SUMMARY

The use of lower limb prostheses has been steadily increasing, which has led to significant research conducted towards developing powered prostheses that can help restore natural gait in amputation patients. Intelligent machine learning algorithms have been implemented with active prosthetics for user intent recognition and context estimation. Although these data-driven methods have been proven successful for high-level control in estimating locomotion modes or real-time gait speed analysis, they often overlook the importance of analyzing gait symmetry in users and are fixated on evaluation metrics focused solely on the prosthetic’s performance. To improve this, this thesis aims to investigate methods of leveraging machine learning with onboard prosthesis sensors to estimate sound-side limb joint mechanics. Such would allow for a way to evaluate the appropriateness of the prosthetic’s assistance level in regard to the energy expenditure on the user’s biological limb. The main hypothesis of the study was that while it may be possible to estimate for joints that are close to the sensors of interest, complete inter-joint estimations may be challenging due to a lack of correlation in sensor signals. CNN and TCN-based models were used to estimate various lower limb joints using data collected from subjects walking on level ground and ramps. The results showed that a properly tuned CNN can achieve reasonable estimates of inter-joint work values in various locomotion modes. This study also evaluated the most minimal and optimal set of sound-side sensors required for accurate sound-side sensing capabilities for each joint.