Location

https://ccse.kennesaw.edu/computing-showcase/cday-programs/spring2021program.php

Streaming Media

Document Type

Event

Start Date

26-4-2021 5:00 PM

Description

Abstract Hand mentor robotic device is beneficial for stroke patients . This is rehabilitation technique used in stroke therapy. It strengthens and improves the range of motion which ultimately improves the quality of life for severely impaired stroke patients. It is easy to use without assistance and most importantly stroke survivors able to use independently. Usage of hand mentor device is quite expensive for stroke patients on hourly basis . Coming up with most efficient deep learning algorithm for sensor data is motivation to cut down the cost and easy availability usage for stroke patients. EMG signal is recorded using relevant sensors which provides useful information to infer muscle movement. In this study, we utilized publicly available EMG signal datasets recorded from upper limb of human subjects to develop a neural network based model for the prediction of wrist motion intention. Research Question or Motivation The Motivation of this study is to train a simple neural network model to accurately predict three basic wrist motions (extension, flexion and no motion) using optimum number of EMG sensors. This model can be further deployed to augment the capabilities of commercially available robotic-assistive rehabilitation devices. Materials and Methods Sensor-based continuous hand gesture recognition activity requires profound knowledge about gesture activities from multitudes of low-level sensor readings. There are two ways to provide the solutions either to go by handcrafted features from sensor data or use deep learning techniques. The advantage of using deep learning technique is to utilize the automatic high-level feature extraction with outstanding performance. However, sensor data requires signal pre-or post-processing such as feature selection, dimension reduction, denoising, etc. Based on the literature review of many research papers, we found that 1D Convolutional Neural Network have recently become the state-of-the-art technique for crucial signal processing applications. 1D CNN is very effective when we aim to extract features from fixed-length segments of the overall dataset and where the location of the feature within the segment is not of high relevance. In addition to this, real-time and low-cost hardware implementation is feasible using 1D CNN. After a successful literature review on 1D CNN knowing its advantages and benefits of using over signal. We decided to use 1D CNN on raw EMG signal data. Preliminary results: Since it is an application-based project, we planned to work in phases to achieve the long-term goal of benefitting stroke patients using deep learning techniques. In this initial phase of the study, we utilized publicly available EMG dataset for hand gestures from UCI Machine Learning Repository to test the performance of the 1D CNN algorithm on gesture classification. We used only 3 labels (hand at rest, wrist flexion, wrist extension) out of 8 labels in the dataset for our particular application requirement. This dataset contains 8 EMG channels collected from commercial MYO Thalmic bracelet device. We first performed an initial analysis to investigate the optimum number of sensor/channels based on the highest gesture classification accuracy using KNN, Decision Tree and Naïve Bayes algorithms. As a result of this analysis, we obtained the optimum channel combination (Ch1, Ch4, Ch5, Ch8) data which generates the best classification accuracy. We used these 4 sensor datasets to train a 1D CNN with 78/22 train/test split. Dataset contains total 36 subjects. Data with subject number less than or equal to 28 is considered as training set and data with subject number greater than 28 is considered as test set. We also performed an optimization study on finding the optimum time signal window and overlap sizes of 100 ms and 50 ms . We achieved test accuracy of 97% for the classification accuracy of 3 gestures (hand at rest, wrist flexion, wrist extension).Advisors(s): Supervisor : Dr. Coskun Tekes Email id : ctekes@kennesaw.eduTopic(s): Artificial IntelligenceCS7992

Share

COinS
 
Apr 26th, 5:00 PM

GR-29 Wrist Intent Recognition for Stroke Rehabilitation

https://ccse.kennesaw.edu/computing-showcase/cday-programs/spring2021program.php

Abstract Hand mentor robotic device is beneficial for stroke patients . This is rehabilitation technique used in stroke therapy. It strengthens and improves the range of motion which ultimately improves the quality of life for severely impaired stroke patients. It is easy to use without assistance and most importantly stroke survivors able to use independently. Usage of hand mentor device is quite expensive for stroke patients on hourly basis . Coming up with most efficient deep learning algorithm for sensor data is motivation to cut down the cost and easy availability usage for stroke patients. EMG signal is recorded using relevant sensors which provides useful information to infer muscle movement. In this study, we utilized publicly available EMG signal datasets recorded from upper limb of human subjects to develop a neural network based model for the prediction of wrist motion intention. Research Question or Motivation The Motivation of this study is to train a simple neural network model to accurately predict three basic wrist motions (extension, flexion and no motion) using optimum number of EMG sensors. This model can be further deployed to augment the capabilities of commercially available robotic-assistive rehabilitation devices. Materials and Methods Sensor-based continuous hand gesture recognition activity requires profound knowledge about gesture activities from multitudes of low-level sensor readings. There are two ways to provide the solutions either to go by handcrafted features from sensor data or use deep learning techniques. The advantage of using deep learning technique is to utilize the automatic high-level feature extraction with outstanding performance. However, sensor data requires signal pre-or post-processing such as feature selection, dimension reduction, denoising, etc. Based on the literature review of many research papers, we found that 1D Convolutional Neural Network have recently become the state-of-the-art technique for crucial signal processing applications. 1D CNN is very effective when we aim to extract features from fixed-length segments of the overall dataset and where the location of the feature within the segment is not of high relevance. In addition to this, real-time and low-cost hardware implementation is feasible using 1D CNN. After a successful literature review on 1D CNN knowing its advantages and benefits of using over signal. We decided to use 1D CNN on raw EMG signal data. Preliminary results: Since it is an application-based project, we planned to work in phases to achieve the long-term goal of benefitting stroke patients using deep learning techniques. In this initial phase of the study, we utilized publicly available EMG dataset for hand gestures from UCI Machine Learning Repository to test the performance of the 1D CNN algorithm on gesture classification. We used only 3 labels (hand at rest, wrist flexion, wrist extension) out of 8 labels in the dataset for our particular application requirement. This dataset contains 8 EMG channels collected from commercial MYO Thalmic bracelet device. We first performed an initial analysis to investigate the optimum number of sensor/channels based on the highest gesture classification accuracy using KNN, Decision Tree and Naïve Bayes algorithms. As a result of this analysis, we obtained the optimum channel combination (Ch1, Ch4, Ch5, Ch8) data which generates the best classification accuracy. We used these 4 sensor datasets to train a 1D CNN with 78/22 train/test split. Dataset contains total 36 subjects. Data with subject number less than or equal to 28 is considered as training set and data with subject number greater than 28 is considered as test set. We also performed an optimization study on finding the optimum time signal window and overlap sizes of 100 ms and 50 ms . We achieved test accuracy of 97% for the classification accuracy of 3 gestures (hand at rest, wrist flexion, wrist extension).Advisors(s): Supervisor : Dr. Coskun Tekes Email id : ctekes@kennesaw.eduTopic(s): Artificial IntelligenceCS7992