Human-Centered Deep Learning Neural Network Trained Myoelectric Controller for a Powered Wheelchair

Abstract

The powered wheelchair is one of the most widely used devices that assists users in activities of daily living. Although powered wheelchairs have improved the quality of life for most people, some users with certain disabilities, such as Spinal Cord Injury, may have difficulties with common types of controllers like joysticks. Because of this, other types such as head, mouth, vision, and speech controls have been employed, but each control type has its own limitations in terms of safety and accuracy. This research presents a human-centered approach that detects hand gestures using non-invasive surface electromyography (sEMG) signals from the human forearm by an artificial intelligence in the form of a Deep Learning Neural Network (DLNN). An Institutional Review Board application was approved to recruit participants without muscular disability between 18 to 50 years of age to evaluate the proposed controller mechanism. Each participant s sEMG signals were acquired at 200 Hz sampling frequency followed by DLNN training and validation. Methods such as an ultrasonic sensor to avoid large obstacles and a proportional-integral (PI) controller to produce smooth wheelchair motor movements were integrated with DLNN-based hand gesture recognition in this research to ensure the user s safety. The trained Bayesian Regularization DLNN had an average accuracy of 98.4% across all subjects and hidden layers. All subjects were able to successfully navigate the path using the proposed controller in an average of 4.85 minutes and touching an average of 2.3 obstacles out of 13. Being that so many people require the use of a wheelchair but oftentimes have difficulties controlling it, there is a clear need for the development of a safer and easier method of control and this research strives to fulfil that need.