Convolutional Neural Network Controller for Autonomous Variable Stiffness ESR Foot Prosthesis: The MyFlex-θ

Human ankles naturally adjust stiffness to accommodate various activities and terrains—an ability that conventional passive energy-storing-and-returning (ESR) prostheses cannot replicate due to their fixed stiffness. To address this limitation, this study introduces the MyFlex- \theta prosthesis, an autonomous ESR foot equipped with a Variable Stiffness System (VSS) and a Human Activity Recognition (HAR) control system based on a Convolutional Neural Network (CNN) algorithm. The HAR-CNN system classifies seven activities—walking, fast walking, standing, ramp ascent/descent, and stair ascent/descent— using linear acceleration data from two embedded Inertial Measurement Units (IMUs). When three consecutive predictions are identical, the VSS actuator adjusts foot stiffness in real time. Offline evaluation of the HAR-CNN, following the optimization of nine input channels, algorithm hyperparameters, and sliding window and hop size configurations, achieved a classification accuracy of 97.2% (loss: 0.103). Real-time testing with an non-disabled participant yielded an accuracy of 91.5%, confirming robust performance under dynamic conditions. Furthermore, real-time ankle torque estimation was implemented using a third-order response surface derived from sixth-order polynomial Finite Element Analysis (FEA) models, correlating stiffness settings with IMU-recorded joint rotations. This integration of adaptive stiffness control and embedded torque estimation enables continuous, portable gait analysis. The MyFlex-θ represents a significant advancement toward intelligent prosthetic devices capable of autonomously adapting to user activity and biomechanical demands.