Machine Learning for Trajectory Planning of Active Lower Limb Prosthetic Devices

: (Alternative Format Thesis)

Abstract

The increasing prevalence of lower limb amputation highlights the need for improved walking assistive devices, such as powered prostheses, that can support mobility and quality of life. A key challenge in prosthesis control lies in generating reference trajectories that adapt to varied environments and individual gait styles. Many existing methods rely on predefined trajectories, discrete gait-phase segmentation, or simplified terrain assumptions, which limit their generalizability. Recent advances in machine learning, particularly in time-series modeling, offer promising directions for more flexible and user-oriented trajectory generation.

This research explores two attention-based deep learning models for trajectory planning. A feed-forward attention model is developed to estimate ankle and hip accelerations, while a self-attention model predicts ankle and knee angles. Both approaches provide smoother and more accurate trajectories compared with their baseline models, and have the potential to operate continuously across activities without requiring explicit activity transitions.

To better understand sensor contributions, feature selection and transformation are investigatedthrough Recursive Feature Elimination (RFE) and Principal Component Analysis (PCA). RFE provides a ranking of sensor signals for more cost-effective deployment, while PCA reduces dimensionality by dropping redundant signal components. Their effectiveness is evaluated using statistical tests, an aspect often overlooked in related studies. In addition, the biomechanical plausibility of predicted trajectories is examined using OpenSim musculoskeletal modeling. This allows the early identification of potentially adverse biomechanical loads, which helps mitigate discomfort and injury risks.

Overall, the findings suggest that data-driven trajectory generation, combined with systematic sensor analysis and biomechanical validation, may contribute to more adaptive and
physiologically consistent prosthetic control. While further investigation is required, this study points to the potential of machine learning-based methods to enhance the personalization and usability of powered lower limb prostheses.

Date of Award	8 Oct 2025
Original language	English
Awarding Institution	University of Bath
Supervisor	Ioannis Georgilas (Supervisor) & Andrew Plummer (Supervisor)