Paraphrasing

There are many technologies being introduced in today’s world of artificial intelligence yet some of the things still remain undiscovered. The purpose of this paper is to classify the surface EMG Signals using stationary wavelet decomposition and another technology known as ensemble classifiers. The ensemble classifier is very easy and working if the individual classifiers are in agreement. The error rates may vary. The research will focus on the use of ensemble classifiers for Electromyographic technology. The research question and problem statement are mentioned in the introduction chapter and the research methods and sample method will be explained in the later chapters.

Keywords: EMG, Ensemble classifiers, Machine learning.

I

There are many people who helped to make my years at the graduate school most valuable. First, We thank Prof. Dr. Abdulhamit SUBASI, for her guidance and help. He was not only served as our supervisor but also encouraged, challenged and shared us his valuable time throughout this thesis . And we thank all the Information System Department staff for their efforts through the years to reach graduation fulfilling.

VIII

APPROVAL PAGE I ABSTRACT II ACKNOWLEDGEMENTS III DECLARATION IV TABLE OF CONTENTS V LIST OF TABLES VIII LIST OF FIGURES IX CHAPTER ONE : INTRODUCTION TO THE STUDY I INTRODUCTION 2 STATEMENT OF THE PROBLEM 2 AIMS AND OBJECTIVES 3 AIM 3 OBJECTIVES 3 RESEARCH QUESTION/HYPOTHESIS 3 SCOPE 3 MOTIVATION 3 METHODOLOGY 4 EXPECTED OUTCOMES/DELIVERABLES 4 RESEARCH CONTRIBUTION TO KNOWLEDGE IN IS FIELD 4 MILESTONE 4 SUMMARY 6 CHAPTER TWO : BACKGROUND AND LITERATURE REVIEW 7 CHAPTER THREE : REQUIREMENTS SPECIFICATION AND SYSTEM ANALYSIS 8 INTRODUCTION 17 FUNCTIONAL REQUIREMENTS 17 RESOURCE REQUIREMENTS 17 BUDGET REQUIREMENTS 18 DATA FLOW DIAGRAM 19 DATA FLOW DESCRIPTION 19 SUMMARY 19 CHAPTER FOUR : SYSTEM DESIGN AND DEVELOPMENT 17 INTRODUCTION 21 COLLECTION OF DATA 21 FEATURE EXTRACTION 24 Wavelet Transform 24 Multi-scale Principal Component Analysis (MSPCA) 25 Tunable Q-Factor Wavelet Transform (TQWT) 26 Dual Tree Complex Wavelet Transform 27 ALGORITHMS FOR SEMG CLASSIFICATION 28 Artificial Neural Networks (ANN) 28 K-Nearest Neighbour (k-NN) 29 Support Vector Machine (SVM) 29 Naïve Bayes 30 CART (classification and regression tree) 30 REPTree 30 NBTree 30 LADTree 30 C4.5 (J48) Decision Tree 31 RandomTree Classifiers 31 Random Forests (RF) 31 Rotation Forest (RoF) 31 BAGGING 32 BOOSTING 32 AdaBoost 32 MultiBoosting: A Technique for Combining Boosting and Wagging 33 The Random Subspace Method 33 DATA PROCESSING 33 CLASSIFICATION IMPLEMENTING 34 SUMMARY 34 CHAPTER FIVE : RESULTS AND EVALUATION 21 INTRODUCTION 35 PERFORMANCE EVALUATION 35 EXPERIMENTAL RESULTS 35 SUMMARY 59 CHAPTER SIX : DISCUSSION 35 INTRODUCTION 53 ACHIEVEMENTS 53 COMPARATIVE STUDY 53 CHALLENGES AND DIFFICULTIES 57 SUMMARY 57

Classification of Surface EMG Signals Using Ensemble Classifiers Dept of IS , EFFAT UNIVERSITY

Biomedical signals are observations of physiological activities of organisms using electrodes attached to skin or muscles tissues

Usually, the electromyographic signals are composed of action potentials from the group of muscles fibers into motor units. The signal can be detected with sensors placed on the surface of the skin or with needle or wire sensors introduced into the muscle tissue. When two of the muscle motor units in the vicinity of the sensors are working, it can be possible to identify and analyze individual motor units action potential.

Signal processing aims at extracting significant information and record it in organized datasets to handle it later in many forms.

And with the advancement of medical measurements and process in what serve patients and humans, we think about acting in that service by linking the information system aspects with the medicine by using data mining techniques to organize the previous mentioned data sets to obtain useful patterns or information.

The study will be basically inclined to the way of classifying of EMG signals by using the latest technical elements of stationary wavelet decomposition and ensemble classifiers.

The implementing of this idea will be by using data mining tool such as WEKA: Waikato Environment for Knowledge Analysis which is a suite of machine learning software written in Java programing language and supports several standard data mining tasks

With the increasing advancements in the medicine and other human life sectors besides technology advancements every day, the need of getting usage from information systems also increased.

Classification of Surface EMG Signals Using Ensemble Classifiers Dept of IS , EFFAT UNIVERSITY

4

So we saw that applying classification methods to classify surface EMG signals using velvets decomposition and ensemble classifiers to increase EMG signal quality to where the signal becomes much more accurate, simple, reliable and steady.

And getting the improvements that we can propose to enhance the quality of the results.

Improve the classification of accuracy of surface EMG signals using different feature extraction and classification techniques for prosthesis control.

. The aim is broken into these objectives which are

· To identify the working of EMG signals

· To identify and implement decomposition methods and ensemble classifiers

· To understand how both technologies can play a great role in classifying EMG signals

· How can stationary wavelet decomposition and ensemble classifiers help in EMG classification?

The assumptions will be made later by choosing the best methodologies in order to continue the research in the field.

The scope of this thesis is to classify the surface EMG Signals using ensemble classifiers applied to datasets taken from amputees using data mining tool such as WEKA.

The main motivation of this work is handling the collected data sets using data mining techniques and tools to get highest accuracy levels. Besides linking Information system technologies with medicine and sectors related to humans and health. And although there are many

previous researches on the same topics. But many gaps were found and we aim to avoid these gaps in our thesis .

To achieve our aim we decide to use an appropriate methodology for the implementation especially designed for Data Mining analysis. For this we followed an approach based on the gathered data to analyze it and reach the required output.

At the end of the thesis we intend to deliver an output for the processed data including classified instances, classifier accuracy and interpretation for that output so it can be take in consideration and apply in many useful fields in the future.

Besides, these deliverables are expected:

· To obtain the best data mining classifier that gives the highest accuracy.

· To publish a conference paper.

Clients who will avail from our research are:

· Medical analysts

· Security Investigations Team.

Since we are IS senior students, It is very important to follow organized approach in our thesis where we link the medicine aspects with the information system by applying data mining techniques and ensemble classifiers on Surface EMG signals data sets to help in classifying the huge data related to amputees in efficient way and getting accurate result in relatively short time and less efforts.

The scheduled work is a key of the successful projects, so we will organize the project phases according to a specific time using project plan as follows :

Proposed steps schedule is planned like in the table.

Classification of Surface EMG Signals Using Ensemble Classifiers Dept of IS , EFFAT UNIVERSITY

5

We will implement Surface EMG signal classification using ensemble classifiers to help classifying data in huge data sets using data mining techniques. This classification methods will save time and effort, give accurate results and above of all link Information system technologies with medicine and sectors related to humans and health.

Classification of Surface EMG Signals Using Ensemble Classifiers Dept of IS , EFFAT UNIVERSITY

6

Classification of Surface EMG Signals Using Ensemble Classifiers Dept of IS , EFFAT UNIVERSITY

Gaudet, et al. (Gaudet, 2018 ) conducted a research entitled “Classification of upper limb phantom movements in transhumeral amputees using electromyographic and kinematic features” which was quite similar to the current research in which Gaudet analyzed that the ability to generate surface electromyography can improve the myoelectric prostheses control with multiple DoFs i.e. degrees of freedom. The study evaluated the accuracy level of classification of main eight upper limb movements of phantom and zero class of movement in transhumeral amputees dependent on surface electromyography data which was recorded exclusively on the residual limb. Gaudet further evaluated the impact and effect on the accuracy of classification by kinematic data. In the study, five transhumeral amputees took part and the accuracy of classification achieved with an artificial neural network had a range between 60.9 % and 93 %. The accuracy level might have minimized if the amount of DoF in the classification so far had been increased or if the movements of phantom had become distal. An increase of 4% in accuracy was due to the kinematic feature added in it. Finally, the study led to the production of a novel myoelectric control procedure for multi-DoF prostheses based on phantom movements of the both amputee and kinematic data of the prostheses.

Another study conducted in this regard was conducted by (A. W. Waris,2018 ) and title was “The effect of time on EMG classification of hand motions in able-bodied and trans-radial amputees” in which the importance of long term investigation in terms of EMG classification was highlighted. There are very less long-term investigations. The study investigated the changes in the performance of classification with the passing of time. Six of the amputees and ten individuals played their part in the study. From intramuscular electrodes and EMG signals were concurrently recorded and these were kept into the muscles for consecutive seven days. Through the use of linear discriminant analysis, the seven motions of hand were

Classification of Surface EMG Signals Using Ensemble Classifiers Dept of IS , EFFAT UNIVERSITY

34

analyzed. The analysis was also done by the classification error numbered between WCE and Days of BCE. For all the possible combinations between the days, BCE was computed. There was a regression between WCE and seven days were very insignificant because of the fact that performance was considered similar each day. The only significance was the regression between BCE and the difference of time in days. The slope was entirely different, and the result of the research was that performance continuously degraded and became less as the difference of time between testing and training day increased. Moreover, the performance in amputees was directly proportional to the size of the residual limb.

( wang-T. Shi, 2018) also did a research on the same subject entitled “A bionic hand controlled by hand gesture recognition based on surface EMG signals: A preliminary study” which basically bridged the gap between the features of Semg Signals of fingers and the postures of a bionic hand. A bionic hand with fine motor ability can be a favorable option for replacement of a human hand while performing multiple operations. Past studies haven’t been able to focus much on the bionic hand so mostly people are unaware of its use. Aim of the study is construction of a prototype system for recognition of hand postures. Controlling of bionic hand via analysis of sEMG signals which were measured at the flexor digitorium superficialis. The results showed through experiments that bionic hand could easily be used in place of proper normal human hand.

Timemy et al. (Timemy, 2016) improvised a study entitled “Improving the performance against force variation” where she differentiated between the use of force and its different variations and also the multiple ways of taking in account its performance. The result was that performance needed to be maintained in order to produce better results.

Next study in this regard was “Resolving the adverse impact of mobility on myoelectric pattern recognition in upper-limb multifunctional prostheses” conducted by (Samuel, 2017)

, which dealt with the potential improving of clinical robustness of the already available multifunctional prostheses. The effect and outcomes of the ability of mobility of the performance of EMG-PR motion classifier were highly investigated on the basis of myoelectric and accelerometer signals. Almost three methods were kept into consideration. Mobility importantly led to about 8.98% increase of classification.

(Samuel, 2017) did a research on “Improving control of Dexterous Prostheses Using Adaptive Learning” in which there was a clinical setting, data was attained from two electrodes which guided the movements of hand using many postures. It was decided that shortening of time was an easy way to use prostheses. The general method to experience the past experience in the form of models synthesized from previous subjects, so that adaptively of the prostheses could be availed.

(Abdel Maseeb, 2016) conducted a research on “Extraction and Classification of Multichannel Electromyographic Activation Trajectories for Hand Movement Recogninition” in which the writer proposed a well-defined system for hand movement recognition using many multichannel myographic signals obtained from the forearm surface. System can be used to provide inputs for a wide range of human computer interface systems. The hand movement recognition problem is seen as a multi-class distance-based classification of multi-dimensional sequences. The designed methods are evaluated using NINAPro database which comprises of 40 different hand movements. These movements are performed by 40 subjects in total.

(Iqbal, 2017) did a study on “Hand movement Recognition Based on singular value decomposition of Surface EMG Signal” which highlighted the effective techniques to classify the typical hand movements from the EMG signals based on principal component analysis. Other than the extracted singular values, the statistical parameters of the first five main components are seen as features in the projected values of the original data and characteristics for multiple hand movements. Hence, it was assumed that the proposed technique was offering high classification accuracy for classification of multiple hand movements.

Another study in this regard was an exploratory, computationally movement classification method for prosthetic hand application which was, “Hand movements classification for myoelectric control system using adaptive resonance theory” conducted by (Fariman, 2016). It was evident that from the four muscles namely flexor, extensor, biceps and triceps, myoelectric signals were desired which were undergone through segmentation and the respective features were extracted with a newly combined time-domain method of future extraction. The theory-based neural network which was highly adaptive resonance one had a famous improved and responsible computation period.

(Geng, 2018) recently did his study on “A Robust Representation Based Pattern Recognition Approach for Moelectric Control” in which they explained the process of intermittent interference and Robust pattern scheme which was invariant to noise contamination. The proposed scheme and its robustness and accuracy were investigated using high-density surface EMG recordings from the patients suffering from trauma. The findings tend to prove the comparison between proposed RMS-SRC pattern recognition scheme over the other schemes in the myoelectric control.

(Resnik, 2018) did a study to compare performance outcomes and self -report of a transradial amputee after the training and then one week training of EMG pattern recognition in order to examine the outcomes of PR and DC which showed decline and fall in the dexterity and highlighted a need for additional research quantifying the clinical and functional benefits for the upper limb protheses. The research was “Evaluation of EMG pattern recognition for upper limb protheses”.

(Li, 2013) did a research on “Boosting-based EMG patterns classification”in which the robust of classifier and its scheme was introduced and in order to make it easy to use was told. The outcome of the study was that the proposed scheme had better results than the previous one.

(Young, 2013) did a study on “Classification of Simultaneous movements using surface EMG pattern recognition” which focused on using amputee subjects through DoFS at the same time. Bayesian theory was the basis and low error rates was the outcome.

(Ray, 2011) conducted a research on “Identification of motion from multi-channel EMG signals for control of prosthetic hand” in which efficient and effective pattern recognizing ways and techniques from four of the channel electromyogram signals for control of multifunction prosthetic hand was made and constructed. Absolute value, zero crossings number, waveform lengths and slope sign changes were considered as parameters for pattern recognition. Using simple logistic regression and decision tree, it was modified much more. Keeping in view, the dominant features, the processing time as well as memory space were found to be less comparative.

Another sophisticated study in the context was “Decoding Motor unit activity from forearm muscles, perspectives for myoelectric control” conducted by (Kapalner, 2018) which

focused on recording surface EMG signals during motions of three degrees of freedom of wrist. The discharge timings were calculated in accuracy. The findings proved that accurate identification of neural drive to muscles can allow the development of new generation of myocontrol methods. These myocontrol methods will be based on motor unit spike trains.

In many cases the patients with amputation faces a lot of problems due to upper cut amputation which makes it very difficult to use electronic prostheses. The use of such device can restore a lot of actions and daily life activity that a normal person can perform like opening a bottle, doors, typing and other complex functions that normal prostheses cannot do. This is very crucial for the patients as the nerves in the upper limb must be attached with the active prostheses in order to transmit the neural link command to the device. The upper limbs are more crucial in daily life and their amputation must be kept very clear and up to the point where the electronic neural link can be established. In most cases the amputation is not up to the order to people are mostly using the conventional prostheses which is limited in the daily life use (Gaudet, Raison, & Achiche, 2018).

The generation of surface electromyography (SEMG) patterns associated with the distinct phantom limb movement of the whole hand, fingers, elbow and shoulder. It is very important that prostheses must be integrated with the SEMG so that it will be easier for the patient to use. But mostly studies shows that it is not easy to attached and integrate which creates a lot of challenges. The neural network between the muscle androgen and the device must be smooth in order to generate enough signals to move the parts of the prostheses use (Gaudet et al., 2018).

The prostheses accuracy obtained with an artificial neural network ranged between 60.9% and 93.0%. Device accuracy may decrease if the number of DoF considered in the classification increased, or the movement becomes more slow and complex. With a kinematic feature in it, it

can produce an increase of 4.8% in total. This study may lead to the development of a new myoelectric control method for electronic prostheses based on neural-muscle movements of the limbs and kinematics of prosthesis use (Gaudet et al., 2018).

There are many problems related to the upper limb amputation control and there are many cases where people are not able to control many daily life works. There are over 2 million people in United with upper limb amputation with every year the number are getting higher. There are many prostheses with electromyogram in the market that can do many problems solving works with the use of electric signal coming from muscles to the device. But the robustness, clarity in work and problem solving thing is far away in those devices because of the distance between the industry and the practical life design. The use of different techniques has been used in the past to overcome this but no one has clearly solved the problem (Al-Timemy, Khushaba, Bugmann, & Escudero, 2016, p.).

The research is about finding different variables, factors and problems that about the gap between the concept and the working devices. The use of such electric signal can provide a lot of robustness in the device but after utilizing it still it is not a practical design. This is because of the pattern detection and EMG controls. The pattern detection works like an independent A.I system that can sense the signals and movement of hand and provide more control and kinetics to the movement. It is like a pattern recognizing machine that can help the person in performing a task well because the machine has already recognized it. The use of such techniques can truly change the prostheses concept where the device has knowledge of the pattern, daily life movements and working ability of the person (Al-Timemy et al., 2016, p.).

The performance of the proposed features was tested on EMG. The data was collected from nine trans-radial amputees which performs different movements. The results indicate that the proposed features can achieve 8% more clarity in works and can provide a more complex

working device to the people who can perform complex tasks in their daily life (Al-Timemy et al., 2016, p.).

In the conventional myoelectric prostheses which are available in the market the device gets the signal from muscle using the two rods attached to the muscle and after it senses the muscle movement, it sends the signal to move the part. The signals delivered from the muscle are very crucial in generating the right kind of movement using prostheses. The signals which are delivered will then be decoded by the transmitter in the device to allow the right part to move. The use of prostheses in different complex work is often difficult and users have complained about it. The system is not very smooth and reliable and often receives negative reviews because of lack of not performing well in complex works (Samuel et al., 2017)

The researchers have studied six different subjects with upper limb amputation and they have tested three basic movements with them to help them conclude the result. The first one is the ascending on the chair, second is descending and third is sitting on the chair. They have deduced that over the time the performance degraded and the only solution is related to three separate solutions that they have mentions. The three possible solutions include the Dual-stage sequential strategy, Multi-scenario classification strategy, and Hybrid training method. These solutions were proposed to address this problem, and from the result the researches see that a meaningful reduction in classification error was achieved. The study also suggests that the proposed solutions, especially the Dual-stage strategy would be more practical in solving this problem once and for all. The contraction of the muscle is directly related to the signals that the device receives up on making the device more sensitive towards receiving the signal from contraction the prostheses shows a significant downfall in the classifier data. The use of such proposed strategies can solve the problem up to a great extant (Samuel et al., 2017).

The electromyography signals (EMG) are being used heavily in the prostheses now days and they are becoming very clinical in providing help to the people with limb amplitudes. People are using such devices to solve many basic problems now days but they have faced difficulty while wearing these for longer period of time. The researchers are trying to analyze the devices in a longer time period that how they perform in the longer period of time. There are many studies which show promising results about using such prostheses while there are very few studies that have been done in order to highlight the use of such devices in longer period of time. The researcher wants to know if the performance of the prostheses degrades or not after using for several days (A. Waris et al., 2018)

They have gathered a group of six people with limp amplitude who uses the prostheses for over sever days while performing many tasks in the daily life. There are EMG rods that connect with the muscles and calculate the contraction of the muscle in order to generate enough signals for the movement of the device in prostheses. What they have seen is that over the course of time the EMG rods performs way less as compare to the starting., The signal which they are generating is slowly degrading after the first use and it is a common pattern among all other candidates. They have analyzed that after a certain degree use the EMG rods degrade in the performance that is causing the delay in signals and it impacts on the performance of tasks. (A. Waris et al., 2018)

The researchers have carefully analyzed the results for all subjects and what they have deduced that the surface SEMG (7.2 ± 7.6%), IEMG (11.9 ± 9.1%) and CEMG (4.6 ± 4.8%) were significantly different (P < 0.001) from one another. The performance level decreases over several days from training to the performance of the task (A. Waris et al., 2018)

The use of myoelectric is the trending thing in the market and now days the devices are using the same technology in order to produced prostheses which can help people with amplitudes.

The research is using a new concept where the use of bionic hand will be theorized and analyzed. The use of motorized mechanical hand is the main course of this research which will take a new concept and way in order to solve the problem. The previous researchers have worked on the movement of the devices and muscles but not on the motion itself. The use of such bionic arms and hands that can generate the surface EMG signals can outperform all other prostheses available in the market (W.-T. Shi, Lyu, Tang, Chia, & Yang, 2018)

The researchers have studied various ways to construct such bionic hand that can detect the hand postures with the aim of controlling the hand by analyzing the SEMG signals coming from the muscles. The muscles will transfer the signals when the will try to think and moves the hand and in doing so it will generate the movement and neural connection between muscles and the bionic hand. The EMG receiver will receive the signals and then decode it to the movement SEMG signals in order to make the device work. The conversion here is very important as it will be the base of the smooth movement of the bionic hand (W.-T. Shi et al., 2018)

The use of similar features is utilized in the research for the development of the prototype. The features such as absolute value zero crossing, slope sign change, and waveform length in the algorithm for extracting hand-posture movements, and the k-nearest-neighbors (KNN) algorithm has been used as the main identifier of the hand posture movement. The microprocessor that they have utilized here is the Arduino microprocessor which can convert the signals coming from the muscle to the one which hand can utilize. The SEMG pattern recognition is also utilized in order to speed up the movement of the hand in recognized movements (W.-T. Shi et al., 2018).

The researchers are proposing an effective method of identification of the movement of hands and muscles that will provide a smooth flow in the overall functions in daily life. The use of such methods in which four channel EMG signals will be used in order to control the

multifunction prosthetic hand. The use of separate EMG signals will provide a better performance as the EMG signals degrade in the daily life usage. In order to avoid this degrade and a smooth flow a multi-channel EMG will be used. The use of hand motion muscular movement can be predicted and recorded in the use of prosthetic hand. The EMG signals are used in many other fields like in clinical field it can be used for the diagnoses of the muscular molecular diseases. In the electronic field it can be used to detect the motion of the muscle through neural impulses and can provide devices like prosthetics in order to assist people in their life (Geethanjali & Ray, 2011)

The hand can detect the intended movement and assist in performing the task more easily. This technique can provide more smoothness and reliability in the work where a multi-channel EMG signal is proving to be clinical. The pattern of the movement needs to be classified here so that the device can detect it much earlier where it can perform a better decision action. Time domain features such as mean absolute value, number of zero crossings, number of slope sign changes and waveform length are considered for pattern recognition. This recognition method will not only reduce the time to act up on the movement by the user but also it will be long lasting as compare to single channel EMG (Geethanjali & Ray, 2011)

After running test based on the dominant features for pattern recognition, the processing time as well as space of memory of the SLR and DT factors is found to be less in comparison with neural network (NN), k-nearest neighbor model 1 (kNNModel- 1), k-nearest neighbour model 2 (kNN-Model-2) and linear discriminant analysis. The overall accuracy of SLR classifier factor is found to be 91 ± 1.9% (Geethanjali & Ray, 2011)

There is a dire need of using prosthetics that are very reliable and their use is very smooth in daily life. Unfortunately 40-50 % people do not use the prosthetics in daily mainly due to the difficult use of such things. The people are in dire need for such a thing that is mainly doing

almost all of their work that they can perform in daily life. The industry is trying to achieve the robotic prosthetics that can use the Surface Electromyography (SEMG) in order to get the device working. Now days SEMG is the technology that is being use, in almost all the prosthetics available in the industry. The two rods connects the upper limb muscles with the device from where the device can get the signals and based on the algorithm and movement it performs. The problem here is the quality of the signal and movement that cannot be accepted by the users in their daily life (Tommasi, Orabona, Castellini, & Caputo, 2013)

The use of machine learning has gained a lot of ground in couple of years where the machine can act and reason like human mind and it can be utilized in various fields and machines. The machine learning can be applied to the processor of the hand processing unit where it can receive the SEMG signal which can improve the quality of the movement. The proposed system can improve the learning of prosthetics by the user simply by implying the recorded algorithm of machine learning which will contain the data collected from previous patients where they can perform different tasks. By doing this the processor which decodes the SEMG can simply learn how to use the movement based on the SEMG signals. This will lessen the time to act and perform much better in the case of movements (Tommasi et al., 2013)

The researchers have concluded extensive tests on databases recorded from healthy subjects in controlled and no controlled conditions. This data reveal that the method significantly improves the results over the period of extensive time. This promising approach might be employed to train the prosthesis itself before shipping it to a patient, leading to a shorter training phase.

The use of pattern recognition EMG has been researched heavily in past couple of years due to its extensive degree use and approach that is totally different from others. The convention EMG prosthesis uses the signals coming from the neural link up between the processor and the muscle using the electrodes that gets the muscle movement information and transmit it to the processor

where it is decoded based on the algorithm and then transmitted to the part. The pattern recognition is the recoded movement processor that can record the signals based on its frequency and value calculated by the processor itself. This can give a more suitable movement gesture that can be supported well by the mechanical parts of the prosthesis (Geng et al., 2018) The researchers have emphasizes on the use of pattern learning algorithm that can be problem solving in the prosthesis field. As mostly prosthesis are having problems of longevity and cannot perform well after a certain point of time. The use of pattern learning can also improve their life in the long run where the signals can be easily decoded by the processor based on their value. The electronic signals coming from the electrode are totally different from one another and from there the processor will not only record them but also learn how to use them in the future. If the user is using the certain muscle movement to grab a bottle of water than the processor will record this pattern in the memory. Next time when the user will try performing the same task, then the prosthesis will initiate the action without decoding it. This will not only save a significant amount of time but also it can help to provide long life to the SEMG signals. The use of WGN signals are very crucial in this whole proposed system where an extra 5 to 30 DB extra signal to noise ratio (SNG) is added. These will also amplify the signals coming from rods and it will help perform better in the overall action and the smoothness in kinetic movements (Geng et al., 2018).

The use of electric signals like SEMG are very crucial now days in the treatment of such people with amplitudes and it is providing great help in achieving the great feats with it. The use of these prostheses are getting advance and there are many other devices like these that can perform multiple movements at same time. This is what the researchers called as the degree of freedom (DOF) which is providing to be a game changer in the coming years. But unfortunately it has performing not up to the mark as there is only single DOF at one time. The working of

such devices is based on the pattern learning SEMG where the signals are already configured by the algorithm to perform the various tasks. The problem here is that the multiple SEMG must be recognized at same time in order to provide multiple movements by the prostheses ((Young, Smith, Rouse, & Hargrove, 2013)

The researchers are proposing a multiple classifier based on the Bayesian theory to provide classification simultaneously in the device. The simultaneous movement can be recorded by the algorithm at same time coming from different SEMG signals from the rods. The results of the proposed device is very promising as the devices were given to the amputee and non-amputee in order to check its design and working in real life. The three-DOF classification will be used in the overall proposed method(Young et al., 2013)

The new approach which they have proposed is based on a set of conditional parallel classifiers algorithms. It is the most promising with errors significantly less than 0.05, as compare to a single linear discriminant analysis (LDA) or another proposition. For the three-DOF classification, the conditional parallel approach had error rates of 6.6% on discrete and 10.9% on combined motions, while the single LDA had error rates of 9.4% on discrete and 14.1% on combined motions. These results show that the approach greatly reduces the errors and increase the functionality up to a great margin. (Young et al., 2013)

The proposed method is decoding the signals coming from the muscles and then transmitting it to the forearm non-isometric watch that can perform will in the movement. The researchers are proving the feasibility of decomposing the high density surface EMG signals from forearm muscles that is connected with the rods. It transmits it to the watch on the wrist that is motorized and from there the decoding process begins. The decoding of the EMG signal during motions of the hand gives three degree freedom of the writs in sever normally limbed subjects and two patients with limb deficiency. The test runs on them proves to be very efficient and provides a

lot of useful data that can show the feasibility and success of the matter ((Kapelner, Negro, Aszmann, & Farina, 2018)

The signals were recorded in another separate motor unit where they were decomposed and decoded to the use. The algorithm which uses to separate the signals based on the movement of the muscles. As the muscles from each category generates its own separate EMG signals and that is needed to be separated by using the algorithm. If the signals are not separated than the overall classification cannot work and the result will not be precise. The motor can take all the signals as one or cannot differentiate between signals so that is what the algorithm is being used in order to decode and separate (Kapelner et al., 2018)

The result shows that for each subject, 16 ± 7 motor units were identified in motor task. The discharge timings of these motor units were estimated with an accuracy of more than 85%. Also it is noted that the activity of 5-6 motor units per motor task was regularly detected in all repetitions of the same task. These findings prove the feasibility of accurate identification of the neural signals to muscles when it contracts. The relevant for myoelectric control, allowing the development of a new generation of myocontrol methods are based on motor units (Kapelner et al., 2018)

The pattern learning algorithm and SEMG are researched heavily in the past and researchers have seen the difference between the tests in the laboratory and in real life cases. The problem here is that the test which is conducted in the laboratory has special environment, movements and limited usage capability that allow the developers to believe that it can solve the problem or they have innovated something. But the real life scenario is changed where the use if very different, movements are different and the overall environment where it is being used is also very different. The main thing here is that the EMG based patterns are not up to the mark and they require a new approach in order to get the required innovation in prosthesis. The SEMG

boosting is the one thing that the researchers are trying to implement in it where the overall signal boosts to a new level. This will give better signals to the processor where the movements will be more robust and smoother. The use of such scheme and classification may also provide enough data that can be used in the implementation. The researchers have implemented this on two different classes of people in order to test it fully functionaly (Li, Wang, Yang, Xie, & Su, 2013)

In order to make the classifiers more robust and smooth to untrained classes, a classifier scheme is developed by the researchers in order to boost the signals. The SEMG are labeled with basic sever isometric movements and all are collected from healthy people in order to get the better result of muscle movement and overall SEMG signals. According to the results it is shown that the proposed classification scheme can reach up to about 92% accuracy in recognizing trained classes and 20% for untrained classes. But after adjusting the threshold of SEMG signals, the accuracy of rejecting untrained classes reaches up to around 80%, with small decrease in recognizing trained classes (down to 80%). It is also seen in the result that the proposed scheme has a better result in collecting the error distribution among different classes (Li et al., 2013)

The use of electromyogram (EMG) with pattern recognition (PR) is not new and it has been there in the field for many years. It is one of the main solutions present there for the people with limb amputee. The control and its working have been researched a lot by the researchers but there are very few studies for its medical benefits. The researchers are trying to research on various dimensions and factors connected to its clinical usage and how it is working in real life. The researchers have taken the test from the amputees after their training of the prosthesis and it gives them valuable data in recognizing the whole usage and its impact. The second thing that they have worked on is the use of the prosthesis after one week. As we know that the quality

of the signal degraded after the usage so they are here to classify the system based on the first hand use. (Resnik et al., 2018)

The research shows the outcomes, including measures of dexterity with and without self- learning things, activity performance, self-reported function, and prosthetic satisfaction was recorded immediately and 1 week after training. Speed of skill acquisition was assessed hourly by the researchers which included the first-hand experience of the people who were in the test. One subject completed training under both pattern recognition control and DC conditions. Both subjects completed pattern recognition training and testing (Resnik et al., 2018)

Outcomes of pattern recognition and DC for operating 2-DOF prosthesis in a single subject cross-over study were similar for 74% ratings, and favored DC in 26% rating. The two subjects who completed pattern recognition training showed decline in doing complex movement working one week after training ended. This truly highlights a need for additional research working on the functional and clinical benefits of pattern recognition control for upper limb prostheses (Resnik et al., 2018)

The human hand is very difficult to analyze in the movement of muscles because of abundance of nerves and muscles. The test is very difficult in order to get the right movement while using prostheses with pattern recognition. The main thing is that the prosthesis can perform only a couple of movements based on the muscles that can be decoded for the signals. There are many muscles in the hand and in the skin of forearm that is not easy for the EMG signals to be decoded and transmitted in order to start movement in the prosthesis. The researchers are working on a proposed method for the hand movement that works on the multichannel EMG signal obtained from the surface of the forearm. This proposed method will use more than one signal channel to get a variety of different movements from the muscles. This will provide a more complex

group of tasks that people can perform using the prostheses and performing different movements (AbdelMaseeh, Chen, & Stashuk, 2016).

The researchers have formulated the method for hand movement recognition problem in which a multi class distances based classification of multi-dimensional sequences will be utilized. The extraction of multi-channel EMG activation trajectories which underlies the hand movement and classifying the extracted trajectories using a metric based algorithm which works on multi- dimensional time wrapping is developed in it. The researchers have evaluated their method using the publically available NINApro database which is comprised of 40 different hand movements performed by 40 subjects (AbdelMaseeh et al., 2016).

This gives a huge group of movements of hands and testing through this mechanism provides a lot of boost to the development of the system. The result shows very lower rate of the errors in the movements of hands when testes. This shows that the proposed trajectory extraction method and the approach of distance metric are working better than the conventional EMG signal method (AbdelMaseeh et al., 2016).

The electromyogram is the activity of the neural network inside the body that transmits the neural signal from brain to the muscles in order to move the body parts. It is very essential thing in the body movements and without these electric signals the body parts cannot move. The same technique is taken from the body and implemented in the prosthesis where the prosthesis are connected with the body muscles using a needle of the electrode that takes the SEMG signals from the muscles and transmit it to the prosthesis in order to move the parts. The researchers are classifying some of the basic movements that the prosthesis can perform using these signals. They have classified the movements based on the SEMG, singular value decomposition (SVD) and principal component analysis (PCA). This will give more variety of the result in the testing phase as recommended by the researchers. (Iqbal, Fattah, & Zahin, 2017)

In order to classify the SVD frame on the SEMG data the first thing is to apply the short duration overlapping sub frames which create a matrix. The researchers have employed the SVD on the sub frame matrix to extract the singular value as well as the components computation which are involved in the PCA (Iqbal et al., 2017)

After the extraction of singular values there are more things to consider. As some statistical parameters of the first five principal components are proposed in the research too. The researchers have worked in the Eigen-space; the projected values of the original data are going to provide more complex movements for the hands in the daily life. With a view to performing the classification, the K-nearest neighborhood (KNN) classifier is applied on it. The researchers have validated their scheme and data with the help SEMG database consisting of six different hand movements obtained from three females and two males. The result has found that the proposed technique offers regularly high classification accuracy in classifying various hand movements with lower computational complexity (Iqbal et al., 2017)

As we have discussed above that hand movements are very complex because of the presence of different muscle group’s nerves and their working order. The daily life use of the hand is very crucial and with the people who are suffering with upper limb amputee it is very difficult to follow up the normal works in the daily life. The use of prostheses provides a degree of satisfaction in the life but that is also not the final product that can be utilized by the people. The use of EMG and pattern learning has been theorized before and it has been the main course of approach in this field for many years. The use of EMG surely provided certain degree of working capability but there is no surety over complex working scenario where the human hand can be involved in more than one movement at same time. For example: if a person wants to move a ball and at same time he wants to grip and press it than it will be very difficult from the current EMG prosthesis approach (Fariman, Ahmad, Marhaban, Ghasab, & Chappell, 2016)

This research proposes an exploratory study of a simple, accurate, and complex set of efficient movement classification technique for prosthesis mainly for the hands. For the research they have acquired the myoelectric from the four muscles, namely, flexor carpi ulnaris, extensor carpi radialis, biceps brachii, and triceps brachii, of four normal-limb subjects. The signals were decomposed, and the features were taken out with a new combined time-domain feature extraction method. The feature that they used is “Fuzzy C” which means clustering method and scatter plot was used to evaluate the performance of the proposed multi-feature method. The movements were classified with a multi-dimensional Adaptive Resonance Theory-based neural network. The results show promising data that the accuracy of the movement of the parts rose to 89%. It also shows decline in the errors that users may face in the complex hand movements which are not achievable by using the conventional prosthesis (Fariman et al., 2016).

The use of EMG has been started clinically since the 1960 and after that there have been many innovations and experimentation done with the technology. The paper presents the multi- channel EMG usage activity detection algorithm which uses histogram. This histogram will be obtained from recently developed one dimensional local binary pattern method which is being used in some fields. The algorithm here helps a lot in providing the activity/inactivity decision for multiple channel in setting up few parameters. These are windows length histogram type and the number of histogram bins. Both of these are utilized heavily in the process in order to increase the accuracy level and decrease the number of errors faced by the users in daily life. These formations are set once for the entire data set of EMG which will improve the energy Bonato methods which are being used in it (McCool et al., 2014)

The implementation requires no such preprocessing where whitening is needed. It can simply utilize multi concurrent EMG channel, LBP codes, accord the concurrent windows of each channel which will be simply shown into a single histogram instead of requiring a subjective

majority of vote mechanism. In future work, real-time testing would assess the applicability of the algorithm for use in prostheses as SNR changes. The implementation of fitting is currently available in commercial myoelectric prosthetic hands. The patients require manual calibration of threshold settings by clinicians according to each patient’s ability to move the muscles and produce EMG signals. It is common for patients to return to the clinic to change the threshold of the EMG signals because of the change in the muscles tone and ability to work. Therefore, the amount of noise during inactivity and the signal-to-noise ratio (SNR) changes from the previous one. In future work, the work needs to be done on the recalibration of the signals that are being used in the prosthesis (McCool et al., 2014)

After any accident or stroke which can affects the brain and nerves in any part of the body the motor neurons gets badly affected and their movement in the nerves collapsed. The impaired vision neuros are not able to communicate with the body and cells and thus this causes the amputation. The use of prosthesis is not new and it has many advantages and disadvantages to the whole process of providing support to the patient. The firing of the motor neurons stops from the nerves to the muscles and thus the muscle and whole body part is virtually disconnected from the overall system of the body. The neurons are the main cause of the communication between the brain and the parts which provides the main information to move or not. The researchers are working on the changes in the current and frequency of the EMG signals while flexing the elbows. They have taken the samples from both patients and normal people in order to get the precise data that can help the overall working capability of the prosthesis (Angelova, Ribagin, Raikova, & Veneva, 2018).

The test shows the involvement of fifteen normal and 10 patients with stroke were chosen for the test. Electromyography data from 6 muscles of the upper limbs during elbow flexion were examined by the researchers and normalized to the amplitudes of EMG signals during maximal

isometric tasks. When the motion starts to happen and people started flexing their elbow the highest flexing value was obtained during the process. Equal intervals of 0.3407 second were defined between these two moments of muscles and one extra interval before the start of the flexing of the elbow movement. For each of these time intervals the power/frequency function of EMG signals, the mean (MNF) and median frequencies (MDF), the maximal power (MPW) and the area under the power function (APW) were calculated. The result shows that the MNF was always higher than MDF and the power frequency was always higher during flexing of elbow (Angelova et al., 2018)

SEMG prosthetic hand has grasped attention of many due to its autonomous nature of control. As of now, include extraction from SEMG, particularly the multi-channel SEMG signals, is as yet a difficult issue due to its multifaceted nature. Different SEMG highlights have been proposed for SEMG-based prosthetic control. They can be commonly ordered into three classes: time area highlights, recurrence space highlights, and time recurrence space highlights. A research paper by (Chen, Zhang, Cheng, & Xi, 2018) proposed a scheme which consisted of two novel components to recognize multiple hand motions from SEMG (Surface Electromyography). Initially, they utilize the cumulative residual entropy (CREn), a proportion of vulnerability in a random variable, as the element. Second, we utilize the extreme learning machine (ELM), a quick and viable classifier utilizing single-concealed layer feedforward neural network with added substance neurons, to recognize diverse motions. To assess execution of the proposed framework, we contrast CREn and fluffy entropy, test entropy, and inexact entropy, and a best in class time-area highlight; and ELM with direct discriminant investigation and bolster vector machine. They are tried on four channel SEMG signals gained from ten typical subjects. Exploratory outcomes demonstrate that the grouping similarities of CREn are not just superior to those of different entropies with every one of the classifiers, yet

in addition similar to the time-area include for all the section lengths of two hundred, two hundred fifty and one thousand ms with all classifiers that are assessed. Moreover, the computational multifaceted nature of CREn is lower than those of different highlights, and ELM performs altogether quicker than different classifiers without relinquishing any execution. It recommends that the proposed CREn-ELM plot can possibly be connected to ongoing control of SEMG-based multifunctional prosthesis.

Surface electromyography (EMG) signals have been generally utilized in velocity studies and human-machine interface applications. In this paper, a regression model which relates the multichannel surface EMG signals to human lower limb flexion/expansion (FE) joint edges is built. In the exploratory worldview, three dimensional trajectories of 16external markers on the human lower limbs were recorded by optical movement catch framework and surface EMG signals from 10 muscles legitimately worried about the lower limb movement were recorded synchronously. With the crude information, the joint edges of hip, knee and lower leg were determined precisely and the time arrangement of power for surface EMG signals were extricated. At that point, a profound conviction networks (DBN) that comprises of confined Boltzmann machines (RBM) was worked, by which the multi-channel prepared surface EMG signals were encoded in low dimensional space and the ideal highlights were extricated. At long last, a back spread (BP) neural network was utilized to outline ideal surface EMG highlights to the FE joint points. The outcomes demonstrate that, the highlights extricated from multichannel surface EMG signals utilizing DBN technique (Chen et al., 2018)

As an electrical indication of muscle constriction, the myoelectric flag (electromyogram, EMG) contains rich data with respect to the action of neurons gushing from the spinal rope to muscle strands. Explicit examples of neural motioning to muscles can be precisely decoded from surface EMG (sEMG) signals with example acknowledgment methods. Customary

electromyogram (EMG) design acknowledgment does not consider the impact of frustrating factors, forestalling its viable clinical application. In this paper, we researched EMG design recognition of compound frustrating variables including anode move, compel variety, limb act and temporary float. We began with highlight extraction, to recognize reasonable portrayals to determine classification debasement brought about by cathode doffing. At that point, we continued to analyze an assortment of classifiers, to accomplish a great grouping exactness (CA) when exchanging the control arms (predominant and non-dominant) without extra adjustment. In conclusion, we researched versatile learning methodologies, to relieve the decrease in CA over a moderately long length (at some point) of utilization. Two information gathering were connected, and EMG datasets gathered in these two conventions were masterminded into three situations for dynamic approval of the highlights, classifiers, and versatile learning techniques. Our outcomes demonstrated that the highlights extricated from the recurrence space improved the CA by 30% in examination with their unique time-space portrayals (14 motions, 67.5 ± 9.8%, p < .001). While exchanging the control arms, the execution accomplished by the classifiers all declined and no technique had all the earmarks of being more dependable than the others (14 motions, the best 45.5 ± 8.4%, probabilistic neural network; 7 motions, the best 70.4 ± 10.6%, direct discriminant investigation). Finally, the new versatile learning technique using delegate tests could successfully moderate characterization corruption by refreshing the classifier's preparation tests progressively (Gu, Yang, Huang, Yang, & Liu, 2018).

Electromyography (EMG) has been broadly used to extricate a control motion in various human-machine interfaces. These interfaces have applications in clinical and restoration drug and have been contemplated broadly over the most recent two decades. EMG signals are stochastic in nature and proportions of its abundancy can be utilized to appraise muscle

initiation level and power. EMG signals are one of the major neural control hotspots for myoelectric prostheses, furnishing semi regular control modalities in spite of the fact that with constrained functionalities. EMG signals were recorded simultaneously from surface and intramuscular cathodes, with intramuscular terminals kept in the muscles for seven days. Seven hand motions were assessed day by day utilizing straight discriminant examination and the order mistake quantified inside (WCE) and between (BCE) days. BCE was processed for every conceivable mix between the days. For all subjects, surface sEMG (7.2 ± 7.6%), iEMG (11.9 ± 9.1%) and cEMG (4.6 ± 4.8%) were fundamentally unique (P < 0.001) from one another. A regression among WCE and days (1– 7) was by and large not noteworthy suggesting that execution might be viewed as comparable inside every day. Regression among BCE and time difference (Df) in days was critical. The incline among BCE and Df (0– 6) was altogether not quite the same as zero for sEMG (R2 = 89%) and iEMG (R2 = 95%) in amputees. Results show that execution consistently corrupts as the time distinction among preparing and testing day increments. Moreover, for iEMG, execution in amputees was straightforwardly corresponding to the span of the residual limb (A. Waris et al., 2018)

As of late, a few techniques have been created to dissect and de-clamor the EMG signal. The regular procedures dependent on Fourier investigation (e.g., IIR flters) are generally utilized for EMG-based filtering. In any case, Fourier investigation is absolutely founded on pre- defined premise capacities, which lessens the clamor as well as weaken the EMG flag. As an option in contrast to the typical Fourier change strategy, wavelet examination is additionally promoted because of its favorable circumstances as far as the time– recurrence portrayal. Te wavelet- based methodologies, in any case, are additionally problematic in light of the fact that the pre- chosen wavelet work is regularly not appropriate for coordinating the common property of an EMG flag. Past investigations have likewise presented the Empirical Mode Decomposition

(EMD) way to deal with handle EMG signals. Rather than those revealed in written works, the EMD is a completely information driven versatile time– recurrence examination technique, and offers no earlier presumption through the general information preparing system (Chen et al., 2018)

Analysis

Classification of Surface EMG Signals Using Ensemble Classifiers Dept of IS , EFFAT UNIVERSITY

It is required from the thesis to classify data that is obtained from sensors and measures and stored in different datasets according to data mining approach of ensemble classifiers so we can get results and compare them using many measurements tools of accuracy and other to get the best classifier and highest accuracy.

· Laptop.

· iPad or some application for people have a special case.

· Need WEKA software to see the result of data.

· Zotero software for citations the needed and used references.

Classification of Surface EMG Signals Using Ensemble Classifiers Dept of IS , EFFAT UNIVERSITY

19

Budget/Funding

Using the WEKA software specialized for data mining and applying classifying techniques and methods on the identified data sets, the process will flow from the start point which is loading the datasets files into the WEKA to finding the output details and visualization as the next diagram shows:

The data flow diagram displayed above shows how the process of classifying Surface EMG signals will be done. The process starts by reading the data from the specified datasets, the applying specific classifier then after the process ended, the result will be displayed so we can analyses it based on accuracy value. If the accuracy level is acceptable we will consider it, else reprocessing data will be applied with another classifier.

Displaying the previous explanation of the process of classifying Surface EMG signals, we got overall idea about the plan we will use through the process and the requirements of the process.

Classification of Surface EMG Signals Using Ensemble Classifiers Dept of IS , EFFAT UNIVERSITY

Data is downloaded from the following web site:

One of the classes of basic functions which are placed in time and frequency are described as wavelets and can be symbolized as the following;

Ψ ( t ) = 1 𝜓 ((𝑡 − 𝑢)/𝑠) (1)

√𝑆

In that formula, the dilation parameter is presented as s and the translation parameter is presented as u.

W (u,s) = 1 +∞ 𝑥(𝑡)ψ*((𝑡 − 𝑢)/𝑠) d𝑡 (2)

x √𝑆 ∫−∞

Dilation and translation parameters of the wavelet can be separated dyadically.

mk(t) = 2-m/2 (2-mt - k)(3)

At this equation, the symbol ( (t)) can be considered as the mother wavelet and m symbolizes dilation parameter while k symbolizes translation parameter. Since the dilation has a dyadic discretization and translation parameters, at rougher scales wavelet numbers and scaling function coefficients fall (Clerc & Mallat, 2002)

The capacity of PCA is integrated by Multi-scale Principal Component Analysis (MSPCA) in order to remove the relationship between variables by extracting a linear relationship with the ability of wavelet analysis in order to extract deterministic features and to remove nearly the relationship between auto-correlated measurements. Because of the innate ability of latent variable regression methods errors can be decreased through keeping only latent variables which obtain the relationship between the variables. As noise relied on the relationship between variables is discarded by the former, wavelet thresholding of the latent variable and noise removal by getting rid of the latent variables that are of less importance supplement each other while noise relied on extracting features from measurements is discarded by the latter approach. The relationship between wavelet thresholding and latent variable models is made use of by present multi-scale empirical modeling methods (Han, J., Kamber, 2011)The multi-scale modeling approach is considered as a kind of error-in-variables modeling as both methods take a step to discard errors and decide the model parameters at the same time. Multiscale PCA integrates the features of wavelet analysis and PCA through decomposing every variable on a chosen wavelet which is pursued by calculating by PCA of the matrix of coefficients (Lee, Park, & Vanrolleghem, 2005). Establishing different PCA model for the coefficients at each scale will allow better removal of the noise employing scale dependent threshold values if the noise is non-white or scale-dependent. The modeling and thresholding steps are not combined truly and a rigorous statistical basis is missed by Multiscale modeling as Multiscale modeling by thresholding the latent variables can function better than the present methods DWT For instance, a discrete-time signal is separated into a set of signals or wavelet coefficients by Discrete Wavelet Transform (DWT) in order to scale and shift mother wavelet

Dj [i]   x[k ] h[2  i  k ]

k

Aj [i]   x[k ]  l[2  i  k ]

k

(1)

(2)

Until j does not exceed jm , the generation process of Dj and Aj that is showed above repeats (Kumar, Alam, & Siddiqi, 2017)The wavelet transform is used to have a better time resolution of a signal's decomposition because it serves to decompose a signal into a set of basic functions. WPD is taken into consideration as a continuous time wavelet decomposition that is tested at various frequencies at each level or scale. While WPD produces sets of wavelet coefficients, DWT provides wavelet coefficient sets for j-levels (Gokhale & Khanduja, 2010).However, WPD provides a better frequency resolution than DWT does for the signal that is decomposed since it might cause to lack important information in higher frequency components (Constable & Thornhill, 1993).

(Gkoktsi & Giaralis, 2015) A novel numerical study is undertaken to assess the influence of the frequency domain (FD) attributes of wavelet analysis filter banks for vibration- based structural damage detection and localization using the relative wavelet entropy (RWE): a damage-sensitive index derived by wavelet transforming linear response acceleration signals from a healthy/reference and a damaged state of a given structure subject to broadband excitations. Four different judicially defined energy-preserving wavelet analysis filter banks are employed to compute the RWE pertaining to two benchmark structures via algorithms which can efficiently run on wireless sensors for decentralized structural health monitoring. It is shown that filter banks of compactly supported in the FD wavelet bases (e.g., Meyer wavelets and harmonic wavelets) perform significantly better than the commonly used in the literature dyadic Haar discrete wavelet transform filter banks since they achieve enhanced frequency selectivity among scales (i.e., minimum overlapping of the frequency bands corresponding to adjacent scales) and, therefore, reduce energy leakage and facilitate the interpretation of numerical results in terms of scale/frequency dependent contributors to the RWE. Moreover, it is

demonstrated that dyadic DWT filter banks with large constant Q values (i.e., ratio of effective frequency over effective bandwidth) are better qualified to capture damage information associated with high frequencies. Finally, it is concluded that wavelet analysis filter banks achieving non-constant Q analysis are most effective for RWE-based stationary damage detection as they are not limited by the dyadic DWT discretization and can target the structural natural frequencies in cases these are a priori known.(Gkoktsi & Giaralis, 2015)In order to analyze oscillatory signals the tunable-Q wavelet transform (TQWT) is considered as a strong transform (Bayram & Selesnick, 2011).For instance, comparatively high Q-factor ought to be had by the wavelet transform as making use of wavelets in order to analyze and process oscillatory signals such as speech, EEG, ECG and so on. Most wavelet transforms do not possess much skill to adjust the wavelet's Q- factor rather than the continuous wavelet transform. On the other hand, they are different from each other because the RADWT is complex as a concept, and RADWT cannot be applied sufficiently like TQWT making use of radix-2 FFTs, the Q-factor of the transform does not relate to the parameters of RADWT easily unlike the parameters of TQWT. The transform, which depends on scaling of discrete-time signals by arbitrary scaling factors making use of DTFT, is alike the continuous wavelet transform, however, is enhanced specially for discrete-time signals. In the case of making use of this filter bank in order to implement a wavelet transform, it will not also be possible to locate these wavelets well (J. R. DELLER, JR., 1994)The difference of the TQWT from others is that it is considered as a result of the exploration of how well a tunable Q-factor wavelet transform that depends on the structure of the discrete\dyadic wavelet transform (DWT) might be enhanced (Gkoktsi & Giaralis, 2015)(Bayram & Selesnick, 2011)

The huge theoretical, practical, and computational resources of the classic DWT is also used by the CWT as the DT-CWT's real and imaginary parts are real wavelet transforms. So what makes dual-tree wavelet design a challenging design is the FB's shared design in order to give an intricate wavelet and scaling function which are very similar to analytic. Therefore, the DT-CWT approaches very near to mirroring the good features of the Fourier transform, like being gentle waveless weight, a nearly shift-invariant magnitude that has a basic near-linear phase encoding of signal shifts, a drastically lowered

misidentification, and directional wavelets present in dimensions with higher levels. It is more difficult to design an intricate nearly analytic wavelet basis than a real wavelet basis. The real part of the transform is given by DWT whereas the imaginary part of the transform is given by the second DWT. We are going to show the two real wavelets which are related to each of the two real wavelet transforms as being h(t ) and g(t ). As the DT- CWT is implemented to a real signal, the result of the upper and lower FBs that might be stocked one by one will be considered as the real and imaginary parts of the complex coefficients. Additionally, the present theory and practice of real wavelet transforms give information about the use of the DT-CWT due to the implementation of DT-CWT made use of two real wavelet transforms. For instance, it is possible to implement principles like vanishing moments for wavelet design and like thresholding of wavelet coefficients for wavelet-based signal processing algorithms which might be enhanced for real wavelet transform to the DT-CWT. In the case of application of the dual-tree wavelet transform with filters that do not match the necessity, then the all benefits of analytic wavelets that are explained above will not be ensured by the transform (Bayram & Selesnick, 2011)

This feature means that a unique z transfer function can define the signal path through any single subband of the transform and its inverse and it is not influenced by the down and up sampling in the transform (Bayram & Selesnick, 2011)

The main idea in our thesis depends on the classification. And there are different methods that can be used. Following we will explain the most important classification methods to be used in out thesis to achieve efficient control using EMG signal classification. These methods are classified into the following main categories:

Is a set of connected input/output units in which each connection has a weight associated

with it. A multilayer feed-forward neural network consists of an input layer, one or more hidden layers, and an output layer. These inputs pass through the input layer and are then weighted and fed simultaneously to a second layer of "neuronlike" units, known as a hidden layer. The outputs of the hidden layer units can be input to another hidden layer, and so on. The weighted outputs of the last hidden layer are input to units making up the output la..;yer, which emits the network's prediction for given tuples. It is a feed-forward network since none of the weights cycles back to an input unit or to a previous layer's output unit. Each output unit takes, as input, a weighted sum of the outputs from units in the previous layer.

The k-closest neighbor method is work serious when given vast preparing sets. It has been generally utilized as a part of the range of example acknowledgment. Closest neighbor classifiers depend on learning by relationship, that is, by contrasting a given test tuple and preparing tuples that are like it. The preparation tuples are portrayed by n properties. Each tuple speaks to a point in a n-dimensional space. Along these lines, all the preparation tuples are put away in a n-dimensional example space. At the point when given an obscure tuple, a k-closest neighbor classifier looks the example space for the k preparing tuples that are nearest to the obscure tuple. These k preparing tuples are the k "closest neighbors" of the obscure tuple. "Closeness" is defined as far as a separation metric, for example, Euclidean separation. As it were, for each numeric property, we take the distinction between the relating estimations of that characteristic in tuple X1 and in tuple X2, square this distinction, and gather it. The square root is taken of the aggregate collected separation tally. For k-closest neighbor classification, the obscure tuple is relegated the most well-known class among its k-closest neighbors. Whenever k=1, the obscure tuple is doled out the class of the preparation tuple that is nearest to it in design space. All in all, the bigger the quantity of preparing tuples, the bigger the estimation of k will be (Han, J., Kamber, 2011)

Support vector machines (SVMs) is a method for the classification of both linear and nonlinear data. Within this new dimension, it searches for the linear optimal separating hyperplane (i.e., a "decision boundary" separating the tuples of one class from another).

The good news is that the approach described for linear SVMs can be extended to create nonlinear SVMs for the classification of linearly inseparable data. Once the data have been transformed into the new higher space, the second step searches for a linear separating hyperplane in the new space. (Rosenbaum, Hinselmann, Jahn, & Zell, 2011)

Dependencies between attributes inevitably reduce the power of Naive Bayes to discern what is going on. The normal-distribution assumption for numeric attributes is another restriction on Naive Bayes as we have formulated it here.

Regression trees were introduced in the CART system of Breiman et al. Using model trees for generating rule sets (although not partial trees) has been explored by Hall et al. A comprehensive description (and implementation) of model tree (Hall, Thomsen, Henriksen, & Lohse, 2011)

REPTree fabricates a choice or relapse tree utilizing data pick up/difference lessening and prunes it utilizing decreased blunder pruning. Improved for speed, it just sorts esteems for numeric qualities once. It manages missing esteems by part cases into pieces, as C4.5 does. You can set the base number of cases per leaf, most extreme tree profundity (helpful while boosting trees), least extent of preparing set difference for a split (numeric classes just), and number of folds for pruning. (Abdel Maseeb, 2016)

NBTree is a cross breed between choice trees and Naïve Bayes. It makes trees with leaves that are Naïve Bayes classifiers for the occasions that achieve the leaf. While developing the tree, cross-approval is utilized to choose whether a hub ought to be part further or a Naïve Bayes display utilized rather (Hall et al., 2011)

LADTree is a rotating choice tree calculation that can deal with multiclass issues in view of the LogitBoost. Like ADTree, the quantity of boosting cycles is a parameter that can be tuned for the current information and decides the measure of the tree built. (Han, J.,

Kamber, 2011)

Another heuristic in C4.5 is that candidate splits on numeric attributes are only considered if they cut off a certain minimum number of instances. The information gain may be negative after subtraction, and tree growing will stop if there are no attributes with positive information gain a form of prepruning (Hall et al., 2011)

The classification works as follows: the random trees classifier takes the input feature vector, classifies it with every tree in the forest, and outputs the class label that received the majority of "votes". Random Trees are essentially the combination of two existing algorithms in Machine Learning: single model trees are combined with Random Forest ideas. Secondly, when growing a tree, instead of always computing the best possible split for each node only a random subset of all attributes is considered at every node, and the best split for that subset is computed.(Han, J., Kamber, 2011)

More formally, each tree depends on the values of a random vector sampled independently and with the same distribution for all trees in the forest. Random forests are a combination of tree predictors such that each tree depends on the values of a random vector sampled independently and with the same distribution for all trees in the forest. The generalization error for forests converges as to a limit as the number of trees in the forest becomes large. The generalization error of a forest of tree classifiers depends on the strength of the individual trees in the forest and the correlation between them.(Han, J., Kamber, 2011)

Is a classifier gathering in view of highlight extraction. For a base classifier, we make the preparation information by haphazardly part the list of capabilities into K subsets and Principal Component Analysis (PCA) is connected to every subset. Keeping in mind the end goal to protect the fluctuation data in the information, all foremost parts are held. The idea of the revolution approach is to energize at the same time singular precision and decent variety inside the troupe. Decent variety is developed through the element

extraction for each base classifier. Because of their affectability to revolution of the component hub, choice trees were picked here and, subsequently, named "timberland". Precision is looked for by holding all central segments, and furthermore utilizing the entire dataset to prepare each base classifier (Hall et al., 2011)

In bagging, numerous training datasets of the same size are selected at random from the problem domain to build a decision tree for each dataset. But, bagging yields a combined model which mostly achieves better results than the single model created from the original training data. Bagging can be used for numeric prediction in which instead of voting, the outcomes of the individual predictions are averaged (Hall et al., 2011)

The boosting which joins numerous models to accomplish a mindfulness by searching for learning models which is a partner of another. Like packing, boosting likewise utilizes voting or averaging to join the yield of individual models. However, boosting is iterative in which each made model is influenced by the execution of those assembled beforehand, though stowing constructs singular models independently. Boosting urges new models to wind up specialists for cases took care of mistakenly by prior ones by doling out more noteworthy weight to those occurrences. Another contrast amongst sacking and boosting is that boosting weights a model's effect by its execution as opposed to giving equivalent weight to all models.(Li, 2013)

(Hall et al., 2011)The boosting calculation, starts by allotting measure up to weight to all occasions in the preparation information. It at that point calls the learning calculation to frame a classifier for this information and reweights each occurrence as per the classifier's yield. The heaviness of effectively characterized cases is diminished, and that of misclassified ones is expanded. This creates an arrangement of "simple" examples with low weight and an arrangement of "hard" ones with high weight. In the following cycle— and every single resulting one—a classifier is worked for the reweighted information, which thusly concentrates on ordering the hard examples effectively. At that point the examples' weights are expanded or diminished by the yield of this new classifier. Thus,

some hard cases may turn out to be significantly harder and less demanding ones much less demanding; then again, other hard occasions may wind up noticeably less demanding, and simpler ones harder—all conceivable outcomes can happen by and by. After every cycle, the weights reflect how frequently the cases have been misclassified by the classifiers delivered up until this point. By keeping up a measure of "hardness" with each case, this strategy gives an exquisite method for creating a progression of specialists that supplement each other (Hall et al., 2011)

Wagging (Webb, 2000)is variant of bagging, that requires a base learning algorithm that can utilize training cases with differing weights. there is evidence that bagging is more effective than AdaBoost at reducing variance their combination may be able to retain AdaBoost's bias reduction while adding bagging's variance reduction to that already obtained by AdaBoost. In addition to the bias and variance reduction properties that this algorithm may inherit from each of its constituent committee learning algorithms, MultiBoost has the potential computational advantage over AdaBoost that the sub- committees may be learned in parallel, although this would require a change to the handling of early termination of learning a sub-committee.(Webb, 2000)

When the number of training objects is relatively small compared with the data dimensionality, by constructing classifiers in random subspaces one may solve the small sample size problem. When data have many redundant features, one may obtain better classifiers in random subspaces than in the original feature space. The combined decision of such classifiers may be superior to a single classifier constructed on the original training set in the complete feature space.(Ho, 1998)

The processing of that required uploading the datasets into the WEKA Then we will apply the following steps

It is necessary to select an appropriate model to use in the classification process. The main models that we will use in our research are the ensemble classifiers to find out the results which

displays many important statistics using the SVM , ANN , RF, C4.5, Random Tree, REPTREE, LAD Tree and NB.

After uploading the datasets and choosing the proper classifiers, the WEKA will implement statistical analysis to display later the output through output screen that contain all the information about the used dataset , used classifier and the statistical results and observing the changes in the output.

After applying that methodology, the classification and data processing becomes clear and know we can say that the results are ready to be obtained in order to be displayed and evaluated under many considerations and statistical measures in the next section to determine the best classifier used in our datasets.

Classification of Surface EMG Signals Using Ensemble Classifiers

Dept of IS , EFFAT UNIVERSITY

Following are the results for the classification process apply on S3, S4 data sets. Data are summarized in the tables according the major four measures:

Taking advantage of the results required well understanding of the presented results. These results are about statistical information includes:

· accuracy,

· F-measure and

· mean area under the ROC curve (AUC)

· Kappa Statistics

· Classification Results for S1 :

Classification of Surface EMG Signals Using Ensemble Classifiers Dept of IS , EFFAT UNIVERSITY

60

Classification results for S2 :

· Classification results for S3:

· Classification results for s4

As a new work we apply the classification methods on a new datasets and the results are as the follows :

Figure 17 : S1 Accuracy for different classifiers

Figure 18 :S2 Accuracy for different classifiers

Figure 19 : S3 Accuracy for different classifiers

Figure 20 : S4 Accuracy for different classifiers

S4

Figure 21 : S5 Accuracy for different classifiers

Figure 22 : S6 Accuracy for different classifiers

Figure 23 : S7 Accuracy for different classifiers

Figure 24 : S8 Accuracy for different classifiers

Figure 25 : S9 Accuracy for different classifiers

After applying the different ensemble classifiers we get the previous results and after observing the accuracy for the acted classifiers we find that :

For s1 data set the rotation forest with the simple logistic gave the highest accuracy with 79.33

%

And for s2 Random sub space with simple logistic gave the highest accuracy with 80.67 And for s3 the Bagging with simple logistic also gave the highest accuracy with 93.333% And for s4 rotation forest with ANN also gave the highest accuracy with 88%

So and as a result the simple logistic classifier is very good classifier and it give high accuracy comparing with other types of classifiers.

And later we will do more studies on other data sets to observe more results to ensure these points.

Classification of Surface EMG Signals Using Ensemble Classifiers

Dept of IS , EFFAT UNIVERSITY

And our thesis as any other one and through the different phases it passed faced some challenges. On the other hand it has many achievements.

The next parts display these sides in addition to the future work that will be done for the next term of this thesis.

The main achievement of this thesis is the literature review and methodology that was applied under organized approach to get the aimed results.

The study investigates the rule of classifier from the EMG signal parameters to differentiate the EMG signal coming from different patients.

According to our experimental results, the suitable parameters were determined to successful implemented to complete system. The performance of the system which is the ability to identify the EMG signals coming from different people was verified.

Using the machine learning algorithms add many features from accuracy to samples independence.

Every study has their different objectives, methods and results. The following table shows a compression between all these studies and our study:

Classification of Surface EMG Signals Using Ensemble Classifiers Dept of IS , EFFAT UNIVERSITY

58

The challenges that the thesis had summarized as follows:

· The classification of an SEMG becomes difficult when the level of muscle contraction is low and when there are multiple active muscles.

· The presence of noise and crosstalk from closely located and simultaneously active muscles is exaggerated when muscles are weakly active such as during sustained wrist and finger flexion and of people with neuropath logical disorders or who are amputees.

After the previous chapters, the idea, aim and objectives of the thesis became clear for all. And as a summary we can say that the using of the classification of the Surface EMG signals using ensemble classifiers serve the users in handling the huge data sets in the perfect way that save their time and effort resulting accurate results with the ability to analyse and visualize it in many forms.

Classification of Surface EMG Signals Using Ensemble Classifiers Dept of IS , EFFAT UNIVERSITY

After the previous chapters, we identified the classification process that we used for classifying Surface EMG signals using the ensemble classifiers under a well-studied methods and techniques.

This process is very important in the scientific field. Where analysts benefit from the study of data sets and the preparation of research and reports on them, taking advantage of the high accuracy that classifiers gave during the classification process.

The obtained results for applying classification with WEKA software was evaluated and we got the evaluation that shows the Bagging with simple logistic gave the highest accuracy with 93.333% .

After introducing the current work which includes the overview of the research, literature review, design and methods and a results of applying ensemble classification methods on the datasets, the future work will be a proceeding in the processing of the data sets to get higher accuracy percentage results. Besides to publishing our research paper in a journal.

As we finish the work on the current research paper about Classification of Surface EMG Signals Using Ensemble Classifiers using WEKA data mining software, we recommend to continue this research in the future and testing different datasets with another data mining techniques to discover other effective tools and functions.

Classification of Surface EMG Signals Using Ensemble Classifiers Dept of IS , EFFAT UNIVERSITY

62

Abdel Maseeb. (2016). Extraction and Classification of Multichannel EMG for Hand Movement Recognition.

AbdelMaseeh, M., Chen, T.-W., & Stashuk, D. W. (2016). Extraction and classification of multichannel electromyographic activation trajectories for hand movement recognition. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 24(6), 662–673.

Al-Timemy, A. H., Khushaba, R. N., Bugmann, G., & Escudero, J. (2016). Improving the Performance Against Force Variation of EMG Controlled Multifunctional Upper- Limb Prostheses for Transradial Amputees. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 24(6), 650–661. https://doi.org/10.1109/TNSRE.2015.2445634

Angelova, S., Ribagin, S., Raikova, R., & Veneva, I. (2018). Power frequency spectrum analysis of surface EMG signals of upper limb muscles during elbow flexion–A comparison between healthy subjects and stroke survivors. Journal of Electromyography and Kinesiology, 38, 7–16.

Bayram, I., & Selesnick, I. W. (2011). A dual-tree rational-dilation complex wavelet transform. IEEE Transactions on Signal Processing, 59(12), 6251–6256. https://doi.org/10.1109/TSP.2011.2166389

Chen, J., Zhang, X., Cheng, Y., & Xi, N. (2018). Surface EMG based continuous estimation of human lower limb joint angles by using deep belief networks. Biomedical Signal Processing and Control, 40, 335–342.

Clerc, M., & Mallat, S. (2002). The texture gradient equation for recovering shape from texture. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(4), 536– 549. https://doi.org/10.1109/34.993560

Constable, R., & Thornhill, R. J. (1993). Using the discrete wavelet transform for time- frequency analysis of the surface EMG signal. Biomedical Sciences Instrumentation, 29, 121–127.

Fariman. (2016). Hand movements classification for myoelectric control system using adaptive resonance theory.

Fariman, H. J., Ahmad, S. A., Marhaban, M. H., Ghasab, M. A., & Chappell, P. H. (2016).

Hand movements classification for myoelectric control system using adaptive resonance theory. Australasian Physical & Engineering Sciences in Medicine, 39(1), 85–102.

Gaudet. (2018 ). Classification of Upper limb phantom movements in transhumeral. Gaudet, G., Raison, M., & Achiche, S. (2018). Classification of Upper Limb Phantom

Movements in Transhumeral Amputees Using Electromyographic and Kinematic Features. Eng. Appl. Artif. Intell., 68(C), 153–164. https://doi.org/10.1016/j.engappai.2017.10.017

Geethanjali, P., & Ray, K. (2011). Identification of motion from multi-channel EMG signals for control of prosthetic hand. Australasian Physical & Engineering Sciences in Medicine, 34(3), 419–427.

Geng. (2018). A Robust Sparse Representation Based Pattern Recognition Approach for Myoelectric Control.

Geng, Y., Ouyang, Y., Samuel, O. W., Chen, S., Lu, X., Lin, C., & Li, G. (2018). A robust sparse representation based pattern recognition approach for myoelectric control.

IEEE Access, 6, 38326–38335.

Gkoktsi, K., & Giaralis, A. (2015). Effect of frequency domain attributes of wavelet analysis filter banks for structural damage localisation using the relative wavelet entropy index. International Journal of Sustainable Materials and Structural Systems, 2(1/2), 134. https://doi.org/10.1504/IJSMSS.2015.078365

Gokhale, M. Y., & Khanduja, D. K. (2010). Time Domain Signal Analysis Using Wavelet Packet Decomposition Approach. International Journal of Communications, Network and System Sciences, 03(03), 321–329. https://doi.org/10.4236/ijcns.2010.33041

Gu, Y., Yang, D., Huang, Q., Yang, W., & Liu, H. (2018). Robust EMG pattern recognition in the presence of confounding factors: Features, classifiers and adaptive learning.

Expert Systems with Applications, 96, 208–217.

Hall, V., Thomsen, R. W., Henriksen, O., & Lohse, N. (2011). Diabetes in Sub Saharan Africa 1999-2011: Epidemiology and public health implications. a systematic review. BMC Public Health, 11(1), 564. https://doi.org/10.1186/1471-2458-11-564

Han, J., Kamber,. (2011). Data Mining: Concepts and Techniques.

Ho, T. K. (1998). The random subspace method for constructing decision forests. IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(8), 832–844. https://doi.org/10.1109/34.709601

Iqbal. (2017). Hand Movement Recognition Based on Singular.

Iqbal, O., Fattah, S. A., & Zahin, S. (2017). Hand movement recognition based on singular value decomposition of surface EMG signal. 2017 IEEE Region 10 Humanitarian Technology Conference (R10-HTC), 837–842. IEEE.

J. R. DELLER, JR. (1994). The story of three students’ discovery of the elationship between continuous and discrete transforms.

Kapalner. (2018). Decoding Motor Unit Activity From Forearm.

Kapelner, T., Negro, F., Aszmann, O. C., & Farina, D. (2018). Decoding motor unit activity from forearm muscles: perspectives for myoelectric control. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 26(1), 244–251.

Kumar, N., Alam, K., & Siddiqi, A. H. (2017). Wavelet Transform for Classification of EEG Signal using SVM and ANN. Biomedical and Pharmacology Journal, 10(4), 2061– 2069.

Lee, D. S., Park, J. M., & Vanrolleghem, P. A. (2005). Adaptive multiscale principal component analysis for on-line monitoring of a sequencing batch reactor. Journal of Biotechnology, 116(2), 195–210. https://doi.org/10.1016/j.jbiotec.2004.10.012

Li. (2013). Boosting-Based EMG Patterns Classification.

Li, Z., Wang, B., Yang, C., Xie, Q., & Su, C.-Y. (2013). Boosting-based EMG patterns classification scheme for robustness enhancement. IEEE Journal of Biomedical and Health Informatics, 17(3), 545–552.

McCool, P., Chatlani, N., Petropoulakis, L., Soraghan, J. J., Menon, R., & Lakany, H. (2014).

Lower arm electromyography (EMG) activity detection using local binary patterns. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 22(5), 1003– 1012.

Ray. (2011). Identification of motion from multi-channel EMG signals for control of prosthetic hand.

Resnik. (2018). Evaluation of EMG pattern recognition for upper limb prosthesis control_ a case study in comparison with direct myoelectric control.

Resnik, L., Huang, H. H., Winslow, A., Crouch, D. L., Zhang, F., & Wolk, N. (2018).

Evaluation of EMG pattern recognition for upper limb prosthesis control: a case study in comparison with direct myoelectric control. Journal of Neuroengineering and Rehabilitation, 15(1), 23.

Rosenbaum, L., Hinselmann, G., Jahn, A., & Zell, A. (2011). Interpreting linear support vector machine models with heat map molecule coloring. Journal of Cheminformatics, 3, 11. https://doi.org/10.1186/1758-2946-3-11

Samuel. (2017). Resolving the adverse impact of mobility on myoelectric pattern recognition in upper-limb multifunctional prostheses.

Samuel, O. W., Li, X., Geng, Y., Asogbon, M. G., Fang, P., Huang, Z., & Li, G. (2017).

Resolving the adverse impact of mobility on myoelectric pattern recognition in upper- limb multifunctional prostheses. Computers in Biology and Medicine, 90, 76–87. https://doi.org/10.1016/j.compbiomed.2017.09.013

Shi, wang-T. (2018). A bionic hand controlled by hand gesture recognition based on surface EMG signals_ A preliminary study.

Shi, W.-T., Lyu, Z.-J., Tang, S.-T., Chia, T.-L., & Yang, C.-Y. (2018). A bionic hand controlled by hand gesture recognition based on surface EMG signals: A preliminary study. Biocybernetics and Biomedical Engineering, 38(1), 126–135.

Timemy. (2016). Improving the Performance Against Force Variation.

Tommasi, T., Orabona, F., Castellini, C., & Caputo, B. (2013). Improving control of dexterous hand prostheses using adaptive learning. IEEE Transactions on Robotics, 29(1), 207–219.

Waris, A., Niazi, I. K., Jamil, M., Gilani, O., Englehart, K., Jensen, W., … Kamavuako, E. N. (2018). The effect of time on emg classification of hand motions in able-bodied and transradial amputees. Journal of Electromyography and Kinesiology, 40, 72–80.

Waris, A. W. (2018). The effect of time on EMG classification of hand motions in able-bodied and transradial amputees.

Webb, G. I. (2000). MultiBoosting: A Technique for Combining Boosting and Wagging.

Machine Learning, 40(2), 159–196. https://doi.org/10.1023/A:1007659514849 Young. (2013). Classification of Simultaneous Movements Using.

Young, A. J., Smith, L. H., Rouse, E. J., & Hargrove, L. J. (2013). Classification of simultaneous movements using surface EMG pattern recognition. IEEE Transactions on Biomedical Engineering, 60(5), 1250–1258.