Development and Validation of a CNN Human Activity Recognition based Control System applied to a Variable Stiffness Prosthetic Foot: The MyFlex-θ

Lower-limb amputation has a major impact on mobility and quality of life. Modern energy storing and returning prosthetic feet can improve gait, but their mechanical properties are fixed and cannot adapt to different daily activities. Powered prostheses can respond better to changes in terrain and speed, but they are still limited by high weight, power consumption, and cost. This thesis addresses this gap through the development of an intelligent control system for a variable stiffness prosthetic foot. The work was carried out on the MyFlex-θ platform, a research prototype developed at the University of Bologna. The thesis covers the full pipeline, from electronic integration and software development to dataset creation, CNN design, real-time deployment, and experimental validation. The system combines a Raspberry Pi 4, two IMU sensors, and a motorized stiffness adjustment mechanism. To support data collection, a custom adaptor was designed to let an able-bodied subject wear the prosthesis and reproduce key features of prosthetic gait. A dataset was collected across seven daily activities: standing, walking, fast walking, ramp and stair ascent/descent. At the core of the system is a convolutional neural network (CNN) trained on IMU data to recognize the current activity and select the appropriate stiffness setting in real time. The model and the full processing pipeline were optimized for embedded use. Experimental results showed 97.2% classification accuracy offline and 91.5% during real-time outdoor testing. These results show that CNN-based activity recognition can support autonomous stiffness adaptation on embedded hardware, bringing variable stiffness prosthetic feet closer to practical real-world use. To the best of current knowledge, the MyFlex-θ platform with this control system is the first variable stiffness prosthetic foot to demonstrate autonomous real-time adaptation, a capability not available in current commercial devices.