An EMG Controlled Bionic Arm with Machine

Learning

Mehmet Ilker Uygun Berk Recber Melisa Aleyna Celikli

Department of Mechatronics Department of Mechatronics Department of Mechatronics
Engineering Engineering Engineering
Istanbul Bilgi University Istanbul Bilgi University Istanbul Bilgi University
2023 7th International Symposium on Multidisciplinary Studies and Innovative Technologies (ISMSIT) | 979-8-3503-4215-4/23/$31.00 ©2023 IEEE | DOI: 10.1109/ISMSIT58785.2023.10304962

Istanbul, Turkey Istanbul, Turkey Istanbul, Turkey

[Link]@[Link] berk.recber02@[Link] [Link]@[Link]

Yesim Oniz
Department of Mechatronics
Engineering
Istanbul Bilgi University
Istanbul, Turkey
[Link]@[Link]

Abstract—This study uses electromyography (EMG) signals This study used machine learning to build a bionic arm that
and machine learning techniques to create a bionic arm picks up new skills from data and EMG signals. By attaching
specifically made for amputees. The user’s intended movements electrodes to the surface of a few selected muscles in the user's
can be decoded and converted into orders for the bionic arm by residual limb, EMG signals are first collected. These
observing and analyzing these signals. An EMG sensor was electrodes record the electrical activity that the muscles
initially placed on the surface of particular muscles in the lower generate during different movements. After the EMG signals
arm as part of the project’s data-collection phase. As the are acquired, they go through a number of preprocessing steps
muscles performed various movements, the electrodes captured to improve the data's quality. The signals may then be
the bioelectric signals they produced. These recorded
amplified and filtered to eliminate noise and unwanted
bioelectrical signals are classified according to each movement
by going through a series of processes, and it was aimed to
artifacts, ensuring that they accurately reflect the user's muscle
increase the functionality of the bionic arm by gaining many activity. The machine learning model needs to be trained using
movements control. The results showed promising the gathered EMG data. This can be accomplished by using a
advancements in the field and highlighted the potential for variety of classification algorithms, even more deep learning
improving the quality of living for people with limb loss. architectures. Establishing a relationship between the input
signals and the corresponding intended actions, the model
Keywords—Prosthetics, Bionic, Limb loss, Machine learning, learns the patterns and characteristics of the EMG signals
Electromyography (EMG), Bioelectrical signals, Mechanical arm linked to various movements. Each EMG signal is linked to a
design particular hand movement during the model's exposure to the
labeled dataset during the training phase. The model gains the
I. INTRODUCTION ability to identify signal patterning and link it to movement
There have been significant improvements in the control correlation. By using the newly arriving EMG signals, the
and design of prosthetic devices over the past few years. model is able to precisely predict the intended movement.
Researchers have devoted their research efforts to creating After training, the machine learning model is incorporated into
novel replacements for lost limbs that can be used to perform the bionic arm's control system. The model can continuously
tasks that were previously only possible with a biological arm, analyze and interpret the user's muscle activity thanks to real-
both in academic and industrial settings. These developments time processing of the EMG signals. The bionic arm's
have placed a lot of emphasis on being able to replicate actuators receive commands from the model that are used to
complex hand movements, including independent finger move the hand. The bionic arm develops the ability to
control. replicate a biological hand's natural movements by learning
from the data and EMG signals. By simply contracting certain
Electromyography (EMG) signals are a critical component muscles, the user can carry out a variety of tasks, and the
in the design of a bionic arm that offers the user intuitive and bionic arm translates those contractions into the
natural control. These bioelectric signals, which the muscles corresponding hand movements [2-3].
generate during contraction, can be captured and analyzed in
order to decode the user's intended movements. Traditionally, Compared to traditional control techniques, this strategy
mechanical switches or manually activated sensors have been has a number of benefits. It offers a more intuitive and organic
used to control prosthetic devices. These techniques, though, control experience, enabling the user to move the bionic arm
can be time-consuming and sometimes don't give the user a very much like they would move their own hand. The use of
comfortable and natural feeling of control. Using EMG EMG signals also enables more seamless integration between
signals, the user's own muscle activity can be captured and the user and the prosthetic device, improving their sense of
transformed in accordance with the instructions of the embodiment and lowering the cognitive effort needed for
designed prosthetic device [1]. control.

979-8-3503-4215-4/23/$31.00 ©2023 IEEE

Authorized licensed use limited to: National Institute of Technology - Agartala. Downloaded on August 09,2024 at [Link] UTC from IEEE Xplore. Restrictions apply.
 There are many studies in the literature in the field of section, the results are mentioned. The last section contains
motion control of prostheses with bioelectric signals and, the concluding remarks and future work.
prosthetic arms often have a very limited number of degrees
of freedom, largely due to the limitations of myoelectric II. METHODS
technology, but of course some designs using this technology The development of a bionic arm using electromyography
are much closer to the human hand than others. (EMG) signals and machine learning for intuitive and natural
In [4], a portable robotic hand controlled by EMG signals control is the goal of this project, which makes use of a variety
through an ANN classifier has been developed. The aim of of techniques. Data acquisition, amplification, machine
the study was to develop a portable robotic hand for learning, and integration are the different categories that apply
individuals with physical disabilities, enabling them to to the methods as in Fig. 1.
perform basic hand movements. After using surface
electromyography signals from the muscles in the human
forearm to determine the subject's intentions, the researchers
selected six different types of gestures for discussion. ANN
is trained and used to distinguish the desired motion based on
the signals received from the sensor. The outcome was a basic
robotic hand with seven degrees of freedom, accompanied by
hardware circuits for signal collection, power management,
and a microprocessor, all without the need for a wired
connection to the computer. This design makes the hand
small and practical for use.
The objective of [5] was to create an affordable robotic
prosthesis and manage it using the Arduino platform. The
development of the robotic arm's control has been made
possible using EMG/ECG electromyographic surface sensors.
The system tried to simulate the natural movements of the arm
and hand.
Fig. 1: Block diagram of process.
The study [6] presents a highly integrated, low-powered
prosthetic hand equipped with a feedback system to perform
human-machine intervention. The prosthetic hand consists of
five fingers. The hand is driven by six DC motors, one for each A. Data Acquisition
finger plus one for thumb opposition. While the motors are in In order to create a bionic arm that is controlled by
the palm, the sensors are spread over the entire structure. The electromyography (EMG) signals, data acquisition is an
integrated control system consists of the motion control essential step. It entails observing and recording the
subsystem and the sensory subsystem. bioelectric signals produced by the user's muscles during
various movements. Electrode placement and signal recording
Dextra is an ongoing printable human-sized robotic hand
are both parts of the data acquisition process [8].
project to develop an open-source and affordable robotic hand
prosthesis [7]. Adaptive grip, compact size, mechanical • Electrode Placement
simplicity, and simplicity of replication are the main design
elements of Dextra. The finger module is the fundamental To achieve dependable signal acquisition, it is crucial to
component of Dextra. The hand is modular in that the four make sure that the electrodes are positioned correctly [9-
fingers are interchangeable, and the thumb is a finger module 10].Electrode placement has been carried out with regard to
variation. The printable mechanical finger, the actuator, and the targeted muscles involved in controlling hand movements.
an encoder are all part of the finger module. The small actuator To accurately record EMG signals, two electrodes are
rotates a spool that winds a fishing line with a DC micro positioned on the surface of the user's residual limb wrist and
gearmotor, turning the motor's rotational action into a linear the other two electrodes placed on flexor capri radialis, i.e,
motion.
The current study aims to develop an affordable, light-
weight prosthetic arm specifically made for amputees. The
main advantage of the proposed design over the previous ones
seen in literature is that the developed bionic arm with 15
degrees-of-freedom (DOF) can perform most of the
movements that a human arm can do. Three electrodes have
been used per EMG sensor, and in total two EMG sensors
have been utilized. Python programming language has been
used in data visualization, signal processing, and in the
generation of the control signals for the actuators.
The paper is organized as follows: The next section
includes the methods used. These methods are discussed
under 3 subheadings: Data acquisition, machine learning and
mechanical design of the bionic arm respectively. In the third

Authorized licensed use limited to: National Institute of Technology - Agartala. Downloaded on August 09,2024 at [Link] UTC from IEEE Xplore. Restrictions apply.
 Fig. 4: Amplified signals from ring finger when the arm is
in fatigue state. (50 times in a set)

The signals recorded for each different movement of the

arm has a unique character, i.e. different movements result in
different signal patterns. To be able to distinguish the
corresponding movement from the recorded data, the signals
recorded from EMG sensors must first be displayed.
Matplotlib, seaborn and plotly from the Python library were
used to display these signals as a plot and heatmap. There are
Fig 2: 7 main gestures used when gathering data. approximately 50 digital readings in the signal formed from a
flexion or extension movement as the sets are formed. To
surface of the forearm. Remaining electrodes have been make this move unique, 20 digital readings are selected as
placed on the elbow for establishing a ground signals required shown in Fig. 5. Each of these sets includes 50 movements, a
in measurements. Usually, straps or adhesive patches are used total of 12 sets. This indicates that 600 extension and flexion
to affix the electrodes to the skin. movements are performed on only one finger.
• Signal Recording
As shown in Fig. 2, a total of 7 main movements were
selected at the beginning and their signals were amplified.
Once the electrodes are properly placed, the EMG signals
generated by the user's muscles are recorded. These signals
represent the electrical activity produced by the muscles
during contraction [11].
The recorded EMG signals are typically sampled at a high
frequency to capture the fast changes in muscle activity during
movements. A sampling rate of 77 Hz is chosen to ensure an
accurate representation of the signals without introducing
excessive data volume. Therefore, the selected movements
were recorded not only from the arm at rest, but also taking
into account the noisy signals of the arm in the fatigued state.
The resulting signals of the arm in two different states are Fig. 5: Amplified ring finger signal when flexing.
shown in Fig. 3 and Fig. 4. B. Machine Learning
Before training the machine learning model, the recorded
EMG signals undergo preprocessing using the Python
programming language to enhance their quality and facilitate
effective analysis. Preprocessing steps include hz filtration,
high-pass, and low-pass filtering, and rectification to remove
the unwanted noise in signals.
Hz filtration, also referred to as frequency filtering, is used
to improve the recorded signals' quality by eliminating
unwanted noise and artifacts that could obstruct the accurate
interpretation and analysis of the EMG data. To see all the sets
together, the frequency value is fixed to 77 Hz as in Fig. 6.
Fig. 3: Amplified signals from ring finger when the arm
moved from the rest. (50 times in a set)

Authorized licensed use limited to: National Institute of Technology - Agartala. Downloaded on August 09,2024 at [Link] UTC from IEEE Xplore. Restrictions apply.
 Fig. 6: Resulting signal after frequency filtering.

Fig. 10: Feature extraction for ring finger.

C. Mechanical Design of The Bionic Arm
Fig. 7: Resulting signal of high-pass filter with a cut-off
frequency of 20 Hz. The main goal of this study was to create a bionic arm
that will give amputees the illusion of having arms. The
bionic arm must enable the user to perform a broad range of
movements. This is crucial so that various daily tasks can be
carried out. In the design, ergonomics and comfort should be
taken into account. Furthermore, the materials and
components should be selected such even after extended use,
the user can easily control the bionic arm and is at ease. The
bionic arm must also have a sturdy construction and be simple
to maintain and repair [12].
Realistic constraints were used to establish the design
Fig. 8: Resulting signal of low-pass filter with a cut-off
criteria. The bionic arm is intended to function with roughly
frequency of 10 Hz.
the same strength and power as a real arm. For the bionic arm
to be used for extended periods of time and to support daily
activities, energy efficiency is also crucial. The usage time is
also significantly influenced by battery life. To meet all these
requirements, a number of different designs have been
considered. Then each design has been evaluated with regard
to the given specifications, and the final design has been
generated.
SolidWorks has been use for the design of the bionic
arm. The prosthetic arm, which weighs approximately 814
Fig. 9: Rectification on ring finger signals.
grams and has dimensions of 460 mm x 270 mm for whole
After all sets are combined, a high pass filter is applied to arm and 210 mm x 95 mm for hand, was obtained with a 3D
each set to remove frequencies lower than 20 Hz. Resulting printer using PLA filament, and steel bearings and pins were
signal is presented in Fig. 7. Fig. 8. shows the signal of the used in fingers, wrists and areas that require many rotations
next phase, which involves the application of a low pass filter to ensure ease of movement. To increase the range of motion,
with a cut-off frequency of 10 Hz to all of them to remove 15 servo motors and two Li-Ion rechargeable batteries of two
higher noise. Then lastly rectification is applied to get rid of cells were used to enable these motors to perform the
the negative parts in the signal. Fig.9 presents the resulting necessary tasks during the day. The video of the bionic arm
signal. in motion is available at [Link]
After obtaining the preprocessed EMG signals, pertinent
features are extracted to represent the signals' fundamental
properties. In order to distinguish between various hand
movements, feature extraction attempts to capture
discriminatory information. Features captured for 600
different movements of ring finger have been presented in Fig.
10.

Authorized licensed use limited to: National Institute of Technology - Agartala. Downloaded on August 09,2024 at [Link] UTC from IEEE Xplore. Restrictions apply.
 Predicted
0 1 2 3 4
0 464 12 5 9 0
1 26 382 20 42 0

Actual
2 33 40 367 22 0
3 10 32 39 392 0
4 0 0 0 0 505

The performance of the created bionic arm was evaluated

against previously published research. The outcomes showed
significant improvements in dexterity, functionality, and
motion control. Users were able to complete a variety of tasks
precisely and accurately, closely mimicking the abilities of a
biological hand.
IV. CONCLUSION
This study used electromyography (EMG) signals and
Fig. 11: The developed bionic arm. machine learning to create a specialized bionic arm for
amputees. The study effectively dealt with the difficulties of
controlling bionic arms by combining machine learning
algorithms with an EMG sensor. The EMG sensor acquired
III. RESULTS bioelectric signals from the targeted lower arm muscles,
Real-time processing of the classified signals was made which were amplified, filtered, and categorized for precise
possible by the incorporation of the trained machine learning interpretation. In order to predict the intended movements of
model into the bionic arm's control system. This made it easier the bionic arm, a machine-learning model was trained using
to create the corresponding commands needed to operate the the classified signals, and it was able to do so with an accuracy
bionic arm. The bionic arm's mechanical layout was adjusted rate of about 89 to 90. Particular attention was paid to, and
to accommodate the few movements derived from the repeated practice was given to finger movements to increase
classified signals, producing the best possible functionality accuracy. The machine learning model was incorporated into
and usability. the bionic arm’s control system, enabling real-time processing
of the EMG signals. This made it possible for the bionic arm
GridSearchCV tests the parameters of all models to be controlled precisely and naturally, enabling users to
according to the range specified for each parameter and carry out various tasks with accuracy. To maximize usability
reports the parameters that give the best results. As a result, and functionality, the mechanical design of the bionic arm was
models were established according to the parameters given by modified to accommodate the movements that could be
GridSearchCV. After training the models, the predictions of deduced from the classified signals.
the models were taken, and then, by using the performance
evaluation metrics F1 score, recall, and precision, which are The work’s results showed how well EMG signals and
commonly used in Python, the accuracy of the models was machine learning could be combined for bionic arm control.
calculated, and the predicted values were compared with the In comparison to earlier literature, the bionic arm’s motion
real values, and 90% results were obtained for XGBClassifier, control and functionality were greatly enhanced. Overall, this
89% for the RFclassifier. Table 1 and Table 2 give the work offered a thorough method for improving the
confusion matrices obtained with RFClassifier and performance of prosthetic arms using machine learning and
XGBClassifier, respectively. EMG signals. The accurate interpretation of bioelectric
signals made it possible for organic control, which could
TABLE 1. CONFUSION MATRIX FOR RF CLASSIFIER significantly enhance the quality of life for amputees by
Predicted boosting independence and productivity in daily tasks.

0 1 2 3 4 Encouraged by the results of this work, a new study is

about to be launched, in which a haptic feedback sensor along
0 463 10 10 7 0 with a heat sensor will be integrated into the palm to increase
the level of biological realism.
1 16 412 24 18 0
Actual

2 39 41 359 23 0 REFERENCES
[1] Rupom, F. F., Jannat, S., Tamanna, F. F., Al Johan, G. M., & Islam, M.
3 7 20 33 413 0 M. (2020). EMG Controlled Bionic Robotic Arm using Artificial
Intelligence and Machine Learning. 2020 IEEE Region 10 Symposium
4 0 0 0 0 505 (TENSYMP)
[2] Sattar, N. Y., Kausar, Z., Usama, S. A., Farooq, U., & Khan, U. S.
(2021). EMG Based Control of Transhumeral Prosthesis Using
Machine Learning Algorithms. International Journal of Control,
TABLE 2. CONFUSION MATRIX FOR XGB CLASSIFIER Automation and Systems.

Authorized licensed use limited to: National Institute of Technology - Agartala. Downloaded on August 09,2024 at [Link] UTC from IEEE Xplore. Restrictions apply.
 [3] Md, L. W. D., Md, J. W. M., & Md, J. S. K. (2022). Easy EMG: A [7] Villoslada, A. (2016, February). Dextra - Open-source myoelectric
Guide to Performing Nerve Conduction Studies and Electromyography hand prosthesis. [Link]. [Link]
(3rd ed.). Elsevier. [8] Gianluca De Luca (2003) “Fundamental Concepts in EMG Signal
[4] J. Wang, H. Ren, W. Chen and P. Zhang, "A portable artificial robotic Acquisition”, Delsys Incorporated.
hand controlled by EMG signal using ANN classifier," 2015 IEEE [9] D.J. Hewson, J.Y. Hoqrel and J. Duchene (2003) “Evolution in
International Conference on Information and Automation, Lijiang, impedance at the electrode-skin interface of two types of surface EMG
China, 2015, pp. 2709-2714 electrodes during long-term recordings” Journal of Electromyography
[5] J. S. Artal-Sevil, A. Acón, J. L. Montañés and J. A. Domínguez, and Kinesiology, Vol. 13, Issue 3, pp. 273-279.
"Design of a Low-Cost Robotic Arm controlled by Surface EMG [10] Gilroy, A. M., MacPherson, B. R., Wikenheiser, J., Schuenke, M.,
Sensors," 2018 XIII Technologies Applied to Electronics Teaching Schulte, E., & Schumacher, U. (2020). Atlas of Anatomy (4th ed.).
Conference (TAEE), La Laguna, Spain, 2018, pp. 1-8 Thieme
[6] Borisov, I. I., Borisova, O. V., Krivosheev, S. V., Oleynik, R. V., & [11] Carlo J. De Luca (2002) “Surface Electromyography: Detection and
Reznikov, S. S. (2017). Prototyping of EMG-Controlled Prosthetic Recording”, Delsys Incorporated.
Hand with Sensory System * 074-U01 . IFAC-PapersOnLine, 50(1),
16027–16031 [12] Arshi, A. R. (2019). Introduction. Handbook of Biomechatronics, 3–
29.

Authorized licensed use limited to: National Institute of Technology - Agartala. Downloaded on August 09,2024 at [Link] UTC from IEEE Xplore. Restrictions apply.