International Journal of Electrical and Computer Engineering (IJECE) Vol.
No.
October 2025, pp.
ISSN: 2088-8708.
DOI: 10.
11591/ijece.
Decomposition and multi-scale analysis of surface electromyographic signal for finger movements Afroza Sultana1.
Md.
Tawhid Islam Opu1.
Md.
Shafiul Alam2.
Farruk Ahmed3 Department of Electrical and Electronic Engineering.
Independent University.
Dhaka.
Bangladesh Department of Electrical and Electronic Engineering.
University of Dhaka.
Dhaka.
Bangladesh Department of Computer Science and Engineering.
Independent University Bangladesh.
Dhaka.
Bangladesh
Article Info
ABSTRACT
Article history:
Decomposition of the surface electromyography .
EMG) signal is vital for separating the composite, complex, noisy signals recorded from muscles into their integral motor unit action potentials (MUAP.
By precisely identifying each motor unitAos activity, this method offers greater insights into the functioning of the neuromuscular system, which helps isolate each motor unit's contribution, making it essential for understanding muscle coordination and diagnosing neuromuscular disorders.
In this study, we employ the maximal overlapping discrete wavelet transform (MODWT), which is well-suited for analyzing signals in the time-frequency domain.
The study decomposed the sEMG signal into six levels to identify the neural activity of finger movements and analyzed the motor unit action potential (MUAP).
In the frequency range of 30.
2 and 64.
6 Hz, the signal exhibits the highest MUAP which is independent of movement.
Using inverse MODWT, it was rebuilt from the decomposed levels.
With 95.
8% accuracy, the similarity between the reassembled signal and the original signal was determined using correlation analysis to assess the efficacy of the method.
Received Dec 21, 2024 Revised Jun 23, 2025 Accepted Jun 30, 2025 Keywords:
Decomposition Finger movements sEMG signal Wavelet transform This is an open access article under the CC BY-SA license.
Corresponding Author:
Afroza Sultana Department of Electrical and Electronic Engineering.
School of Engineering Technology and Sciences.
Independent University Bangladesh (IUB) Plot-16.
Block B.
Aftabuddin Ahmed Road.
Bashundhara R/A.
Dhaka-1229.
Bangladesh Email: afroza@iub.
INTRODUCTION
Surface electromyography .
EMG) is a flexible and non-invasive method for estimating muscle force, which has been utilized in neuromuscular physiology, movement disorder investigation, control of assistive devices such as prosthetic hands, and for diagnosing neuromuscular diseases.
Information about muscle movements and neural activity is obtained by measuring the electric currents produced in muscles during contraction.
The raw EMG signal is a mixture of overlapping motor unit action potentials (MUAP.
from multiple muscle fibers.
Decomposing this complex signal helps isolate individual MUAPs, allowing researchers and medical professionals to analyze the contribution of specific motor units for a more accurate assessment of muscle health and function.
In physiological research, motor unit features and muscle motor control mechanisms are studied using both the MUAP waveform characteristics and the statistics of the interpulse intervals .
The unique characteristics of a degraded MUAP can yield important information about the state of the nervous systemAiinformation that is necessary for clinically diagnosing myopathies and neuropathies .
, stroke patients .
, research into the neuromuscular control loop .
, and the prediction of human movements in prosthetics and exoskeletons .
, .
Decomposition of sEMG data into motor unit Journal homepage: http://ijece.
ISSN: 2088-8708
discharge patterns provides information on the recruitment and discharge behavior of motor neurons for investigating the neural control of movement.
Decomposition of EMG signals requires a technique that can handle four levels of complexity:
changing action potentials due to sensor movement, similar shapes at different times, superimposed action potentials, and a wide dynamic range of amplitudes, making the decomposition process more challenging .
Various decomposition approaches, such as blind source separation (BSS), convolution kernel compensation (CKC), independent component analysis (ICA), empirical mode decomposition (EMD), and the Fourier decomposition method (FDM) that have been used for several decades to decode motor neuron activities in the sEMG-based system.
Negro et al.
and Mohebian et al.
applied the convolutive BSS method in decomposition algorithms to segregate individual motor unit action potentials.
The BSS method of sEMG enhances motor unit study with non-invasive recordings, but challenges persist, including varying success rates across conditions, muscles, and individuals .
Besides, it brings substantial challenges due to its unique characteristics, including low signal-to-noise ratio, high similarity, and severe superposition of MUAP waveforms .
The EMD base decomposition was used by Wei et al.
for the recognition of lower limb movements, but EMD has limitations due to the mix-mode effect brought on by intermittent signal components .
The mix-mode effect was addressed by an enhanced version called ensemble EMD (EEMD), which also brought the difficulty of including residual supplemental noises during signal reconstruction .
Fatimah et al.
applied the Fourier decomposition method to decompose the surface EMG signal for the recognition of hand gestures.
Fourier analysis's assumption of signal stationarity results in inaccurate frequency representation over time and lacks time resolution, making it challenging to track transient features like muscle activation patterns.
Chen et al.
applied the CKC method for individual segments to decode the motor unit discharges from each motor neuron.
According to the studies .
when more motor activities are involved, the traditional CKC approach is unable to find enough MUs for myoelectric control.
The wavelet transform decomposition method was used by Liu et al.
and Duan et al.
and Phinyomark .
to recognize different hand motions for prosthetic hands.
Wavelet-based methods have advanced, but have drawbacks like dependency on wavelet function selection, inability to combine smoothness with numerical characteristics, and difficulty handling non-stationary EMG signals, limiting precise denoising and reconstruction .
Ae.
This study uses a multiresolution decomposition method based on the maximal overlapping discrete wavelet transform (MODWT) to offer adequate denoising and reconstruction of multi-class EMG signals.
Because of its improved noise reduction capabilities using the wavelet coefficient, the MODWT is a suitable method for more accurate multiresolution analysis of complex, noisy data .
The main contribution of this research is:
The work proposes a novel technique based on multiresolution decomposition using MODWT for the appropriate denoising decomposition and reconstruction of multi-class EMG signals.
The method potentially identified the specific frequency band where the motor neurons activate during different movements of single and multiple fingers.
Identified the dominant channels from eight-channel sEMG data by effectively measuring the average relative energy.
MATERIALS AND METHODS
Work flowchart The research work was completed in several phases.
The different phases of the work are presented by a flow diagram in Figure 1.
Figure 1.
Flow diagram of the work Int J Elec & Comp Eng.
Vol.
No.
October 2025: 4593-4604
Int J Elec & Comp Eng
ISSN: 2088-8708
Data acquisition This study made use of the dataset obtained from Khushaba and Kodagoda .
where fifteen classes of movement data were obtained from eight volunteers .
ix males and two females, aged between 20 to .
using 8 channels.
Fifteen classes of movements were collected during the Cexion of each of the individual Angers, i.
, thumb (TT), index (II), middle (MM), ring (RR), little (LL), and the combined fingers-thumbindex (TI), thumb-middle (TM), thumb-ring (TR), thumb-little (TL), index-middle (IM), middle-ring (MR), ring-little (RL), index-middle-ring (IMR), middle-ring-little (MRL), and hand close class (HC).
Preprocessing and segmentation The sEMG signals are often tainted by background noises caused by electronic equipment, subject movements, and physiological factors.
Proper detection and processing using efficient and cutting-edge techniques can be a basic prerequisite for its use in various domains.
The EMG power spectrum can be shaped using a variety of band-pass, notch, high-pass, and low-pass filters.
Typically, surface EMG signals are band-pass filtered between 20 Hz and 500 Hz to remove noise at frequencies below 20 Hz and above 500 Hz .
Ae.
In this study, the collected data was sampled at 4000 Hz, amplified to a total gain of 1000 dB, and to eliminate the 50 Hz line interference, the signal was band-pass filtered with a 20 Hz to 450 Hz filter.
To reduce processing time, the preprocessed signal must be extracted based on a threshold to determine the part of the signal that corresponds to each movement .
Therefore, the collected signal is sectioned into four non-overlapping segments of equal length, and the segment with the highest PSD is considered for decomposition.
Decomposition using maximal overlapping discrete wavelet transform The time-frequency analysis method evaluates time-varying non-stationary signals using the wavelet transform, which uses orthogonal bases with different resolutions.
It distributes signal decomposition in narrow frequency bands, filters without altering patterns, and handles time domain data without compromising frequency domain precision .
The MODWT is a DWT variation that uses a high-pass and low-pass filter to decompose a time signal into detailed and approximation signals, allowing for multiresolution analysis of smooth and detailed coefficients .
MODWT performs multi-resolution analysis of a signal, which is a scale-based additive decomposition like DWT.
Within precise and approximative components, the MODWT module does not generate phase changes .
The MODWT filters the input signal according to the number of levels at which it eliminates the noise coefficients from the signal.
MODWT offers flexibility in signal starting points, can handle any sample size, and is more efficient than DWT as its smooth and detailed coefficients are associated with zero-phase filters .
For noisy data analysis.
MODWT frequently uses the Daubechies wavelet function, and it provides a balance in the time-frequency localization .
The Daubechies wavelet function in MODWT yields the least error and allows for more coherent structure extraction compared to the Haar.
Coiflet.
Symmlet, and Biorthogonal functions .
, .
, and is therefore used in this analysis.
MODWT
splits the frequency spectrum of the input signal into scaling and wavelet coefficients, as shown in Figure 2.
Any segmented signal N that is an integer multiple of 2j, for yc = 1, 2, 3.
A ya can be implemented using the MODWT, where J is the level and j is the scale of the decomposition.
Figure 2.
Wavelet decomposition The MODWT scaling filters hEl and wavelet filters gEl are represented as .
, .
Eayco EaEyco = .
Oo2 yciEyco = Oo2 Decomposition and multi-scale analysis of surface electromyographic A (Afroza Sultan.
A ISSN: 2088-8708
The quadrature mirror filters used in MODWT are expressed as .
EaEyco = (Oeyc.
yco 1 EayaOe1Oeyco yciEyco = (Oeyc.
yco 1 yciyaOe1Oeyco where l=0, 1, 2, .
, ya Oe 1, and ya is the length of the wavelet filter.
The nth element of the first-stage wavelet coefficients (WE l,.
and scaling coefficients (VEl,.
of MODWT with the input time series signal X .
is presented as .
, .
Eyco,ycu = Ocya1Oe1 yco=0 Eayco ycUycuOeyco ycoycuycc ycA ycOEyco,ycu = Ocya1Oe1 Eyco ycUycuOeyco ycoycuycc ycA yco=0 yci where n=1, 2, 3, .
N, and N is the length of the signal in a sample to be analyzed.
The equations .
can be used to calculate the first-level approximations and details.
yaEyco,ycu = Ocya1Oe1 Eyco ycUycuOeyco ycoycuycc ycA yco=0 yci Eyco,ycu = Ocya1Oe1 Eyco,ycu yco ycoycuycc ycA ya Eyco ycO yco=0 yci Eycycu ) and scaling For a time, series X of random sample size N, the jth level MODWT wavelet coefficients .
cO coefficients .
cOEycycu ) are defined in .
ycOEycycu = Ocya1Oe1 Eyc,yco ycUycuOeyco ycoycuycc ycA yco=0 yci Eycycu = Ocya1Oe1 yco=0 Eaycyco ycUycuOeyco ycoycuycc ycA Similarly, equation .
give the approximations Aj and the details yayc of the nth element of the jth stage MODWT.
yaEyc,ycu = Ocya1Oe1
Eyco ycUycuOeyco ycoycuycc ycA
yco=0 yci
E E
Eyc , ycu = Ocya1Oe1 ya yco=0 Eayco ycOyco,ycu yco ycoycuycc ycA where gEl is the MODWT wavelet filter periodized to length N and hEl is the MODWT scaling filter periodized to length N.
Thus, the original time series signal can be expressed using the following estimates and details:
yc Eyc yaEyc ycU.
= Ocyco=0 ya .
The fundamental block diagram of MODWT is shown in Figure 3, and the complete procedure is explained in Algorithm 1.
Reconstruction using inverse maximal overlapping discrete wavelet transform (MODWT) Restoring a continuous-time signal from discrete samples is known as signal reconstruction.
reconstruct the signal, the approximation and detail coefficients at each level must be combined again.
The inverse transform reconstructs the detail and approximation coefficients for every level j and is expressed as:
Eyc .
) yaycuycyceycycyceycAycCyaycOycN.
aEyc .
)) X.
=Ocyayc=1.
aycuycyceycycyceycAycCyaycOycN.
The inverse MODWT uses the scaling filters hE l and wavelet filters gE l to reverse the decomposition using:
E yc.
yciEycOeyco .
= Ocyco yaEyc.
EaEyaOeyco .
Ocyayc=1 Ocyco ya .
By recombining all detail coefficients and approximation levels, the inverse MODWT reconstructs the original signal while preserving its length and resolution .
Int J Elec & Comp Eng.
Vol.
No.
October 2025: 4593-4604
Int J Elec & Comp Eng
ISSN: 2088-8708
Figure 3.
Decomposed and reconstructed signal using MODWT .
Algorithm 1.
Algorithm of sEMG signal decomposition and reconstruction Initialization A Input the sEMG signal X.
of length N.
A Specify the level of decomposition J.
A Select the wavelet filter Decomposition using MODWT For l=1, 2.
Apply the MODWT scaling filter hEl to obtain approximation coefficients AEl Apply the MODWT wavelet filter gEl to extract detail coefficients DE Repeat steps 1 to 3 for all levels l until the maximum level L is reached Store the approximation coefficients AEl from the highest level and the detail coefficients DE l from each level l Reconstruction using Inverse MODWT Initialize the reconstructed signal Xr.
as a zero vector of length N.
Reconstruct the signal using the approximation and detail coefficients from all levels:
Xr.
=AEl.
For each level l=1, 2.
L, reconstruct the contribution from the detail coefficients:
Xr.
= Xr.
End Decomposition and multi-scale analysis of surface electromyographic A (Afroza Sultan.
A ISSN: 2088-8708
RESULTS AND ANALYSIS
The sEMG signal was decomposed with multiresolution analysis methods using a 4th-order Daubechies filter .
up to level 6.
Upon increasing the levels from 4 to 7, we found that the optimal results were obtained at level 6.
Inverse MODWT was used to reconstruct the original signal.
The frequency range that matched the detail and approximation coefficients at each wavelet level of decomposition is displayed in Table 1, and the result of decomposition is presented in Figure 4, where Figure 4.
represents the original signal and Figure 4.
shows the decomposed signals.
To identify the possible frequency range of MU firings and find the relative energy at each level, we decomposed the dataset for each of the fifteen classes of movement.
Table 2 presents the results of decomposition for each class of movement signal from different channels, and Table 3 displays the results of the average MUAP for all 15 classes, considering all subjects.
Table 1.
Frequency band corresponding to each wavelet level Decomposition level Level-1 Level-2 Level-3 Level-4 Level-5 Level-6 Approx.
Frequency range (H.
1000- 2000
241 - 517
121 - 258
Bandwidth (H.
Overlapping frequency (H.
Figure 4.
Decomposition using MODWT .
original sEMG signal and .
signal after decomposition into 6 levels Int J Elec & Comp Eng.
Vol.
No.
October 2025: 4593-4604
Int J Elec & Comp Eng
ISSN: 2088-8708
Table 2.
The percentage of the relative energy of the decomposed signal for different movements (Sub-.
Freq levels Level-1 Level-2 Level-3 Level-4 Level-5 Level-6 Approx.
Freq levels Level-1 Level-2 Level-3 Level-4 Level-5 Level-6 Approx.
IMR
MRL
Table 3.
Average of the relative energy for different motor neuron activities (Ch-.
IMR
MRL
According to the decomposition result, level 6, or the frequency range of 30.
2 to 64.
6 Hz, showed the maximum concentration of action potentials as shown in Figure 5.
Furthermore, we found that neither the subject, as illustrated in Figure 6, nor the specific movement of any subject, as illustrated in Figure 7, affects the average relative energy of any level For each class of movement, the signal was reconstructed by combining all the decomposed levels using inverse MODWT.
the original signal is shown in Figure 8, and the reconstructed signal of the corresponding movement is shown in Figure 9.
The signal was also reconstructed from the level-5 and level-6 coefficients, as shown in Figure 10, because the highest energy was found at these levels, and we observed that the resulting signal appears close to the signal reconstructed from the coefficients of all levels.
Average MUAP Level-1 Level-2 Level-3 Level-4 Level-5 Level-6 Apprx Figure 5.
Average MUAP at different frequencies for levels Average Relative Energy Level-1 Level-2 Level-3 Level-4 Level-5 Level-6 Apprx Subject Figure 6.
Average MUAP of different frequency levels of different subjects Decomposition and multi-scale analysis of surface electromyographic A (Afroza Sultan.
A ISSN: 2088-8708
Average Relative Energy Level-1
Level-2
Level-3
Level-4
Level-5
Level-6
Apprx
IMR
MRL
Figure 7.
Average MUAP of different frequency levels for different movements Figure 8.
The original surface EMG signal Figure 9.
The reconstructed signal with inverse MODWT Figure 10.
The signal reconstructed from the coefficients of level 5 and level 6 After decomposition and reconstruction of the sEMG signals, we perform a correlation analysis between the original and reconstructed signals to evaluate the accuracy.
Correlation analysis is one of the popular techniques for determining the similarities or differences between two sets of multidimensional data.
If the shapes of two curves are identical, the cross-correlation coefficient R=1.
if not, it will be any value Int J Elec & Comp Eng.
Vol.
No.
October 2025: 4593-4604
Int J Elec & Comp Eng
ISSN: 2088-8708
between 0 and 1.
Tables 4 and 5 display the value of the correlation coefficients for each of the fifteen classes of data from subjects 4 and 5, and Table 6 shows the average value of the correlation for all subjects.
The average accuracy of decomposition and reconstruction was found to be 95.
Figure 11 shows the decomposition accuracy over the fifteen classes of movement.
Figure 12 shows the decomposition accuracy of all fifteen classes of movement for all the subjects, respectively.
We investigated the decomposition and reconstruction accuracy results for various movements and About different subjects as shown in Figure 14 and diverse movements as shown in Figure 15, the investigation discovered that channel 2 had the highest average accuracy of 95.
The outcome confirms the earlier finding from .
where Ch-2 and 4 were found to be dominating for the 15 classes of movements during the classification of the movements.
This leads to a reduction in overhead and dimensionality, which are crucial to wearable and real-time prosthetic applications.
Table 4.
Correlation between the original and reconstructed signals (S-.
Ch-1
Ch-2
Ch-3
Ch-4
Ch-5
Ch-6
Ch-7
Ch-8
Ch-1
Ch-2
Ch-3
Ch-4
Ch-5
Ch-6
Ch-7
Ch-8
S-1
S-2
S-3
S-4
S-5
S-6
S-7
S-8
IMR
MRL
Table 5.
Correlation between the original and reconstructed signals (S-.
IMR
MRL
Table 6.
Average correlation between the original and reconstructed signals IMR
MRL
Decomposition Accuracy % of Accuracy IMR
MRL
Figure 11.
Average decomposition accuracy of different classes of movement (Ch-.
% of Accuracy S-1
S-2
S-3
S-4
S-5
S-6
S-7
S-8
Subjects Figure 12.
Average decomposition accuracy of different subjects Decomposition and multi-scale analysis of surface electromyographic A (Afroza Sultan.
A Channel Accuracy Channel Accuracy Average Accuracy Ch-1
Ch-2
Ch-3
Ch-4
Ch-5
Ch-6
Ch-7
S-1
S-2
S-3
S-4
S-5
S-6
Ch-8
IMR
MRL
Average Accuracy ISSN: 2088-8708 Channel Figure 14.
Average decomposition accuracy of different channels Movements Figure 15.
Average decomposition accuracy of different channels for different movement
CONCLUSION
Decomposition of EMG signals is challenging due to the complexity and unpredictable nature of muscle activity patterns.
Therefore, a comprehensive assessment of signal processing techniques and validation methodologies is required to ensure the correctness and dependability of the results.
This paper presents a successful method for multiresolution decomposition and denoising of surface EMG data from fifteen finger movements using the MODWT and a 4th-order Daubechies filter.
The most significant finding of this work is the identification of particular frequency bands with the highest levels of motor neuron This finding shows that the quality and interpretability of sEMG signals linked to finger motions are successfully improved by the suggested MODWT-based multiresolution decomposition technique.
also found that the average relative energy of each level remains independent of the movement of any individual or subject.
Additionally, the channel selection technique based on average relative energy reduces computational complexity without sacrificing performance, which is essential for embedded and wearable systems that require real-time applications.
A correlation examination between the original and reconstructed signals revealed that the average reconstruction and decomposition accuracy was 95.
The detection of particular frequency bands that correlate with the firing of motor neurons during finger motions is a remarkable result.
This knowledge has physiological ramifications since it can help with the creation of neuromuscular models and neuroprosthetic adaptive control algorithms.
As interest in biomedical signal processing and human-centered computing grows, we hope that this research will contribute to improving the functionality of control systems in gesture-based interfaces, rehabilitation equipment, and prosthetics.
FUNDING INFORMATION
This research work was supported by the Startup Fund 2024 by Independent University.
Bangladesh (IUB), achieved for the Research Project Number: SU-SETS:24-007.
REFERENCES