A Machine Learning Approach to Predicting Distal Leg Muscle Activation for Real-time Control

This work explores three machine learning approaches to predict muscle activations in the lower limbs during walking. This was a group project for our Machine Learning Course (EECS 545) in Spring 2021.

Our approach is motivated by the potential to integrate this prediction into real-time control for above-knee robotic prostheses, allowing the user to control the prosthesis by activating intact muscles. First, we preprocess the data so that each feature instance includes enveloped muscle activation data over a certain history window, gait cycle phase, and walking speed. Then, we apply Kernel Ridge Regression and Support Vector Regression and evaluate their performance given the computational limits with our very large dataset. Then, we design and fit a Temporal Convolutional Network and again evaluate performance. Due to the TCN’s high performance, we also test its accuracy on additional tasks such as ramp ascent. All methods were able to predict the EMG data with somewhat reasonable accuracy (> 60% VAF minimum), and the TCN model was highest performing with > 92% VAF across all tasks. Finally, we discuss the different methods, including accuracy trends seen across muscles, and forward propagation complexity in regards to real-time control feasibility.

The full report is available here.