Description

This thesis investigates how machine learning-based human action recognition can be used to enhance the control of soft robotic exoskeletons in real-world industrial tasks. Using IMU sensor data, various neural network architectures were developed, optimized, and evaluated to detect user actions in real time.

The study examines the effects of key hyperparameters such as window size, overlap, L2 regularization, and data set composition, revealing how these influence classification accuracy and control stability. A custom dataset was collected with actuated and non-actuated exoskeleton data, enabling two control strategies to be implemented and tested. Results show that providing support only during physically demanding actions can significantly reduce air consumption and improve user comfort. The findings offer practical insights for future exoskeleton systems, including passive and electrically powered variants, by demonstrating how human action recognition can inform adaptive support and prevent user overload.