data accuracy issues affecting convergence and optimality.
• Existing RL algorithms struggle with control channel disturbances and actuator faults, limiting their real-world applicability.

Purpose of the Study:

• To develop a data-based feedback relearning (FR) algorithm for uncertain nonlinear systems susceptible to disturbances and actuator faults.
• To enhance the convergence and optimality of RL algorithms using empirical data and advanced data processing techniques.

Main Methods:

• Implemented a feedback relearning (FR) algorithm for online strategy updates using empirical data.
• Utilized experience replay technology for improved data utilization efficiency and algorithm convergence.
• Employed a neural network (NN)-based fault observer for model-free fault compensation and redesigned polynomial activation functions for improved generalization.

Main Results:

• The proposed FR algorithm demonstrated improved convergence and optimality by continuously approaching the optimal solution.
• The NN-based fault observer effectively compensated for actuator faults and system disturbances.
• Comparative simulations confirmed the robustness and effectiveness of the strategy in guaranteeing control performance.

Conclusions:

• The developed data-based FR algorithm offers a robust solution for controlling uncertain nonlinear systems with disturbances and faults.
• The integration of experience replay and NN-based fault compensation significantly enhances algorithm performance and data efficiency.
• The redesigned activation functions in NNs simplify design and improve generalization for unknown nonlinear systems.