Neural network decoupling technique and its application to a powered wheelchair system

Area of Science:

• Control Engineering
• Artificial Intelligence
• System Dynamics

Background:

• Multivariable systems often exhibit complex input-output interactions, leading to coupling effects.
• Controlling uncertain systems poses challenges due to unpredictable dynamics.
• Existing decoupling methods may require extensive system identification or Jacobian calculations.

Purpose of the Study:

• To propose a novel neural network decoupling technique for uncertain multivariable systems.
• To develop a method that avoids explicit calculation of the plant Jacobian.
• To enhance control performance by minimizing coupling effects.

Main Methods:

• A linear diagonalization technique is used to design a reference model with nominal parameters.
• A neural network decoupler is trained using signals from the reference model.
• A separate neural network learns the uncertain system dynamics, trained with the Lavenberg-Marquardt algorithm and Bayesian regularization.
• Real-time recurrent learning is employed for gradient information acquisition.

Main Results:

• The proposed neural network decoupling technique effectively minimizes coupling effects.
• The method demonstrates robustness under system uncertainties.
• Experimental results in powered wheelchair control validate the technique's efficacy.

Conclusions:

• The neural network decoupling approach offers an effective solution for controlling uncertain multivariable systems.
• The technique successfully mitigates undesirable input-output interactions.
• This method provides a viable alternative for applications requiring precise control of complex systems.