See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/370870215

Deep Learning Enabled Perceptive Wearable Sensor: An Interactive Gadget for

Tracking Movement Disorder

Article in Advanced Materials Technologies · April 2023

DOI: 10.1002/admt.202300046

CITATION READS

1 63

4 authors, including:

Spandan Ranpariya Anand Babu

Indian Institute of Information Technology Vadodara INST-IISER Mohali
9 PUBLICATIONS 4 CITATIONS 17 PUBLICATIONS 117 CITATIONS

SEE PROFILE SEE PROFILE

Dipankar Mandal
Institute of Nano Science and Technology
177 PUBLICATIONS 6,387 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

COVID and Wearable Sensors View project

Fish 'biowaste' converted to piezoelectric energy harvesters View project

All content following this page was uploaded by Anand Babu on 21 June 2023.

The user has requested enhancement of the downloaded file.

RESEARCH ARTICLE
www.advmattechnol.de

Deep Learning Enabled Perceptive Wearable Sensor: An

Interactive Gadget for Tracking Movement Disorder
Anand Babu, Spandan Ranpariya, Dhirendra Kumar Sinha, and Dipankar Mandal*

movement monitoring such as embed-

Biosignals of diverse body posture contain a key information about the ded cameras, inertial measurement units
physiological, kinesiological, and anatomical status of human body that can (IMU), and electromyography (EMG)
facilitate in early assessing the neurological maladies such as Parkinson’s have the limitations of low latencies,
disease, multiple sclerosis, and other neurological disorders. Early detection bulky system, expensive, and more power
consumption to operate. Here, wearable
and timely intervention of specific diagnosis for curing the disorders is healthcare devices/sensors play a crucial
considered as one of the effective step of diagnosis. Existing technologies of role. Owing to its flexibility, confirma-
capturing human movements have the limitations of low latencies, bulky bility, and compactness, it can easily be
counterpart, cost-intensive, and power consuming. A highly sensitive worn on different parts of the body.[1,2]
(≈440 mV N−1 ), confirmable, and flexible wearable gadget is introduced for It can be fabricated by utilizing the dif-
ferent transduction mechanisms such as
addressing such complexities; moreover, the gadget is fabricated by recycling
piezoelectric, piezoresistive, piezocapac-
waste material. Subsequently, it is interfaced with various deep/machine itive, and triboelectric effects. The for-
learning algorithms for classification/prediction of different hand poses; mer needs a specific material that must
particularly unsupervised k-means clustering is used for observing discrete be non-centrosymmetric in nature, while
classes, and different supervised algorithms such as k-nearest neighbor, the latter two require an external bias to
support vector machine, deep neural network (DNN), and pattern recognition operate. However, the triboelectric effect
is universal to all materials, thus provid-
that provide the high degree of prediction accuracy (up to ≈98%) for ing a broad selectivity of the materials.
classification of different postures. Thus, artificial intelligence-aided wearable In view of their cost effectiveness, sim-
gadget can not only offer an unique solution for autonomously tracking ple fabrication techniques, comprehen-
various body movements but also assist in reducing waste materials, which is sive materials choice, and self-powered
otherwise a threat to the environment. abilities, triboelectric nanogenerator is
one of the promising candidates to ful-
fill the requirements of the wearable
gadgets.[3–9] It has been extensively used
1. Introduction in energy-harvesting devices to biomedical sensors that can help
to detect, diagnose numerous diseases by its early detection and
Biosignals of different movements of the human body contain
substantially cut down the huge expenditure spending on health-
vital information about the physiological status of the body that
care monitoring.[9–11]
helps to assess the various neurological maladies such as Parkin-
Another notable point is the massive production of plastic ma-
son’s disease, multiple sclerosis, rapid eye movement (REM)
terials, a major threat to the environment, the major trouble is
sleep disorder, and other neurological disorders. The early de-
the lack of feasible technology available so far for the degrada-
tection of these disorders facilitates timely intervention of the
tion of plastic.[12,13] Polyamide (Nylon 66) contributes a big chunk
diagnosis that can tailor a specific remedy for curing it before
to this plastic waste (≈35% share in manufacturing and trans-
becoming fatal. The existing technologies of capturing human
port industries); it is significantly used in our daily life activities
and industries in various forms such as toothbrush bristles, ny-
A. Babu, D. Mandal lon clothes, wear pads, wheels, gloves, medical implants, cable
Quantum Materials and Devices Unit ties, engineering plastics, and impaired machine tools to men-
Institute of Nano Science and Technology tion a few.[14] Its high toughness, excellent abrasion resistance,
Knowledge City, Sector 81, Mohali 140306, India high mechanical damping ability, and good resistance to energy
E-mail: [email protected]
radiation make polyamide one of the most favorable materials for
S. Ranpariya, D. K. Sinha
Department of Science and Humanities making machine tools, circuit insulation board, gears, and other
Indian Institute of Information Technology Vadodara engineering stuff.[15] However, the degradation of nylon 66 is a
Gandhinagar, Gujarat 382028, India cumbersome and cost-intensive task that poses a serious threat
to the environment. Simultaneously, with the rapid growth in the
The ORCID identification number(s) for the author(s) of this article population, it is hard for everyone to reach to the doctors, the sit-
can be found under https://doi.org/10.1002/admt.202300046
uation seems more worse in case of developing countries.
DOI: 10.1002/admt.202300046

Adv. Mater. Technol. 2023, 2300046 2300046 (1 of 10) © 2023 Wiley-VCH GmbH
 2365709x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/admt.202300046 by Bordeaux INP, Wiley Online Library on [05/04/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmattechnol.de

On the other hand, the quantum leap in deep learn- and ≈24.6°, a signature of the 𝛼 phase (triclinic unit cell) in nylon
ing/machine learning (ML) algorithms from the last decade 66, corresponding to the reflection plane of (100) and (010)/(110),
makes them one of the advanced tools for providing intelligence respectively.[21,22] In FTIR spectra (Figure S1c, Supporting Infor-
to the systems for learning, improvement, and predict the out- mation), vibrational bands at ≈1717 cm−1 associated with the
comes without any human intervention.[16,17] Although it has C═O stretching, ≈1650, ≈1540, and 1360 cm−1 correspond to
taken roots around the 1960s, still in the developing stage, be- the signature of amide I, amide II, and amide III, respectively.
cause of the introduction of new and powerful algorithms day Likewise, ≈1145, 945, and 560 cm−1 vibrational band attribute
by day. In the current scenario, machine learning has been per- to CO─CH symmetric stretching, C─C stretching, and O═C─N
vasive in almost every field from financial market to the plan- bending modes, respectively. A comprehensive table of IR active
ning of smart cities and cyber securities to mention a few.[18–20] modes is provided in Table S1 (Supporting Information) and the
Deep learning in healthcare has been extensively used in differ- FTIR spectra recorded from different waste materials are shown
ent domains from medical image segmentation to crowdsource in Figure S2 (Supporting Information).[23,24] It indicates that W-
data collection and clinical trial research. TENG is possible to fabricate from a wide range of waste materi-
In this work, we have fused all the above three steps: waste als that are subsequently used to acquire the different postures of
nylon 66 materials has been utilized to fabricate a wearable body motions and the data were collected in the form of dataset,
sensor that works based on contact electrification, termed as and afterward used as an input of the deep learning (DNN)/ML
wearable triboelectric nanogenerator (W-TENG). This could be models. The ML/DNN has been executed to train the dataset
treated as one of the approach of recycling the waste nylon 66 that predicted/classified different discrete classes (Figure 1). Fig-
(polyamide). Subsequently, it is integrated with the deep learn- ure 2a demonstrates the typical structure of W-TENG; the po-
ing/machine learning algorithms for classification and predic- tential difference has been generated by the contact-separation
tion of human motion monitoring particularly hand postures. between the two dielectric layers, namely, nylon 66 and PDMS,
We have employed two different categories of machine learn- where Ag works as an electrode and W-TENG operates on the
ing: unsupervised and supervised learning. Primarily the dataset single electrode mode.[25] That leads to produce the electrical out-
has been characterized by unsupervised machine learning (k- put corresponding to input mechanical stimuli. For further elu-
means clustering) which indicates the classes are discrete and cidating the generation of triboelectric potential between two tri-
data are well separated by clusters. Subsequently, various super- boelectric layers, scanning probe microscopy based Kelvin probe
vised machine learning algorithms have been implemented such force microscopy (KPFM) has been deployed. The surface poten-
as support vector machine (SVM) and k-nearest neighbor (KNN), tial images of nylon 66 and PDMS are provided in Figure 2b,c,
and deep learning algorithms such as deep learning neural net- and their corresponding morphology images are shown in Fig-
work (DNN) and pattern recognition, recurrent neural network ure S3a,b (Supporting Information), respectively. Surface poten-
(RNN) which shows the prediction accuracy of ≈93%, ≈97%, tial of nylon 66 and PDMS was found to be ≈310 and ≈−55 mV,
≈94%, and ≈98%, respectively. Thus, wearable gadget, i.e., ar- respectively.[6,26,27] The work function of the two triboelectric layer
tificial intelligence-aided wearable sensor, not only facilitates in was estimated with the help of the following equation[28,29]
reducing the waste materials but also can be in different next-
generation applications such as development of artificial prosthe- (𝜑tip − 𝜑sample )
ΔVCPD = (1)
sis, automatic wheel chair for critically ill and its implementation e
in augmented reality (AR)/virtual reality (VR). Through this work
where ΔVCPD is the contact potential difference; 𝜑tip and 𝜑sample
we have endeavored to utilize the plastic-based materials as an
are the work functions of the tip and the sample, respectively; and
intelligent wearable gadget that can early detect the various phys-
e is the electronic charge. For the calibration of the KPFM, highly
iological disorders so that it can be cured before transforming
oriented pyrolytic graphite (HOPG) is used as a reference sample
into fatal disease. It not only helps to fabricate the intelligent dis-
and work function of the AFM cantilever is found to be ≈4.7 eV
order tracking devices but also assists to reduce the plastic waste
(𝜑tip ≈ 4.7 eV).[29]
prevailing in the environment, eventually aid to clean, green, and
The work function of nylon 66 and PDMS was found to be
circular economy.
≈4.3 and ≈4.8 eV, respectively, as demonstrated in their corre-
sponding band diagrams (Figure 2d,e, respectively). It is clearly
2. Results and Discussion evident that there is ≈0.5 eV work function difference between
the two triboelectric layers that translated into the electrical out-
2.1. Characterization of the Materials
put under the principle of contact electrification. As the funda-
mental principle of triboelectricity suggests, open-circuit voltage
Primarily, waste nylon 66 material was collected from various
produced between the two triboelectric layers is directly propor-
sources such as toothbrush bristles, cable ties, impaired machine
tional to the work function difference between them, see the fol-
tools, nylon cloths, fibers, etc. Furthermore, it was cleaned and
lowing equations[30]
processed into a film by using solution casting method. The
enlarged region displayed in Figure S1a (Supporting Informa- ( )
C 𝜑1 − 𝜑2 X
tion) represents the molecular structure of monomer of nylon 66. Voc = (2)
The phase identification of the material was carried out by X-ray eA𝜀0
diffraction (XRD) pattern and Fourier transform infrared (FTIR) [ )
spectra (Figure S1b,c, Supporting Information). As depicted in C (𝜑nylon 66 ) − (𝜑PDMS − ]X CΔ𝜑X
VOC = = (3)
the XRD spectra, it consists of two diffraction peaks at ≈20.56° eA𝜀0 eA𝜀0

Adv. Mater. Technol. 2023, 2300046 2300046 (2 of 10) © 2023 Wiley-VCH GmbH
 2365709x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/admt.202300046 by Bordeaux INP, Wiley Online Library on [05/04/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmattechnol.de

Figure 1. A detailed view of the study: Data have been acquired from the various movements of the body, collected data employed as an input for the
execution of the different algorithms, further classified and predict the class of an unknown input.

Figure 2. a) Schematic view of W-TENGs, where enlarged region demonstrates the layered structure of the device. Surface potential images of b) nylon
66 and c) PDMS (scale bar ≈ 5 μm). Band diagram indicating the values of work function of d) nylon 66 and e) PDMS; the diagonal arrow between the
two band diagrams represented the triboelectric potential difference between the two materials. f) The triboelectric series with the position of nylon 66
and PDMS along with some standard reference materials.

Adv. Mater. Technol. 2023, 2300046 2300046 (3 of 10) © 2023 Wiley-VCH GmbH
 2365709x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/admt.202300046 by Bordeaux INP, Wiley Online Library on [05/04/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmattechnol.de

Figure 3. a) Short-circuit charge transfer (QSC) variation with gap distance, while insets represent the theoretical model of W-TENG. b) Maximum
energy density (Emax ) and open-circuit voltage (Voc ) variation with distance between two tribolayers, insets displayed the equivalent circuit of TENG.
c) FEA-based simulation of triboelectric potential, generated by contacting between the two tribolayers at different gap distances.

where Voc is the open-circuit voltage, C is capacitance, X is the S𝜎x (t)

Qsc = (4)
distance between two tribolayers, A is the surface area of the elec- d0 + x (t)
trode, 𝜑nylon 66 is the work function of nylon 66, and 𝜑PDMS is the
work function PDMS (Figure 2f). Owing to its flexibility, robust- 𝜎x (t)
VOC = (5)
ness, and confirmability with the body (Figure S4, Supporting 𝜀
Information), W-TENG exhibits superior performance as a wear-
where Qsc is the short-circuit charge transfer, S is the surface area,
able sensor with a high sensitivity of ≈440 mV N−1 .
𝜎 is the surface charge density, x(t) is the gap distance as a func-
tion of time, VOC is the open-circuit voltage, 𝜖 is the relative per-
∑n d
2.2. Finite Element Analysis Based Simulation mittivity, and d0 is the effective thickness ( dO = i = 1 𝜀𝜀i ).
ri

Finite element analysis (FEA) based simulation was carried out to 2.3. Critical Biomedical Signal Monitoring
visualize the generation of potential difference, open-circuit volt-
age, and short-circuit charge transfer in W-TENG (Figure 3a), in Owing to its high mechanosensitivity (≈440 mV N−1 ), W-TENG
which ≈200 μm thickness of each dielectric layer, electrode thick- is qualified to capture the numerous key features of biomedical
ness is ≈1 μm, and charge density (𝜎) of 8 μC/m2 was taken.. signals; for instance, it is tested for monitoring breathing rate and
Short-circuit charge transfer (Qsc ) is attributed to the maximum arterial pulse. Breathing is a fundamental biological process that
charge transfer between two tribolayers, which shoots up to a very consists of two modes: inspiration (air enters into the lungs) and
small distance (≈μm) between the two tribolayers, and becomes expiration (air is blown out from the lungs), both are precisely
saturated (Figure 3a). captured by the W-TENG (Figure 4a). Fundamentally, there are
Insets of Figure 3a demonstrate the theoretical model of W- three forces involved in the process of breathing, namely, elas-
TENG. Maximum energy density (Emax ) and open-circuit voltage tic force of chest-lung system (PEle ), air distance and turbulence
(Voc ) are the two critical parameters to assess the electrical per- (PTurb ), and nonelastic tissue deformation resistance (PDef ). The
formance of a TENG; variation of these parameters with gap dis- total force required for respiration can be estimated from the fol-
tance for W-TENG is shown in Figure 3b; as depicted from the lowing equation[34–36]
figure Voc increase consistently with the gap distance, while Emax ( ) ( )2
increase abruptly up to a certain distance, further saturating. In- dV dV
PT = KV + K1 + K2 (6)
sets of Figure 3b demonstrate the equivalent circuit of TENG. Fol- dt dt
lowing equations have been utilized for theoretical calculations.
The visualization of the triboelectric potential with the gap dis- where K is the elastance, K1 and K2 are the constants, dV dt
is the
tance (0–10 mm) is shown in Figure 3c.[31–33] rate of change of lungs volume, and V is the volume of the lung. If

Adv. Mater. Technol. 2023, 2300046 2300046 (4 of 10) © 2023 Wiley-VCH GmbH
 2365709x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/admt.202300046 by Bordeaux INP, Wiley Online Library on [05/04/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmattechnol.de

Figure 4. a) Schematic of the mechanism of human breathing. b) Measurements of the breathing at different breathing conditions from the W-TENG. c)
STFT plot of the different breathing rates, showing the frequency of the breathing with time. d) Schematic of the mechanism of arterial pulse monitoring
from W-TENG. e) Arterial pulse monitoring from W-TENG and f) corresponding STFT plot.

the change in the volume is assumed as a first approximation to originating from breathing problems such as asthma, allergies,
be sinusoidal in nature, which is close to real scenario, the output and chronic obstructive pulmonary disease (COPD) in the earlier
mechanical work per respiration can be calculated as stage.[10,37,38]
Figure 4c represents the short-time Fourier transform (STFT)
V 1 1 2 plots of different breathing rates that can clearly distinguish the
W = ∫0 T PT dV = K VT2 + K1 𝜋 2 fVT2 + K2 𝜋 2 f 2 VT3 (7)
2 4 3 different breathing rates in terms of frequency spectrum. An-
other key parameter of healthcare monitoring is the arterial pulse
where VT and f are the tidal volume and respiration frequency of that contains vital information about the physiological status of
respiration, respectively. This work done is utilized as an input body. Arterial pulse is the manifestation of the pumping of the
stimulus for contact electrification. Subsequently, it is deployed blood, it embodies the two modes, namely, a systolic mode (oxy-
to examine the different breathing conditions, such as normal genated blood supplied from the heart to the different parts of the
breathing, after jogging and after fast running and W-TENG can body through arteries) and diastolic mode (influx of deoxygenated
effectively monitor the breathing at different rates (Figure 4b), blood into the heart) consolidately known as the heartbeat (Fig-
indicating that it has immense potential to recognize the diseases ure 4d).[39] W-TENG can record the arterial pulse as displayed in

Adv. Mater. Technol. 2023, 2300046 2300046 (5 of 10) © 2023 Wiley-VCH GmbH
 2365709x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/admt.202300046 by Bordeaux INP, Wiley Online Library on [05/04/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmattechnol.de

Figure 5. a) (i) Six determinants model and (ii) inverted pendulum model of human motion. b) Schematic of different modes (heel-strike, foot flat,
midstance, heel-off, toe-off) of human walking pattern. c) Voltage versus force plot, quantified the value of voltage with the input force. d) Periodic
pattern of human walking, enlarged region of the one peak shows the corresponding modes of walking.

Figure 4e consisting of the crucial information about the physi- and AIr are found at 0.375 s, 4.72 m s−1 , and 0.52, respectively,
ological health. As an interpretation of the wave form it consists which indicates the human health of a healthy person.[42]
of two modes: systolic and diastolic mode; in systolic mode, P
represents the peak pressure of blood flow from the heart to the
arteries known as early systolic pressure peak, and Q represents 2.4. Gait Analysis/Monitoring
the pressure that arises due to the reflection of blood through the
arteries known as late systolic pressure, while R represents the Gait analysis is another one of the critical fields of human health
pressure that arises due to diastolic pressure. There are different monitoring and growth analysis. The medical examination of
parameters to assess the vascular aging and cardiovascular risks; any patient based on his/her clinical assessment has three main
however, we have analyzed them by taking two important param- steps: history of the patient, physical examination, and specific
eters into consideration[37,40,41] investigation. For the patient who has a problem with walking,
both the history and physical examination are focused on the
h cause of the walking disorder. Here, gait analysis assumes to be
Reflection index (RI) = (8) a special investigation.[43,44] Other relevant techniques may be X-
Δt
rays, blood biochemistry, and electrodiagnostic tests. Gait analy-
( ) P sis better can be understood by the two famous models, first is
Radial augmentation index AIr = 2 (9) six determinants model and the other one is inverted pendulum
P1
model.[45] In six determinants model, the center of mass (COM)
where h is the height of the subjects, Δt is the delay between the moves in a horizontal path with a constant velocity ⃗vCOM . The
systolic and diastolic peak, and P1 and P2 are the systolic and di- stance leg generates a force F⃗ which supports body weight and
astolic pressure, respectively. Figure 4f represents the STFT spec- balance the COM by doing positive and negative work. During
trogram of the arterial pulse, clearly distinguishing the frequency double support (ds) both legs produce forces of F ⃖⃖⃖⃗2 [Figure
⃖⃖⃖⃗1 and F
spectrum and amplitudes of different modes in the spectrogram. 5a(i)]. The flattened structure requires a large amount of work
The time of delay between the systolic and diastolic peak (Δt), RI, done in single support (ss) and double support (ds). The variables

Adv. Mater. Technol. 2023, 2300046 2300046 (6 of 10) © 2023 Wiley-VCH GmbH
 2365709x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/admt.202300046 by Bordeaux INP, Wiley Online Library on [05/04/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmattechnol.de

Figure 6. Schematic overview of working mechanism of deep learning neural network/machine learning.

h is the height of COM from the base, s is the distance of step, When it is coupled with machine learning/deep learning,
and r is the radius of arc. In the inverted pendulum, the stance it becomes more powerful for utilization as Internet of Med-
leg is straight and COM moves in perpendicular arc. ical Things (IoMT), and helpful in producing the artificial
intelligence-driven medical equipments. Although here we have
The work rate can be estimated by the following equation[46,47] integrated different ML/DNN algorithms with the gesture recog-
nition signals generated by W-TENG and prospects the possibil-
S2 f
WR = √ (10) ity to use the gesture recognition with ML/DNN.[50–52] Different
4 4 − S2 classification models are used to train the model for classifying
the input datasets, i.e., k-means clustering (associate discussion
where S and f are the step length and step frequency, respec- S1, Figure S5, Supporting Information), DNN (associate discus-
tively. The features of six determinants model are that it pro- sion S2, Figure S6, Supporting Information), SVM (associate dis-
posed to minimize the energy expense by curtailing the vertical cussion S3, Figure S7, Supporting Information), KNN (associate
displacement of the COM, while in inverted pendulum model it discussion S4, Figure S8, Supporting Information), and pattern
follows more circular arc rather than horizontal path for COM recognition (associate discussion S5, Figure S9, Supporting In-
[Figure 5a(ii)]. These dynamic walking models provide a new ap- formation).
proach to explain the human gait features in a worthwhile man-
ner. In general, gait analysis consists of different steps in normal
walking, which are given in Figure 5b such as heel strike, foot 2.6. DNN/ML Results for the classification/Prediction of
flat, midstance, heel-off, and toe-off. W-TENG is calibrated by a Different Movements
voltage versus force plot (also known as sensitivity plot) as shown
in Figure 5c,d which demonstrates the force versus time plot A schematic of different finger postures is shown in Figure 7a,
that distinctively displays the force involved in different modes which is used as an input dataset for ML/DNN model.[53–56] We
of walking. The enlarged region of one cycle shows the different have used different machine learning algorithms, for instance,
peaks of the pattern with real-time walking. unsupervised k-means clustering (Figure 7b) which indicates fea-
tures are well separated from each other (as each cluster is vis-
ible far away from each other) and also well organized around
2.5. Development of Deep Learning/Machine Learning the centroid forming the splitted clusters. The deep neural net-
Algorithms for Tracking Different Movements work contains several layers of nodes, and these are like neurons
which are grouped into layers.[57,58] Accuracy is one of the impor-
Different deep learning and machine learning models are fur- tant metric that is used to quantify the measurement of correct
ther developed to classify different body movements. The gen- predictions/classifications. It can be calculated by the following
eral overview with the salient process of the machine learning equation
algorithms is represented in Figure 6, the workflow can be un-
derstood as input data collection, targeted model selection, spe- Number of correct predictions
cific algorithm execution followed by the output results. Gesture Accuracy =
Total number of predictions
recognition is one of the vital areas for advanced technological ∑
developments such as artificial limbs, automotive industry, self- TP + TN
= ∑ (11)
driving cars, and smart homes to name a few.[48,49] TP + TN + FP + FN

Adv. Mater. Technol. 2023, 2300046 2300046 (7 of 10) © 2023 Wiley-VCH GmbH
 2365709x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/admt.202300046 by Bordeaux INP, Wiley Online Library on [05/04/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmattechnol.de

Figure 7. a) Schematic of different hand gestures. b) k-means clustering plot of the gesture data, represents the features are well separated. Confusion
matrices related to different ML/DNN-based algorithms: c) DNN, d) SVM, e) KNN, and (f) pattern recognition.

The classification using the deep neural network with artifi- tion is the data analysis technique to recognize the pattern for the
cial neurons provides an efficiency of ≈94% (Figure 7c). The classification of the datasets which recognizes familiar patterns
support vector machine is the supervised learning algorithm, quickly and accurately which are used to build the confusion ma-
which selects the extreme points to create the hyperplane in an trix.
N-dimensional space that best fits to the classification of the data More specifically, we have used RNN as a pattern recognition
points. These extreme points are called support vectors, and the algorithm. It works on the principle of the recurrent connections
dimensions depend on the number of features. The support vec- that allows the network to take the previous data as an input for
tor machine classifies the datasets with an accuracy of ≈93% (Fig- predicting the next data which is the property of remembering
ure 7d). k-nearest neighbor is nonparametric based on the super- the states and analyze the pattern of the data based on the
vised learning technique, which assigns the class label based on remembrance. It shows an accuracy of ≈98% (Figure 7f).[59,60]
the similarity between the new case and the available cases, which For visualizing the different steps of the algorithm development
classifies with an accuracy of ≈97% (Figure 7e). Pattern recogni- such as training, testing, and validation, receiver operating

Adv. Mater. Technol. 2023, 2300046 2300046 (8 of 10) © 2023 Wiley-VCH GmbH
 2365709x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/admt.202300046 by Bordeaux INP, Wiley Online Library on [05/04/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmattechnol.de

characteristic (ROC) plots has also been calculated (Figure Characterization Techniques: FTIR spectroscopy measurement was
S10, Supporting Information). There are several measurement carried out by Thermofisher iS20 in attenuated total reflection (ATR) mode
uncertainties that affect the overall assessment of the algorithm; with resolution of 4 cm−1 from wavenumber range of 1000–3500 cm−1 .
Surface potential measurements were carried out by atomic force mi-
a comprehensive discussion on the measurement uncertainties
croscopy (CoreAFM, Nanosurf) with a cantilever of force constant 3 N m−1
and their overcoming techniques has been provided in the and resonance frequency 75 kHz. Open-circuit voltage and short-circuit
associate discussion S6 (Supporting Information). We have used current were acquired by Digital storage oscilloscope (DSOX 1102G,
different ML/DNN algorithms to classify the dataset, in which Keysight) and source measure unit (B29028A, Keysight), respectively.
pattern recognition based deep learning neural network (RNN) Development of Deep Learning Neural Network/ Machine Learning Algo-
gives the highest accuracy,[61] which is expected to be the fact rithms: All the machine learning and deep learning algorithms were cre-
ated and implemented by using python 3 and MATLAB 2021a. The input
that, it is working on the principle of pattern analysis of the
dataset was prepared by acquiring the different body movement signals
input data, extracts the crucial features of the data and learns with W-TENG, which was supplied as an input for the algorithms. Before
from it, also a counterproof of data reproducibility/reliability of execution of the algorithm, the data were preprocessed with the help of
W-TENG. The stability of W-TENG over time under repeated python libraries such as seaborn, matplotlib, numpy, and pandas to name
use is also assessed which is shown in Figure S11 (Supporting a few. A DNN network with 100 neurons was found optimal for attain-
Information), indicating the long-term stability of the device by ing the optimum accuracy. For visualizing the performance of the algo-
providing almost constant electrical response (current ≈0.4 μA) rithms during training, testing, and validation, ancillary results were also
produced such as receiver operating characteristics, error histogram, loss
over a longer time under constant pressure (≈4 kPa). function, and classification accuracy. Whole dataset was divided into three
parts, e.g., training, testing, and validation, that was splitted in 70%, 15%,
3. Conclusion and 15%, respectively.

An artificial intelligence-enabled highly sensitive perceptive

wearable sensor, i.e., gadget, has been fabricated by recycling the Supporting Information
waste material of nylon 66. The gadget is able to capture the nu- Supporting Information is available from the Wiley Online Library or from
merous vital biosignals such as breathing monitoring, gait analy- the author.
sis, cardiovascular measurements and also records different pos-
ture of the body movements, which was assisted in the early de-
tection of the neurological disorders such as Parkinson’s disease, Acknowledgements
REM sleep disorder, multiple sclerosis, and other neurological
The authors are grateful to the Science and Engineering Research Board
disorders. Early detection of the biomarkers of the disease as- (CRG/2020/004306), the Government of India for financially supporting
sists in timely diagnosis, which helps to avoid any fatality in the this work. A.B. and S.R. are thankful to the University Grants Commission
future. Numerous deep learning/machine learning algorithms (UGC) for providing the fellowship [1354/ (CSIR-UGC NET DEC. 2018)
have been employed to train the sensor for tracking different and 501800/ (CSIR-UGC NET June 2017)].
movements, particularly unsupervised and supervised learning
has been utilized, in which pattern recognition-based deep learn-
ing neural network shows the highest accuracy of ≈98%, while Conflict of Interest
DNN, SVM, and KNN show accuracy of ≈94%, ≈93%, and ≈97%, The authors declare no conflict of interest.
respectively. Thus, the wearable gadget equipped with artificial
intelligence could open up a new pathway to utilize the waste
material for acquiring the key indicators of human health mon- Author Contributions
itoring, which not only helps to reduce the waste materials that
A.B. performed the entire work, analyzed the data, and wrote the
provides a unique solution for healthcare monitoring and simul-
manuscript. S.R. and D.K.S. have helped to build the machine learn-
taneously it is also the impetus to the clean, green, and circular ing/deep neural network model and its associated calculations, D.M.
economy. planned the work and wrote the manuscript. All the authors read and ap-
prove before submission.
4. Experimental Section
Materials: Nylon 66 waste materials such as toothbrush bristles, cable Data Availability Statement
ties, defective machine tools, etc., were collected and cleaned with deion-
ized (DI) water followed by isopropyl alcohol. A solution of 10 wt% was The data that support the findings of this study are openly available
prepared by mixing the solvent mixture of trifluoroacetic acid and acetone in GitHub at https://github.com/ABnano/Gest-RECOG, reference num-
in a ratio of 6:4. The transparent solution was obtained by stirring the ny- ber 1.
lon waste material at 80 °C, overnight. Finally, a nylon film of 120 ±10 μm
was prepared by casting the solution on the clean glass slides at 60 °C.
Fabrication of W-TENG: To fabricate the sensor, silver paste was used Keywords
on both sides of the film, connected it with the electrical wires (0.5 mm).
For enhancing the flexibility of the sensor, it was encapsulated with poly- biomedical sensor, deep neural network, machine learning, movement dis-
dimethylsiloxane (PDMS), by mixing the elastomer and curing agent in the order, waste material, triboelectric nanogenerator
ratio of 10:1 and heated it 80 °C for 4 h. All the measurements with the W-
TENG were performed by the author himself as a volunteer in non-invasive Received: January 10, 2023
mode. Noteworthy to mention that the written consent from author vol- Revised: March 1, 2023
unteer was also taken in prior to the research. Published online:

Adv. Mater. Technol. 2023, 2300046 2300046 (9 of 10) © 2023 Wiley-VCH GmbH
 2365709x, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/admt.202300046 by Bordeaux INP, Wiley Online Library on [05/04/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmattechnol.de

[32] S. Niu, S. Wang, L. Lin, Y. Liu, Y. S. Zhou, Y. Hu, Z. L. Wang, Energy

[1] C. Cui, X. Wang, Z. Yi, B. Yang, X. Wang, X. Chen, J. Liu, C. Yang, ACS Environ. Sci. 2013, 6, 3576.
Appl. Mater. Interfaces 2018, 10, 3652. [33] J. Fu, X. Xia, G. Xu, X. Li, Y. Zi, ACS Nano 2019, 13, 13257.
[2] K. Meng, J. Chen, X. Li, Y. Wu, W. Fan, Z. Zhou, Q. He, X. Wang, [34] J. Wan, H. Bin Wang, L. M. Miao, H. Guo, H. T. Chen, X. L. Cheng,
X. Fan, Y. Zhang, J. Yang, Z. L. Wang, Adv. Funct. Mater. 2019, 29, H. X. Zhang, Proc. 14th Annu. IEEE Int. Conf. Nano/Micro, Eng. Mol.
1806388. Syst. NEMS 2019 2019, 229, 229.
[3] H. Yao, Z. Wang, Y. Wu, Y. Zhang, K. Miao, M. Cui, T. Ao, J. Zhang, D. [35] A. B. Otis, W. O. Fenn, H. Rahn, J. Appl. Physiol. 1950, 2, 592.
Ban, H. Zheng, Adv. Funct. Mater. 2022, 32, 2112155. [36] R. M. Peters, Ann. Thorac. Surg. 1969, 7, 51.
[4] Y. C. Lai, J. Deng, S. L. Zhang, S. Niu, H. Guo, Z. L. Wang, Adv. Funct. [37] J. F. Fie, M. Hendryx, C. M. Helms, J. Gen. Intern. Med. 1990, 8, 354.
Mater. 2017, 27, 1604462. [38] J. Li, Y. Long, F. Yang, X. Wang, EcoMat 2020, 2, e12045.
[5] A. Babu, P. Malik, N. Das, D. Mandal, Small 2022, 18, 2201331. [39] W. W. Nichols, Am. J. Hypertens. 2005, 18, 3.
[6] M. Wang, J. Zhang, Y. Tang, J. Li, B. Zhang, E. Liang, Y. Mao, X. Wang, [40] A. P. Avolio, M. Butlin, A. Walsh, Physiol. Meas. 2010, 31, R1.
ACS Nano 2018, 12, 6156. [41] S. K. Ghosh, D. Mandal, ACS Sustainable Chem. Eng. 2017, 5, 8836.
[7] T. Bhatta, S. Sharma, K. Shrestha, Y. Shin, S. Seonu, S. Lee, D. Kim, [42] A. Babu, S. Ranpariya, D. K. Sinha, A. Chatterjee, D. Mandal, Adv.
M. Sharifuzzaman, S. S. Rana, J. Y. Park, Adv. Funct. Mater. 2022, 32, Mater. Technol. 2023, 2202021.
2202145. [43] A. Saboor, T. Kask, A. Kuusik, M. M. Alam, Y. Le Moullec, I. K. Niazi,
[8] J. Xiong, P. Cui, X. Chen, J. Wang, K. Parida, M. F. Lin, P. S. Lee, Nat. A. Zoha, R. Ahmad, IEEE Access 2020, 8, 167830.
Commun. 2018, 9, 4280. [44] R. Alkhatib, M. O. DIab, C. Corbier, M. El Badaoui, IEEE Sensors Lett.
[9] Y. Su, G. Chen, C. Chen, Q. Gong, G. Xie, M. Yao, H. Tai, Y. Jiang, J. 2020, 4, 4.
Chen, Adv. Mater. 2021, 33, 2101262. [45] J. Li, Z. Wang, Z. Zhao, Y. Jin, J. Yin, S. L. Huang, J. Wang, Ubi-
[10] G. Chen, C. Au, J. Chen, Trends Biotechnol. 2021, 39, 1078. Comp/ISWC 2021 – Adjun. Proc. 2021 ACM Int. Jt. Conf. Pervasive
[11] X. Pu, S. An, Q. Tang, H. Guo, C. Hu, iScience 2021, 24, 102027. Ubiquitous Comput. Proc. 2021 ACM Int. Symp. Wearable Comput.
[12] O. O. Ayeleru, S. Dlova, O. J. Akinribide, F. Ntuli, W. K. Kupolati, P. F. 2021, https://doi.org/10.1145/3460418.3480410
Marina, A. Blencowe, P. A. Olubambi, Waste Manage. 2020, 110, 24. [46] A. D. Kuo, Hum. Mov. Sci. 2007, 26, 617.
[13] W. C. Li, H. F. Tse, L. Fok, Sci. Total Environ. 2016, 566, 333. [47] A. D. Kuo, J. M. Donelan, Phys. Ther. 2010, 90, 157.
[14] M. Shams, I. Alam, M. S. Mahbub, Environ. Adv. 2021, 5, 100119. [48] K. Dermitzakis, A. H. Arieta, R. Pfeifer, Proc. Annu. Int. Conf. IEEE Eng.
[15] J. M. margolis, Engineering Plastics Handbook, McGraw-Hill, Canada Med. Biol. Soc. EMBS 2011, 4530, 22255345.
2006. [49] F. Wen, Z. Sun, T. He, Q. Shi, M. Zhu, Z. Zhang, L. Li, T. Zhang, C.
[16] Y. Lecun, Y. Bengio, G. Hinton, Nature 2015, 521, 436. Lee, Adv. Sci. 2020, 7, 2000261.
[17] M. I. Jordan, T. M. Mitchell, Science 2015, 349, 255. [50] S. Khan, Ethem Alpaydin. Introduction to Machine Learning, Adaptive
[18] A. L. Samuel, IBM J. Res. Dev. 2000, 44, 206. Computation and Machine Learning Series, The MIT Press, Cambridge,
[19] B. M. Henrique, V. A. Sobreiro, H. Kimura, Expert Syst. Appl. 2019, USA 2008.
124, 226. [51] P. Tan, X. Han, Y. Zou, X. Qu, J. Xue, T. Li, Y. Wang, R. Luo, X. Cui,
[20] J. Yang, A. Mannodi-Kanakkithodi, MRS Bull. 2022, 47, 940. Y. Xi, L. Wu, B. Xue, D. Luo, Y. Fan, X. Chen, Z. Li, Z. L. Wang, Adv.
[21] L. A. Díaz-Alejo, E. C. Menchaca-Campos, J. Uruchurtu Chavarín, R. Mater. 2022, 34, 2200793.
Sosa-Fonseca, M. A. García-Sánchez, Int. J. Polym. Sci. 2013, 2013, [52] M. Wang, T. Wang, Y. Luo, K. He, L. Pan, Z. Li, Z. Cui, Z. Liu, J. Tu, X.
323854. Chen, Adv. Funct. Mater. 2021, 31, 2008807.
[22] C. García-Pérez, C. Menchaca-Campos, M. A. García-Sánchez, E. [53] Y. Al-Handarish, O. M. Omisore, J. Chen, X. Cao, T. O. Akinyemi, Y.
Pereyra-Laguna, O. Rodríguez-Pérez, J. Uruchurtu-Chavarín, Open J. Yan, L. Wang, Appl. Sci. 2021, 11, 7264.
Compos. Mater. 2017, 7, 146. [54] V. Moysiadis, D. Katikaridis, L. Benos, P. Busato, A. Anagnostis, D.
[23] X. Hu, X. Wang, ICMREE 2013 – Proc. 2013 Int. Conf. Mater. Renew. Kateris, S. Pearson, D. Bochtis, Appl. Sci. 2022, 12, 8160.
Energy Environ. 2013, 2, 585. [55] H. Zhou, W. Huang, Z. Xiao, S. Zhang, W. Li, J. Hu, T. Feng, J. Wu, P.
[24] M. Nasr-Esfahani, M. Mehrabanian, Int. J. Nanomed. 2011, 6, 1651. Zhu, Y. Mao, Adv. Funct. Mater. 2022, 32, 2208271.
[25] Y. Fang, Y. Zou, J. Xu, G. Chen, Y. Zhou, W. Deng, X. Zhao, M. Roustaei, [56] ACS Appl. Mater. Interfaces 2022, 14, 25629.
T. K. Hsiai, J. Chen, Adv. Mater. 2021, 33, 2104178. [57] A. Moin, A. Zhou, A. Rahimi, A. Menon, S. Benatti, G. Alexandrov, S.
[26] H. T. Baytekin, A. Z. Patashinski, M. Branicki, B. Baytekin, S. Soh, B. Tamakloe, J. Ting, N. Yamamoto, Y. Khan, F. Burghardt, L. Benini, A.
A. Grzybowski, Science 2011, 333, 308. C. Arias, J. M. Rabaey, Nat. Electron. 2021, 4, 54.
[27] J. H. Kim, B. K. Yun, J. H. Jung, J. Y. Park, Appl. Phys. Lett. 2016, 108, [58] Y. Djenouri, J. C. Lin, A. Yazidi, Expert Syst. 2022, e113093.
133901. [59] J. Wang, Y. Chen, S. Hao, X. Peng, L. Hu, Pattern Recognit Lett 2019,
[28] R. S. Blacker, A. W. Birley, Polym. Test. 1991, 10, 241. 119, 3.
[29] W. N. Hansen, G. J. Hansen, Surf. Sci. 2001, 481, 172. [60] I. Sutskever, O. Vinyals, Q. V. Le, Proceedings of the 27th International
[30] P. Taylor, J. Lowell, Adv. Phys. 1980, 29, 947. Conference on Neural Information Processing Systems, 2014, 2, 3104.
[31] Y. Chu, R. Han, F. Meng, Z. Cao, S. Wang, K. Dong, S. Yang, H. Liu, [61] D. Liu, D. Zhang, Z. Sun, S. Zhou, W. Li, C. Li, W. Li, W. Tang, Z. L.
X. Ye, F. Tang, Nano Energy 2021, 89, 106383. Wang, Adv. Funct. Mater. 2022, 32, 2113008.

Adv. Mater. Technol. 2023, 2300046 2300046 (10 of 10) © 2023 Wiley-VCH GmbH

View publication stats