This paper aims to classify grasp types from electromyography (EMG) data using artificial neural networks. EMG data was collected from six grasps and decomposed into intrinsic mode functions using empirical mode decomposition. Seven features were extracted from the frequency and time domains. Various feature subsets were used to train a neural network classifier, with the best results achieved using all features except variance from the EMG data and the first three intrinsic mode functions. The paper seeks to recognize intended grasps from EMG input data using neural networks in order to improve prosthetic control.
EEG based Brain Computer Interface (BCI) establishes a new channel between human brain and the
surrounding environment in order to disseminate instructions to the outside world. It is based on the
recording of temporary EEG changes during different types of motor imagery such as imagination of
different hand movements. The spatial pattern of activated cortical areas during motor imagery is similar
to that of real time executed movement. Time domain features and frequency domain features are extracted
with emphasis on recognizing discriminative features representing EEG trials recorded during imagination
of different hand movements. Then, classification into different hand movements is carried out.
OPTIMIZATION OF NEURAL NETWORK ARCHITECTURE FOR BIOMECHANIC CLASSIFICATION TA...ijaia
Electromyogram signals (EMGs) contain valuable information that can be used in man-machine interfacing between human users and myoelectric prosthetic devices. However, EMG signals are
complicated and prove difficult to analyze due to physiological noise and other issues. Computational
intelligence and machine learning techniques, such as artificial neural networks (ANNs), serve as powerful
tools for analyzing EMG signals and creating optimal myoelectric control schemes for prostheses. This
research examines the performance of four different neural network architectures (feedforward, recurrent,
counter propagation, and self organizing map) that were tasked with classifying walking speed when given
EMG inputs from 14 different leg muscles. Experiments conducted on the data set suggest that self
organizing map neural networks are capable of classifying walking speed with greater than 99% accuracy.
he main idea of the current work is to use a wireless Electroencephalography (EEG) headset as a remote control for the mouse cursor of a personal computer. The proposed system uses EEG signals as a communication link between brains and computers. Signal records obtained from the PhysioNet EEG dataset were analyzed using the Coif lets wavelets and many features were extracted using different amplitude estimators for the wavelet coefficients. The extracted features were inputted into machine learning algorithms to generate the decision rules required for our application. The suggested real time implementation of the system was tested and very good performance was achieved. This system could be helpful for disabled people as they can control computer applications via the imagination of fists and feet movements in addition to closing eyes for a short period of time
Functional magnetic resonance imaging-based brain decoding with visual semant...IJECEIAES
The activity pattern of the brain has been activated to identify a person in mind. Using the function magnetic resonance imaging (fMRI) to decipher brain decoding is the most accepted method. However, the accuracy of fMRI-based brain decoder is still restricted due to limited training samples. The limitations of the brain decoder using fMRI are passed through the design features proposed for many label coding and model training to predict these characteristics for a particular label. Moreover, what kind of semantic features for deciphering the neurological activity patterns are unclear. In current work, a new calculation model for learning decoding labels that is consistent with fMRI activity responses. The approach demonstrates the proposed corresponding label's success in terms of accuracy, which is decoded from brain activity patterns and compared with conventional text-derived feature technique. Besides, experimental studies present a training model based on multi-tasking to reduce the problems of limited training data sets. Therefore, the multi-task learning model is more efficient than modern methods of calculation, and decoding features may be easily obtained.
SUITABLE MOTHER WAVELET SELECTION FOR EEG SIGNALS ANALYSIS: FREQUENCY BANDS D...sipij
Wavelet transform (WT) is a powerful modern tool for time-frequency analysis of non-stationary signals such as electroencephalogram (EEG). The aim of this study is to choose the best and suitable mother wavelet function (MWT) for analyzing normal, seizure-free and seizured EEG signals. Several MWTs can be used, but the best MWT is the one that conserves the quasi-totality of information of the original signal on wavelet coefficients and gather more EEG rhythms in terms of frequency. In this study, Daubechies, Symlets and Coiflets orthogonal families were used as bsis mother wavelet functions. The percentage rootmeans square difference (PRD), the signal to noise ratio (SNR) and the simulated frequencies as the selection metrics. Simulation results indicate Daubechies wavelet at level 4 (Db4) as the most suitable MWT for EEG frequency bands decomposition.Furthermore, due to the redundancy of the extracted features, linear discriminant analysis (LDA) is applied for feature selection. Scatter plot showed that the selected feature vector represents the amount of changes in frequency distribution and carries most of the discriminative and representative information about their classes. Then, this study can provide a reference for the selection of a suitable MWT and discriminativefeatures.
EEG based Brain Computer Interface (BCI) establishes a new channel between human brain and the
surrounding environment in order to disseminate instructions to the outside world. It is based on the
recording of temporary EEG changes during different types of motor imagery such as imagination of
different hand movements. The spatial pattern of activated cortical areas during motor imagery is similar
to that of real time executed movement. Time domain features and frequency domain features are extracted
with emphasis on recognizing discriminative features representing EEG trials recorded during imagination
of different hand movements. Then, classification into different hand movements is carried out.
OPTIMIZATION OF NEURAL NETWORK ARCHITECTURE FOR BIOMECHANIC CLASSIFICATION TA...ijaia
Electromyogram signals (EMGs) contain valuable information that can be used in man-machine interfacing between human users and myoelectric prosthetic devices. However, EMG signals are
complicated and prove difficult to analyze due to physiological noise and other issues. Computational
intelligence and machine learning techniques, such as artificial neural networks (ANNs), serve as powerful
tools for analyzing EMG signals and creating optimal myoelectric control schemes for prostheses. This
research examines the performance of four different neural network architectures (feedforward, recurrent,
counter propagation, and self organizing map) that were tasked with classifying walking speed when given
EMG inputs from 14 different leg muscles. Experiments conducted on the data set suggest that self
organizing map neural networks are capable of classifying walking speed with greater than 99% accuracy.
he main idea of the current work is to use a wireless Electroencephalography (EEG) headset as a remote control for the mouse cursor of a personal computer. The proposed system uses EEG signals as a communication link between brains and computers. Signal records obtained from the PhysioNet EEG dataset were analyzed using the Coif lets wavelets and many features were extracted using different amplitude estimators for the wavelet coefficients. The extracted features were inputted into machine learning algorithms to generate the decision rules required for our application. The suggested real time implementation of the system was tested and very good performance was achieved. This system could be helpful for disabled people as they can control computer applications via the imagination of fists and feet movements in addition to closing eyes for a short period of time
Functional magnetic resonance imaging-based brain decoding with visual semant...IJECEIAES
The activity pattern of the brain has been activated to identify a person in mind. Using the function magnetic resonance imaging (fMRI) to decipher brain decoding is the most accepted method. However, the accuracy of fMRI-based brain decoder is still restricted due to limited training samples. The limitations of the brain decoder using fMRI are passed through the design features proposed for many label coding and model training to predict these characteristics for a particular label. Moreover, what kind of semantic features for deciphering the neurological activity patterns are unclear. In current work, a new calculation model for learning decoding labels that is consistent with fMRI activity responses. The approach demonstrates the proposed corresponding label's success in terms of accuracy, which is decoded from brain activity patterns and compared with conventional text-derived feature technique. Besides, experimental studies present a training model based on multi-tasking to reduce the problems of limited training data sets. Therefore, the multi-task learning model is more efficient than modern methods of calculation, and decoding features may be easily obtained.
SUITABLE MOTHER WAVELET SELECTION FOR EEG SIGNALS ANALYSIS: FREQUENCY BANDS D...sipij
Wavelet transform (WT) is a powerful modern tool for time-frequency analysis of non-stationary signals such as electroencephalogram (EEG). The aim of this study is to choose the best and suitable mother wavelet function (MWT) for analyzing normal, seizure-free and seizured EEG signals. Several MWTs can be used, but the best MWT is the one that conserves the quasi-totality of information of the original signal on wavelet coefficients and gather more EEG rhythms in terms of frequency. In this study, Daubechies, Symlets and Coiflets orthogonal families were used as bsis mother wavelet functions. The percentage rootmeans square difference (PRD), the signal to noise ratio (SNR) and the simulated frequencies as the selection metrics. Simulation results indicate Daubechies wavelet at level 4 (Db4) as the most suitable MWT for EEG frequency bands decomposition.Furthermore, due to the redundancy of the extracted features, linear discriminant analysis (LDA) is applied for feature selection. Scatter plot showed that the selected feature vector represents the amount of changes in frequency distribution and carries most of the discriminative and representative information about their classes. Then, this study can provide a reference for the selection of a suitable MWT and discriminativefeatures.
Brain-computer interface of focus and motor imagery using wavelet and recurre...TELKOMNIKA JOURNAL
Brain-computer interface is a technology that allows operating a device without involving muscles and sound, but directly from the brain through the processed electrical signals. The technology works by capturing electrical or magnetic signals from the brain, which are then processed to obtain information contained therein. Usually, BCI uses information from electroencephalogram (EEG) signals based on various variables reviewed. This study proposed BCI to move external devices such as a drone simulator based on EEG signal information. From the EEG signal was extracted to get motor imagery (MI) and focus variable using wavelet. Then, they were classified by recurrent neural networks (RNN). In overcoming the problem of vanishing memory from RNN, was used long short-term memory (LSTM). The results showed that BCI used wavelet, and RNN can drive external devices of non-training data with an accuracy of 79.6%. The experiment gave AdaDelta model is better than the Adam model in terms of accuracy and value losses. Whereas in computational learning time, Adam's model is faster than AdaDelta's model.
5. detection and separation of eeg artifacts using wavelet transform nov 11, ...IAESIJEECS
Bio-medical signal processing is one of the most important techniques of multichannel sensor network and it has a substantial concentration in medical application. However, the real-time and recorded signals in multisensory instruments contains different and huge amount of noise, and great work has been completed in developing most favorable structures for estimating the signal source from the noisy signal in multichannel observations. Methods have been developed to obtain the optimal linear estimation of the output signal through the Wide-Sense-Stationary (WSS) process with the help of time-invariant filters. In this process, the input signal and the noise signal are assumed to achieve the linear output signal. During the process, the non-stationary signals arise in the bio-medical signal processing in addition to it there is no effective structure to deal with them. Wavelets transform has been proved to be the efficient tool for handling the non-stationary signals, but wavelet provide any possible way to approach multichannel signal processing. Based on the basic structure of linear estimation of non-stationary multichannel data and statistical models of spatial signal coherence acquire through the wavelet transform in multichannel estimation. The above methods can be used for Electroencephalography (EEG) signal denoising through the original signal and then implement the noise reduction technique to evaluate their performance such as SNR, MSE and computation time.
Wavelet-based EEG processing for computer-aided seizure detection and epileps...IJERA Editor
Many Neurological disorders are very difficult to detect. One such Neurological disorder which we are going to discuss in this paper is Epilepsy. Epilepsy means sudden change in the behavior of a human being for a short period of time. This is caused due to seizures in the brain. Many researches are going onto detect epilepsy detection through analyzing EEG. One such method of epilepsy detection is proposed in this paper. This technique employs Discrete Wave Transform (DWT) method for pre-processing, Approximate Entropy (ApEn) to extract features and Artificial Neural Network (ANN) for classification. This paper presented a detailed survey of various methods that are being used for epilepsy detection and also proposes a wavelet based epilepsy detection method
EEG S IGNAL Q UANTIFICATION B ASED ON M ODUL L EVELS sipij
This article proposes a contribution to quantify EE
G signals outline. This technique uses two tools fo
r EEG
signal characteristics extraction. Our tests were r
ealized on the basis of 32 canals EEG canals using
Neuroscan software. EEG example demonstration is re
ferenced CZ and is sampled at 1000HZ. The
principal aim of this technique is to reduce the im
portant volume of EEG signal data Without losing an
y
information. EEG signals are quantified on the basi
s of a whole predefined levels The obtained results
show that an EEG alignment can be posted in a quant
ified form.
Recognition of new gestures using myo armband for myoelectric prosthetic appl...IJECEIAES
Myoelectric prostheses are a viable solution for people with amputations. The chal- lenge in implementing a usable myoelectric prosthesis lies in accurately recognizing different hand gestures. The current myoelectric devices usually implement very few hand gestures. In order to approximate a real hand functionality, a myoelectric prosthesis should implement a large number of hand and finger gestures. However, increasing number of gestures can lead to a decrease in recognition accuracy. In this work a Myo armband device is used to recognize fourteen gestures (five build in gestures of Myo armband in addition to nine new gestures). The data in this research is collected from three body-able subjects for a period of 7 seconds per gesture. The proposed method uses a pattern recognition technique based on Multi-Layer Perceptron Neural Network (MLPNN). The results show an average accuracy of 90.5% in recognizing the proposed fourteen gestures.
This document will examine issues pertaining to feature extraction, classification and prediction. It will
consider the application of these techniques to unlabelled Electroencephalogram (E.E.G.) data in an
attempt to discriminate between left and right hand imagery movements
INHIBITION AND SET-SHIFTING TASKS IN CENTRAL EXECUTIVE FUNCTION OF WORKING ME...sipij
Understanding of neuro-dynamics of a complex higher cognitive process, Working Memory (WM) is
challenging. In WM, information processing occurs through four subsystems: phonological loop, visual
sketch pad, memory buffer and central executive function (CEF). CEF plays a principal role in WM. In this
study, our objective was to understand the neurospatial correlates of CEF during inhibition and set-shifting
processes. Thirty healthy educated subjects were selected. Event-Related Potential (ERP) related to visual
inhibition and set-shifting task was collected using 32 channel EEG system. Activation of those ERPs
components was analyzed using amplitudes of positive and negative peaks. Experiment was controlled
using certain parametric constraints to judge behavior, based on average responses in order to establish
relationship between ERP and local area of brain activation and represented using standardized low
resolution brain electromagnetic tomography. The average score of correct responses was higher for
inhibition task (87.5%) as compared to set-shifting task (59.5%). The peak amplitude of neuronal activity
for inhibition task was lower compared to set-shifting task in fronto-parieto-central regions. Hence this
proposed paradigm and technique can be used to measure inhibition and set-shifting neuronal processes in
understanding pathological central executive functioning in patients with neuro-psychiatric disorders.
Robot Motion Control Using the Emotiv EPOC EEG SystemjournalBEEI
Brain-computer interfaces have been explored for years with the intent of using human thoughts to control mechanical system. By capturing the transmission of signals directly from the human brain or electroencephalogram (EEG), human thoughts can be made as motion commands to the robot. This paper presents a prototype for an electroencephalogram (EEG) based brain-actuated robot control system using mental commands. In this study, Linear Discriminant Analysis (LDA) and Support Vector Machine (SVM) method were combined to establish the best model. Dataset containing features of EEG signals were obtained from the subject non-invasively using Emotiv EPOC headset. The best model was then used by Brain-Computer Interface (BCI) to classify the EEG signals into robot motion commands to control the robot directly. The result of the classification gave the average accuracy of 69.06%.
Smart element aware gate controller for intelligent wheeled robot navigationIJECEIAES
The directing of a wheeled robot in an unknown moving environment with physical barriers is a difficult proposition. In particular, having an optimal or near-optimal path that avoids obstacles is a major challenge. In this paper, a modified neuro-controller mechanism is proposed for controlling the movement of an indoor mobile robot. The proposed mechanism is based on the design of a modified Elman neural network (MENN) with an effective element aware gate (MEEG) as the neuro-controller. This controller is updated to overcome the rigid and dynamic barriers in the indoor area. The proposed controller is implemented with a mobile robot known as Khepera IV in a practical manner. The practical results demonstrate that the proposed mechanism is very efficient in terms of providing shortest distance to reach the goal with maximum velocity as compared with the MENN. Specifically, the MEEG is better than MENN in minimizing the error rate by 58.33%.
Tactile Brain-Computer Interface Using Classification of P300 Responses Evoke...Takumi Kodama
Kodama T, Makino S, Rutkowski TM. Tactile Brain-Computer Interface Using Classification of P300 Responses Evoked by Full Body Spatial Vibrotactile Stimuli. In: Asia-Pacific Signal and Information Processing Association, 2016 Annual Summit and Conference (APSIPA ASC 2016). APSIPA. Jeju, Korea: IEEE Press; 2016.
Evolutionary Algorithm for Optimal Connection Weights in Artificial Neural Ne...CSCJournals
A neural network may be considered as an adaptive system that progressively self-organizes in order to approximate the solution, making the problem solver free from the need to accurately and unambiguously specify the steps towards the solution. Moreover, Evolutionary Artificial Neural Networks (EANNs) have the ability to progressively improve their performance on a given task by executing learning. An evolutionary computation gives adaptability for connection weights using feed forward architecture. In this paper, the use of evolutionary computation for feed-forward neural network learning is discussed. To check the validation of proposed method, XOR benchmark problem has been used. The accuracy of the proposed model is more satisfactory as compared to gradient method.
Motor Imagery Recognition of EEG Signal using Cuckoo Search Masking Empirical...ijtsrd
Brain Computer Interface BCI aims at providing an alternate means of communication and control to people with severe cognitive or sensory motor disabilities. Brain Computer Interface in electroencephalogram EEG is of great important but it is challenging to manage the non stationary EEG. EEG signals are more vulnerable to contamination due to noise and artifacts. In our proposed work, we used Cuckoo Search Masking Empirical Mode decomposition to ignore such vulnerable things. Initially, the features of EEG signals are taken such as Energy, AR Coefficients, Morphological features and Fuzzy Approximate Entropy. Then, for Feature extraction method, Masking Empirical Mode Decomposition MEMD is applied to deal with motor imagery MI recognition tasks. The EEG signal is decomposed by MEMD and hybrid features are then extracted from the first two intrinsic mode functions IMFs . After the extracted features, Cuckoo Search algorithm is used to select the significant features. Different weights for the relevance and redundancy in the fitness function of the proposed algorithm are used to further improve their performance in terms of the number of features and the classification accuracy and finally they are fed into Linear Discriminant Analysis for classification. This analysis produces models whose accuracy is as good as more complex method. The results show that our proposed method can achieve the highest accuracy, maximal MI, recall as well as precision for Motor Imagery Recognition tasks. Our proposed method is comparable or superior than existing method. Jaipriya D ""Motor Imagery Recognition of EEG Signal using Cuckoo-Search Masking Empirical Mode Decomposition"" Published in International Journal of Trend in Scientific Research and Development (ijtsrd), ISSN: 2456-6470, Volume-4 | Issue-2 , February 2020,
URL: https://www.ijtsrd.com/papers/ijtsrd30020.pdf
Paper Url : https://www.ijtsrd.com/engineering/electronics-and-communication-engineering/30020/motor-imagery-recognition-of-eeg-signal-using-cuckoo-search-masking-empirical-mode-decomposition/jaipriya-d
The paper is based on feed forward neural network (FFNN) optimization by particle swarm intelligence (PSI) used to provide initial weights and biases to train neural network. Once the weights and biases are found using Particle swarm optimization (PSO) with neural network used as training algorithm for specified epoch, the same are used to train the neural network for training and classification of benchmark problems. Further the approach is tested for offline signature classifications. A comparison is made between normal FFNN with random weights and biases and FFNN with particle swarm optimized weights and biases. Firstly, the performance is tested on two benchmark databases for neural network, The Breast Cancer Database and the Diabetic Database. Result shows that neural network performs better with initial weights and biases obtained by Particle Swarm optimization. The network converges faster with PSO obtained initial weights and biases for FFNN and classification accuracy is increased.
An Entropy-based Feature in Epileptic Seizure Prediction Algorithmiosrjce
Epilepsy prediction is a vital demand for people suffering from epileptic onset. Prediction of seizure
onsets could be very useful for drug-resistant epileptic patients. We propose an epileptic seizure prediction
algorithm to predict an onset of epilepsy and discriminate between pre-seizure periods and seizure free periods.
The proposed algorithm is based on entropy features of 60 (1 hour segmented into 60 periods) with free seizure
periods and repeated for 24 hour, and 60 (pre-seizure periods) of the CHB-MIT Scalp EEG Database (Female
less or equal 12 age). Critical values of the sample entropy and approximate entropy are estimated to locate
starting of the seizure onset. These values are taken as warning to a probably seizure starts within a specific
time. The prediction time in order of 1min- 49min is achieved in 60 seizure periods under study in this task.
SVM is used to classify pre-seizure periods from seizure free periods for the mentioned data. The performance
is evaluated and analysed
Classification of EEG Signal for Epileptic Seizure DetectionusingEMD and ELMIJTET Journal
Abstract—This paper proposes the classification of EEG signal for epilepsy diagnosis. Epilepsy is a neurological disorder which occurs due to synchronous neuronal activity in brain. Empirical Mode Decomposition (EMD), Extreme Learning Machine (ELM) are the techniquedelivered in the proposed method.Input EEG signal, which is available in online as Bonn Database is decomposed into five Intrinsic Mode Functions (IMFs) using EMD.Higher Order Statistical moments such as Variance, Skewness and Kurtosis are drawn out as features from the decomposed signals. Extreme Learning Machine is used as a classifier to classify the EEG signals with the taken features, under various categories that include healthy and ictal, interictal and ictal, Non seizure and seizure, healthy, interictal and ictal. The proposed method gives 100%accuracy, 100%sensitivity in discriminating interictal and ictal, non seizure and seizure, healthy and ictal, healthy, interictal and ictal, 100% specificity in classifying healthy and ictal, interictal and ictal and 100% and 99%accuracy in case of discriminating interictal and ictal, non seizure and seizure.
Brain-computer interface of focus and motor imagery using wavelet and recurre...TELKOMNIKA JOURNAL
Brain-computer interface is a technology that allows operating a device without involving muscles and sound, but directly from the brain through the processed electrical signals. The technology works by capturing electrical or magnetic signals from the brain, which are then processed to obtain information contained therein. Usually, BCI uses information from electroencephalogram (EEG) signals based on various variables reviewed. This study proposed BCI to move external devices such as a drone simulator based on EEG signal information. From the EEG signal was extracted to get motor imagery (MI) and focus variable using wavelet. Then, they were classified by recurrent neural networks (RNN). In overcoming the problem of vanishing memory from RNN, was used long short-term memory (LSTM). The results showed that BCI used wavelet, and RNN can drive external devices of non-training data with an accuracy of 79.6%. The experiment gave AdaDelta model is better than the Adam model in terms of accuracy and value losses. Whereas in computational learning time, Adam's model is faster than AdaDelta's model.
5. detection and separation of eeg artifacts using wavelet transform nov 11, ...IAESIJEECS
Bio-medical signal processing is one of the most important techniques of multichannel sensor network and it has a substantial concentration in medical application. However, the real-time and recorded signals in multisensory instruments contains different and huge amount of noise, and great work has been completed in developing most favorable structures for estimating the signal source from the noisy signal in multichannel observations. Methods have been developed to obtain the optimal linear estimation of the output signal through the Wide-Sense-Stationary (WSS) process with the help of time-invariant filters. In this process, the input signal and the noise signal are assumed to achieve the linear output signal. During the process, the non-stationary signals arise in the bio-medical signal processing in addition to it there is no effective structure to deal with them. Wavelets transform has been proved to be the efficient tool for handling the non-stationary signals, but wavelet provide any possible way to approach multichannel signal processing. Based on the basic structure of linear estimation of non-stationary multichannel data and statistical models of spatial signal coherence acquire through the wavelet transform in multichannel estimation. The above methods can be used for Electroencephalography (EEG) signal denoising through the original signal and then implement the noise reduction technique to evaluate their performance such as SNR, MSE and computation time.
Wavelet-based EEG processing for computer-aided seizure detection and epileps...IJERA Editor
Many Neurological disorders are very difficult to detect. One such Neurological disorder which we are going to discuss in this paper is Epilepsy. Epilepsy means sudden change in the behavior of a human being for a short period of time. This is caused due to seizures in the brain. Many researches are going onto detect epilepsy detection through analyzing EEG. One such method of epilepsy detection is proposed in this paper. This technique employs Discrete Wave Transform (DWT) method for pre-processing, Approximate Entropy (ApEn) to extract features and Artificial Neural Network (ANN) for classification. This paper presented a detailed survey of various methods that are being used for epilepsy detection and also proposes a wavelet based epilepsy detection method
EEG S IGNAL Q UANTIFICATION B ASED ON M ODUL L EVELS sipij
This article proposes a contribution to quantify EE
G signals outline. This technique uses two tools fo
r EEG
signal characteristics extraction. Our tests were r
ealized on the basis of 32 canals EEG canals using
Neuroscan software. EEG example demonstration is re
ferenced CZ and is sampled at 1000HZ. The
principal aim of this technique is to reduce the im
portant volume of EEG signal data Without losing an
y
information. EEG signals are quantified on the basi
s of a whole predefined levels The obtained results
show that an EEG alignment can be posted in a quant
ified form.
Recognition of new gestures using myo armband for myoelectric prosthetic appl...IJECEIAES
Myoelectric prostheses are a viable solution for people with amputations. The chal- lenge in implementing a usable myoelectric prosthesis lies in accurately recognizing different hand gestures. The current myoelectric devices usually implement very few hand gestures. In order to approximate a real hand functionality, a myoelectric prosthesis should implement a large number of hand and finger gestures. However, increasing number of gestures can lead to a decrease in recognition accuracy. In this work a Myo armband device is used to recognize fourteen gestures (five build in gestures of Myo armband in addition to nine new gestures). The data in this research is collected from three body-able subjects for a period of 7 seconds per gesture. The proposed method uses a pattern recognition technique based on Multi-Layer Perceptron Neural Network (MLPNN). The results show an average accuracy of 90.5% in recognizing the proposed fourteen gestures.
This document will examine issues pertaining to feature extraction, classification and prediction. It will
consider the application of these techniques to unlabelled Electroencephalogram (E.E.G.) data in an
attempt to discriminate between left and right hand imagery movements
INHIBITION AND SET-SHIFTING TASKS IN CENTRAL EXECUTIVE FUNCTION OF WORKING ME...sipij
Understanding of neuro-dynamics of a complex higher cognitive process, Working Memory (WM) is
challenging. In WM, information processing occurs through four subsystems: phonological loop, visual
sketch pad, memory buffer and central executive function (CEF). CEF plays a principal role in WM. In this
study, our objective was to understand the neurospatial correlates of CEF during inhibition and set-shifting
processes. Thirty healthy educated subjects were selected. Event-Related Potential (ERP) related to visual
inhibition and set-shifting task was collected using 32 channel EEG system. Activation of those ERPs
components was analyzed using amplitudes of positive and negative peaks. Experiment was controlled
using certain parametric constraints to judge behavior, based on average responses in order to establish
relationship between ERP and local area of brain activation and represented using standardized low
resolution brain electromagnetic tomography. The average score of correct responses was higher for
inhibition task (87.5%) as compared to set-shifting task (59.5%). The peak amplitude of neuronal activity
for inhibition task was lower compared to set-shifting task in fronto-parieto-central regions. Hence this
proposed paradigm and technique can be used to measure inhibition and set-shifting neuronal processes in
understanding pathological central executive functioning in patients with neuro-psychiatric disorders.
Robot Motion Control Using the Emotiv EPOC EEG SystemjournalBEEI
Brain-computer interfaces have been explored for years with the intent of using human thoughts to control mechanical system. By capturing the transmission of signals directly from the human brain or electroencephalogram (EEG), human thoughts can be made as motion commands to the robot. This paper presents a prototype for an electroencephalogram (EEG) based brain-actuated robot control system using mental commands. In this study, Linear Discriminant Analysis (LDA) and Support Vector Machine (SVM) method were combined to establish the best model. Dataset containing features of EEG signals were obtained from the subject non-invasively using Emotiv EPOC headset. The best model was then used by Brain-Computer Interface (BCI) to classify the EEG signals into robot motion commands to control the robot directly. The result of the classification gave the average accuracy of 69.06%.
Smart element aware gate controller for intelligent wheeled robot navigationIJECEIAES
The directing of a wheeled robot in an unknown moving environment with physical barriers is a difficult proposition. In particular, having an optimal or near-optimal path that avoids obstacles is a major challenge. In this paper, a modified neuro-controller mechanism is proposed for controlling the movement of an indoor mobile robot. The proposed mechanism is based on the design of a modified Elman neural network (MENN) with an effective element aware gate (MEEG) as the neuro-controller. This controller is updated to overcome the rigid and dynamic barriers in the indoor area. The proposed controller is implemented with a mobile robot known as Khepera IV in a practical manner. The practical results demonstrate that the proposed mechanism is very efficient in terms of providing shortest distance to reach the goal with maximum velocity as compared with the MENN. Specifically, the MEEG is better than MENN in minimizing the error rate by 58.33%.
Tactile Brain-Computer Interface Using Classification of P300 Responses Evoke...Takumi Kodama
Kodama T, Makino S, Rutkowski TM. Tactile Brain-Computer Interface Using Classification of P300 Responses Evoked by Full Body Spatial Vibrotactile Stimuli. In: Asia-Pacific Signal and Information Processing Association, 2016 Annual Summit and Conference (APSIPA ASC 2016). APSIPA. Jeju, Korea: IEEE Press; 2016.
Evolutionary Algorithm for Optimal Connection Weights in Artificial Neural Ne...CSCJournals
A neural network may be considered as an adaptive system that progressively self-organizes in order to approximate the solution, making the problem solver free from the need to accurately and unambiguously specify the steps towards the solution. Moreover, Evolutionary Artificial Neural Networks (EANNs) have the ability to progressively improve their performance on a given task by executing learning. An evolutionary computation gives adaptability for connection weights using feed forward architecture. In this paper, the use of evolutionary computation for feed-forward neural network learning is discussed. To check the validation of proposed method, XOR benchmark problem has been used. The accuracy of the proposed model is more satisfactory as compared to gradient method.
Motor Imagery Recognition of EEG Signal using Cuckoo Search Masking Empirical...ijtsrd
Brain Computer Interface BCI aims at providing an alternate means of communication and control to people with severe cognitive or sensory motor disabilities. Brain Computer Interface in electroencephalogram EEG is of great important but it is challenging to manage the non stationary EEG. EEG signals are more vulnerable to contamination due to noise and artifacts. In our proposed work, we used Cuckoo Search Masking Empirical Mode decomposition to ignore such vulnerable things. Initially, the features of EEG signals are taken such as Energy, AR Coefficients, Morphological features and Fuzzy Approximate Entropy. Then, for Feature extraction method, Masking Empirical Mode Decomposition MEMD is applied to deal with motor imagery MI recognition tasks. The EEG signal is decomposed by MEMD and hybrid features are then extracted from the first two intrinsic mode functions IMFs . After the extracted features, Cuckoo Search algorithm is used to select the significant features. Different weights for the relevance and redundancy in the fitness function of the proposed algorithm are used to further improve their performance in terms of the number of features and the classification accuracy and finally they are fed into Linear Discriminant Analysis for classification. This analysis produces models whose accuracy is as good as more complex method. The results show that our proposed method can achieve the highest accuracy, maximal MI, recall as well as precision for Motor Imagery Recognition tasks. Our proposed method is comparable or superior than existing method. Jaipriya D ""Motor Imagery Recognition of EEG Signal using Cuckoo-Search Masking Empirical Mode Decomposition"" Published in International Journal of Trend in Scientific Research and Development (ijtsrd), ISSN: 2456-6470, Volume-4 | Issue-2 , February 2020,
URL: https://www.ijtsrd.com/papers/ijtsrd30020.pdf
Paper Url : https://www.ijtsrd.com/engineering/electronics-and-communication-engineering/30020/motor-imagery-recognition-of-eeg-signal-using-cuckoo-search-masking-empirical-mode-decomposition/jaipriya-d
The paper is based on feed forward neural network (FFNN) optimization by particle swarm intelligence (PSI) used to provide initial weights and biases to train neural network. Once the weights and biases are found using Particle swarm optimization (PSO) with neural network used as training algorithm for specified epoch, the same are used to train the neural network for training and classification of benchmark problems. Further the approach is tested for offline signature classifications. A comparison is made between normal FFNN with random weights and biases and FFNN with particle swarm optimized weights and biases. Firstly, the performance is tested on two benchmark databases for neural network, The Breast Cancer Database and the Diabetic Database. Result shows that neural network performs better with initial weights and biases obtained by Particle Swarm optimization. The network converges faster with PSO obtained initial weights and biases for FFNN and classification accuracy is increased.
An Entropy-based Feature in Epileptic Seizure Prediction Algorithmiosrjce
Epilepsy prediction is a vital demand for people suffering from epileptic onset. Prediction of seizure
onsets could be very useful for drug-resistant epileptic patients. We propose an epileptic seizure prediction
algorithm to predict an onset of epilepsy and discriminate between pre-seizure periods and seizure free periods.
The proposed algorithm is based on entropy features of 60 (1 hour segmented into 60 periods) with free seizure
periods and repeated for 24 hour, and 60 (pre-seizure periods) of the CHB-MIT Scalp EEG Database (Female
less or equal 12 age). Critical values of the sample entropy and approximate entropy are estimated to locate
starting of the seizure onset. These values are taken as warning to a probably seizure starts within a specific
time. The prediction time in order of 1min- 49min is achieved in 60 seizure periods under study in this task.
SVM is used to classify pre-seizure periods from seizure free periods for the mentioned data. The performance
is evaluated and analysed
Classification of EEG Signal for Epileptic Seizure DetectionusingEMD and ELMIJTET Journal
Abstract—This paper proposes the classification of EEG signal for epilepsy diagnosis. Epilepsy is a neurological disorder which occurs due to synchronous neuronal activity in brain. Empirical Mode Decomposition (EMD), Extreme Learning Machine (ELM) are the techniquedelivered in the proposed method.Input EEG signal, which is available in online as Bonn Database is decomposed into five Intrinsic Mode Functions (IMFs) using EMD.Higher Order Statistical moments such as Variance, Skewness and Kurtosis are drawn out as features from the decomposed signals. Extreme Learning Machine is used as a classifier to classify the EEG signals with the taken features, under various categories that include healthy and ictal, interictal and ictal, Non seizure and seizure, healthy, interictal and ictal. The proposed method gives 100%accuracy, 100%sensitivity in discriminating interictal and ictal, non seizure and seizure, healthy and ictal, healthy, interictal and ictal, 100% specificity in classifying healthy and ictal, interictal and ictal and 100% and 99%accuracy in case of discriminating interictal and ictal, non seizure and seizure.
Correlation Analysis of Electromyogram SignalsIJMTST Journal
An inability to adapt myoelectric interfaces to a user’s unique style of hand motion. The system also adapts
the motion style of an opposite limb. These are the important factors inhibiting the practical application of
myoelectric interfaces. This is mainly attributed to the individual differences in the exhibited electromyogram
(EMG) signals generated by the muscles of different limbs. In this project myoelectric interface easily adapts
the signal from the users and maintains good movement recognition performance. At the initial stage the
myoelectric signal is extracted from the user by using the data acquisition system. A new set of features
describing the movements of user’s is extracted and the user’s features are classifed using SVM
classification. The given signal is then compared with the database signal with the accuracy of 90.910 %
across all the EMG signals.
Survey analysis for optimization algorithms applied to electroencephalogramIJECEIAES
This paper presents a survey for optimization approaches that analyze and classify electroencephalogram (EEG) signals. The automatic analysis of EEG presents a significant challenge due to the high-dimensional data volume. Optimization algorithms seek to achieve better accuracy by selecting practical features and reducing unwanted features. Forty-seven reputable research papers are provided in this work, emphasizing the developed and executed techniques divided into seven groups based on the applied optimization algorithm particle swarm optimization (PSO), ant colony optimization (ACO), artificial bee colony (ABC), grey wolf optimizer (GWO), Bat, Firefly, and other optimizer approaches). The main measures to analyze this paper are accuracy, precision, recall, and F1-score assessment. Several datasets have been utilized in the included papers like EEG Bonn University, CHB-MIT, electrocardiography (ECG) dataset, and other datasets. The results have proven that the PSO and GWO algorithms have achieved the highest accuracy rate of around 99% compared with other techniques.
Using deep neural networks in classifying electromyography signals for hand g...IAESIJAI
Electromyography (EMG) signals are used for various applications, especially in smart prostheses. Recognizing various gestures (hand movements) in EMG systems introduces challenges. These challenges include the noise effect on EMG signals and the difficulty in identifying the exact movement from the collected EMG data amongst others. In this paper, three neural network models are trained using an open EMG dataset to classify and recognize seven different gestures based on the collected EMG data. The three implemented models are: a four-layer deep neural network (DNN), an eight-layer DNN, and a five-layer convolutional neural network (CNN). In addition, five optimizers are tested for each model, namely Adam, Adamax, Nadam, Adagrad, and AdaDelta. It has been found that four layers achieve respectable recognition accuracy of 95% in the proposed model.
Overview of Machine Learning and Deep Learning Methods in Brain Computer Inte...IJCSES Journal
Research under the field of Brain Computer Interfaces is adapting various Machine Learning and Deep
Learning techniques in recent times. With the advent of modern BCI, the data generated by various devices
is now capable of detecting brain signals more accurately. This paper gives an overview of all the steps
involved in the process of applying Machine Learning as well as Deep Learning methods from Data
Acquisition to application of algorithms. It aims to study techniques currently employed to extract data,
features from brain data, different algorithms employed to draw insights from the extracted features, and
how it can be used in various BCI applications. By this study, I aim to put forward current Machine
Learning and Deep Learning Trends in the field of BCI.
METHODS OF COMMAND RECOGNITION USING SINGLE-CHANNEL EEGSijistjournal
This work proposes to recognize a user's commands by analysing his/her brainwaves captured with single channel electroencephalogram (EEG). Whenever a user intends to issue one of the pre-defined commands, the proposed system prompts him/her all the candidate commands in turn. Then, the user is asked to be concentrated as possible as he/she can, when the desired command is shown. It is assumed that the concentration will present a certain pattern of “Yes” in the captured EEG, as opposed to a certain pattern of “No” when the user is relaxed. Accordingly, the task is to determine that the captured EEG is “Yes” or not. This work compares three recognition methods, respectively, based on Gaussian mixture models, hidden Markov models and recurrent neural network, and conducts experiments using 2400 test EEG samples recorded from 10 subjects.
METHODS OF COMMAND RECOGNITION USING SINGLE-CHANNEL EEGSijistjournal
This work proposes to recognize a user's commands by analysing his/her brainwaves captured with single channel electroencephalogram (EEG). Whenever a user intends to issue one of the pre-defined commands, the proposed system prompts him/her all the candidate commands in turn. Then, the user is asked to be concentrated as possible as he/she can, when the desired command is shown. It is assumed that the concentration will present a certain pattern of “Yes” in the captured EEG, as opposed to a certain pattern of “No” when the user is relaxed. Accordingly, the task is to determine that the captured EEG is “Yes” or not. This work compares three recognition methods, respectively, based on Gaussian mixture models, hidden Markov models and recurrent neural network, and conducts experiments using 2400 test EEG samples recorded from 10 subjects.
METHODS OF COMMAND RECOGNITION USING SINGLE-CHANNEL EEGSijistjournal
This work proposes to recognize a user's commands by analysing his/her brainwaves captured with single
channel electroencephalogram (EEG). Whenever a user intends to issue one of the pre-defined
commands, the proposed system prompts him/her all the candidate commands in turn. Then, the user is
asked to be concentrated as possible as he/she can, when the desired command is shown. It is assumed
that the concentration will present a certain pattern of “Yes” in the captured EEG, as opposed to a
certain pattern of “No” when the user is relaxed. Accordingly, the task is to determine that the captured
EEG is “Yes” or not. This work compares three recognition methods, respectively, based on Gaussian
mixture models, hidden Markov models and recurrent neural network, and conducts experiments using
2400 test EEG samples recorded from 10 subjects.
METHODS OF COMMAND RECOGNITION USING SINGLE-CHANNEL EEGSijistjournal
This work proposes to recognize a user's commands by analysing his/her brainwaves captured with single channel electroencephalogram (EEG). Whenever a user intends to issue one of the pre-defined commands, the proposed system prompts him/her all the candidate commands in turn. Then, the user is
asked to be concentrated as possible as he/she can, when the desired command is shown. It is assumed
that the concentration will present a certain pattern of “Yes” in the captured EEG, as opposed to a
certain pattern of “No” when the user is relaxed. Accordingly, the task is to determine that the captured EEG is “Yes” or not. This work compares three recognition methods, respectively, based on Gaussian mixture models, hidden Markov models and recurrent neural network, and conducts experiments using
2400 test EEG samples recorded from 10 subjects.
METHODS OF COMMAND RECOGNITION USING SINGLE-CHANNEL EEGSijistjournal
This work proposes to recognize a user's commands by analysing his/her brainwaves captured with single
channel electroencephalogram (EEG). Whenever a user intends to issue one of the pre-defined
commands, the proposed system prompts him/her all the candidate commands in turn. Then, the user is
asked to be concentrated as possible as he/she can, when the desired command is shown. It is assumed
that the concentration will present a certain pattern of “Yes” in the captured EEG, as opposed to a
certain pattern of “No” when the user is relaxed. Accordingly, the task is to determine that the captured
EEG is “Yes” or not. This work compares three recognition methods, respectively, based on Gaussian
mixture models, hidden Markov models and recurrent neural network, and conducts experiments using
2400 test EEG samples recorded from 10 subjects.
A Comparative Study of Machine Learning Algorithms for EEG Signal Classificationsipij
In this paper, different machine learning algorithms such as Linear Discriminant Analysis, Support vector
machine (SVM), Multi-layer perceptron, Random forest, K-nearest neighbour, and Autoencoder with SVM
have been compared. This comparison was conducted to seek a robust method that would produce good
classification accuracy. To this end, a robust method of classifying raw Electroencephalography (EEG)
signals associated with imagined movement of the right hand and relaxation state, namely Autoencoder
with SVM has been proposed. The EEG dataset used in this research was created by the University of
Tubingen, Germany. The best classification accuracy achieved was 70.4% with SVM through feature
engineering. However, our prosed method of autoencoder in combination with SVM produced a similar
accuracy of 65% without using any feature engineering technique. This research shows that this system of
classification of motor movements can be used in a Brain-Computer Interface system (BCI) to mentally
control a robotic device or an exoskeleton.
A Comparative Study of Machine Learning Algorithms for EEG Signal Classificationsipij
In this paper, different machine learning algorithms such as Linear Discriminant Analysis, Support vector
machine (SVM), Multi-layer perceptron, Random forest, K-nearest neighbour, and Autoencoder with SVM
have been compared. This comparison was conducted to seek a robust method that would produce good
classification accuracy. To this end, a robust method of classifying raw Electroencephalography (EEG)
signals associated with imagined movement of the right hand and relaxation state, namely Autoencoder
with SVM has been proposed. The EEG dataset used in this research was created by the University of
Tubingen, Germany. The best classification accuracy achieved was 70.4% with SVM through feature
engineering. However, our prosed method of autoencoder in combination with SVM produced a similar
accuracy of 65% without using any feature engineering technique. This research shows that this system of
classification of motor movements can be used in a Brain-Computer Interface system (BCI) to mentally
control a robotic device or an exoskeleton.
A Comparative Study of Machine Learning Algorithms for EEG Signal Classificationsipij
In this paper, different machine learning algorithms such as Linear Discriminant Analysis, Support vector
machine (SVM), Multi-layer perceptron, Random forest, K-nearest neighbour, and Autoencoder with SVM
have been compared. This comparison was conducted to seek a robust method that would produce good
classification accuracy. To this end, a robust method of classifying raw Electroencephalography (EEG)
signals associated with imagined movement of the right hand and relaxation state, namely Autoencoder
with SVM has been proposed. The EEG dataset used in this research was created by the University of
Tubingen, Germany. The best classification accuracy achieved was 70.4% with SVM through feature
engineering. However, our prosed method of autoencoder in combination with SVM produced a similar
accuracy of 65% without using any feature engineering technique. This research shows that this system of
classification of motor movements can be used in a Brain-Computer Interface system (BCI) to mentally
control a robotic device or an exoskeleton.
A COMPARATIVE STUDY OF MACHINE LEARNING ALGORITHMS FOR EEG SIGNAL CLASSIFICATIONsipij
In this paper, different machine learning algorithms such as Linear Discriminant Analysis, Support vector
machine (SVM), Multi-layer perceptron, Random forest, K-nearest neighbour, and Autoencoder with SVM
have been compared. This comparison was conducted to seek a robust method that would produce good
classification accuracy. To this end, a robust method of classifying raw Electroencephalography (EEG)
signals associated with imagined movement of the right hand and relaxation state, namely Autoencoder
with SVM has been proposed. The EEG dataset used in this research was created by the University of
Tubingen, Germany. The best classification accuracy achieved was 70.4% with SVM through feature
engineering. However, our prosed method of autoencoder in combination with SVM produced a similar
accuracy of 65% without using any feature engineering technique. This research shows that this system of
classification of motor movements can be used in a Brain-Computer Interface system (BCI) to mentally
control a robotic device or an exoskeleton.
Brain computer interface is a technique used to capture the emotions and thoughts of a brain activity using Electroencephalogram(EEG). So it is useful to communicate with humans.In this paper, it deals with a Neuro sky mind wave to detect the signals for physically challenged people. If the brain activity of a signal and already attained signals are matched it displayed on the PC then it converted into an audible signal.
International Journal of Engineering Research and Applications (IJERA) is an open access online peer reviewed international journal that publishes research and review articles in the fields of Computer Science, Neural Networks, Electrical Engineering, Software Engineering, Information Technology, Mechanical Engineering, Chemical Engineering, Plastic Engineering, Food Technology, Textile Engineering, Nano Technology & science, Power Electronics, Electronics & Communication Engineering, Computational mathematics, Image processing, Civil Engineering, Structural Engineering, Environmental Engineering, VLSI Testing & Low Power VLSI Design etc.
Similar to rob 537 final paper(fourth modify) (20)
1. ROB 537 Final Project 1
Abstract— this paper proposes an approach to classify grasp
types from Electromyography (EMG) data by extracting features
from the data and using those features to train a neural network
classifier. EMG data was used from six different grasps, and this
data was broken down into Intrinsic Mode Functions (IMFs)
using Empirical Mode Decomposition (EMD). Seven features
were then extracted in the frequency and time domain. Various
subsets of these features were then used to train a feedforward,
single hidden layer neural network classifier using scaled
conjugate gradient backpropagation as a training function and
cross entropy as a performance measure. A highly successful
classification rate of that was 1% higher using all features except
for variance from the EMG data and the first 3 IMFs.
I. INTRODUCTION
Grasping, like many other areas of robotics, is still an activity
where human beings exceed their artificial counterparts in
both robustness and consistency. Humans rarely fail when
attempting to pick something up, whereas robots still struggle
with multiple aspects of grasping, such as object recognition,
finding the correct grasp for an object, and providing the
correct amount of pressure to hold the object securely while
not damaging it. It can clearly be seen that the human brain is
still a better planner for grasping than the current robotic
methods.
Electromyography (EMG) sensors are used for controlling
prosthetics, and will be the focus of this project. Machine
learning will be used to analyze time series EMG data. As
stated above, robots still struggle with grasping, making it
difficult to use robotic manipulators and prosthetics
efficiently. Therefore the goal of this paper is to use EMG
sensor data and machine learning to recognize a grasp based
on the input EMG data, with the goal of performing more
accurate brain controlled grasping with prosthetics.
The idea of this paper is training a neural network system,
which draws its inspiration from the human brain, to take
The paper is final paper of final project of ROB 537. It also contains six
aspects --- introductions, background, related work, method, results and
conclusions.
In this paper, the group concentrate on the data of Artificial Neural
Networks in EMG to find the relation between the input data and Intrinsic
Mode Function. Then vary the features of input.
S. K. Allani, Master student in Oregon State University. He is now with
the Department of Robotics (e-mail: allanis@oregonstate.edu).
Alexander Zatopa, Master student with Robotic Department in Oregon
state University.(e-mail:zatopaa@oregonstate.edu).
Huanchi Cao, Master student in Robotic Department in Oregon State
University (e-mail : caohu@oregonstate.edu)
EMG data as an input and recognize the intended grasp. To do
this, the input data must be recorded from EMG sensors on the
arm muscles of a person performing grasping tasks, and the
outputs will be the different intended grasps. The main
difficulty presented is how to find the features from the input
electric signal and translate these features to make the neural
networks, the “brain” of the robotic hand, understand what
grasp output is intended by the EMG data input. First,
amplifying and filtering of the EMG signal is needed to
improve the ability to recognize the important features and
ignore the noise.
In addition to processing the EMG signal, a mathematical
model is needed using signal processing and machine learning
to match the features of EMG signal to the correct output. Part
of the difficulty is in manipulating the time varying signal of
the EMG data into a form that can be used for machine
learning. The other difficulty is finding a machine learning
algorithm that will be able to train the data well, and correctly
map EMG inputs to the appropriate grasp output. With the
ultimate goal of real time control, the calculation and the data
should be simple to process and reduced to a minimum
necessary. However, this paper is focused on predicting
grasps, which is an important step to judge whether the model
is correct and may eventually be used for real time control.
Correct modelling is the key to achieve the required results
and explore the underlying process behind the results. There is
little literature dedicated to predicting particular grasps from
EMG data with two sensors using neural networks, a relatively
simple machine learning algorithm, and it would be beneficial
to be able to recognize a specific grasp with this algorithm and
so few EMG sensors.
As discussed above, the most essential contribution of this
paper is understanding how neural network works and
exploring different architectures of neural networks and how
we can use it to predict time varying signals. Another
contribution is EMG signal classification. Machine learning is
a fundamental element of the whole system to classify the
input electrical signals. The hope is to find a high quality
classification algorithm that is able to understand the EMG
data features correctly. Many current arm prosthetics are
controlled purely by flexing muscles in the arm, and will only
open or close. Many other arm prosthetics are not controllable
at all. Controlling prosthetics by leaning to understand the
brains intent would give amputees much greater dexterity and
allow them to perform more everyday tasks that were
previously time consuming and difficult for them, allowing
them to live more normal and fulfilling lives. To sum up, our
Time Series Analysis of EMG Data Using
Artificial Neural Networks
Sai. K. Allani, Alexander Zatopa , and Huanchi Cao
2. ROB 537 Final Project 2
main contribution is finding the correct model to enhance the
understanding and execution level of EMG signals in
prosthetics controlling areas.
In the above paragraphs, we have looked at the main problem,
the difficulties, possible solutions and contribution to the field.
Using EMG data is based on an signal processing techniques.
In previous work, we find that using Empirical Mode
Decomposition (EMD) can enhance the identification
accuracy or a pattern recognition[4]. In this paper, we further
exploit the use of EMD and extracting the features of EMG
data using Artificial Neural Networks (ANN). Therefore, in
the next paragraphs, more specific information about the
background and some key points in choosing correct data
processing and machine learning algorithm will be discussed
in Section II and III. Then Section IV consists the actual
manner in which the data was processed for this paper, and
finally the machine learning algorithm used. In Section V and
VI, it may analysis the results draw the detailed conclusions.
II. SPECIFICS AND BACKGROUND ABOUT THE
PROBLEM
In this sections we will talk in detail about the things needed
to know to understand the problem being solved. This can be
divided into different sections: understanding how EMG
works, what signal processing is and feature extraction means.
And finally, what is machine learning?
A. What is EMG?
An electromyogram (EMG) measures the electrical activity
of muscles at rest and during contraction. It is the potential
difference between muscles during contraction and expansion
of muscles. When humans do different types of tasks with
their hands, these signals are passed to the hand via neurons
from the brain. Nerve conduction studies measure how well
and how fast the nerves can send electrical signals.
EMG is mainly used in clinical purposes in analyzing the
condition of muscles of the patients. We can measure these
electrical signals and based upon the features of the signals,
like amplitude, we can understand what’s happening in the
muscles. Due to the development in the electronic field EMG
has gained a lot of interest in the recent past. Many
researchers and companies used EMG to control prosthetics
and wearables. Since EMG gives out very small electrical
signal, i.e. micro-millivolts, and since it is also susceptible
very to noise it is quite hard to analyze an EMG signal. This
is where signal processing and feature extraction comes into
play. As data is being used which has already processed,
focus is not being placed on this part very much, even though
this is one of the essential step in EMG analysis.
Fig.1 Raw EMG signal
Most of the time the signals we collect are in raw format
which is basically useless and they cannot be understood.
Signal processing is a technique where mathematical models
are used to extract information from the signals so they can be
analyzed. There are many types of signals and depending upon
the type of signal being dealt with different mathematical
models are used. In this case it is time-series data, which
means the signal changes with respect to time. There are
different types of signal processing techniques such as time
frequency analysis or Hilbert transform to process time
varying data. This can also be helpful in feature extraction.
Feature extraction is one of the main parts of this project.
In this paper, the most efficient technique in transforming
useful information from signals is Empirical Mode
Decomposition (EMD). EMD is a method of breaking down a
signal without leaving the time domain. EMD process is
useful for analyzing natural signals which are most often
nonlinear and non-stationary [14]. The EMD’s functions,
known as Intrinsic Mode Functions (IMF) are sufficient to
describe the signal even they are not necessarily orthogonal.
Therefore, the main process of experiments in this paper is
collecting signals, then through EMD to extract different
features from EMG signal and its functions IMFs. Then, the
output received in this stage is used as input to the learning
algorithm being used. So, it is very important to extract the
features from which the algorithm can learn and be trained
easily.
The next step is choosing the right algorithm which can
predict the output based on the inputs given. There are many
algorithms which can do this, but in this project Artificial
Neural Networks are being focused on. ANN is a part of
machine learning which is mimicked from the human brain. It
has a set of inputs, outputs and many adaptable weights
which are helpful in mapping inputs to outputs. ANN has
different kinds of architects in it, i.e Back-prop, Recursive
ANN, and more [4]. Different kinds of architectures are used
to solve different problems. The kinds of feature extraction
methods and architecture being used in this project in coming
sections will be discussed in great detail.
3. ROB 537 Final Project 3
Fig.2 Simple representation of ANN
In this project, data from UCI machine learning repository is
going to be used. This data is collected from the human hand
using surface EMG sensors. There are total of 3 EMG sensors
of which one is used as datum and another two are used to
measure the electrical activity in the muscles. The data is
collected from six subjects (three male and three female) when
they are performing six different daily grasps, i.e pinch,
cylindrical, spherical, tip, palmar, lateral, and hook. Each
grasp is done 30 times for a duration of six seconds.
MATLAB and Python are then going to be used for the feature
extraction and coding the algorithm. The following section
will describe the work already done in this field and how it
can be used to understand and develop ANN’s better.
III. GENERAL APPROACHES AND RELATED WORK
The problem of muscle actuation prediction for control using
EMG signals is not a new one, and has been approached in
many different ways. The problem can also be split into
different sections, each of which can be solved independent of
how the other sections are solved. The first section is data
acquisition. As stated above, the goal here is to record a signal
that most directly represents the myoelectric signal, using
equipment and filters to remove as much of the noise as
possible. Noise from various sources, such as 50/60 Hz
electromagnetic induction from power lines, can greatly affect
the signal, mainly due to the small amplitude of the EMG
signal (micro-millivolts) [4]. Additionally, the signal has to
travel through the tissue and skin, adding opportunity for noise
and cross contamination between signals to be introduced. To
attempt to mitigate this, the skin should be thoroughly cleaned
before the sensors are applied, and the sensors should be
placed in precise orientation to each other to maintain
consistency.
Fig.3 EMG sensors and example of where they may be
placed on the forearm
The signal can then be amplified and filtered, using a bandpass
filter to remove low and high frequency noise, and the data
collected.
The second challenge is processing the data, and converting it
into a form that can be used for machine learning. This
process usually entails additional filtering or utilization of
transforms, and the choice of how to process the data will be
dictated by the machine learning algorithm used in the next
step, due to the fact that the output of this step will be used as
the input to the algorithm. There are many methods to process
the data, two in particular will be described. First, the data can
be processed as a time series, or second, specific features can
be collected from the data.
One way to perform time series analysis is using the
autoregressive model [4]. This model predicts that the next
value in a time series ŷ[n] is linearly dependent on its previous
outputs (y[n-1], y[n-2],...) and some unpredictable term. The
equation for this model is below, where ŷ(n) is the estimated
signal in a discrete time n, am are the AR-coefficients, e(n) is
the estimation error, and Mis the order of the model.
ŷ(n) = ∑ 𝑎𝑎 𝑚𝑚 (𝑛𝑛)𝑦𝑦(𝑛𝑛 − 𝑚𝑚) + 𝑒𝑒(𝑛𝑛)𝑀𝑀
𝑚𝑚=1 (1)
Fitting the signal to this model requires solving for the weights
defined in the autoregressive model using an optimization
algorithm such as least mean square. These weights defining
the equation for each signal using the autoregressive model are
then used as the inputs for the learning algorithm.
Specific features from the time series may also be found, and
those may be used as the inputs to the learning algorithm. One
method to find the features was performed by Sapsanis,
Georgoulas, and Tzes, using Empirical Mode Decomposition
(EMD) and a set of features calculated using various metrics
from the decomposed signal. EMD decomposes a signal into a
collection of intrinsic mode functions (IMF), which allow for
the calculation of certain features, such as Integrated
Electromyogram (IEMG), zero-crossing, Slope Sign Changes,
waveform length, Willison amplitude, variance, skewness and
kurtosis [3]. These features are then used as the inputs to the
learning algorithm.
The final part of the process, and the topic of greatest interest,
is the learning algorithm. Learning algorithms receive data,
learn from it, and attempt to make predictions based on that
data. Machine learning can be classified into two large camps,
supervised learning and unsupervised learning, with
combinations and offshoots as well. Supervised learning uses
a set of inputs and outputs, processed through an algorithm, to
try to predict new outputs based on new inputs of the same
process. Unsupervised learning (clustering, dimensionality
reduction, recommender systems, self organizing learning) has
no initial set of outputs mapped to learn from, and therefore
has no feedback from the environment [2]. Between
supervised and unsupervised learning are other machine
learning methods, such as semi-supervised learning, which
creates both labeled and unlabeled examples, and
reinforcement learning, where the learning algorithm receives
4. ROB 537 Final Project 4
feedback from its environment to improve the accuracy of its
response.
There are many types of machine learning algorithms that
have been used for training EMG signals for prosthetic
control. One method is fuzzy logic. Fuzzy logic is similar to
boolean algebra or probabilistic models. It is based on the idea
that human thought processes are often imprecise and
uncertain, and they aim to mimic that process. Whereas a
traditional set would contain members that were of that set or
not at all, fuzzy systems allow members to have a “degree of
membership” to the set [2]. This may allow for greater
flexibility and accuracy when modeling real world systems.
Various other probabilistic model algorithms have also been
used for machine learning. These models include Bayesian
networks, Linear Discriminant Analysis, Gaussian Mixture
Model, and Support Vector Machines [2].
Another machine learning method to extract patterns and
predict trends is Artificial Neural Networks (ANN). ANN is
an information processing system that draws its inspiration
from the human brain. It uses a series of interconnected
parallel processing units, called neurons,that connect the
inputs to the outputs, and use interneuron connection weights
to store knowledge and predict the correct output based on the
inputs. The network is trained by introducing a data set to the
network with known inputs and outputs, and training it, by
adjusting the weights between the neurons, to map the given
inputs to the correct outputs. Advantages of ANN include
adaptiveness, real time operation, and fault tolerance [2].
Additionally, hybrid algorithms of the aforementioned
methods, and other methods, may be used. These hybrid
methods include Adaptive Neuro-Fuzzy Inference System, and
Fuzzy Clustering Neural Network [2].
In previous literature, there have been various algorithms used
to attempt to apply machine learning to EMG prosthetic
control. Soares, Dandrade, Lamounier and Carrijo used an
autoregressive model and a backpropagation neural network
algorithm to attempt to classify EMG signals in real time,
however, they did not perform this classification for various
specific grasps. Neural networks has been a popular method
for control using EMG data. On the other hand, Sapsanis,
Georgoulas, and Tzes used feature extraction methods using
various transforms to find unique features for various graphs
to classify them. Other modifications of various learning
algorithms have also been used, such as a modification to
fuzzy ARTMAP networks[5] and support vector machines [1].
IV. METHODS
As seen in the previous section, two common methods for
classification of EMG data is autoregressive model with a
neural network [2], or feature extraction and decomposition to
allow a linear classifier to be used [3]. The following method
uses feature extraction, but instead of decomposing to allow
the use of linear classification, the features are used to train an
ANN for use in classification.
A. Description of set experiments:
In this paper, data from EMG sensors are used as the input,
features are extracted from the data by decomposing into
Intrinsic Mode Functions (IMF) and extracting features in the
time and frequency domain, and then this data will be
classified in an Artificial Neural Network (ANN). In order to
improve the learning ability of the ANN, one of the variables
is the number of hidden layers which can guarantee to process
the data. Each person has his unique muscles activities, and
the input data is different. In this paper, different inputs are set
as an important variable to extract the output data which is a
main standard which evaluating the performance of algorithm.
The following part is a brief description of signal processing
and feature extraction and then analysis using ANN.
Fig.4 A Simple diagram showing the key elements of the
project.
B. Signal Processing and Feature Extraction:
In the background section this paper described previously
what signal processing means. In simple words, signal
processing is the technique where useful information is
extracted from the signal. EMG signals are non-stationary
signals and extracting useful information from it is very
complicated. There are a few techniques like fourier
transforms which help in getting useful information from the
signals. But these techniques are not very efficient when
compared to EMD. If the signals are considered a combination
of fast oscillating and slow oscillating signals EMD attempts
to differentiate the signal into fast and slow oscillating
5. ROB 537 Final Project 5
components called Intrinsic Mode Functions (IMF). This IMF
represents a simple oscillatory function satisfying two
conditions:
1. The number of zero crossings and the number of
local extrema are either equal or differ by one.
2. The local average (defined by average of local
maximum and local minimum envelops) is equal to
zero
If we consider a signal x(t), the EMD algorithms can be
summarized as follows:
1. Find all local minima and local maxima of given
signal x(t). Create an upper emax(t) and a lower emin(t)
envelope interpolating between successive local
maxima and local minima respectively.
2. Calculate the running mean:
m(t)=
1
2
[emin(t)+e max (t)] (2)
3. Subtract the mean from the signal to extract the detail
d(t) = x(t) - m(t) (3)
4.
5. Repeat the whole process replacing x(t) with m(t)
until the final residual is a monotonic function
So, the original signal is eventually decomposed into a sum of
IMFs plus a residual term:
x(t)=∑IMF(t) +r(t) (4)
Fig.5 Raw EMG signal and IMFs of lateral grip.
Fig. 6 Residual term of EMG signal of lateral grip
Once we are done processing the raw EMG signal we will
extract different features from the raw EMG signal and IMFs.
The feature extraction stage is always executed when different
kinds of bio-signals are used. It gives out the relevant
information and also alleviates the problem with high
dimensionality. Features should be select in such a way that
they give out a good classification rate. Most of the
researchers used the following features to classify bio-
signals[3].
These features are also being used in this paper.
1. Integrated Electromyogram(IEMG)
This feature is an average value of the absolute values of
the EMG. It is defined as follows:
IEMG=
1
𝑁𝑁
∑xk (5)
Where xk is the kth sample data out of N samples of EMG
raw data
2. Slope Sign Change (SSC):
SSC counts the number of times the slope of the signal
changes sign. Given three contiguous EMG signals xk-1, xk and
xk+1 the number of slope sign changes can be calculated by
SSC = ∑f(x) where:
f(x) = 1 if xk < xk+1 and xk < xk-1
or xk > xk+1 and xk > xk-1
f(x) = 0 otherwise
3. Variance(VAR):
VAR is a measure of the power density of the EMG signal
and it is given by:
VAR =
1
1−𝑁𝑁
∑(xk -µ)2
(6)
4. Zero Crossing(ZC):
ZC counts the number of times that the signal crosses zero.
A threshold needs to be introduced to reduce the noise induced
at zero crossing. Given two contiguous EMG signals xkand
xk+1, then ZC can be calculated as: ZC = ∑f(x) where
f(x) = 1 if xk >0 and xk+1<0
or xk <0 and xk+1 >0
f(x) = 0 otherwise
5. Wavefrom Length(WL):
WL is a cumulative variation of the EMG that can indicate the
degree of variation about the EMG signal. It is given by
WL = (|xk+1-xk|) (7)
6. Kurtosis:
6. ROB 537 Final Project 6
The kurtosis of a distribution is defined as:
𝑘𝑘 =
𝐸𝐸(𝑥𝑥−µ)4
σ4 (8)
7. Skewness:
The skewness of a distribution is defined as
𝑠𝑠 =
𝐸𝐸(𝑥𝑥−µ)3
𝜎𝜎3 (9)
Where is the mean of x, is the standard deviation of x, and E(t)
represents the expected value of the quantity t.
C. Processing of Extracted Features Using ANN:
As stated in the above section, 7 features were extracted from
each EMG signal, and 7 features were also extracted from 6
IMFs of each signal. Since there were 2 channels for each
grasp, this leads to 98 total features that may be used to
attempt to classify a grasp. These inputs were then be broken
up into subsets in two different ways. First, by number of
IMFs. First, the features from only the EMG signal were used
as input, then features from 1 additional IMF were added at a
time until all features from EMG and all IMFs were being
used as inputs, and the quality of classification was observed
and compared. Then the features were varied. Subsets of
features were used for input to the ANN, and the resulting
classification quality was observed and compared.
The ANN used was a feedforward ANN with one hidden
layer. The number of inputs were varied to match the number
of features being used for classification, and there were six
outputs, one for each grasp. The number of hidden units used
was based on the number of inputs, and was equal to the
average of the number of inputs and outputs. There are many
methods, or rules-of-thumb, on choosing the number of hidden
units. The rule chosen here was based on trial and error, and a
suggested starting point of between the size of the input layer
and output layer by Jeff Heaton in Introduction to Neural
Networks for Java, 2nd Edition [12]. The training method used
was scaled conjugate gradient (SCG) backpropagation, using
performance evaluation based on cross entropy. SCG is part of
the class of Conjugate Gradient Methods, and, on most
problems, shows super-linear convergence[7]. Unlike gradient
descent, which attempts to minimize a global error function
simply using the direction of the gradient, SCG denotes a
quadratic approximation to the error E in the neighborhood of
a point w by:
E'qw = E(w) +E'(w)T
y +
1
2
yT
E"(w)y (10)
To determine the minimum of this equation the critical points
must be found [8]. They can be found by solving the below
linear system [7].
E'qw(y) = E"(w)y + E'(w) = 0 (11)
SCG was used due to its effectiveness and speed relative to
other typically methods, such as standard backpropagation [7].
Cross entropy is an alternative approach to mean square error
(MSE) [13]. In the MSE function:
Em =
1
𝑚𝑚
∑ (𝑡𝑡𝑘𝑘 − 𝑦𝑦𝑘𝑘)𝑚𝑚
𝑘𝑘=1
2
(12)
Where Em is the total error, t is the target, and y is the output
of the ANN. This is then minimized to improve the ANNs
performance. Alternatively, the following cross entropy error
function can be minimized:
Em =
1
𝑚𝑚
∑ (𝑡𝑡𝑘𝑘 𝑙𝑙𝑙𝑙 𝑦𝑦𝑘𝑘 + (1 − 𝑡𝑡𝑘𝑘)𝑙𝑙𝑙𝑙(1 − 𝑦𝑦𝑘𝑘))𝑚𝑚
𝑘𝑘=1 (13)
When using cross entropy error function, the partial derivative
of Em with respect to the weight wjk is found:
∂Em
∂w
=σ(yk − tk) ∗ zj (14)
Therefore, when minimizing the error signal, the entropy
function has a better network performance with a shorter
stagnation period[13]. Since cross entropy is proven to
accelerate the backpropagation algorithm and to provide good
overall network performance, it was chosen for this paper.
The ANN classification network was created using the built in
patternnet function in Mathworks’ Matlab, with the training
function set to scaled conjugate gradient backpropagation,
and the performance function set to cross entropy. The input
data to the network was split into three categories. 70% of the
data was used to train the network, by computing the gradient
and updating the network weights and biases. 15% of the data
was used to validate the training, by applying the weights
found by the training set and the training function and
measuring the performance. Finally, 15% was used as testing
data, which is solely used for monitoring, and comparing
methods if desired.
Matlab’s Neural Network Training method for classification
and the above training and performance inputs have the
following possible stopping criteria. First, Epoch. The
maximum number of iterations was set to 1000.Second,
performance, which is based on cross entropy, and will stop
the algorithm when the system reaches a certain entropy. This
condition was set to 0. Third,the gradient. This was set to 10-6
.
Finally, validation checks, which measure how many times in
a row the validation set of data is not improving in
performance. If the validation set is not improving for more
iterations than this number, the algorithm stops. This was set
to 10. These stopping conditions were picked after trial and
error. It was found that the first stopping condition met was
typically the validation checks, which is beneficial because the
validation data typically begins to show worse performance
when the training is starting to fit the noise in the system, and
this is a good place to stop the training. A validation check of
10 consistently stopped the training when it began to fit the
noise in the system.
D. Varying the Number of IMFs Input into the ANN:
For the first set of tests, the number of IMFs used as input
was varied. The full input data set included 14 features from
7. ROB 537 Final Project 7
the EMG data, 7 for each channel, and 14 for each of the six
IMFs as well. There were 900 sets of these inputs, based on
the number of grasps performed by the 5 test subjects. For
each set of inputs, there was a target grasp vector, which
consists of a 1 for the target grasp, and 0s for all other grasps.
To analyze how the number of IMFs used affected the
classification, first the ANN was trained with just the EMG
features as the input. Then an additional IMF was added and
the network was retrained, until the features from all IMFs
were being used as input to the neural network. This was
repeated for 100 training trials, and the mean of the
classification failure rate for the test data and the full set of
data was recorded for all sets of inputs.
E. Varying the Number Input into the ANN:
There are 7 different features extracted from the EMG data
and each IMF. Those features can be seen in the following
table.
Table 1: Features extracted from data.
In the first set of classification described above, all features
were used to train the ANN. In this part, the number of IMFs
remained constant, and only subsets of features were used to
train the network. First, one feature was removed at a time
from the input data set, and the network was trained 100 times
with the remaining 6 features from the EMG data and the 6
IMFs. Second, the network was trained 100 times with only
one feature at a time from the EMG data and the 6 IMFs, and
finally, the best results from the previous experiments were
analyzed, and a possible optimum subset of input data was
selected to perform training of the network.
V. RESULTS
A. Results for Varying Number of IMFs:
There are two sets of results that are of interest. First, the
performance of the ANN when the number of IMFs is varied.
Below is a plot showing the average percent of inputs that
were incorrectly classified in the test data over 100 iterations
of the full data set for different quantities of IMFs used with
the EMG data.
EMG feature sets and # of IMFs
Fig 7: Average misclassification percentage of test data for
different numbers of IMFs used.
It can be seen that this classification error remained relatively
constant between 6% and 8% with the test data set, although
there is a slight minimum at just the EMG feature set and a
slight maximum at 3 IMFs and the EMG, combining for 4
feature sets.
The average percent of inputs that were incorrectly classified
in the entire set of the data for 100 iterations quantities of
IMFs used with the EMG data was also recorded and plotted.
EMG feature sets and # of IMFs
Fig 8: Average misclassification percentage of all data for
different numbers of IMF used.
It can be seen that this misclassification percentage is not as
consistent. When the EMG features are used with between 3
No. Feature
1 Integrated Absolute Value
2 Slope Sign Change
3 zero crossing
4 skewness
5 Kurtosis
6 Wave length
7 variance
8. ROB 537 Final Project 8
and 6 IMFs, the misclassification percentage remains between
4% and 6%, with a minimum at EMG features plus features
from 4 IMFs.
However, when using just the EMG features with less than 3
IMFs, the number of misclassified inputs begins to rise,
maxing out at approximately 15% when only features from the
EMG data is used for classification.
Below is a confusion matrix for a typical training output with
the EMG features and features from 4 IMFs. These show the
target classification for each input, vs. the actual output from
the ANN. This is shown for the training data set, the validation
data set, the test data set, and the full data set. The values in
the previous two figures correspond to the red numbers in the
blue boxes of the “test” matrix and the “all” matrix. These
matrices can be useful when observing how the inputs were
misclassified, and possibly observe any trends in
misclassification.
Fig 9: Confusion Matrices for EMG features plus 3 IMFs.
The second set of interesting results is varying the features
used to train the network. First, as stated above, the training
was run 100 times using inputs where 1 feature was removed
at a time. Below is a plot showing the average percent of
inputs that were incorrectly classified in the test set when a
feature was removed from the input set.
Feature that was removed (8 is using all features)
Figure.10: Percent of test set that was misclassified when
removing one feature at a time
It can be seen that the percent misclassified from the test set
was relatively constant, staying around 7%. The average
percent of inputs that were incorrectly classified in the entire
set of the data when one feature was removed was also
plotted, and can be seen below.
Feature that was removed (8 is using all features)
Figure.11: Percent of full set that was misclassified when
removing one feature at a time
It can be seen that with the misclassification from the entire
data set, the error was largest, around 7%, when features 1,
Integrated Absolute Value, and 3, zero crossing, were
removed. The error was smallest, around 5%, when features 5
through 7 were removed, as well as with the full data set.
Next, the network was trained with just one of the features at a
time, for 100 training runs. The plot below shows the percent
misclassified from the test set when only one feature was used
at a time.
9. ROB 537 Final Project 9
Feature that was removed (8 is using all features)
Figure.12: Percent of test set that was misclassified using
one feature at a time.
It can be seen that the error varied substantially, but remained
fairly low, below 8%. The percent misclassified for the entire
set was also plotted for the same test, and can be seen below.
Feature that was removed (8 is using all features)
Figure.13 Percent off full set that was misclassified using
one feature at a time.
It can be seen that the full set produced the lowest
misclassification rate, while features 1 and 3 produced the
lowest rate out of the single features at around 20%. Features
6, wave length, and 7, variance, produced the highest
misclassification rate, around 75%.
Finally, the observation was made above in figure 8 that using
only 3 IMFs produced approximately the same
misclassification rate of the full data set as when all the IMFs
are used. Therefore, this subset of 3 IMFs and the EMG
features was used as the new input, and one set of features at a
time was removed from this set of inputs. This setup was then
trained 100 times for each removal of a feature. Below is the
misclassification rate of the full set of data for this setup.
Feature that was removed (8 is using all features)
Figure.14 Percent of full set that was misclassified when
removing one feature at a time from input subset of EMG
features and features from 3 IMFs.
It can be seen that the misclassification rate when feature 7
was removed was the lowest misclassification percentage, at
4.1%.
VI. ANALYSIS AND CONCLUSIONS
A. Analysis:
First, when training the full input data set 100 times, a
misclassification rate of the full set had a mean of 5.1% as
seen in figure 8. This means there was a successful rate of
classification 94.9% of the time, which is a high rate of
success, and is comparable, if not better, than the classification
results in Sapsanis et. All, where similar features were used to
classify grasps with two dimensionality reduction techniques
as opposed to artificial neural networks. Furthermore, in figure
8, it can be seen that using less than 6 IMFs with the EMG
data did not increase the misclassification rate until only two
IMFs were being used. When two or less IMFs were used, the
misclassification began to increase rapidly.
It can also be seen in figures 11 and 13 that using different
features have different effects on the classification rate of the
full set of data. In figure 11, it can be seen that when the 1st
and 3rd features are removed, the misclassification rate
increases the most. We may infer from this that these features
have the greatest effect on classification. This observation is
reinforced when figure 13 is observed. From this figure it can
be seen that when only features 1 and 3 are used by
themselves respectively for training, the network trains better
than with any other feature individually.
After the observation that the EMG features and 3 IMFs were
sufficient for classification as compared to using all the IMFs
calculated, this subset of IMFs was used, and one feature at a