Neural Network Direct Control with Online Learning for Shape Memory Alloy Manipulators

Area of Science:

• Robotics and Control Systems
• Materials Science and Engineering
• Artificial Intelligence in Engineering

Background:

• Industrial processes face challenges with complex actuator behaviors, particularly nonlinear hysteresis in shape memory alloys (SMAs).
• Traditional control systems struggle with the unique response of each SMA system, often requiring complex hysteresis modeling.
• Developing effective control strategies for SMA-based actuators is crucial for improving robotic system performance.

Purpose of the Study:

• To develop a novel neural network direct control (NNDC) system with online learning for SMA manipulators.
• To enable precise position control of SMA manipulators without requiring prior training or explicit hysteresis models.
• To implement a real-time, low-computational cost control solution for SMA actuators.

Main Methods:

• A neural network direct control approach with online weight coefficient updates was employed.
• The controller utilized real-time actuator position data for continuous learning during operation.
• Experimental validation was conducted on a 1-DOF manipulator system actuated by an SMA wire.

Main Results:

• The proposed NNDC scheme effectively controlled the angular position of the SMA manipulator.
• The system demonstrated compensation for the inherent hysteretic behavior of the SMA actuator.
• A maximum static error of 0.83° was achieved using a sine wave reference signal and validated across multiple set-points.

Conclusions:

• The developed online learning neural network controller offers an effective solution for SMA manipulator position control.
• This method successfully addresses the challenge of nonlinear hysteresis in SMAs without complex modeling or pre-training.
• The real-time, low-cost implementation shows promise for practical applications in robotics and industrial automation.