Classification of Five Finger Movement, based on a Low-cost,

Real-time EMG System

Clive Seguna, Adrian Von Brockdorff, Jeremy Scerri and Kris Scicluna

Department of Electrical and Electronics Engineering, MCAST, Corradino Hill Paola, PLA9032, Malta

Keywords: EMG, Electromyography, Biopotential, Myoelectric.

Abstract: Researchers commonly use myoelectric signals to study the electrical activity produced by skeletal muscles

for the control of prosthetic arms, hands and limb replacement devices. Additionally, to the application in

prostheses, a myoelectric control system for multiple finger movements has the potential to develop

commercial products including advanced human-computer interfaces. The objective of this work is to

implement a set of low-cost active electrodes for the decoding of finger movement via time-domain analysis,

with an auto-gain adjustment technique. Different people will have different EMG amplitudes; therefore, it is

difficult to determine the gain required prior performing further signal processing. In this work, an auto-

adjustable gain amplifier circuit processes the maximum EMG signal amplitude and adjusts the gain stage

accordingly, without the need of any user interaction. This ensures that the gain is always automatically

adjusted to get the most effective performance from the data acquisition or analogue to digital converter (ADC)

module since the signal will be neither too low in amplitude to cause inefficient use of the ADC resolution,

nor too high to cause saturation of the signal. Through extensive experiments, the developed low-cost EMG

data acquisition system achieves reproducible and repeatable results for the detection and classification of the

five finger movements.

1 INTRODUCTION

This paper is an extension of the work originally

presented in NGCAS conference (Seguna, 2018).

Electromyography (EMG) signal acquisition is a

medicine technique used for recording and analysing

the electrical activity produced by skeletal muscles.

EMG systems detect the electrical potential generated

by muscles when they are neurologically activated

(Tsuji, 2000). The EMG signal can be used to obtain

several information related to muscle activity, such as

detecting medical abnormalities, muscle activation

levels, or to analyse the biomechanics of the human. It

is also used as a research tool for studying kinesiology,

which can then be used to control prosthetic devices

such as prosthetic arms, hands and lower limbs. This

is possible since muscles in the remaining part of the

limb function in a normal way, enabling the EMG

signals extracted from them to be used in limb

replacement devices. The benefit of this technique is

that the signal is acquired from the patient's remaining

limb muscles, which after appropriate processing can

be used to control motors. These motors can be used

to control several applications, including the control of

motorised wheelchairs and the control of prosthetic

devices which can be worn by amputees and activated

by their own EMG signals (Sudarsan, 2012),

(Osamu, 2003), (Jingpeng, 2013), (Côté, 2015)

provided. EMG studies in general are useful for

assessing the health of the neuromuscular system,

since certain diseases, such as multiple sclerosis,

suppress or even slow down normal nerve and muscle

firing. Surface EMG (sEMG) signal is the product of

all the action potentials which are picked from the

muscles below skin surface electrode. The amplitude

of sEMG signals is stochastic (random) in nature and

hence the reason why appropriate signal processing is

required for interpreting and using the signal.

Although the amplitude is random, it can be

reasonably represented by a gaussian distribution

function. The typical EMG amplitude varies from

microvolts to the low millivolts range (with the

maximum amplitude being around 10mV peak-to-

peak). The amplitude depends on the force applied

since the bigger the force, the more action potentials

will be stimulated which will trigger the contraction of

more muscle fibres (Naeem, 2012).

The more the action potentials are in reach of the

surface electrodes, the bigger the product result and

Seguna, C., Von Brockdorff, A., Scerri, J. and Scicluna, K.

Classiﬁcation of Five Finger Movement, based on a Low-cost, Real-time EMG System.

DOI: 10.5220/0008978901490159

In Proceedings of the 13th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2020) - Volume 1: BIODEVICES, pages 149-159

ISBN: 978-989-758-398-8; ISSN: 2184-4305

149

therefore the higher the amplitude. The frequency of

EMG signals can range from a few hertz up to the

lower kilohertz range, but the frequencies below 20 Hz

and above 200 Hz are usually not considered to

contain any useful physiological information. For this

reason, EMG acquisition systems normally filter these

frequencies out. Since the 50 Hz power line frequency

is within this range and can contribute to interference

in the EMG data being analysed, it is sometimes

recommended to set the cut-off frequency of the low-

pass filter at 50 Hz to attenuate most of the power lines

interference, or else apply a notch filter at that

particular frequency (Khoshaba, 1990), (Ahmad,

2017). The EMG signal can also contain small DC

contents producing an EMG signal with a non-zero

baseline. The DC content is eliminated in the EMG

acquisition circuitry, usually by using an

instrumentation amplifier (IA) or a high-pass filter

(Hong Quach, 2017).

Table 1: Applications of EMG (Côté, 2015).

Medical

research

Rehabilitation Ergonomics

Sports

Science

Orthopedic Post-surgery

Analysis on

Demand

Biomechanics

Surgery

Neurological

Disorders

Risk

Prevention

Movement

Analysis

Functional

Neurology

Physical

Therapy

Ergonomics

Design

Athletes

Strength

Training

Gait &

Posture

analysis

Active

Training

Therapy

Product

Certification

Sports

Rehabilitation

Common EMG analysis techniques include amplitude

analysis, time duration analysis, frequency analysis

and time-frequency analysis. The amplitude of the

EMG signal expresses the level of the muscle activity

and it changes with the amount of electrical activity

detected by the electrodes. EMG acquisition systems

usually make use of techniques to smoothen the raw

EMG signal amplitude and form a better

representation with respect to time. The most common

techniques are the Root Mean Square (RMS) followed

by the Mean Absolute Value (MAV). The RMS

technique is considered to be the most meaningful

since it provides a measure of the power of the raw

EMG signal (Tijssen, 2000).The ability to correlate

EMG amplitude with muscle force allows one to

determine whether the respective muscle is inactive or

active. When a muscle is inactive, the EMG amplitude

is effectively at 0 V and when the muscle is active, the

amplitude gets greater than 0 V. When a muscle is

active, one can also determine the time duration of the

muscle being active. This is achieved by simply

measuring the time when the amplitude exceeds a pre-

set threshold. Frequency analysis applies Fast Fourier

Transform (FFT) technique to obtain meaningful

frequency information, for a fixed stationary time-

domain data segment. This factor makes frequency

analysis not the ideal method when fast data

processing is required, such as for the use of prosthetic

limbs. On the other hand, this type of analysis is ideal

for studying muscle fatigue since in various studies it

has been proven that the mean frequencies of the EMG

signal will decrease with time during tasks that induce

muscle fatigue. The frequency analysis can also be

used for detecting interfering frequencies in the raw

signal, such as power line frequencies. Time-

Frequency analysis comprises the study of EMG

signal in both the time domain and the frequency

domain simultaneously. As already discussed, both

the time domain and frequency domain analysis can

be used to extract specific muscle activity. For this

reason, many researchers have combined the two to

benefit from information the two types of domains can

offer. This type of analysis is sometimes used to

achieve multiple classifications from the same EMG

signal, such as the angle and the force applied at a

joint. This is because the muscle force show more

change in the time domain, while any change in the

joint angle is more visible in the frequency domain of

the EMG signal (Clancy, 2008). During the process of

EMG signal acquisition one must follow certain steps

to prevent unwanted factors that may influence the

process. Although the human body is a good

conductor of electricity, there are still many aspects

that effect the conductivity level. Tissue conductivity

level can vary with the type, thickness, physiological

changes and even with temperature. These conditions

will vary from one person to another and sometimes

may even vary within the same person when the test is

performed at different time. Additionally, the human

body has approximately 640 skeletal muscles which

are close to one another, it is difficult to monitor

signals originating from a single muscle when using

surface electrodes. Neighbouring skeletal muscles

may produce signals which will eventually be picked

from the electrodes together with the wanted signals.

This is known as cross-talk, and normally it does not

exceed 15% of the overall signal contents.

Electrocardiography (ECG) signals can also interfere

the EMG signal recording. This is especially common

when performing EMG monitoring near the upper

trunk or the shoulder muscle. Another factor that may

alter EMG reading is when the distance between the

skeletal muscle belly (origin of the signal) and the

surface electrode changes during the signal acquisition

process. This normally happen when the patient

BIODEVICES 2020 - 13th International Conference on Biomedical Electronics and Devices

150

moves which causes the electrode to change position.

To prevent this from happening, one must secure the

surface electrodes and any wires that may cause them

to move during signal monitoring. The EMG signal

whose amplitude is between 0-10mV, when passing

through various tissues, is contaminated by various

noises (Amrutha, 2017), (De Luca, 2010), (Guohua,

2009).Therefore, it is vital to understand the properties

of various unwanted electric signals. EMG signals are

very sensitive to external noise and artifacts, mainly

due to the signal ranging from a few microvolts.

Inherent noise present in all electronic equipment

cannot be eliminated but can be reduced drastically

through intelligent circuit design. Additionally, the

silver/silver chloride electrode are electrically stable

and as their size increases, the impedance decreases.

Most of these interferences may be filtered out using

active or digital filters, by preparing the skin and

placing the electrodes properly. If proper skin

preparation and proper electrode placing is not

fulfilled signal quality is deteriorated. The electrode

cable and interface will also cause movement

artefacts, where such artifacts can be reduced

significantly using recessed electrodes. Further to this,

between the surface of the skin and the electrode-

electrolyte interface, a conductive gel layer is applied.

Electrical noise causes EMG interference since most

of the electronic components generate electrical noise

(known as Johnson–Nyquist noise) whose frequency

can range from few hertz to thousands of hertz. Such

electrical noise can be reduced drastically by using

quality components and through the implementation

of a well-designed circuit. Ambient noise is the main

source of electromagnetic radiation whose amplitude

is sometimes one to three times greater than the

desired EMG signal.

The surface of the human body is constantly

flooded with electromagnetic radiation. To prevent

these interferences, one must use an IA with a high

CMRR. This will attenuate any common mode noise

at the inputs of the electrodes. Another technique to

reduce ambient noise is to use the shortest possible

leads. If long leads are used, they will serve as an

antenna which will pick any ambient noise in the

vicinity. The leads should also be shielded to reduce

the possibility of noise from being picked. If noise

problems persist, the EMG acquisition circuit can be

covered by a Faraday cage. This will shield the circuit

from any Electromagnetic interference (EMI). When

the Faraday cage is grounded, the electric field energy

is drained away without affecting the circuit

performance. EMG instrumentation can pick various

types of influences that one may not even be aware of,

which include emotions and thoughts. These factors

can cause skeletal muscles to slightly contract since

humans tend to tighten up with certain emotions or

thoughts. These influences are better known as

involuntary activities which are picked by an EMG

measuring equipment (Bekir, 2014). There are

various techniques used to process and classify EMG

signals. Researchers make use from both the

amplitude and the spectral properties of the raw EMG

signals to supplement information on the muscle

activity which is used to increase the classification

accuracy. Following are some of the commonly used

techniques for signal acquisition, processes used and

algorithms for eliminating unwanted artefacts,

process the raw EMG signals and for classifying

different muscle movements. EMG signals can be

picked up using surface electrodes in two different

configurations, these being the monopolar and the

bipolar. The monopolar configuration makes use of

two surface electrodes, where one is placed on the

belly of the muscle and the other electrode is placed

as a reference on an electrically neutral tissue (such

as joints or other bony areas). The difference of the

two electrodes is then compared and processed for

further filtering and smoothing (Hudgins, 1993). The

other technique is the bipolar. This configuration

makes use of two electrodes (known as the detecting

electrodes) which are both placed on the belly of the

muscle. The detecting electrodes are typically kept

one to two centimetres apart. Another electrode is

used as a reference and must be placed on an

electrically neutral tissue. The advantage of using this

configuration is that the common noise can easily be

eliminated, something which is not possible to

achieve with the monopolar configuration. When

eliminating the common noise or any interference,

one will achieve a better signal-to-noise ratio and

hence a clear raw EMG signal can be obtained. The

pre-amplification is one of the most important aspect

when it comes to processing very low signals such

that of EMG. This is because the components used in

this stage must be of high precision and produce the

minimum noise possible, or else the noise can be

interpreted as the wanted signal. The most common

pre-amplification component used in EMG devices is

the instrumentation amplifier. Instrumentation

amplifiers are used to amplify the difference between

two inputs, which are connected to the two detecting

electrodes. They are designed to reject any signals

that are common to both inputs and therefore, are

used where precision and gain accuracy must be

maintained within a noisy environment, and where

large common-mode signals are present. After

reviewing the literature, it was found that the most

commonly used instrumentation amplifiers for EMG

Classiﬁcation of Five Finger Movement, based on a Low-cost, Real-time EMG System

151

devices are the INA126P, INA128, INA141, AD8221,

AD8421, AD623 and the AD642. Some papers

suggest that the IA gain must not be set too high or

else it may amplify the noise components together

with the wanted signal. Most of the EMG devices set

the instrumentation amplifier reference pin to half the

supply voltage (virtual grounding), while other

devices keep the output of the IA at a zero volts

baseline and then rectify the EMG signal prior

entering the ADC input. An experiment conducted by

the University of Utah includes the use of precision

rectifiers (super diodes) to rectify the raw EMG signal

prior inputting it to the IA. Other techniques were

used in other studies, which include the use of three

separate operation amplifiers that form the IA which

permits more flexibility in the selection of parameters.

After the pre-amplification stage, most devices

perform filtering to remove unwanted signal prior

further processing. Different EMG devices divide this

section into different stages, with some using separate

circuit for the low-pass and high-pass filtering, others

make use of a band-pass filter circuit and other

devices perform this task either on a microcontroller

or desktop computer. Digital filtering is usually

performed using Infinite Impulse Response (IIR)

filtering structure or Finite Impulse Response (FIR)

filtering structure, with the latter being the most

popular since it is more stable and less likely to

introduce non-linear phase distortions.

Most of the existing devices which make use of

hardware filtering, achieve this by using active filters

based on operational amplifiers or by using dedicated

filter ICs. Some EMG data acquisition

implementations make use of a combination of a low

order hardware filtering stage, which is then followed

by a higher order software filter. This is usually done

so that the hardware filtering can perform the first

stage filtering, prior the signal is inputted to an

analogue to digital converter (ADC). Another

technique would be the use of an adaptive noise

cancellation. Such technique can be implemented

using the Least Mean Square algorithm and has been

proved to be reliable and efficient (Phinyomark,

2012).

This will contribute to better ADC processing

since it will eliminate any major baseline drift and

high amplitude noise. Further filtering of the EMG

signal is then achieved by a second stage digital filter.

Some existing devices also make use of notch filters

to attenuate any frequencies that may interfere with

the wanted signal, with the most common being the

50 - 60 Hz power line frequency. This type of filtering

is not suggested by some researches since the

frequencies in the 50-60 Hz range can contain useful

information on the muscle contraction. They suggest

that a high-end instrumentation amplifier with a high

CMRR should be used instead. This should attenuate

any common power line distortion picked up by the

human body.

Although many studies agreed that the low-pass

filter (LPF) cut-off frequency should be set to around

15 to 50 Hz, it was noted that when it comes to the

high-pass filter (HPF) cut-off frequency, different

papers used different values with the range varying

from 150 Hz up to 800 Hz. Some of the papers

recommend that a high cut-off frequency for the HPF

is preferred so that any rapid on-off bursts of EMG

activity will not be filtered out. EMG devices which

perform hardware smoothing need to first rectify the

signal. Some existing devices use half-wave

rectification, but the most popular is the full-wave

rectification. Devices that use full-wave rectification

have the advantage of maintaining all of the raw EMG

signal information, unlike half-wave rectification

where the negative cycles are completely blocked. The

common technique used for the signal rectification is

through the use of a precision rectifier (also known as

a super diode) which is a circuit that acts as an ideal

rectifier.

The stage following the EMG rectification, is

usually the signal smoothing stage which is normally

achieved through an integrator circuit or a low-pass

filter. A similar technique which is also commonly

used is the envelope detector circuit, which gives a

similar output effect as the integrator and the low-pass

filter circuits. There are other techniques which are

sometimes used instead of root-mean square (RMS),

these being the Absolute Mean Value (AMV), the

Difference Absolute Mean Value (DAMV) and the

Variant Value (VAR) (Garavito, 2016). A study

entitled “Evaluation of EMG processing techniques

using Information Theory” shows that the RMS

technique provides the most meaningful information

out of the EMG signal.

More complex EMG processing devices can make

use of different algorithms to achieve better results.

Some of the commonly used algorithms are the Neural

Network, the Support Vector Machine and the

Euclidean Distance. The last two algorithms are

typically used when monitoring and recording finger

movements. They are used to isolate individual finger

movements to be able to control individual outputs,

such as prosthetic limbs. On the other hand, Neural

Networks (Subasia, 2006), (Gutiérrez, 2011)

algorithms are artificial intelligence networks that can

acquire any non-linear mapping of trained data

through learning. This algorithm is normally used to

achieve successful classification for non-stationary

BIODEVICES 2020 - 13th International Conference on Biomedical Electronics and Devices

152

EMG signals. Euclidean distance is used to determine

the distance of the input data points from a set of

predefined target points. Based on the distance

acquired, the system will check if the new data input

lies within a pre-defined target border and is used to

classify the data related to a particular muscle activity

into the desired group of channels.

The concept of autoregressive modelling is to

assume that the real EMG can be approximated by

what is known as the AR process. With this

assumption settled, the order and parameters of the

appropriate autoregressive model are chosen in a way

to fit the acquired EMG signals as closely as possible.

In turn, for every particular autoregressive model, the

power spectrum of the corresponding AR process can

be analytically determined. Thus, the AR method

provides an alternative way for EMG spectral

properties estimation. The work entitled “Real-Time

Computer Control using Pattern Recognition of the

Electromyogram” claimed a 95% accuracy in

classification was achieved when using the

Autoregressive modelling technique.

2 ANALYSIS OF EMG SIGNALS

Analyses of various EMG signals was done using a

pair of electrodes placed over the palmaris longus

muscle, which is mostly active when the ring finger

is contracted. The raw EMG signal was processed

through root mean square calculation. Figure 1

illustrates the result of the processed signal where it

is observed that the amplitude increases relatively

proportional with every 10 N of extra force applied.

This signal feature can be utilized in prosthetic hands

to apply variable force depending on the EMG

amplitude.

Figure 1: RMS of the EMG bursts with different forces,

starting from a force of 1 kg and increasing the force by 1 kg

with every burst.

The amplitude and frequency components of the ring

finger being closed at different angles were analysed.

Figure 2 illustrates the EMG signals obtained at

different angles, starting from 0 degrees (finger fully

opened) up to an angle of 180 degrees (fully closed),

with intervals of 45 degrees. The rectified EMG signal

amplitude increases quasi-proportional with the angle

of the finger. This feature can be utilised for prosthetic

hands for adjusting individual finger angle. A test was

conducted to analyse the EMG signal pattern with

respect to muscle fatigue. The setup used is shown in

Figure 3.

Figure 2: Rectified EMG amplitude signal at various finger

angles, starting from angle 0 degrees (finger fully opened)

up to an angle of 180 degrees (fully closed).

Figure 3: Setup for analysing EMG signal pattern with

respect to muscle fatigue.

The EMG signal for this test is shown in Figure 4. The

RMS equivalent is illustrated in Figure 5 showing the

profile of an EMG signal obtained for the ring finger

when a constant force of 5 kg for a period of 60

seconds was exerted. As shown in Figure 5 the

amplitude of the EMG signal increases slightly with

muscle fatigue when applying a constant force of 5 kg.

Therefore, since the difference in amplitude is

minimal the signal was then analysed in the frequency

domain where it was noticed that the frequency of the

EMG signal shifts to the lower side with muscle

fatigue as shown in Figure 6.

The amplitude frequency spectrum was performed

on a raw EMG signal at various angle position of the

finger. Figure 7 illustrates the magnitude frequency

spectrum plots obtained for four different ring finger

angles. From the results obtained, the magnitude of the

Classiﬁcation of Five Finger Movement, based on a Low-cost, Real-time EMG System

153

100 Hz frequency bin increased with angle position.

Frequency domain analysis could be challenging to

apply with accuracy due to problems such as

frequency resolution, magnitude accuracy at steady

state, and more generally, due to data processing.

Figure 4: EMG signal pattern with ring finger exerting a

constant force of 50 N.

Figure 5: RMS of the EMG signal with the ring finger

exerting a constant force of 50 N.

The time domain feature analysis is concerned with

the extraction of various EMG signal features in time

domain. Time domain features such as mean absolute

value, root mean square and wavelength were the most

popular in EMG pattern recognition because of high

processing speed in classification. The mean absolute

value of an EMG signal is defined as the average of

the total absolute value, while root mean square is the

amplitude modulated Gaussian random process

related to muscle force and contraction Time domain

features can easily and efficiently be used for the

recognition of an EMG pattern recognition. On the

other hand, frequency domain features can be used to

estimate the EMG power spectrum in frequency form.

In addition, the frequency domain spectrum is

commonly used in muscle fatigue and muscle force

estimation. Therefore in this work the classification of

finger movement was performed through the time

domain analysis rather than frequency spectrum.

(a)

(b)

(c)

Figure 6: Frequency spectrum shifting to the lower side of

the spectrum with muscle fatigue for various time durations

(a) 5-15 (b) 15-25 (c) 35-45 seconds.

3 CLASSIFICATION OF FINGER

MOVEMENT

H124SG muscle sensor surface electrodes were

placed at a particular area on the hand as shown in

Figure 8. The forearm has nineteen major muscles

responsible for the flexion, extension and other

movements of the fingers, wrist and elbow.

Reviewing the anatomy of the muscles, it was

concluded that the muscles used for the contraction of

the fingers are mostly exposed at the lower part of the

forearm Muscles responsible for finger movement

include the flexor digitorum superficialis (responsible

for

flexing all fingers -primarily at proximal

interphalangeal joints), flexor digitorum profundus

BIODEVICES 2020 - 13th International Conference on Biomedical Electronics and Devices

154

Figure 7: Frequency magnitude spectrum of an EMG signal for (a) 0°, (b) 90°, (c) 135°, and (d) 180° ring finger angle position.

(responsible for flexing the distal and proximal

interphalangeal joints) and flexor pollicis longus

(responsible for flexing the thumb).

This work covered two different experiments. In

the first experiment the electrodes were placed along

various muscles as shown in Figure 8a, while in the

second experiment six electrodes were placed on the

lower part of the forearm. The first experiment did not

show repeatable results from person to person. This

was mainly caused by the fact that not every person

has the same muscle anatomy and not every person

has the same amplitude peaks for the same muscle.

The physical factor of the person also played a big role

in the lack of consistency. When the system was used

on overweight people, it was noticed that it is difficult

to get finger movement classification. This is due to

the constantly changing physical distance between the

surface electrodes and the muscles being monitored.

To avoid the use of complex algorithms and other

additional signal processing for isolating finger

movements from other unwanted muscle activities,

such an area was selected. Calibration process

followed electrode placement. This process consisted

of contracting each finger multiple times one at a time.

With each contraction, the amplitudes acquired

from all electrodes being recorded. This process was

repeated for a pre-defined amount of repetitions so to

establish the required thresholds. EMG bursts were

monitored and processed so to evaluate the upper and

lower thresholds for each finger contraction, with the

highest monitored amplitude being set as the upper

threshold and the lowest amplitude being set as the

lower threshold. The maximum and minimum

amplitudes detected at each electrode after the raw

square algorithm.

(a) First Experiment

(b) Second Experiment

Figure 8: Electrode placement for (a)1

and (b) 2

Experiment.

Ten EMG bursts were recorded during this

calibration procedure. From the plots shown in Figure

9, it is observed that the thresholds for each finger

(c)

(d)

Classiﬁcation of Five Finger Movement, based on a Low-cost, Real-time EMG System

155

contributed to a unique pattern thus enabling EMG

signals were processed using the root mean- the

possibility of classifying various finger movements.

4 ACTIVE ELECTRODES FOR

ECG

Most EMG devices use of passive electrodes. In this

work active electrodes were used because they tend

to perform better for applications where very low

signals need to be acquired. Active electrodes

contribute to a have high input impedance with

minimal stray-capacitances at the inputs and low

output impedance contributing to a low cable

movement artifact. As shown in Figure 10 the main

front-end component for the developed active

electrode module is the LT1167 instrumentation

amplifier (IA). The LT1167 operates with a single or

dual rail supply voltage of ±2.3 V, common-mode

rejection ratio (CMRR) of 126 dB, and input

impedance of 1000 GΩ, thus contributing to less

attenuation in the input signal. Such parameters

satisfy the Surface Electromyography for the Non-

Invasive Assessment of Muscles (SENIAM) standard.

To minimize the gain error and achieve best CMRR

the REF pin of the LT1167 was connected to a 1.25

V supplied by the REF3312AIDBZT voltage

reference IC. This IC required low supply current

(typically 3.6 μA), has low temperature drift and has

an internal accuracy of ±0.15%. The maximum output

impedance does not exceed 0.1Ω, assuming the

output of the REF3312 is not switching at high

frequencies. This integrated circuit is also suggested

for use in medical applications. A 4.7 μF and 1.5 μF

supply bandpass capacitors are connected to the input

and output of the REF3312 respectively for better

stability of the input and output signal. A 604 Ω

resistor is used to set a fixed gain of 83. Note that this

gain will only amplify the raw EMG signal to around

500 mV peak-to-peak as per requirement. The two

5.1 kΩ resistors connected in series to each input of

the instrumentation amplifier input. These resistors

are made from carbon composition which can

withstand large short-term pulses and high voltages

when compared to other resistor types. Although

these resistors will contribute to higher noise at the

inputs of the IA, they are necessary to protect the IC

from any ESD.

The filtering stage consists of a 2

order

Butterworth high-pass filter with a cut-off frequency

of 15 Hz, a 5

order low-pass filter with a cut of

frequency of 500 Hz. A digital amplifier followed

the filtering stage.

Figure 9: RMS and threshold plots for various finger

movements.

The high-pass filter is required so that the baseline

drift will not affect the ADC performance. The

operational amplifier for the active high-pass filter

selected is the MCP604, which is a single supply, rail-

to-rail, unity gain stable CMOS quad op-amp IC.

Such component has a Butterworth response and can

operate at 3.3V, while consuming maximum current

of 1.2mA. The input of the filter comes directly from

the electrode, which pre-amplifies the raw EMG gain

accordingly. DS1804-050 is a 50kΩ potentiometer

BIODEVICES 2020 - 13th International Conference on Biomedical Electronics and Devices

156

that has 100 tap-points. The DS1804-050 can operate

from 3V or 5V.

Figure 10: Schematic diagram for the front-end component

interfaced with active electrodes.

Figure 11: Low-pass filtering stage based on the MAX7414

order LPF.

Figure 12: Baseline drift causing the EMG signal to be

saturated by the ADC references.

Figure 13: Digital Potentiometer.

This digital potentiometer also has a built in

EEPROM to store the wiper position even when the

supply is disconnected. This is useful so that the

system will not have to be re-calibrated every time the

supply is turned off. The DS1804 is specified to

provide an absolute linearity of ±0.60 LSB, which is

irrelevant for this application. It has a -3dB cut-off

frequency of 200 kHz. Since the EMG frequencies are

low, 200 kHz are enough for this application.

Developed software monitors the signal amplitudes

from each electrode and adjust the gain accordingly

through the digital potentiometer. Such technique will

not require the user interaction. The digital amplifier

circuitry consists of the DS1804-050 and the

MCP604 operational amplifier (same op-amp IC used

for the high-pass filter). The op-amp is configured as

an inverting amplifier, which can vary the gain from

1 (unity) up to 25. The MCP604 can be supplied with

a single rail supply between 2.7 V to 5.5 V. The

inverting configuration was used, so to implement a

linear gain amplification, by incrementing the

feedback resistance at equal intervals. The non-

inverting input of the op-amp is connected to a 1.65V

reference supply to offset the output by half the

supply voltage.

5 CONCLUSION

In this work, we proposed the successful

implementation of an active, noise cancelling,

affordable and wearable 6-channel sEMG data

acquisition system for the detection and classification

of finger movement. Such classification feature can

be combined with other systems for myoelectric

control applications. Additionally, unlike other

systems the gain is auto-adjusted using a digital

amplifier. Finger movement can be detected and

classified easily via EMG time domain rather than

frequency domain analysis. The most basic and

effective algorithm for enveloping the raw EMG

signals was found to be root-mean-square (RMS)

with a wide averaging window of 3000 instead of

1000. RMS only requires basic mathematical

calculations, which sums up in a system that requires

less processing power. As a result, a wider selection

of microcontrollers could be used for processing

EMG signals. The classification of finger movement

was done through the placement of six electrodes at

the lower part of the forearm. For this experiment,

there was no need for the electrodes to be placed

precisely in a specific area. The forearm was selected

because it has thin layer of fat, thus reducing the

problem of baseline drift drastically. After studying

the anatomy of the muscles, it was also concluded that

the muscles used for the contraction of the fingers are

Classiﬁcation of Five Finger Movement, based on a Low-cost, Real-time EMG System

157

mostly exposed at the lower part of the forearm. The

developed active electrodes with integrated IA placed

as close as possible to the input electrodes contributes

a better signal to noise ratio. The use of an auto-

adjustable gain stage contributed to a practical user-

friendly system. This circuit monitors the maximum

EMG signal amplitude and adjusts the gain stage

accordingly, without any user interaction. This

ensures that the gain is always adjusted to get the most

effective performance from the ADC module since

the signal will be neither too low in amplitude to

cause inefficient use of the ADC resolution, not too

high to cause saturation of the signal. A comparison

of our active electrode sEMG processing system with

other systems available in the literature and

commercial products in terms of frequency, weight,

supply voltage, wearable and other classification

features is shown in Table 2. Through extensive

experimentation system was tested by ten different

people of various weights, size and genders with

classification results observed to be repeatable and

reproducible.

6 FUTURE WORK

A small footprint prototype board is currently under

development and planned to be finalized by 2020.

This new prototype will enable the extraction of more

finger muscle movement features including finger

angle and muscle fatigue. Additionally, such a

wearable module will enable the processing of EMG

signals wirelessly over the cloud so to help of patients

suffering from conditions such as Carpal Tunnel,

Diabetic Peripheral Neuropathy, Ulnar Neuropathy,

Chronic fatigue syndrome, and Fibromyalgia, among

others.

Table 2: Comparison with other similar systems.

This

work

Myo

Armband

Biometrics

Datalog

Hercules

(Mert, 2018)

Classification

of Finger

Movement

Yes No No No

Contraction

Detection

Yes No No Yes

Wearable Yes Yes No Yes

Bandwidth (Hz) 20-589 - 20-460 20-500

Supply Voltage 2.5V - 3V 3.7V

REFERENCES

Seguna, C., et al., “Development of a New Low-Cost EMG

Monitoring System for the Classification of Finger

Movement”, 2018 New Generation of CAS (NGCAS),

pp. 126-129, 2018.

Tsuji, T., et al., “Pattern classification of time-series EMG

signals using neural networks”, International Journal of

Adaptive Control And Signal Processing, pp. 829-848,

2000.

Sudarsan. S., et al., “Design and Development of EMG

Controlled Prosthetics Limb”, International Conference

On Modelling Optimization And Computing, p. 3547-

3551, 2012.

Osamu, F., “A Human-Assisting Manipulator Teleoperated

by EMG Signals and Arm Motions,” IEEE transactions

on Robotics and Automation, vol. vol 19, no. 2, APRIL

2003.

Jingpeng. W, et al., “Surface EMG Signal Amplification

and Filtering”, International Journal of Computer

Applications, November 2013.

Côté, J., Electromyography (EMG): Intro to The Muscle

Activity Assesment Technique, Sept 2015.

Naeem, U. J., et al., (2012). “EMG-muscle force estimation

model based on back-propagation neural network”,

IEEE International Conference on Virtual

Environments Human-Computer Interfaces and

Measurement Systems (VECIMS) Proceedings, pp.

222-227, 2012.

Khoshaba. T., et al. “EMG Pattern Classification Based On

Back Propagation Neural Network For Prosthesis

Control”, Proceedings of the Twelfth Annual

International Conference of the IEEE Engineering in

Medicine and Biology Society, pp. 1474 – 1475, 1990.

Ahmad, J., Et al., “Multiclass Myoelectric Identification of

Five Fingers Motion using Artificial Neural Network

and Support Vector Machine”, Advances in Science,

Technology and Engineering Systems Journal Vol. 2,

Issue 3, pp. 1026-1033, 2017.

Hong Quach, J, “Surface Electromyography: Use, Design

& Technological Overview”, Concordia University,

2007.

Tijssen, M. A., et al. “Frequency analysis of EMG activity

in patients with idiopathic torticollis”. Brain: A Journal

of Neurology, pp. 677-686, 2000.

https://doi.org/10.1093/brain/123.4.677.

Clancy, E. A., et al. “Time- and frequency-domain

monitoring of the myoelectric signal during a long-

duration, cyclic, force-varying, fatiguing hand-grip

task”, Journal of Electromyography and Kinesiology,

pp. 789–797, 2008.

Amrutha, A., “A Review on Noises in EMG Signal and its

Removal,” International Journal of Scientific and

Research Publications, Vols. Volume 7,, no. issue 5,

May 2017.

De Luca, C. J., et al. “Filtering the surface EMG signal:

Movement artifact and baseline noise contamination”,

Journal of Biomechanics, pp. 1573– 1579, 2010.

BIODEVICES 2020 - 13th International Conference on Biomedical Electronics and Devices

158

Guohua, L., et al., “Removing ECG noise from surface

EMG signals using adaptive filtering”, Neuroscience

Letters, pp. 14-19,2009.

Bekir, K., et al., “Machine Learning Algorithms for

characterization of EMG Signals”, International

Journal of Information and Electronics Engineering,

Vol. 4, No. 3, May 2014.

Hudgins, B., et al., “A new strategy for multifunction

myoelectric control,” IEEE Trans. Biomed. Eng., vol.

40, no. 1, pp. 82–94, Jan. 1993.

Phinyomark, P., et al., “Fractal analysis features for weak

and single-channel upper-limb EMG signals,” Expert

Syst. Appl., pp. 11156–11163, 2012.

Garavito, F., et al, “EMG signal analysis based on fractal

dimension for muscle activation detection under

exercise protocol,” in Proc. 21

Symp. Signal Process.,

Images Artif. Vis., pp. 1–5, 2016.

Subasia. A, et al., “Classification of EMG signals using

wavelet neural network”, Journal of Neuroscience

Methods, pp. 360–367, 2006.

Gutiérrez, J. M., et al. “Wavelet Neural Network as EMG

classifier”, Pan American Health Care Exchanges, pp.

67 – 71, 2011.

Thalmic Labs. (2015). Myo Gesture Control Armband

[Online]. Available: https://www.thalmic.com/en/myo/.

Biometrics Ltd, Newport, U.K., “Datalog MWX8”.

Available: http://www.biometricsltd.com/datalog.htm

Mert, E., et al., “An Embedded, Eight Channel, Noise

Cancelling, Wireless, Wearable sEMG Data

Acquisition System with Adaptive Muscle Contraction

Detection”, IEEE transactions on biomedical circuits

and systems, 2018.

Classiﬁcation of Five Finger Movement, based on a Low-cost, Real-time EMG System

159