A neural observer with time-varying learning rate: analysis and applications

Area of Science:

• Control Systems Engineering
• Artificial Intelligence
• Biochemical Engineering

Background:

• Accurate state estimation is crucial for controlling complex nonlinear systems.
• Traditional observers struggle with parameter variations and external disturbances.
• Recurrent High Order Neural Networks (RHONN) offer a powerful framework for modeling nonlinear dynamics.

Purpose of the Study:

• To propose a Reduced Order Neural Observer (RONO) with a time-varying learning rate.
• To enhance the robustness and learning capability of neural observers in dynamic environments.
• To validate the observer's performance in a challenging nonlinear application.

Main Methods:

• Development of a discrete-time RHONN-based observer.
• Integration of an Extended Kalman Filter (EKF) for training the neural network.
• Design of a time-varying learning rate algorithm to adapt to system uncertainties.
• Mathematical proof of the stability for the time-varying learning rate.

Main Results:

• The proposed RONO demonstrates improved learning performance in the presence of disturbances and parameter variations.
• Stability analysis confirms the robustness of the time-varying learning rate.
• Simulations show the observer's effectiveness for a nonlinear anaerobic digestion process.

Conclusions:

• The developed RONO with a time-varying learning rate is a robust and effective tool for state estimation in nonlinear systems.
• The EKF-trained RHONN provides a flexible platform for observer design.
• This approach offers significant potential for applications in process control and monitoring.