Machine Learning in Human Hand Exoskeleton

Exoskeleton devices have become tools for supporting multiple human activities. As such activity recognition in real-time has become a very important aspect of exoskeleton design. It provides useful information to allow the robot to predict future motions that may be carried out by the human to assist and reduce the chances of injury. Although multiple methods exist for assisting in such predictions a proper catalog of such methods well as their effectiveness does not exist. Due to the various types of data available, it is necessary to focus on a single kind. As multiple works exist based on image-based data, signal-based data is chosen as the focus of this work as in a real-life case such information will be commonly used due to the availability with the help of sensors. Such data also provides better results when used for classification processes. This work attempts to catalog different Machine learning algorithms used for real-time activity recognition systems based on signal data. Multiple machine learning models have been trained and evaluated. These models include Artificial Neural Networks (ANN), Support vector Machines (SVM), Random Forest, and Gradient Boosted Decision Trees. The prediction is done based on four different classes each representing a different position of the human hand for which EMG data was obtained. The best result is given by the Gradient Boosted Decision Trees which gives a prediction accuracy of 96% while the worst results are given by Artificial Neural Network (ANN) with a prediction accuracy of $\mathbf{9 0. 0 2 \%}$. Class-wise prediction is also compared to choose between the algorithms.