The Impact of Muscle Fatigue on Deep Neural Network Performance

The main objective of the project is to prove that fatigue in the muscles affect the performance of deep nets and give low performance as compared to muscles in a fresh state.

Data were collected on three different states of the muscles which were static-fresh, statictired & transitional-fresh. These techniques will be explained in detail in the methodology.

The data was then used to train the models of Deep Neural Network (DNN), Convolutional Neural Network (CNN) & CNN with Feature Extraction (FE). DNN defines as networks that have an input, output and at least one hidden layer in between. CNN is one kind of deep neural network which contains convolutional layers which used filters to convolve an area of an input data to a smaller area detecting the important or specific part within it. CNN can contains normalization & pooling layers as well depending on the structure of the model.

Pooling layers are used to control overfitting by reducing the number of features and computation in the training. It converts the values of multiple neuron as an input into the one neuron as an output to the next layer. There are different kinds of pooling techniques in which Max-pooling & Average-pooling some of them. Max-pooling is a kind in which the maximum value of a neuron from a cluster of neurons becomes an input of a next layer neuron, while an Average-pooling takes an average of values of neurons in cluster and becomes an input of a next layer neuron. We used a feature extraction technique to improve the performance of our network. The feature extraction technique which was used in this project was Mean Absolute Value (MAV), Integrated EMG (IEMG) & Root Mean Square (RMS) proposed in [8]

Research Questions

1. i) Fatigue in the muscles affects the accuracy of EMG signals.
2. ii) Static state muscles have more accuracy than the transitional state.

iii) Comparison of DNN, CNN and CNN + FE with static-fresh muscles, static-tired and transitional-fresh muscles.

Methodology

In this project, we are using raw EMG data to classify four gestures using deep learning, the gestures are shown in Figure(2). This project consists of two main modules which are as follow:

1. Data Gathering with Fresh and Tired Muscles in Static and Transitional State
2. Apply Deep Neural Network on gathered data for accuracies

The first module of the project was to collect the EMG data of four gestures using MYO armband. An online course of Android development was taken to learn how tomodify and add more features to the open-source application. An open-source Android application [2] was modified and implemented according to the requirement of this project, and all of the required functionalities were added into the application. The language used in this project was Kotlin Android Studio 3.1. Kotlin language which was completely new for me learned and integrated into the application. To apply the neural network, I took a course for deep learning on Udemy. An open-source library by google was used for deep nets named TensorFlow including Keras with Python in Jupyter notebook. Working for the first time in Tensorflow, it took a lot of time to learn basic things and implementation of deep neural networks. After  the implementation of DNN, CNN was implemented using Keras and with the help of an online course. Furthermore, the feature extraction functions were implemented and tested with debugging to make sure they are working fine.

Data Gathering Tool (Android)

• User can change the frequency.
• Add or Remove gestures.
• Customize the number of points for collection before going to the next gesture.
• Customize pause duration before going to the next gesture.
• Customize the total number of iterations.
• Users can run the test on all gestures at once. In the end, the data points can be saved in just one file.
• The file can be saved on the phone in the required format.