ABSTRACT

In the field of prosthetic, a lot of research is being done to replicate the functionality of
missing parts. A bionic hand is one of the devices that can be used to replace a missing limb.
Amputees will be able to use their prosthetics in a variety of ways. A great deal of effort has
gone into developing a bionic hand that can replicate the actions of a human hand. This work
discusses the creation of real time bionic hand control based on electromyography signals
gathered from the lower elbow. Amputee who can reposition and do hand opposition.

This survey also explains the Techniques of signal processing which is needed to
classify electromyography data and determine the actions required for controlling a bionic
hand. These results show that for the feature classification of electromyography signals for
controlling of a bionic hand are K Nearest Neighbors (KNN) and Naive Bayesian Pattern
Classification (NBPC) methods were adopted. The results show that for the feature extraction
are Root Mean Square (RMS) and Integrated Absolute Value (IAV). The obtained algorithms
from the above-mentioned methods can achieve 92-94 percent of accuracy.

Electromyography is an electrodiagnostic medical techniques for assessing and

storing skeletal muscle electrical activity. Electromyography is done with the help of an
electromyograph which creates a record called an electromyogram.

To create a collection of classification algorithms, The Baye’s Theorem is used which

is called as Naive Baye’s Pattern Classifier. It’s a family of algorithm that gives a similar
idea, specifically that each pair of features to be categorised is independent of the others. The
K- nearest neighbours’ approach is a simple and best monitored machine learning technique
this may be applied to handle regression and classification tasks. It’s easy to set up and
comprehend, but it has the problem of becoming substantially slower as the amount of data
used increases.

The square root of mean square is known as the root mean square. The quadratic
mean commonly known as the RMS, is a special instance of the generalised mean with
exponent2.
TABLE OF CONTENTS
CHAPTER 4: RESULTS AND DISCUSSIONS

4.1 Summary of Results...........................................................................................20,21

4.2 Advantages...............................................................................................21

4.3 Limitations........................................................................................21

CHAPTER S: FINAL COMMENT S

5.1 Future Trends in Bionic Hands...........................................................................22,23

5.2 Conclusion........................................................................................24

REFERENCES.........................................................................2S

Plagiarism Check Report........................................................................................27—30

LIST OF ABBREVIATIONS
 LIST OF FIGURES

FIGURES PAGE NO.

1.Signal Flow Diagram to Control Bionic Hand 8

2.Hand Skeleton Model and Kinetic Model 12

3.Bionic Hand Model Design using CAD Software 13

4.Block Diagram of Signal Processing Algorithm 14

5.Dual Channel Surface EMG Sensor 15

6.Hand Opposition and Re-position EMG Signal 15

7.Samples of dataset A and B 17

8.Inclusion of Test Sample in the Database 17

9.Microcontroller Interfaced with Bionic Hand 19

10.Word Cloud of Popular Topics in Bionic Hand 19

11.RMS and IAV Extracted Features of EMG Signal 20

12.Confusion Matrix of Naive Bayesian and KNN Classification 20

LIST OF TABLES

TABLE NAME PAGE NO.

1.An Overview in the Field of Bionic Hands 9

2.Feature Extraction and Classification Accuracy for EMG Signal 20

3.Future Trend Division in the Field of Bionic Hands 22

CHAPTER 1: INTRODUCTION

1.1 Preface

The Bionic Hand is a robotic hand which imitates the human hand features such as hand
grasp and release, flexion-extension, elbow flexion-extension, supination-pronation etc.
which is integrated with the nervous system. The purpose of the Bionic hand is to improve
the quality of life for those who last hand for various reasons such as accident, current shock,
paralysis etc., hence there is need for automated bionic hand controlled by Electromyography
(EMG) signal which can able to perform actions like human hand. Recent developments
show that EMG signals are collected from amputee muscles are used to control the bionic
hand by identifying the particular actions.

The Bionic hand is an important extremity that facilitates many of the human hand tasks
and movement with nervous system. The main objective behind developing a Bionic hand, is
to create a prosthetic with multiple DOFs and have an end effecter with capability to change
into different arrangements. This paper presents development of bionic hand provided two
DOF such as hand grasp and release action. This paper also presents control of bionic hand
using EMG signals collected from lower elbow amputee. The main strength of this research
is to control the bionic hand in real time so that amputee can perform required action. The
EMG signals from lower elbow amputee are recorded with the help of EMG data acquisition
(DAQ) system. The DAQ system internally consists of amplifier, Band pass filter and Digital
and Analog converter.

Signal processing algorithms such has feature extraction and classification was carried out
to identify the desired action. Different feature extraction and classification algorithms are
tested on collected EMG signal in order to control Bionic hand accurately for particular
haptic action. Out of all tested algorithms Root Mean Square (RMS) and Integrated Absolute
Value (IAV) are selected for feature extraction. k-Nearest Neighbor (KNN) and Naive
Bayesian Pattern Classification (NBPC) are selected for feature classification. The classified
EMG signal used to control the bionic hand for desired haptic actions. Fig. 1 illustrates the
signal flow diagram of bionic hand control using EMG signal collected from amputee. The
subsequent sections give the detail explanation of the design and implementation process
involved in development of EMG signal based bionic hand.
 Commercially available body-powered and externally powered prostheses have been
helping amputees restore the functionalities of their natural hand. The present article mainly
discusses concepts from a bionic hand’s perspective. The critical steps of the bionic hand
include 1) Bio-signal Acquisition, 2) Control Architecture, and 3) Feedback Interface.

1.1.1 Bio-signal Acquisition

The interface for bio-signal acquisition and processing includes 1) signal acquisition,
2) filtering, 3) feature extraction, and 4) classification/regression. “Signal acquisition”
involves acquiring the bio-signals from the users through invasive or non-invasive
techniques. The hardware in an acquisition system includes electrodes/sensors, filters,
amplifiers, analog-to- digital converters, digital signal processors, and the software mainly
includes microcontroller compiler, waveform visualization tool. “Filtering” involves pre-
processing the bio-signals. “Feature extraction” includes extracting useful features from the
filtered signals. Features extracted include features of time-domain, time-serial, frequency or
spectral domain, and time- frequency domain. A dimensionality reduction step of the feature
sets is often necessary to increase classification accuracy. “Classification/regression” process
includes decoding the extracted features for grasp recognition that the user desires to perform.
Classification algorithm assigns a discrete label (e.g., hook grasp) whereas, a regression
algorithm converts the extracted features into continuous values (e.g., joint velocities).

1.1.2 Control Architecture

After the features are extracted and classified, they are fed into a high-level controller
that decides the desired grasp and the associated force/torque that the prosthetic must apply.
Allow level controller confirms that the desired velocity or torque is correctly applied. In a
regression-based prosthesis, the control architecture does not generally contain the high-level
controller. Classification based prostheses are usually termed “pattern recognition-based” or
“sequential on–off-based” prostheses, and regression-based prostheses are coined
“simultaneous proportional control” or “continuous decoding” prostheses. Regression-based
prostheses have been a captivating research area as they provide independent simultaneous
proportional control of the multi-DOF prostheses. At the low level, a feedback-based
controller helps in generating optimal grasp force and manipulation. Bionic reflex, which is
feedback reflecting emulation of human grasp reflex in prosthetic hand and generally
involves preventing slip and deformation of the grasped object, is essential in a prosthesis.
 1.1.3 Interface for Sensory Feedback

Providing sensory feedback with the aid of tactile sensors is an intuitive way to
control hand prosthesis. Closed loop sensory feedback can be achieved in two ways: 1) by
providing feedback to the controller and 2) by providing invasive/non-invasive feedback to
stimulate the PNS/CNS. In the former case, the sensory loop remains inside the prostheses
and may not include the user. A few commercial prostheses provide sensory feedback, e.g.,
Steeper Bebionic v3 generates audible beeps and vibrates when the grip changes, and Vincent
hand Evolution 2 provides vibrotactile feedback. The various invasive and non-invasive
sensory feedback is summarized in Table I. The idea of bidirectional control of prosthesis
using the PNS is advantageous over other bidirectional communication because it provides
the best compromise between invasiveness and natural embodiment. Visual feedback alone
suffices to correctly and successfully control a prosthetic hand, but added sensory feedback
might give the user a feeling of having a hand again. This article is a state-of-the-art survey of
the progress in bionic hands. The survey is done over a scientific database of publications in
the field of hand prostheses for the last 20 years (1997–2017). We have used the bibliometric
approach of information science, which is defined as “the use of mathematical and statistical
methods to study documents and patterns of publication”. Many interesting reviews and
surveys exist on prosthetic hands. The present survey complements the already present ones
and discerns the research field from the view of a bibliometric study through its information-
theoretic analysis and network-based techniques. To get an overview of research in prosthetic
hand, we feel a novice researcher would be interested in the major research topics, leading
journals and conferences, research subdisciplines, and future trends in hand prostheses.

Figure 1: Signal Flow Diagram to Control Bionic Hand

Table 1: Overview of Different Bio-Signal Acquisition Techniques and Sensory Feedback
Interfaces in The Field of Bionic Hands.
 1.2 A Brief History and Background

The first prosthetic hand application dates back to the early Second Punic War (218–
201 BC) by a Roman general, Marcus Sergius, who, after losing his right arm in the war, had
a prosthesis made of iron. In the early 16th century Goetz Von Berlichingen, a German
mercenary, wore two artificial hands. The first one was a simple device that consisted of a
glove with a thumb and fingers attached to it. The fingers could be brought inward, allowing
him to grip his sword. With the other upgraded version of the hand capable of independent
movement of phalanges and opposition of thumb, the knight could hold his horse’s rein and
pick up a quill. In the mid-16th century came the spring-loaded prostheses developed by the
French military surgeon, Ambroise Pare, in which the fingers could independently operate
with levers and gears. Pare also developed above elbow prosthetic hand, where a lever and
sprocket mechanism controlled the flexion and extension.

Automatic body-powered prosthesis started in the 18th century by Peter Baliff, where the
prosthesis was actuated by the intact muscles of the shoulder and trunk using leather straps as
a transmission mechanism. Another sophisticated mechanical prosthesis by William T.
Carnes was developed and patented in 1911. The above prostheses were controlled by the
overall movements of the healthy limb or chest. In the early 19th century, a breakthrough in
prosthetic hand technology came with the “Cineplastic implant” of Dr. Giuliano Vanghetti,
where the movement of the residual arm could operate the prosthesis. Later, in 1915, his
work was implemented by German surgeon Ferdinand Sauerbruch, where the prosthetic hand
was operated by the muscles of the residual forearm in which ivory pins were surgically
inserted. In the mid-19th century, Bowden cable-based prosthesis was developed, which
replaced the traditional bulky strapped ones with light and durable cables.

In 1948, Reinhold Reiter developed the first externally powered myoelectric prosthetic
hand, which was actuated by electric motors triggered by the voltage potentials obtained from
EMG signals. Then in 1960, Alexander Kobrinski developed the first clinically successful
electric prosthetic hand. According to a review by Biddiss and Chau, the rejection rate is 16–
66% for body-powered arms and 0–75% for electric arms. There is a trade-off between the
use of myoelectric and body-powered ones. Though myoelectric hands could fulfill the
requirements of cosmesis and performs well during activities of daily living tasks under lab
environments, yet it fails to perform robustly in sweat generating conditions during physically
demanding work and remain costly compared to body-powered ones.
 Myoelectric arms, as advanced as they may appear to many, still pose a considerable
functional obstacle in everyday use. Intramuscular EMG, i.e., IMES has been proposed as an
innovative solution for EMG recording. IMES are electrodes that detect and transmit EMG
signals to bionic hands via an electromagnetic coil in the prosthetic socket, providing
intuitive and stable control. With the advancement of surgical techniques, Osseo integrated
prostheses were developed in the mid-19th century by Swedish surgeon Per-Ingvar
Branemark. Then his son, Rickard Branemark, applied it to extremity amputees. Ortiz-
Catalan et al. reported the first-in-human trial of an Osseo integrated upper arm prosthesis,
which attached directly to the bone, nerves, and muscles of the intact limb.

The Osseo integrated prosthesis provided finer motor control and a greater range of
motion compared to its surface myoelectric socket-based counterpart. However, the study
was restricted to only one person, and hence, quantification of improvement is not possible,
preventing generalization of the results. Further, osseointegration has many drawbacks, such
as multiple soft-tissue infections, chronic discharge, and osteomyelitis. Moreover, repeated
fractures are a frequent problem when using osseointegration in very thin bones such as
forearm bones, and metal implants also succumb to fractures. A breakthrough in upper limb
prostheses came through TMR in 2004. With upper limb prostheses based on TMR, amputees
were capable of improved intuitive control as they provide physiologically suitable EMG
signals. Further, a technique called TSR aided in receiving improved somatotopic and
modality matched feedback. Integration of osseointegration and IMES with targeted
innervation proved to provide more intuitiveness eliminating drawbacks of traditional TMR,
which included structural shortfalls due to straps and sockets, limiting the active muscle area
that can be utilized. Another breakthrough came with the development of vascular composite
allograft, which involved transplantation of different tissues as a functional unit of an organ
of the human body. Hand transplant was first attempted by Dr. Robert Gilbert from Ecuador
in 1964. However, the hand transplantation approach may lead to acute organ rejection and
other medical complications. The research field in hand prostheses is still active. Even the
most advanced commercially available prostheses contribute ineffectively to overcome the
limitations of hand amputation and lead to amputees rejecting the prostheses mainly due to
reduced control, functionality, and feedback. Ongoing research includes mitigating issues
related to reduced efficiency of the prosthesis, mainly in the field of efficient underactuated
design, intuitive control interface, multichannel bio-signal decomposition and multimodal
sensor fusion, real-time control, force and somatosensory feedback for bionic reflex, etc.
CHAPTER 2: HAND MODELS

2.1 Skeleton and Kinematic Model of Human Hand

Figure 2: Hand model: Hand skeleton observed from the palmar side (left panel) and
kinematic model (right panel).

The human hand is one of the principals means of interacting with the environment
around us. The hand in its present form is a result of millions of years of evolution. It can
perform a complex repertoire of intricate movements to accomplish a wide variety of tasks. It
is a significant disruption of normal daily living to lose one and affects a person both
physiologically and psychologically. Prosthetic hands have played a significant role by
augmenting amputees who have lost their arm due to disease or disability. Despite the
progress in prosthetic hand technologies, only 50–60% of amputees wear one. The rejection
rate is also estimated to be as high as 40%. The development of bionic hands has not yet
percolated to amputees, mainly due to the high price associated with them. On the other hand,
upper limb loss leads to many patients unable to return to their previous occupation or work
in general [5]. In view of the above, restoration of hand functionalities could lead to
socioeconomic impact. Therefore, research in the area of bionic hand assumes importance.
 The human hand is a sophisticated mechanism. Bionic hands are nearly similar to natural
hand in form and structure; however, replacing the natural bidirectional communication
between the nervous system and the human hand is undoable. Enabling a prosthetic hand to
replicate all the functions of a human hand is impossible. The hand consists of 27 bones:
eight in the wrist, five in the palm, and 14 finger phalanges. The joints of the human hand
help achieve 27 degrees of freedom (DOF) (see Figure 2). There are 34 extrinsic (providing
strength) and intrinsic muscles (providing precise control of fingers), which help in actuation
after receiving signals from the median, radial, and ulnar nerves. These nerves receive signals
from the central nervous system (CNS) for actuation and send somatosensory feedback
received via around 17,000 mechanoreceptors of the peripheral nervous system (PNS).

2.2 Computer Aided Design of Bionic Hand

Figure 3: Bionic Hand Model Design using CAD Software

Using CAD Software, simulation model of bionic hand is designed shown in Fig. 2. The
drawing of the hand is carried out from the parts which are the building blocks of every
model. Each part of the arm, the fingers, palm, forearm and upper arm are designed by
considering the dimensions in millimetre (mm) of a normal human hand. ABS plastic
material used to develop mechanical model of palm, wood and rubber are used to develop
fingers. To control individual fingers of Bionic hand for particular classified action, high
torque servo motor is used for each finger. Each servo motor programmed with suitable
decision logic by 8-bit micro controller.
CHAPTER 3: BIO-SIGNAL ACQUISITION

3.1 Signal Processing Algorithm

Figure 4: The Block Diagram of Signal Processing Algorithm.

The different stages of signal processing are shown in Figure 4. The stage involves signal
acquisition, pre-processing, feature extraction and detection algorithms. The recorded EMG
signals are in millivolts and contain power line interference, hence filters and amplifiers are
used in pre-processing stage to eliminate noise. RMS and IAV are used for feature extraction.
For classification of EMG signal KNN or NBPC are used to identify particular haptic action.

3.1.1 EMG Data/Signal Acquisition

Modern days bionic hand acquires one or multiple bio-signals from the human body
to perform desired haptic actions. The acquired bio-signals are then subjected to feature
extraction and classification techniques in order to identify the haptic action that the amputee
wants to perform with good accuracy. In this paper, non-invasive based EMG signals are
used to control Bionic hand. Two kind of EMG signal actions such as Hand grasp
(Opposition) and release (Re-position) are collected to control Bionic hand.

EMG signals general occur during contraction and relaxation of the muscles. Grasp
action occur when muscles contraction and release action occur when muscles relaxation in
the forehand. The Extensor Digiti Minimi muscles are used to control the thumb and the
Extensor Digitorum and Extensor Indicis muscles are used to control the fingers in hand.
 The EMG signals are collected by placing dual channel surface EMG electrodes on
double amputee lower elbow patient as shown in Fig.5. The EMG signals occurs at a
frequency range of 20–500 Hz with a power line interference at 50 Hz. The Hand opposition
and re- position EMG recorded from amputee shown in Fig.6.

Figure 5: Dual Channel Surface EMG Sensor Placed on Lower Elbow Amputee

Figure 6: Hand Opposition and Re-position EMG signal collected from Amputee

3.2 Techniques for Feature Extraction and Classification of EMG Signal

3.2.1 Feature Extraction

Presence of large bio signals, use of only raw myoelectric signal in control system is
impractical. The success of feature classification system depends on features extraction used
to represent the raw EMG signals. Using of entire raw EMGs signal as input data for feature
classification step is not practical because of huge amount of data for each hand action will
leads to long time of processing. So, different types of feature extractions algorithms have
been used to reduce the dimension of raw EMG signal to find feature vectors.
 The feature vector must contain enough information of signals for fast training and
classification of hand grasp and release data. Feature extraction is one of the most important
attributes of the EMG signal processing which used to achieve optimal feature for EMG
analysis and control. The accessibility and computational simplicity make time domain
method the most popular feature extraction technique for generating desired haptic action for
the myoelectric control system because they do not need a transformation. A brief
explanation of the mathematics involved in time domain feature extraction such as RMS and
IAV methods.

3.2.1.1 Root Means Square (RMS)

RMS is one of the popular feature extraction method used to analyze EMG signal.
RMS is square of all the numbers in given set and then find the arithmetic mean over the time
of a squares which is related to constant force and contraction of the muscle. The equation is
given as:
1
X = √ (x2 + x2 + ⋯ + x2 ) (1)
rms 1 2 n
n

Where 'n' is the total number of samples.

3.2.1.2 Integrated Absolute Value (IAV)

The IAV is, calculated as the summation of the absolute values of the EMG signal
amplitude. Since the pre-processing method rectifies the captured signal before calculating
the mean, it is essentially to employ IAV to determine baseline parameter. The equation is
given as:

X̅i = (∑Ni=1 |xi |) (2)

Where xi is the ith sample in EMG signal and N is the number of samples of a signal.

3.2.2 Signal Classification Algorithm

The myoelectric signals are acquired by placing EMG sensor on selected muscles in
hand. The collected signal contains EMG data for required haptic actions. Signal
representation is achieved by using feature extraction algorithms. The classifier’s role is to
use extracted feature and generate distinctive classes corresponding to the desired action.
From the literature it shows that classification of hand actions is done by sensing the
activities of muscles.

However, the classification accuracy limited to 70-90 percent due to practical

applications of myoelectric control. There are several possible classification techniques.
Among them, the most used are KNN and NBPC are used are discussed below section. This
algorithm has classification rate around 92-94%.

3.2.2.1 Naive Bayesian Pattern Classification (NBPC)

NBPC algorithm classifies the test data based on their posterior probability. Consider
an example of 2 datasets A and B shown in Figure 7. The prior probability of each dataset is
calculated using Eq.3 and Eq.4 respectively.
Number of samples in A
Prior Probability of A α
Total number of samples (3)
Number of samples in B
Prior probability of B α
Total number of samples (4)

Figure 7: Samples of dataset A (Red) and dataset B (Green)

Consider the inclusion of a test sample in the present database with 2 datasets as shown
in Figure 8 whose class has to be found.

Figure 8: Inclusion of Test Sample in the Database

Now, the likelihood of a sample X to belong to dataset A and B is found using Eq. 5
and 6 respectively.

Number of samples A in the given X

Likelihood ratio of X given A α Total number of A cases (5)

Number of samples B in the given X

Likelihood ratio of X given B α Total number of B cases (6)
 The class to which the sample X belongs to is decided by the posterior probability.
The posterior probability of X, being A and being B are found using Eq.7 and Eq.8
respectively. The class of X is decided based upon the largest posterior probability value.

Posterior probability of X being A α (prior probability of A × likelihood of X given A) (7)

Posterior probability of X being B α (prior probability of B × likelihood of X given B) (8)

3.2.2.2 K-Nearest Neighbor Algorithm (KNN)

KNN is one of the most popular and widely used classification algorithms. It is a
simple yet effective classification algorithm which is main reason that it is widely used.
When there is not much information available regarding the distribution of data to be
classified, KNN is one of the first choices for classification. In this classification algorithm
the pattern present in the dataset defines a region called as Voronoi region. Class is assigned
to each pattern based on the Euclidean distance calculated using Eq.9 between the different
patterns.

dist(x1, x2) = √∑n (x1i − x2i)2 (9)

i=1

Where, x1 = point 1 in n-dimensional space, x2 = point 2 in n-dimensional space,

i = attribute number, n = Total number of attributes, x1i = ith attribute of x1 and x2i = ith
attribute of x2. The number of neighbors can be increased in order to increase the
classification accuracy. When a test data is provided to classification, the Euclidean distance
of the data points present in the test data is calculated. The class is determined based on the
closest Euclidean distance to the patterns present in the training data for the given test data.
The classification results are obtained quickly for small database whereas it consumes more
time as the size of the database increases.

3.3 Bionic Hand Interface with Controller

EMG signals are collected from the amputee by placing electrodes on the respective
muscles. The collected EMG signals are amplified using instrumentation amplifier. The
amplified EMG signal is then converted to digital signal using ADC converter. The digitally
converted EMG signal cannot be directly fed as input to control prosthetic hand therefore it
has to be processed through various signal processing algorithms. Signal processing consists
of pre-processing, feature extraction, and feature detection algorithms. The extraction of
EMG signal feature is carried through RMS and IAV techniques. For feature detection, KNN
(or) NBPC method are used.
 The complete signal processing algorithms are implemented in embedded platform
hence, the control of bionic hand is done with the low response time between grasp and
release functions. Figure 9 shows the Bionic hand interfaced to micro-controller and amputee.

Figure 9: Microcontroller Interfaced with Bionic Hand and EMG sensor

Mounted on Amputee

3.4 Information Analysis Methodology

For presenting a state-of-the-art survey of the research in hand prostheses, we chose

the information analysis approach, which comprises mainly of collecting, analyzing,
cleansing, visualizing, and interpreting large datasets. We utilized a combination of
visualization techniques, including word cloud, bar graphs, pie charts, and network graphs to
discern information. The methodology closely follows the one outlined for surveys on
biomimetics and assistive technology for visually impaired.

Figure 10: Word cloud of the most popular topics in bionic hand research extracted from the
titles of the publication.
CHAPTER 4: RESULTS AND DISCUSSIONS
4.1 Summary of Results

The developed prototype of bionic hand model is tested for lower elbow amputee. The
Ag-AgCl electrodes were placed on to the associated muscles with proper inter-electrode
distance. The instrumentation amplifier is connected to lead electrodes to acquire and amplify
the EMG Signal. Several methods of feature extraction and feature classification algorithms
are tested to get best classified data. Out of this RMS and IAV for feature extraction selected.
For feature classification KNN and NBPC are selected which give about 92-94% of accuracy
achieving for two haptic actions such as hand grasp and hand release for lower elbow
amputee and is tabulated as in Table 2.

Table 2: Feature Extraction and Classification Accuracy Table for collected EMG Signal.

Feature
Classification
NBPC KNN
Feature RMS 92.94% 91.01%
Extraction
IAV 94.84% 91.56%

Figure 11: (a) RMS Extracted Features of EMG signal and (b) IAV Extracted Features of
EMG Signal.

Figure 12: Confusion Matrix of (a) Naive Bayesian and (b) KNN Classification.
 The feature extracted EMG signal shown in Figure 11 (a)&(b) which is obtained from
both the RMS and IAV methods. As observed, the data size of raw EMG signal is reduced by
applying feature extracted methods and precision of EMG signal is increased. Feature
extracted EMG signal, the amplitude of hand opposition is higher than the re-position so, it
helps to program a suitable logic on microcontroller to classify the EMG signal and to control
the bionic hand.

The classification of EMG signal in Figure 12 (a)&(b) illustrates the confusion matrix
obtained by NBPC and KNN feature classification respectively. Classification accuracy of
EMG signal is obtained by using confusion matrix. The confusion matrix clearly indicates
that all the correct predicted classes are located in diagonal of the matrix hence, it's easy to
find the accuracy and performance of the classification algorithm.

DAQ collects the EMG signals continuously and monitored by micro-controller. If the
amputee opens or closes hand, controller compares this signal with signal library stored in
microcontroller and provide decision logic to particular motor placed in bionic hand.
Complete operation such as acquiring the signal, making decision command and controlling
the bionic hand take place within 10 Sec of duration. Nonetheless the developed technique
has been demonstrated for only two DOFs such as hand opposition and hand re-position.

4.2 Advantages

 Simple and less costly.

 Common physiological measurement techniques.
 Allows enough feature extraction to control electronics.
 High signal to noise ratio.

4.3 Limitations

 Muscular crosstalk.
 Limb-position effect.
 Temporal effect.
 Nonstationary and signal inconsistency.
CHAPTER 5: FINAL COMMENTS

5.1 Future Trends in Bionic Hands

Table 3: Future Trend Divisions in the Field of Bionic Hand

Bio-signal Acquisition with the help of adaptive signal processing must be able to
take care of disturbances like electrode displacement, artifact’s interference, diverse physical
factors, etc. Improved algorithms of feature selection and nonlinear classifiers for
simultaneous motion detection signify essential research fields, which would relinquish
offline training. Moreover, multimodal sensory fusion shows significant improvement
compared to prostheses that are solely dependent on an individual bio-signal acquisition.
Multimodal sensor fusion signifies the synergistic fusion of sensor data obtained from
multiple sensors to attain inferences with more accurate perception, compared to unimodal
sensors, which suffers from information deficiency and noises. Transferring these signals to
the actuators can help achieve advanced humanlike complex tasks. The latest research trends
acknowledge this by fusing multimodal sensor data coming from muscles, brain, neural
interface, cameras, inertial sensors, force sensors, position encoders, etc., some of which are
briefly discussed in Table 3. A detailed survey of multimodal sensing fusion technologies in
wearable rehabilitation robotics is presented in the literature.

Efficient Design of the prosthetic hand by comprehension of the human hand

biomechanics is fundamental as mimicking the anthropomorphism induces user acceptance.
However, emulating the human hand in restoring all its DOFs by increasing the number of
actuators adds design constraints, eventually prompting a nonanthropomorphic design.
Though postural synergy-driven prostheses can have more DOF with simple myocontrol, yet
amputees have faced difficulty in modulating the grasp force, and research in this area needs
further work. Actuation of force though widely achieved by nonbackdrivable mechanisms
with electric motors, recent focus has started on actuators based on shape memory alloy,
pneumatic air
muscles, electroactive polymers, etc. For transmission of the forces, TDM is likely to gain
immense significance. Recent research shows that actuation based on branched TDM
provides a compact and efficient transmission of grasping force and further mimics the
anatomical model. Optimization of tendon routes in robotic hands can be seen in the recent
literature, and such an approach may be taken up in designing prosthetic hand for enabling
better grasp stability. Issues addressing the number of redundant tendons, tendon breakage,
and switching controller are essential aspects of study in TDM.

Intuitive control is imperative for eradicating poor interaction between the device and
the user. Multi-articulated bionic hand with simultaneous, proportional, and independently
controlled fingers remains a basin of attraction for research. Advanced machine learning
algorithms like deep learning can be applied directly for grasp recognition in real-time control
as they address the problems of feature extraction. When it comes to grasping stability and
manipulation, adaptive grasp control has received immense attention. Bionic reflex is
implemented in many recent works of literature, which involves the prevention of slippage
and deformation of grasped objects. Recent works on grasped object stiffness detection and
slippage prevention using an adaptive controller. For more intuitiveness, haptic feedback by
sensorization of the PNS is a further direction of research that shall empower the
technological advancement of bionic hand. As the human hand is a combination of
proprioceptive and exteroceptive sensors, it is imperative to fuse multimodal sensory
feedback to make the prosthesis more alike to its biological counterpart.

Rehabilitation therapies involve training amputees to get accustomed to their

prostheses. It is found in upper limb amputees with chronic PLP that phantom motor
execution with the help of machine learning and augmented VR game training has
significantly reduced PLP. Further, research in the embodiment of prostheses with hand
sensorization also shows the alleviation of PLP and is a field of immense interest as the goal
of an embodied prosthesis is still limited. Recent literature can be found on combining novel
neuro-prosthetic techniques utilizing a combination of peripheral neurotactile stimulation and
immersive digital technology to reduce telescoping. Recently, researchers have built a
neuromorphic biomimetic computational model (TouchSim), which replicates the peripheral
nerve’s response to sensory stimulation, and the information is then used to stimulate the
nerve for restoring sensory feedback. A comparison of the different encoding strategies viz.,
amplitude, and frequency neuromodulation of stimulus pulses implemented on the nervous
system, is seen in the above research.
 5.2 Conclusion

The need for a multi-DOF bionic hand, which requires less rehabilitation, non-
invasiveness and ability to control the bionic hand in real time using EMG signal. This paper
demonstrates the functionality of a multi-DOF bionic hand which is capable of performing
two degree of freedom with the help of real time EMG data recorded from an amputee.
Collected EMG signal is classified by using developed algorithm which predicts the required
action performed by amputee.

Micro-controller is used to control the bionic hand in real time based on the predicted
action by amputee. The feature extraction algorithms such as RMS and IAV and feature
classification algorithms such as KNN (or) NBPC are implemented in microcontroller which
gives desired haptic functionality with an accuracy around 92-94%. However, the developed
bionic hand limited only to perform two DOFs. Hence, future development can include fine
tuning of the algorithm and increasing number of DOF (individual finger action) with
improved amount of accuracy.

Research integrating sensor fusion with multimodal sensing has been discussed across
recent literature. With the emergence of Internet of Things and knowledge-based artificial
intelligence aided computer vision algorithms, future prostheses may connect to cloud
technology, generating different context-based grasp types without any training. Moreover,
reinforcement learning and heuristically based search algorithms guided by force closure
grasps can contribute to grasping stability. Non-invasive sensory feedback research for
prosthetic embodiment is likely to find immense interest.

Future studies may focus on making hand prostheses more adaptable and integrable to
mitigate the cognitive burden of the user. There should be improved techniques of non-
invasive bio-signal acquisition, which would accomplish robust automatic intent detection.
Providing amputees with prosthetic embodiment should be a significant concern among the
research groups. Modern research may focus on intuitive control by improving the
bidirectional communication across the user and the prosthesis with the help of neural
stimulation. Research must focus on control laws based on robust nonlinear adaptive
controllers with disturbance observers that would provide reliable feedback. Considering the
fusion of the state-of-the-art technologies into current prostheses would immensely bring
down the semantic gap between a bionic prosthesis and a biological hand it ought to replace.
 REFERENCES

[1] Praveen L.S., Nagananda S.N., Dr. Preetham Shankapal, “Design and Development of
Real Time Bionic Hand Control Using EMG Signal”, Dept. of Electronic and
Communication Engineering, Ramaiah University of Applied Sciences, vol. 04, no. 18, pp.
978-1-5386-1112- 8, 2018.

[2] Hirakjyoti Basumatary, Shyamanta M. Hazarika, “State of the Art in Bionic Hands”,
IEEE Transactions on Human-Machine Systems, vol. 50, no. 02, April 2020.

[3] B. G. Singh, A. K. Pithawa and Ravindranath G, “Study of Disabled Treated at Bionic

Limb Centre”, Medical Journal AFI, Vol.65, pp: 232-234, 2009.

[4] Sonrin Herle and Sergiu Man, “Processing Surface EMG Signals for Myoelectric
Control”, Technical University of Cluj Napoca, Romania.

[5] Abu Bakar Yahya, Chong Shin Horng, Mohamad Fani Sulaima, and Rubita Sudirman,
“Features Extraction of Electromyography Signals in Time Domain on Biceps Brachii
Muscle”, International Journal of Modeling and Optimization, Vol. 3, No. 6, December 2013.

[6] M. Zecca, S. Micera, M. C. Carrozza, and Dario, “Control of Multifunctional Prosthetic

hands by Processing the Electromyographic Signal”, Critical Review in Biomedical
Engineering, pp.459-485, 2002.

[7] Bhuvaneswari R. and Kalaiselvi K., “Naive Bayesianian Classification Approach in

Healthcare Applications”. International Journal of Computer Science and
Telecommunications, 106-112, 2012.

[8] Mohammed Z. Al-Faiz, Abduladhem A. Ali, Abbas H. Miry, “A k-Nearest Neighbor

Based Algorithm for Human Hand Movements Recognition Using EMG Signals”, Iraq J.
Electrical and Electronic Engineering, Vol.6 No.2, 2010.

[9] Md. Faisal Kabir, Alamgir Hossain, Keshav Dahal,” Enhanced Classification Accuracy
on Naive Bayesian Data Mining Models”, International Journal of Computer Applications
(0975
– 8887), Volume 28– No.3, August 2011.

[10] M. Fatoorechi, J. Parkison, R.J. Rrancea, H. Prancea, A.K. Seth, D.J. Schwartzman, “A
Comparative Study of Electrical Potential Sensors and Ag/Agcl Electrodes for Characterising
Spontaneous and Event Related Electroencephalagram Signals”, Journal of Neuroscience
Methods, pp-7-16, 2015.
 ANNEXURE

TOPIC: THE BIONIC HAND

ABSTRACT

In the field of prosthetic, a lot of research is being done to replicate the functionality of
missing parts. A bionic hand is one of the devices that can be used to replace a missing limb.
Amputees will be able to use their prosthetics in a variety of ways. A great deal of effort has
gone into developing a bionic hand that can replicate the actions of a human hand. This work
discusses the creation of real time bionic hand control based on electromyography signals
gathered from the lower elbow. Amputee who can reposition and do hand opposition.

This survey also explains the Techniques of signal processing which is needed to
classify electromyography data and determine the actions required for controlling a bionic
hand. These results show that for the feature classification of electromyography signals for
controlling of a bionic hand are K Nearest Neighbors (KNN) and Naive Bayesian Pattern
Classification (NBPC) methods were adopted. The results show that for the feature extraction
are Root Mean Square (RMS) and Integrated Absolute Value (IAV). The obtained algorithms
from the above-mentioned methods can achieve 92-94 percent of accuracy.

Electromyography is an electrodiagnostic medical techniques for assessing and

storing skeletal muscle electrical activity. Electromyography is done with the help of an
electromyograph which creates a record called an electromyogram.

To create a collection of classification algorithms, The Baye’s Theorem is used which

is called as Naive Baye’s Pattern Classifier. It’s a family of algorithm that gives a similar
idea, specifically that each pair of features to be categorised is independent of the others. The
K- nearest neighbours’ approach is a simple and best monitored machine learning technique
this may be applied to handle regression and classification tasks. It’s easy to set up and
comprehend, but it has the problem of becoming substantially slower as the amount of data
used increases.

The square root of mean square is known as the root mean square. The quadratic
mean commonly known as the RMS, is a special instance of the generalised mean with
exponent2.