2023 45th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC) | 979-8-3503-2447-1/23/$31.

00 ©2023 IEEE | DOI: 10.1109/EMBC40787.2023.10340284

Wireless Pressure Sensor Array Module for Sensorized Object

Guan Ming Lim1 , Prayook Jatesiktat1 , Christopher Wee Keong Kuah2 and Wei Tech Ang1

Abstract— To complement rehabilitation assessments that and hinders natural hand movements. Another approach is
involve hand-object interaction with additional information to cover the object with a pressure sensor array to uncover
on the grasping parameters, we sensorized an object with a detailed information such as the shape and size of the contact
pressure sensor array module that can generate a pressure
distribution map. The module can be customized for cylindrical area from the 2D pressure distribution map [5], [6]. This
and cuboid objects with up to 1024 sensing elements and it would allow the therapist to analyze the type of grasp and
supports the efficient transfer of data wirelessly at more than identify the transition between sub-phases (manipulation,
30 Hz. Although the module uses inexpensive materials, it transport, and release) during hand-object interaction. It also
is sensitive to changes in pressure distribution. It can also allows the subject to interact with the object in a natural
depict the shape of various objects with reasonable details as
shown in the small errors for object pose estimation and high context without distraction.
accuracy scores for hand grasp classification. The module’s Due to the unobtrusive nature of the pressure sensor array,
modular design and wireless functionality help to simplify it has been used in various healthcare applications such
integration with existing objects to create a smart sensing as sitting posture classification [7], ulcer monitoring [8],
surface.
exercise monitoring [9], nutrition intake monitoring [10],
Clinical relevance The resulting pressure distribution map sleep monitoring [11], and gait analysis [12].
allows the therapist to analyze grasping parameters that cannot However, the distributed nature of the pressure sensor
be determined from visual observations alone. array implies that a large number of sensing elements is
required and this increases the data processing load, cost
I. INTRODUCTION of fabrication, and wiring complexity. Thus, some works
Our sense of touch plays a crucial role in our interaction [9], [10] use custom-made electronics to handle high data
with the environment. Without it, we would experience acquisition load, but it is not readily available. On the other
difficulties maintaining a stable grasp of an object even hand, some works [4], [7], [11] use a microcontroller such
though we could see what we are doing [1]. This example as Arduino, but it has a low update rate of less than 10 Hz.
shows that relying solely on visual observation is insufficient In addition, most of the works [5], [9], [10], [11] rely on
as tactile sensing is also necessary to guide our movements wired communication, which is not portable.
and perception of the world. In this paper, we leverage the advances in internet of
Similarly, upper limb rehabilitation assessments such as things (IoT) devices to add-on wireless capability to the
the Action Research Arm Test (ARAT) usually involve sensing module. Therefore, we develop a wireless pressure
the therapist observing the subject’s ability to manipulate sensor array module that can be easily customized for
various objects. To provide a more objective assessment, cylindrical and cuboid objects. Although the module uses
inertial and discrete force sensors can be placed inside the inexpensive materials, it is sensitive to changes in pressure
object to generate real-time feedback of motion and force distribution and it can transfer data efficiently with software
data [2]. For instance, an increase in grasp duration with optimization. To demonstrate the feasibility and capability
a decrease in acceleration could indicate an improvement of the sensor module, we experiment with two applications:
in manipulation capability due to smoother movements and object pose estimation and grasp classification. These two
reduced acceleration [3]. examples highlight the flexibility of the sensor module and
To further observe the grasping pattern when the hand present a novel sensing element to complement upper limb
is in contact with the object, one approach is to use a rehabilitation assessments. The video demonstrations are
sensorized glove [4], but it can be cumbersome to put on made available at https://gmntu.github.io/psa/.

*This research is supported by the Rehabilitation Research Institute

of Singapore (RRIS) through its fourth Rehabilitation Research Grant
II. METHOD
(RRG4/2201). RRIS is a joint research institute by the Nanyang Techno-
logical University, Singapore (NTU), Agency for Science, Technology and The sensing module consists of a pressure sensor array
Research (A*STAR) and National Healthcare Group (NGH). and a data acquisition (DAQ) device. The former acts as
1 Guan Ming Lim, Prayook Jatesiktat, and Wei Tech an interface to capture the pressure distribution between
Ang are with the Rehabilitation Research Institute of
Singapore, 11 Mandalay Road, Clinical Science Building, two mating surfaces. The latter measures the change in
308232, Singapore [email protected]; the sensor’s physical property and transmits the data to
[email protected]; [email protected] a computer for processing. They are connected using flat
2 Christopher Wee Keong Kuah is with the Centre for Advanced Re-
habilitation Therapeutics, Tan Tock Seng Hospital, 308440, Singapore ribbon cables and insulation-displacement contact connectors
christopher [email protected] for robust electrical contacts.

979-8-3503-2447-1/23/$31.00 ©2023
Authorized licensed use limited IEEE
to: UNIVERSITY OF SOUTHAMPTON. Downloaded on July 03,2025 at 14:02:42 UTC from IEEE Xplore. Restrictions apply.
 Vin …
Insulating material n n-1

D2 Data m
Copper strips Sensor
D3 CLK
Pressure-sensitive D4 S0 Array
material D5 S1 m-1
D6 S2
…
Copper strips D7 S3
D8 INH Shift
Connector Insulating material register
A0 Vmn

Microcontroller Multiplexer
S0 S1 S2 S3 INH
Vin
Ribbon cable
(a) R1 R2
Force Vmn
Vout
Vin R2 Fig. 2. Simplified schematics for a M × N sensor array module. When
Vout
the shift register activates the first row and the multiplexer is set to read
(b) (c) from the first column, the voltage signal Vmn will correspond to the sensing
element mn. The power supply wires are left out to simplify the diagram.
A shift register requires two pins for data and clock signals. A 16-Channel
Fig. 1. Overview of a pressure sensor array: (a) Pressure-sensitive material multiplexer requires six pins: one for analog voltage signal, four for signals
is sandwiched between two layers of copper strips; (b) Cross-sectional view S0 to S3, and one for INHIBIT signal.
of a sensing element with force applied; (c) Voltage divider circuit where
the variable resistor R1 is equivalent to the pressure-sensitive material, and
the fixed resistor R2 determines the sensor’s output range.
supports wireless transmission of data to make the module
portable. It is also possible to combine Arduino Uno with
A. Pressure Sensor Array the ESP8266 Wi-Fi module for wireless capability.
The pressure sensor array shown in Fig 1 (a) consists of Due to limited I/O pins on a microcontroller, shift registers
a pressure-sensitive material sandwiched between two layers and multiplexers are used to handle a large number of I/O
of conductive material that are arranged in an orthogonal signals. The shift register toggles the supply of voltage to a
direction so that each crossing point forms a pressure sensing selected row of the sensor array, while the analog multiplexer
element. To protect and prevent the inner conductive layers selects a column of the array to read the analog output signal.
from direct contact with the external environment, an insulat- Thus, only two I/O pins are required to control multiple shift
ing material is included at the uppermost and lowest layers. registers as they can be chained together. One more pin is
The sensing material uses Velostat as it is a standard required to control an additional 16-Channel multiplexer by
technology that is widely available and cheap [13]. The using its INHIBIT pin property to toggle between multiple
Velostat used in this work has a thickness of 0.2 mm, volume multiplexers. Therefore, the total number of I/O pins required
resistivity < 500 Ωcm, and surface resistivity < 31 kΩ/cm2 . for a sensor array with 1024 (32 × 32) sensing elements is
nine, where one pin is used to read the analog signal, two
When a force is applied to a sensing element, it com-
pins to control four 8-bit shift registers, and the remaining
presses the semi-conductive sensing material, increases its
six pins to control two 16-Channel multiplexers.
density of conductive particles, and decreases its electrical
resistance [6]. Since resistance is inversely proportional to
the applied force, it can be converted to a voltage signal by C. Software Optimization
a voltage divider circuit as shown in Fig 1 (c). The main bottleneck in using ATmega-based boards (e.g.
The array structure with a scanning approach is selected Arduino Uno or Mega) is the default analogRead() function
as the cost and complexity of the design scales well with the that takes around 100 μs to read an analog input. This slows
number of sensing elements. For example, a M by N sensor down the update rate significantly especially for a sensor
array requires M ×N connections and input/output (I/O) pins array with 1024 sensing elements, as it would take more
to connect to each sensing element, which is not practical than 100 ms to read all the signals. By reducing the original
for a large-scale sensor array. With the array structure it only prescale value from 128 to 16, the time taken for each analog
requires M +N connections, and with the scanning approach read can be shortened to less than 15 μs.
the number of I/O pins is greatly reduced as at any point in Another bottleneck is the digitalWrite() function that has
time only a single row m and column n are activated to read various overheads for safety checks and makes it portable
the output signal Vmn from a sensing element. across different boards. By manipulating the digital ports
directly, we can further reduce the time taken by a few μs.
B. Data Acquisition Device To speed up the transfer of data, we could also reduce
The DAQ device consists of a microcontroller, shift regis- the length of each data to one byte (8-bit) instead of the
ters, multiplexers, and other miscellaneous parts as shown in original 10-bit and 12-bit ADC resolution for Arduino Uno
Fig 2. To allow reproducibility, we use commonly available and ESP32 respectively. Furthermore, only a portion of the
electronic parts and microcontroller development boards with sensor array is under loading most of the time while the other
open-source designs such as the ESP32 Thing Plus from sensing elements would output zero value. Thus, lossless
SparkFun. The board comes in a small form factor and data compression techniques such as run-length encoding

Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on July 03,2025 at 14:02:42 UTC from IEEE Xplore. Restrictions apply.
 Overall C1: Large C2: Ring C3: Tripod C4: Adducted C5: Precision
dimensions diameter thumb disk
7 cm 7 cm 3.5 cm
5 cm 7 cm

10 cm 7 cm 7 cm 3.5 cm

3.5 cm 3.5 cm 3.5 cm

5.5 cm 12 cm
(a) (b)

Raw data Upsampled data Binary threshold Result

32 × 32 128 × 128 128 × 128 128 × 128
7.5 cm
(c)

Fig. 3. Object pose estimation experiment (best viewed in color): (a)

Example of experimental setup where an arc-shaped object is placed on a
32 × 32 sensor array; (b) Dimensions of all the eight objects being tested;
(c) Overview of the image processing steps.
Fig. 4. Grasp classification experiment. The leftmost column shows the
dimensions of the cylindrical and cuboid-shaped objects covered with the
pressure sensor array. The other columns show the five classes of grasp
can be used to send consecutive zero values as a single zero adapted from the GRASP taxonomy [15].
followed by the count of zeros [14].
With software optimization, the overall update rates for
both the Arduino Uno and ESP32 Thing Plus boards can be upsampled to increase the spatial resolution by interpolating
increased by at least 50% as shown in Table I. intermediate pixels using cubic spline interpolation. A binary
threshold is then applied to the upsampled image to extract
TABLE I
the object’s contour. Finally, the iterative closest point (ICP)
U PDATE RATES BEFORE AND AFTER SOFTWARE OPTIMIZATION
algorithm is used to find the best three degrees-of-freedom
(dof) pose to transform the model’s contour to match the ob-
Arduino Uno ESP32 Thing Plus
Sensor array size 16 × 16 32 × 32 16 × 16 32 × 32 ject’s contour. The 3 dof pose contains horizontal translation
tx , vertical translation ty , and orientation angle θ.
Before opt. (Hz) 30 7 45 11
After opt. (Hz) 50 20 67.5 30
B. Grasp Classification
% increase 66 185 50 172
The second experiment on grasp classification shows the
advantage of having a customizable wireless sensor module
III. EXPERIMENT that is flexible and portable for wireless sensing applications.
A. Object Pose Estimation As shown in Fig 4, we sensorized a plastic tumbler and a 7.5
cm block used in the ARAT by covering the object’s surface
The first experiment aims to investigate whether the qual- with the pressure sensor array. We then classify five types
ity of the pressure distribution map is sufficient for a precise of grasping patterns from the GRASP taxonomy [15] that
task of pose estimation that demands a clear representation differ in the position and number of fingers in contact with
of the object’s shape. This is because a better quality pressure the object.
distribution map leads to a clearer representation of the object From the pressure distribution map, grasp classification
and more accurate pose estimation. can be done using an efficient k-nearest neighbors (KNN)
Fig 3 (a) shows an object placed at the center of the sensor algorithm. For each object and its corresponding grasping
array and its pressure distribution map is captured to estimate pattern, a subject is tasked to grasp the object 50 times to
the object’s pose. A total of eight objects with four different gather the training data for KNN. Testing is done on another
shapes (triangle, arc, semicircle, and rectangle) and two sizes set of 50 grasps from the same subject to test the repeatability
(large and small) are tested. For each object, it is rotated of the sensor array and the accuracy of grasp classification.
eight times with 45◦ interval to get eight different poses to
be compared against the ground truth poses. IV. RESULTS & DISCUSSIONS
To evaluate the effect of sensor resolution, we prepare two
sets of sensor arrays with the same sensing area of 20 cm A. Qualitative Evaluation of Pressure Distribution Map
× 20 cm, but one of the arrays has a lower resolution (16 Fig 5 shows qualitative examples of the pressure distribu-
× 16) with bigger sensing elements 7.5 mm × 7.5 mm, the tion map. Although the ADC resolution is only 8-bit, it is
other array has a higher resolution (32 × 32) with smaller sufficient to reveal detailed features and shapes of the object.
sensing elements 3 mm × 3 mm. For instance, individual fingers and toes can be differentiated
A series of 2D image processing techniques are applied with higher resolution (32 × 32) and upsampling the original
to extract the object’s pose. The raw pressure map is first data helps to generate a more detailed pattern.

Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on July 03,2025 at 14:02:42 UTC from IEEE Xplore. Restrictions apply.
 Hand Foot
Original

Raw data Upsampled data Binary threshold Result

16 × 16 64 × 64 64 × 64 64 × 64
16 × 16 32 × 32 16 × 16 32 × 32
Upsampled 2×

Raw data Upsampled data Binary threshold Result

32 × 32 64 × 64 32 × 32 64 × 64
32 × 32 128 × 128 128 × 128 128 × 128
Upsampled 4×

Fig. 6. Comparison of object pose estimation results for large semicircle

objects using sensor arrays with different resolutions (best viewed in color).
The top row shows a poor match between the estimated pose (red outline)
and the ground truth pose (blue outline). The bottom row shows an improved
result, due to a clearer representation of the semicircle shape using the
64 × 64 128 × 128 64 × 64 128 × 128
higher-resolution sensor array.

Fig. 5. Examples of pressure distribution maps for hand and foot. The red
region indicates an area of higher pressure and the dark blue region indicates
an area with zero or low pressure (best viewed in color). For each object, C. Evaluation of Grasp Classification
the top row shows the original resolution taken from two sets of sensor
arrays with the same sensing area of 20 cm × 20 cm, but with different Fig 7 shows that the pressure distribution map for hand-
resolutions (16 × 16 vs 32 × 32). The pressure maps are then upsampled object interaction contains multiple features such as the loca-
two times and four times to generate the next two rows of pressure maps.
In general, the higher the sensing resolution, the clearer the representation
tion and number of fingertips in contact with the object. This
of the object, and upsampling helps to increase the spatial resolution by allows the classification algorithm to distinguish between
interpolating the pressure distribution map. the five types of grasp with high accuracy of 98.0% and
97.6% for the cylindrical and cuboid objects respectively.
An overview of the classification results is shown in the
confusion matrices in Fig 8. Most of the errors in classi-
B. Evaluation of Object Pose Estimation fication occur between C2:Ring and C5:Precision disk, as
both classes share similar patterns of sparse pressure regions
located across a row.
The accuracy of object pose estimation is highly dependent
To keep the DAQ device as compact as possible, the
on the resolution of the pressure distribution map. As shown
maximum number of sensing elements on the sensorized
in Fig 6, the sensor array with a smaller resolution is unable
object is kept at 256, and the size of each sensing element is
to capture the shape of a semicircle object correctly. This
5 mm × 5 mm. Considering the average width of an adult
affects the extraction of the object’s contour resulting in a
finger that is typically greater than 10 mm, the size of the
poor estimation of the object’s pose. On the other hand, the
sensing element is just sufficient for the pressure map to
sensor array with higher resolution can capture the curved
capture the pattern of individual fingers as shown in Fig 7.
shape of the object with greater precision, resulting in better
Apart from classifying the different types of grasp, the
pose estimation.
pressure distribution map could also be used to reveal subtle
The above observation is further supported by the results differences for the same grasping pattern which is not
in Table II, where the sensor array with higher resolution differentiable by a camera. For example, a cylindrical object
performs consistently better with lower mean absolute trans- can be grasped sparsely with the fingertips and thumb (top
lation and orientation errors. In addition, the upsampling row of Fig 7), or it can be grasped densely with the whole
process can help to create a smoother spatial representation hand including the palm (bottom row of Fig 7).
of the object and may further increase the accuracy of pose
estimation to a certain extent. This is because upsampling can V. CONCLUSIONS
only help to refine key features, and if the original resolution
We present a customizable pressure sensor array that
fails to capture important details like the curved surface for
generates a 2D pressure distribution map for feature extrac-
the semicircle object (refer to the top row of Fig 6), the
tion purposes. Its modular design and wireless capability
feature will not be observable in the upsampled image.
simplify integration and complement existing applications
It is also expected that smaller objects will have larger with pressure mapping information that is not available from
errors as there are a lesser amount of sensing elements to visual observation. Although the sensor is inexpensive, we
represent the object accurately. In this work, the minimum demonstrate that it is possible to obtain a pressure distribu-
area of each sensing element is 3 mm × 3 mm as it is limited tion map with satisfactory quality and depict the shapes of
by the width of copper tape. various objects with reasonable details.

Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on July 03,2025 at 14:02:42 UTC from IEEE Xplore. Restrictions apply.
 TABLE II
C OMPARISON OF MEAN ABSOLUTE ERRORS FOR TRANSLATION AND ORIENTATION FOR OBJECT POSE ESTIMATION

16 × 16 sensor array 32 × 32 sensor array

Obj. size Mean absolute errors Original Upsampled 2× Upsampled 4× Original Upsampled 2× Upsampled 4×
Translation (mm) 11.8 8.83 6.31 5.47 3.15 2.24
Large
Orientation (deg) 14.5 7.67 6.39 5.59 2.91 2.33
Translation (mm) 11.3 11.8 6.66 6.77 4.17 2.89
Small
Orientation (deg) 18.3 12.8 13.1 12.1 6.56 4.34

pp. 277–284, Jan. 1984.

[2] A. Hussain, S. Balasubramanian, N. Roach, J. Klein, N. Jarrass,
M. Mace, A. David, S. Guy, and E. Burdet, “SITAR: a system for
independent task-oriented assessment and rehabilitation,” Journal of
Rehabilitation and Assistive Technologies Engineering, vol. 4, pp. 1–
16, Sept. 2017.
Color image of a hand Original Upsampled 2× Upsampled 4×
grasping an object 16 × 16 32 × 32 64 × 64 [3] G. Passetti, F. Cecchi, I. Baldoli, G. Sgandurra, E. Beani, G. Cioni,
C. Laschi, and P. Dario, “Sensorized toys for measuring manipulation
capabilities of infants at home,” in 37th Annual International Con-
ference of the IEEE Engineering in Medicine and Biology Society
(EMBC), Milan, Italy, Aug. 2015, pp. 7390–7393.
[4] S. Sundaram, P. Kellnhofer, Y. Li, J.-Y. Zhu, A. Torralba, and W. Ma-
tusik, “Learning the signatures of the human grasp using a scalable
tactile glove,” Nature, vol. 569, no. 7758, pp. 698–702, May 2019.
Color image of a hand Original Upsampled 2× Upsampled 4×
grasping an object 16 × 16 32 × 32 64 × 64
[5] P. G. Kry and D. K. Pai, “Grasp recognition and manipulation with the
tango,” in 10th International Symposium on Experimental Robotics,
Rio de Janeiro, Brazil, July 2006, pp. 551–559.
Fig. 7. Pressure distribution maps of a hand grasping a cylindrical object. [6] J. Cheng, M. Sundholm, B. Zhou, M. Hirsch, and P. Lukowicz,
Information such as the position and number of fingertips in contact with the “Smart-surface: large scale textile pressure sensors arrays for activity
object can be visualized from the pressure map. Although the hand grasping recognition,” Pervasive and Mobile Computing, vol. 30, pp. 97–112,
patterns from the color images look similar, the pressure map could reveal Aug. 2016.
differences in the parts of the hand that are in contact with the object. [7] W. Xu, M.-C. Huang, N. Amini, L. He, and M. Sarrafzadeh, “eCush-
ion: a textile pressure sensor array design and calibration for sitting
(a) Cylindrical object C2: Ring C5: Precision disk posture analysis,” IEEE Sensors Journal, vol. 13, no. 10, pp. 3926–
C1 3934, Oct. 2013.
C2
[8] P. Chung, A. Rowe, M. Etemadi, H. Lee, and S. Roy, “Fabric-based
pressure sensor array for decubitus ulcer monitoring,” in 35th Annual
Actual

C3
International Conference of the IEEE Engineering in Medicine and
C4 Biology Society (EMBC), Osaka, Japan, July 2013, pp. 6506–6509.
C5 [9] M. Sundholm, J. Cheng, B. Zhou, A. Sethi, and P. Lukowicz, “Smart-
C1 C2 C3 C4 C5
mat: recognizing and counting gym exercises with low-cost resistive
Predicted pressure sensing matrix,” in Proceedings of the 2014 ACM Inter-
(b) Cuboid object C2: Ring C5: Precision disk
national Joint Conference on Pervasive and Ubiquitous Computing,
Seattle, USA, Sept. 2014, pp. 373–382.
C1
[10] B. Zhou, J. Cheng, M. Sundholm, A. Reiss, W. Huang, O. Amft,
C2
and P. Lukowicz, “Smart table surface: a novel approach to pervasive
Actual

C3 dining monitoring,” in IEEE International Conference on Pervasive

C4
Computing and Communications (PerCom), St. Louis, USA, Mar.
2015, pp. 155–162.
C5
[11] W. Li, C. Sun, W. Yuan, W. Gu, Z. Cui, and W. Chen, “Smart mat
C1 C2 C3 C4 C5
system with pressure sensor array for unobtrusive sleep monitoring,”
Predicted
in 39th Annual International Conference of the IEEE Engineering in
Medicine and Biology Society (EMBC), Jeju, South Korea, July 2017,
Fig. 8. Left column shows the confusion matrices for the grasp classifi- pp. 177–180.
cation experiment. Right columns show similar patterns of sparse pressure [12] D. Chen, Y. Cai, and M.-C. Huang, “Customizable pressure sensor
regions located across a row for the C2:Ring and C5:Precision disk. array: design and evaluation,” IEEE Sensors Journal, vol. 18, no. 15,
pp. 6337–6344, Aug. 2018.
[13] D. Giovanelli and E. Farella, “Force sensing resistor and evaluation of
technology for wearable body pressure sensing,” Journal of Sensors,
To expand its potential uses, some modifications to the vol. 2016, pp. 1–13, Feb. 2016.
sensor array could be done. For example, the sensing ma- [14] M. Carrozzino, G. Avveduto, F. Tecchia, P. Gurevich, and B. Cohen,
terial can be changed to a stretchable conductive fabric to “Navigating immersive virtual environments through a foot controller,”
in Proceedings of the 20th ACM Symposium on Virtual Reality
accommodate a deformable object. The copper tape can also Software and Technology, Edinburgh, UK, Nov. 2014, pp. 23–26.
be changed to a conductive thread that is more suitable [15] T. Feix, J. Romero, H.-B. Schmiedmayer, A. M. Dollar, and D. Kragic,
for wearable applications such as pressure mapping for “The GRASP taxonomy of human grasp types,” IEEE Transactions on
Human-Machine Systems, vol. 46, no. 1, pp. 66–67, Feb. 2016.
prosthesis fitting [16]. [16] Y. Cai, J. Chen, D. Chen, G. Qu, H. Zhao, R. Ansari, and M.-C. Huang,
“Smart prosthesis system: continuous automatic prosthesis fitting
R EFERENCES adjustment and real-time stress visualization,” in IEEE Biomedical
[1] G. Westling and R. S. Johansson, “Factors influencing the force control Circuits and Systems Conference (BioCAS), Cleveland, USA, Oct.
during precision grip,” Experimental Brain Research, vol. 53, no. 2, 2018, pp. 1–4.

Authorized licensed use limited to: UNIVERSITY OF SOUTHAMPTON. Downloaded on July 03,2025 at 14:02:42 UTC from IEEE Xplore. Restrictions apply.