processes

Review
Prospects and Challenges of AI and Neural Network
Algorithms in MEMS Microcantilever Biosensors
Jingjing Wang *, Baozheng Xu , Libo Shi, Longyang Zhu and Xi Wei

School of Electronic and Information Engineering, Tiangong University, Tianjin 300380, China
* Correspondence: [email protected]

Abstract: This paper focuses on the use of AI in various MEMS (Micro-Electro-Mechanical System)
biosensor types. Al increases the potential of Micro-Electro-Mechanical System biosensors and opens
up new opportunities for automation, consumer electronics, industrial manufacturing, defense,
medical equipment, etc. Micro-Electro-Mechanical System microcantilever biosensors are currently
making their way into our daily lives and playing a significant role in the advancement of social
technology. Micro-Electro-Mechanical System biosensors with microcantilever structures have a num-
ber of benefits over conventional biosensors, including small size, high sensitivity, mass production,
simple arraying, integration, etc. These advantages have made them one of the development avenues
for high-sensitivity sensors. The next generation of sensors will exhibit an intelligent development
trajectory and aid people in interacting with other objects in a variety of scenario applications as
a result of the active development of artificial intelligence (AI) and neural networks. As a result,
this paper examines the fundamentals of the neural algorithm and goes into great detail on the
fundamentals and uses of the principal component analysis approach. A neural algorithm application
in Micro-Electro-Mechanical System microcantilever biosensors is anticipated through the associated
application of the principal com-ponent analysis approach. Our investigation has more scientific
study value, because there are currently no favorable reports on the market regarding the use of AI
Citation: Wang, J.; Xu, B.; Shi, L.; with Micro-Electro-Mechanical System microcantilever sensors. Focusing on AI and neural networks,
Zhu, L.; Wei, X. Prospects and this paper introduces Micro-Electro-Mechanical System biosensors using artificial intelligence, which
Challenges of AI and Neural greatly promotes the development of next-generation intelligent sensing systems, and the potential
Network Algorithms in MEMS applications and prospects of neural networks in the field of microcantilever biosensors.
Microcantilever Biosensors. Processes
2022, 10, 1658. https://doi.org/ Keywords: microcantilever; AI; neural network; biosensors; MEMS
10.3390/pr10081658

Academic Editor: Zhiwei Gao

Received: 21 July 2022

1. Introduction
Accepted: 19 August 2022
Published: 21 August 2022
Sensor technology is evolving as a result of advances in science technology and the
popularization of information technology. Sensor development has gone through three
Publisher’s Note: MDPI stays neutral generations thus far. Structural sensors make up the initial generation of sensors. By
with regard to jurisdictional claims in utilizing the pertinent sensor system variables, this kind of sensor can convert the signal
published maps and institutional affil-
and detect changes in the signal. A solid-state sensor is the second kind of sensor and is
iations.
made of semiconductors, electrolytes, and magnetic components. In the 1970s, this kind of
sensor started to evolve. This generation of sensors, which currently account for about 75%
of the sensor industry, are often characterized by low cost, high accuracy, and outstanding
Copyright: © 2022 by the authors.
performance. The smart sensor that this paper primarily introduces is the third-generation
Licensee MDPI, Basel, Switzerland.
sensor. Miniaturization and intelligence are currently the two major requirements for
This article is an open access article
sensor systems due to the growth of the Internet and the improvement in human demands.
distributed under the terms and Microelectromechanical systems (MEMS) are integrated at the nanoscale and microscale.
conditions of the Creative Commons As a result, MEMS sensors differ from conventional sensors in that they are miniaturized,
Attribution (CC BY) license (https:// integrated with microelectronics, and manufactured in parallel with great precision. The
creativecommons.org/licenses/by/ potential use of MEMS and nanotechnology in the sensing sector is typically thought of
4.0/). due to their lower size, simpler integration into systems, better portability, and mobility.

Processes 2022, 10, 1658. https://doi.org/10.3390/pr10081658 https://www.mdpi.com/journal/processes

Processes 2022, 10, 1658 2 of 25

Since the atomic force microscope in 1986 [1], its advantages of label-free, high sensi-
tivity, portable, low cost, and fast response have attracted a large number of researchers to
explore and study. Since then, MEMS-based biosensors have continued to develop; in the
meantime, many types of MEMS biosensors such as optical acoustics have emerged [2–5].
Efficient, fast, and sensitive, it can quickly acquire and process information, just the same
as the human body’s sensory organs; it can perceive the external environment and perceive
important physical information, such as sound, light, pressure, and temperature. Due
to these advantages, MEMS biosensors are widely used in automation, aerospace, con-
sumer electronics, defense, industrial manufacturing, medical devices, life sciences, and
telecommunications [6–8].
AI is being used increasingly for connecting and interacting with humans and com-
puters. Users get a more immersive experience with this interactive system, which contains
features of future sensors. It is utilized in a variety of application scenarios, including
entertainment, medical rehabilitation, sports training simulation, and more (as depicted in
Figure 1) [9].

Figure 1. The development trend of future biosensors.

2. Research State of AI Applications in Biosensors

Artificial intelligence has advanced quickly and is currently a popular field of study
in technology. It uses densely connected networks to process information in a manner
that mimics how the human brain processes information. It is capable of self-learning,
parallel processing, and powerful information storage. The fields of application for artificial
intelligence are growing as the field’s theory and technology mature. People are already
able to see how artificial intelligence and MEMS sensors work together.
 Artificial intelligence has advanced quickly and is currently a popular field of study
in technology. It uses densely connected networks to process information in a manner that
mimics how the human brain processes information. It is capable of self-learning, parallel
processing, and powerful information storage. The fields of application for artificial intel-
Processes 2022, 10, 1658 ligence are growing as the field’s theory and technology mature. People are already 3 ofable
25
to see how artificial intelligence and MEMS sensors work together.

2.1.Gas
2.1. GasSensing
Sensing Field
Field
Gasmonitoring
Gas monitoringsystems
systemsare areemployed
employedininaavariety
varietyofofsettings,
settings,including
includingcommercial
commercial
and residential settings, especially in the detection of dangerous
and residential settings, especially in the detection of dangerous gases. The tiny size gases. The tiny sizeand
and
great sensitivity of semiconductor-type resistive gas sensors make
great sensitivity of semiconductor-type resistive gas sensors make them appealing. They them appealing. They
arealso
are alsoinexpensive
inexpensiveand andsimple
simpletotocreate.
create.These
Thesebenefits
benefitsimply
implythat
thatsemiconductor-based
semiconductor-based
gassensors
gas sensorsare areananexcellent
excellentoption
optionforforInternet
InternetofofThings
Thingsapplications.
applications.
Suhpublished
Suh publishedaafully fullyintegrated
integratedportable
portablemulti-gas
multi-gassensorsensormodule
modulefor forInternet
Internetofof
Thingsapplications
Things applications[10]. [10].For
ForInternet
InternetofofThings
Things(IoT)
(IoT)applications
applicationssuch suchasasmultiple
multiplegas gas
sensor reading and data analysis, analog/digital signal processing,
sensor reading and data analysis, analog/digital signal processing, heater control, and heater control, and
wirelesscommunication,
wireless communication,Suh Suhoffers
offersa aportable
portablegas gassensing
sensingmodule
modulewithwithananultra-compact
ultra-compact
MEMS
MEMSgas gassensor
sensordevice.
device.AAmicroprocessor
microprocessoralgorithm’s
algorithm’ssimplified
simplifiedschematic
schematicdiagram
diagramisis
shown
shownininFigure
Figure2.2.(ESP32).
(ESP32).The TheAFE
AFEcircuit
circuitelements
elementsundergo
undergoaanumber
numberofofinitializations
initializations
when
whenthe thesystem
systeminitially
initiallypowers
powersup, up,and
andaaramprampvoltage
voltageisisgenerated
generatedtotobring
bringthetheheater
heater
power to its goal level (VDAC). The microprocessor can compute
power to its goal level (VDAC). The microprocessor can compute the power and maintain the power and maintain
the
thedesired
desiredpowerpowerby bymodifying
modifyingthe theVDAC
VDACfor forthe
theheater
heatercontrol
controlalgorithm
algorithmby bykeeping
keeping
track
trackofofthe
thecurrent
currentpassing
passingthrough
throughthe theheater
heater(Iheat).
(Iheat).Rf Rfought
oughttotobe
beautomatically
automaticallychosen
chosen
ininorder
ordertoto read
readthethe
voltage
voltagein real time.
in real Do Do
time. not not
let the
let TIA output
the TIA voltage
output become
voltage saturated
become satu-
(Vout). Select aSelect
rated (Vout). loweraRF when
lower RFthe voltage
when is almost
the voltage is zero
almostvolts.
zeroSelect
volts.2.1 V as2.1
Select theVupper
as the
limit
upperforlimit
RF change
for RF instead.
change instead.

Figure 2. Schematic diagramFigure Schematicprogrammed

2. algorithm
of the diagram of the
onalgorithm
the ESP32programmed on the
microprocessor. TheESP32 microprocessor.
first procedure Thea first
generates
ramp voltage to initialize the circuit components in the AFE and set each heating power to the target value (VDAC).
procedure generates a ramp voltage to initialize the circuit components in the AFE and set Byeach
heating power to the target value (VDAC). By monitoring Iheat and subsequently changing VDAC,
the heater control keeps the heating power at the desired level. In order to prevent the TIA output
from becoming saturated, the proper Rf is automatically chosen for the Rsens readout control.

No intermediary hardware is needed, because the system fully incorporates RF, AFE,
digital signal processing, and sensors, making it suitable for IoT application right now. The
technology is portable and powerful, and it can be used to monitor the air quality in places
like homes, cars, and manufacturing facilities, as shown in Figure 3.
 order to prevent the TIA output from becoming saturated, the proper Rf is automatically chosen for the Rsens readout
control.

No intermediary hardware is needed, because the system fully incorporates RF, AFE,
digital signal processing, and sensors, making it suitable for IoT application right now.
Processes 2022, 10, 1658 The technology is portable and powerful, and it can be used to monitor the air quality4 in
of 25
places like homes, cars, and manufacturing facilities，as shown in Figure 3.

Figure 3. Data transmission by Figure 3. Data transmission

transmitting raw data from bymultiple
transmitting rawover
sensors dataWi-Fi.
from multiple sensors over
(a) Data transfer over Wi-Fi.
a Wi-Fi(a) Data
access point that is connected transfer over a Wi-Fi
to the Internet. In thisaccess pointthe
instance, that is connected
sensor to the Internet.
data is recorded In thisSpreadsheet
in a Google instance, the sensor
file and data is
displayed as graphs. (b) Data recorded in a Google
transmission via a BLESpreadsheet file and
to an Android displayedthat
application as graphs.
creates(b) Data of
graphs transmission
the sensor via
dataa in
BLE to an
real-time. Android application that creates graphs of the sensor data in real-time.

2.2.
2.2. SoundDetection
Sound DetectionField
Field
Speech
Speech recognition
recognition is is
thethe most
most user-friendly
user-friendly interface
interface forfor two-way
two-way communication
communication
between people and smart devices; however, voice user interfaces (VUI) andbiometric
between people and smart devices; however, voice user interfaces (VUI) and biometricsys-
tems based on artificial intelligence (AI) and the Internet of Things (IoT) are
systems based on artificial intelligence (AI) and the Internet of Things (IoT) are also gar- also garnering
a lot of
nering attention.
a lot In terms
of attention. of sound
In terms detection,
of sound Han etHan
detection, al. proposed a platform
et al. proposed for speaker
a platform for
recognition utilizing machine learning-assisted approaches, as shown
speaker recognition utilizing machine learning-assisted approaches, as shown in Figure in Figure 4. The test
data consisted of 40 individuals, and the machine learning speaker recognition
4. The test data consisted of 40 individuals, and the machine learning speaker recognition algorithm
used was based on the Gaussian Mixture Model (GMM). There are 2800 training data in it.
algorithm used was based on the Gaussian Mixture Model (GMM). There are 2800 train-
They assert that, using the most and second-most sensitive data from the multi-channel
ing data in it. They assert that, using the most and second-most sensitive data from the
output, Gaussian Mixture Models (GMMs) in machine learning obtain a speaker recog-
multi-channel output, Gaussian Mixture Models (GMMs) in machine learning obtain a
nition rate of 97.5% and a 75% reduction in error when compared to a reference MEMS
microphone [11].
Processes 2022, 10, x FOR PEER REVIEW 5 of 26

Processes 2022, 10, 1658 5 of 25

speaker recognition rate of 97.5% and a 75% reduction in error when compared to a refer-
ence MEMS microphone [11].

Figure
Figure4. 4.
AA platform
platform forfor
speaker
speakerrecognition
recognitionusing machine
using machinelearning-assisted
learning-assisted methods.
methods. (a)(a)
Speaker
Speaker
testing and training utilizing the common TIDIGITS dataset and the GMM algorithm
testing and training utilizing the common TIDIGITS dataset and the GMM algorithm (20 men (20 men and
and
20 female speakers, 77 speeches per each speaker, and a total of 3080 voice data). Ninety percent of
20 female speakers, 77 speeches per each speaker, and a total of 3080 voice data). Ninety percent of
the TIDIGITS dataset is utilized for training purposes, and ten percent is used for testing purposes.
the TIDIGITS dataset is utilized for training purposes, and ten percent is used for testing purposes.
(b) The trained STFT features from 2800 training data from 40 individuals are shown in an t-SNE
(b) The t-SNE
plot. The trainedgraphic
STFT features from
integrates 2800 training
similar data from 40 individuals
objects’ high-dimensional data intoarea shown in an t-SNE
low-dimensional
plot. The t-SNE graphic integrates similar objects’ high-dimensional data into a low-dimensional
space associated with probability distribution. (c) In the event that a 12th speaker is discovered, the
space associated
majority with probability
voting approach is used to distribution. (c) In
test the speaker the event mechanism
recognition that a 12th speaker
throughoutis discovered,
the frames.the
majority voting approach is used to test the speaker recognition mechanism throughout the frames.
Machine learning-based training on TIDIGITS speech data (40 people, 2800 speeches)
Machine using
was performed learning-based
the GMMtraining on TIDIGITS
algorithm, which hasspeech data (40 people,
been modified 2800 speeches)
for multi-signal pro-
was performed using the GMM algorithm, which has been modified
cessing. A randomly selected person’s speech is compared with the trained speech data for multi-signal
setprocessing.
for speakerArecognition.
randomly selected person’s speech is compared with the trained speech data
set In
foraddition,
speaker recognition.
as illustrated in Figure 5, Jung et al. reported a flexible piezoelectric
In addition, as illustrated in Figure 5, Jung et al. reported a flexible piezoelectric
acoustic sensor and machine learning for speech processing. Flexible piezoelectric acous-
acoustic sensor and machine learning for speech processing. Flexible piezoelectric acoustic
tic sensors vibrate in response to the speaker’s voice to store electrical impulses, providing
sensors vibrate in response to the speaker’s voice to store electrical impulses, providing
information for preprocessing. To train data and extract linguistic information from
information for preprocessing. To train data and extract linguistic information from voices,
voices, use machine learning-based models. The transition of electronic systems from
use machine learning-based models. The transition of electronic systems from touch to
touch to voice operation will be assisted by this technique. They assert that new user in-
voice operation will be assisted by this technique. They assert that new user interfaces for
terfaces for utilizing artificial intelligence (AI) services will be created by speech recogni-
utilizing artificial intelligence (AI) services will be created by speech recognition systems
tion systems built on cutting-edge sound sensors and optimized machine learning algo-
built on cutting-edge sound sensors and optimized machine learning algorithms [12].
rithms [12].
Due to its simplicity and two-way communication, voice user interface (VUI), the
Due to its simplicity and two-way communication, voice user interface (VUI), the
foundational technology of the Internet of Things (IoT) and Artificial Intelligence (AI), has
foundational technology of the Internet of Things (IoT) and Artificial Intelligence (AI), has
garnered significant interest. Smart acoustic sensors can be applied to a variety of industries,
garnered
including significant interest.
biometrics, smartSmart
home acoustic
products,sensors can berecognition.
and speaker applied to a Speech
varietyrecognition
of indus-
tries, including
software biometrics,
transforms audiosmart home
data into products,
a binary and
digital speaker
format for recognition. Speech
machine learning recog-
algorithms.
nition software transforms audio data into a binary digital format for
Recent developments in deep learning have considerably enhanced the voice processing machine learning
algorithms. Recent developments
task performance, outperforminginthat deepof learning have
traditional considerably
machine learningenhanced
methods.the voice
However,
processing task performance, outperforming that of traditional machine learning
these systems still have poor identification rates because of sensitive hardware problems meth-
ods.
andHowever, thesedata.
a lack of audio systems still
Future have technologies
speech poor identification rates on
should focus because of sensitive
the synergy between
smart acoustic sensors and AI algorithms to overcome the fundamental weakness of speech
recognition [13–15].
Processes 2022, 10, x FOR PEER REVIEW 6 of 26

Processes 2022, 10, 1658 hardware problems and a lack of audio data. Future speech technologies should focus on25
6 of
the synergy between smart acoustic sensors and AI algorithms to overcome the funda-
mental weakness of speech recognition [13–15].

Figure 5. Schematic diagram of the application prospect of the voice user interface platform. By
Figure 5. Schematic diagram of the application prospect of the voice user interface platform. By
vibrating in response to the speaker’s voice, flexible piezoelectric acoustic sensors convert utterances
vibrating in response to the speaker’s voice, flexible piezoelectric acoustic sensors convert utterances
to electrical multi-signals, which can give digitalized data for preprocessing. Language information
istocollected
electricalfrom
multi-signals,
the voice which candata
once the giveare
digitalized data for
trained using preprocessing.
a machine Languagemodel.
learning-based information
This
is collected
procedure from
will aidthe voice
in the once theofdata
transition are trained devices
touch-operated using a to
machine learning-based
sound-operated ones. model. This
procedure will aid in the transition of touch-operated devices to sound-operated ones.
2.3. Wearable Sensing Field
2.3. Wearable Sensing Field
Smart wearables and IoT-based clothes have entered our field of view with the rise
Smart wearables and IoT-based clothes have entered our field of view with the rise of
of the Internet of Things (IoT) and Artificial Intelligence (AI), and they are becoming more
the Internet of Things (IoT) and Artificial Intelligence (AI), and they are becoming more
and more vital to us due to the unparalleled coordination and convenience, as well as the
and more vital to us due to the unparalleled coordination and convenience, as well as the
enjoyment given by fashion. Life has a profound effect. Today, seamless and widespread
enjoyment given by fashion. Life has a profound effect. Today, seamless and widespread
integration of sensors into textiles is made possible by the quick union of textiles and elec-
integration of sensors into textiles is made possible by the quick union of textiles and
tronics. The Internet
electronics. era ofera
The Internet smart clothes
of smart has arrived
clothes with with
has arrived the development of smart
the development fab-
of smart
rics that can communicate with smartphones to process physiological data
fabrics that can communicate with smartphones to process physiological data like heart like heart rate,
temperature, respiration,
rate, temperature, pressure,
respiration, movement,
pressure, acceleration,
movement, and even
acceleration, hormone
and even levels.
hormone levels.
According to Tiago M., a garment must contain the fundamental subsystems
According to Tiago M., a garment must contain the fundamental subsystems depicted de-
picted in Figure 6 in order to be a part of the future smart clothing Internet. Wireless
in Figure 6 in order to be a part of the future smart clothing Internet. Wireless technology tech-
nology or conductive
or conductive fabrics
fabrics can bothcan both for
be used be subsystem
used for subsystem communication
communication [16].
[16]. The first The
option
first option often costs more in terms of technology and uses more energy,
often costs more in terms of technology and uses more energy, but it does away with the but it does
away
need to with the and
create needincorporate
to create and incorporate
conductive yarn conductive yarn into
into smart clothes. Newsmart clothes.fabrics
conductive New
conductive fabrics and printed electronics will make it possible to seamlessly
and printed electronics will make it possible to seamlessly integrate sensors on a broad integrate
sensors
scale intoontextiles
a broad[17,18].
scale into textiles [17,18].
The architecture consists of the following main components:
• Communication gateway, exchanging information with smart clothing in order to
send information to cloud server or blockchain via internet or intranet [19].
• Cloud servers that collect and store data and provide certain remote services for smart
clothing and remote users.
• A blockchain. It is not essential to the basic functions of the smart clothing system.
Additionally, the Internet of Things (IoT) and wearable technology, in conjunction with
device-to-device communication (D2D) [20], virtual/augmented reality (VR/AR) [21,22],
cyber-physical systems (CPS) [23], artificial intelligence (AI) [24], and smart textiles [25], as
well as other developments in 5G communication networks, can enhance human-to-human
and human-to-machine connections and interactions.
 Processes
Processes 2022, FOR10,
10, x2022, 1658REVIEW
PEER 7 of 26 7 of 25

Figure 6.Figure 6. General

General architecture
architecture of IoT
of IoT smart smart clothing
clothing system. system.

2.4. Body Sensor Field

The architecture consists of the following main components:
Body sensors have drawn a lot of interest from scientists lately because of their useful
• Communication gateway, exchanging information with smart clothing in order to
uses in the area of smart medicine. Body sensors are now widely used in a variety of
send information to cloud server or blockchain via internet or intranet [19].
real-world settings, including entertainment, security, health, and healthcare. The ability of
• Cloud servers that collect and store data and provide certain remote services for
body sensors to protect and enhance people’s healthy lifestyles is a significant benefit of
smart clothing and remote users.
employing them to monitor individuals. Body sensor-based human motion detection yields
• A blockchain. It is not essential
insightful information aboutto a the basic functions
person’s functioningof and
the smart
way ofclothing system.
life. Figure 7 illustrates
Additionally, the Internet
Uddin’s proposed bodyofsensor-based
Things (IoT)behavior
and wearable technology,
identification in conjunction
system, in which a person
with device-to-device
wears various body communication (D2D) [20],
sensors on various body virtual/augmented
areas, including thereality (VR/AR)
wrist, ankle, and chest.
[21,22], Sensor
cyber-physical
data is acquired via wireless media and saved to a computer. The basictex-
systems (CPS) [23], artificial intelligence (AI) [24], and smart flow of the
tiles [25], as well
system is as otherindevelopments
shown Figure 8, whichin 5G
hascommunication
two basic steps:networks, cantesting
training and enhance hu-
[26].
man-to-human Theand human-to-machine
authors also employ the connections
promisingand interactions.
deep learning method known as the deep
recurrent neural network (RNN), which is based on sequence data. Research interest
2.4. Bodyin Sensor Field
recurrent neural networks (RNNs) for evaluating rhythmic occurrences in time series
applications
Body sensors have is on the rise.
drawn a lotCompared to other
of interest from deep learning
scientists techniques
lately because of theircurrently
useful used,
it can
uses in the offer
area stronger
of smart discrimination.
medicine. As aare
Body sensors result,
nowvarious
widely activities are trained
used in a variety and recog-
of real-
nized using
world settings, humanentertainment,
including sensor-based RNNs
security,[27–31]. The
health, andfollowing table
healthcare. Theis used to of
ability show the
comparison of AI applied to different types of biosensors, as shown
body sensors to protect and enhance people’s healthy lifestyles is a significant benefit ofin Table 1.
employing them to monitor individuals. Body sensor-based human motion detection
yields insightful information about a person’s functioning and way of life. Figure 7 illus-
trates Uddin’s proposed body sensor-based behavior identification system, in which a
person wears various body sensors on various body areas, including the wrist, ankle, and
chest. Sensor data is acquired via wireless media and saved to a computer. The basic flow
of the system is shown in Figure 8, which has two basic steps: training and testing [26].
 Processes2022,
Processes
Processes 2022, 10,xx1658
2022,10,
10, FORPEER
FOR PEERREVIEW
REVIEW 88 8of
ofof26
25
26

Figure
Figure 7.
Figure7. A
7.A schematic
Aschematic device
schematicdevice of
ofaaahuman
deviceof human activity
humanactivity prediction
activityprediction system
predictionsystem based
systembased on
basedon human
onhuman sensors.
humansensors.
sensors.

Figure 8.
Figure Thebasic
8. The basicflow
flow of
of the
the proposed
proposed body
body sensor-based
sensor-based physical
physical activity
activity recognition
recognition system.
system.
Figure 8. The basic flow of the proposed body sensor-based physical activity recognition system.
 recurrent
parison of neural network
AI applied (RNN), which
to different types ofis biosensors,
based on sequence
as shown data. Research
in Table 1. interest in
recurrent neural networks (RNNs) for evaluating rhythmic occurrences in time series ap-
plications
Table 1. Theis on the rise.ofCompared
comparison AI applied to different
other deep learning
types techniques currently used, it
of biosensors.
can offer stronger discrimination. As a result, various activities are trained and recognized
Type Schematic Application Advantages
using human sensor-based RNNs [27–31]. The following table is used to show the com-
Processes 2022, 10, 1658of AI applied to different types of biosensors, as shown in Table 1.
parison 9 of 25

Table 1. The comparison of AI applied to different types of biosensors

Table 1. The comparison of AI applied to different types of biosensors.
Type Multi-sensor data read-
Table 1. Schematic Application
The comparison of AI applied to different types of biosensors. Advantages
Type Schematic Application
ing and analysis, Advantages
Type Schematic Analog/Digital
Application small size, Advantages
Gas Sensor
Signal Processing, powerfulWidely
Wireless Communica-
Multi-sensor data reading and
Multi-sensor data read-
tion small size,
ing analysis,
and analysis,
Multi-sensor data reading and
Gas Sensor analysis, small size,
Analog/Digital
Gas Sensor
Gas Sensor Analog/Digital
Analog/Digital powerful,
small size,
Signal Processing, powerfulWidely powerfulWidely
Signal Processing,
Wireless Communica-
Signal Processing,
Wireless Communication Widely
tion
Wireless Communication

Speech Recognition
Algorithm Improve- Convenient
Acoustic Sen-
sor Speechment
Recognition Bidirectional
Convenient
Speech
Multiple SignalRecognition
Pro- Widely Convenient
Acoustic Sensor
Acoustic Sensor Algorithm
cessingImprovement
Algorithm Improvement Bidirectional
Bidirectional
Multiple
Speech Signal Processing
Recognition Widely
Algorithm Improve-
Multiple ConvenientWidely
Signal Processing
Acoustic Sen-
ment Bidirectional
sor
Multiple Signal Pro- Widely
cessing
Improve Connec-
Improve Connec-
Human-computer in- tion and Interac- tion and Connection
Improve Interac- and
teraction
Human-computer tion
interaction tion
Interaction
Wearable
Wearable Sensor Human-computer
RemoteRemote
Service interaction
Service Great Business Value
Great Business
Sensor Wireless communication Health Check
Wireless communica- Value Great Business
Wearable Sensor Remote Service
Improve Convenient
Connec-
tion Value
Health Check
2, 10, x FOR PEER REVIEW Human-computer in- tion and Interac-
Convenient10 of 26
Wireless communication
teraction tion Health Check
Wearable
Remote Service Great Business
Sensor
Wireless communica- Value Convenient
tion Health Check
Convenient
Health Check
IdentifyIdentify
activities
activities Health Check
Body Protection
Body Sensor
Body Sensor body detection
body detection Body Protection
ConvenientConvenient
Signal transmission stor- storage
Signal transmission Health Check
age
Identify activities

Body Sensor Body Protection

body detection

Convenient stor-
Signal transmission
Due to the complicated signal, it is impossible to manually measure the
Due to the complicated signal, it is impossible to manually agemeasure
content in- the content in-
formation of each chemical, which is necessary for MEMS microcantilever biosensors
formation of each chemical, which is necessary for MEMS microcantilever to biosensors to
detect many substances simultaneously and more accurately. The issues with MEMS mi-
detect many substances simultaneously and more accurately. The issues with MEMS mi-
crocantilever biosensors can be efficiently
crocantilever biosensorsfixed
can with AI, which
be efficiently alsowith
fixed increases the detection
AI, which also increases the detection
precision. precision.

3. Research Status
3. of MEMSStatus
Research Microcantilever Biosensors Biosensors
of MEMS Microcantilever
Micro-Electro-Mechanical Systems, sometimesSystems,
Micro-Electro-Mechanical known sometimes
as MEMS, are electronic
known me- are electronic me-
as MEMS,
chanical systems. It is a microelectromechanical
chanical system that incorporates
systems. It is a microelectromechanical interface
system that cir- interface circuits,
incorporates
cuits, signal processing and control circuits, microsensors, and microactuators. The micro-
signal processing and control circuits, microsensors, and microactuators. The microcan-
cantilever biosensor operates
tilever on theoperates
biosensor theory that, when
on the thethat,
theory material
whenbeing tested adheres
the material being tested adheres to or
to or remains on its surface, the mass of the microcantilever changes, which causes the
micro-resonant cantilever’s frequency to change. The quality of the test object can be de-
termined by measuring the size of the frequency offset.
The benefits of the miniaturization, integration, intelligence, low cost, and mass pro-
 precision.

3. Research Status of MEMS Microcantilever Biosensors

Micro-Electro-Mechanical Systems, sometimes known as MEMS, are electronic me-
Processes 2022, 10, 1658 chanical systems. It is a microelectromechanical system that incorporates interface 10 ofcir-
25
cuits, signal processing and control circuits, microsensors, and microactuators. The micro-
cantilever biosensor operates on the theory that, when the material being tested adheres
to or remains on its surface, the mass of the microcantilever changes, which causes the
remains on its surface, the mass of the microcantilever changes, which causes the micro-
micro-resonant
resonant cantilever’s
cantilever’s frequencyfrequency
to change.to change. The of
The quality quality
the testof object
the test object
can can be de-
be determined
termined by measuring the size of
by measuring the size of the frequency offset.the frequency offset.
The benefits
The benefits of ofthe
theminiaturization,
miniaturization,integration,
integration, intelligence,
intelligence, lowlowcost,cost,
and andmassmasspro-
production of MEMS microcantilever biosensors have made them popular in a variety in-
duction of MEMS microcantilever biosensors have made them popular in a variety of of
dustries, including
industries, includingwireless
wirelesscommunication
communication[32], [32],biomedicine
biomedicine[33], [33], military
military defense
defense [34],
[34],
consumer electronics
consumer electronics [35],
[35], and
and many
many more.
more. It It may
may bebe claimed
claimed that that MEMS
MEMS microcantilever
microcantilever
biosensors, to a certain extent, represent the future development of sensor technology,
biosensors, to a certain extent, represent the future development of sensor technology,
because these
because these advantages
advantages are are consistent
consistent with
with thethe path
path ofof future
futuresensor
sensordevelopment.
development.
One of
One of the
theMEMS
MEMScomponents
componentsthat thatisisfrequently
frequentlyutilized
utilized is is
thethe microcantilever
microcantilever sen-
sensor.
sor. Sensors based on microcantilevers have been employed in every
Sensors based on microcantilevers have been employed in every part of life since Binning part of life since Bin-
ning
et et al. created
al. created an improved
an improved atomicatomic force microscope
force microscope throughthrough
a cantilevera cantilever in 1986. Re-
in 1986. Researchers
searchers
now have now haveknowledge
a deeper a deeper knowledge of the microcantilever
of the microcantilever sensor thanks sensor thanks totheoretical
to extensive extensive
research and application exploration. In order to provide lateral stresses to thetocells
theoretical research and application exploration. In order to provide lateral stresses the
cells adhering to the material, Akiko [36], a researcher at the National
adhering to the material, Akiko [36], a researcher at the National Institute of Metal Science Institute of Metal
Science
in Japan,inassessed
Japan, assessed the separation
the separation force
force in theincell
theculture
cell culture
media,media, as shown
as shown in in Figure
Figure 9
9 utilizing
utilizing opticalfiber
optical fibersensors.
sensors.AAresolution
resolutionofof0.250.25nmnmand andoutput
output voltage
voltage accuracy
accuracy of of
0.0005m/mV
0.0005 m/mVcan canbebereached
reachedwhenwhen measuring
measuring the the deviation
deviation value of the microcantilever
beam, respectively.
beam, respectively. However,
However,the theoptical
opticalfiber
fibersensing
sensing inin the
the liquid
liquid has has some
some influence
influence on on
the entire
the entire detecting
detecting process.
process. ItIt is
is difficult
difficult and
and inconvenient
inconvenient to to use
usethe theequipment.
equipment.

Figure 9. Reading out the deviation value of the microcantilever through the optical fiber.

Using the optical lever readout method, as shown in Figure 10, to achieve a detection
accuracy of 0.1 ng/mL, Wu [37], a researcher at the University of Berkeley in the United
States, used the bending change of the microcantilever to perform prostate-specific antigen
(prostate-specific antigen, PSA) bioassays and used the microcantilever to perform DNA
hybridization, DNA–protein interactions, and protein–protein binding.
Numerous innovative micro-nano-biosensors have arisen in recent years as a result of
the cross-integration of electronic information technology and biology, MEMS technology,
and nanotechnology. Three groups can be made out of them, depending on the various con-
version mechanisms: optical sensors, electrochemical sensors, and mass-sensitive sensors,
as shown in Table 2.
 Using the optical lever readout method, as shown in Figure 10, to achieve a detection
accuracy of 0.1 ng/mL, Wu [37], a researcher at the University of Berkeley in the United
States, used the bending change of the microcantilever to perform prostate-specific anti-
gen (prostate-specific antigen, PSA) bioassays and used the microcantilever to perform
Processes 2022, 10, 1658 DNA hybridization, DNA–protein interactions, and protein–protein binding. 11 of 25

Figure 10. Microcantilever sensor optical lever readout detection system.

Numerous innovative micro-nano-biosensors have arisen in recent years as a result

of the cross-integration of electronic information technology and biology, MEMS technol-
ogy, and nanotechnology. Three groups can be made out of them, depending on the var-
ious conversion mechanisms: optical sensors, electrochemical sensors, and mass-sensitive
Microcantilever sensor
Figure 10. Microcantilever sensoroptical
opticallever
leverreadout
readoutdetection
detectionsystem.
system.
sensors, as shown in Table 2.
TableNumerous innovative
2. Three types micro-nano-biosensors have arisen in recent years as a result
of micro-nano-biosensors.
Table 2. Three types of micro-nano-biosensors.
of the cross-integration of electronic information technology and biology, MEMS technol-
ogy, and nanotechnology. Three groups canDetection
be made out
Detection of them, depending
Advantages on the
Advantages
and andvar-
Detection type type
Detection Schematic
Schematic
ious conversion mechanisms: optical sensors, Principle
electrochemical
Principle Disadvantages
sensors, and
Disadvantages mass-sensitive
sensors, as shown in Table 2.

Optical
22, 10, x FOR PEER REVIEW sensor Table 2. Three types of micro-nano-biosensors. high sensitivity,
12 of high
26 sensitivity,
Optical sensor Optical properties
(suchplasmon
as surface Optical properties such as
real-time real-time
(such
22, 10, x FOR PEER as surface
REVIEW such Detection12 of 26Advantages and
plasmon Detection type lightasabsorption
Schematic
light absorp-
and reflection complex
resonance SPR [38])reso- complex
tion and reflection Principle Disadvantages
bulky device
nance SPR [38]) bulky device

Electrochemical Optical sensor Redox ReactionOptical properties high sensitivity,

(such as surface high sensitivity, real-time
sensors (such as
Electrochemical chemical
RedoxRedoxreaction
Reaction such as light absorp-
plasmon reso- easysensitivity,
Reactionhigh to array. complex
microelectrodes
sensors (such as nance SPR [38]) Dielectricreaction
chemical changetion and reflection high sensitivity,
Electrochemical sensors (such chemical reaction Unstable bulky device
[39]) between electrodes easy to array. easy to array.
as microelectrodes [39])
microelectrodes Dielectric
Dielectric change
change Unstable
[39]) between
between electrodes Unstable
electrodes

After adsorption of
the biological
After adsorption sam-of ofhigh sensitivity,
Mass-sensitive After adsorption the
high sensitivity,
ple to be
the biologicaldetected,
sam- to
biological sample high mature
be sensitivity,
sensors
Mass-sensitive (such
sensors (suchas
Mass-sensitive mature
the
ple resistance,
detected, thevolt-
to be detected, easy
resistance, to array
mature
microcantilevers
as microcantilevers [40])
sensors (such as easy to array
voltage,
age, and and frequency
frequency of the
[40]) the resistance, volt- Easy easytotointegrate
array
Easy to integrate with IC;
microcantilevers of the sensing
sensingunit change with IC;
unit
age, and frequency Easy to integrate
[40])
of the change
sensing unit with IC;
change

In termstime,
In terms of the detection of thesensitivity,
detection time, sensitivity,
safety, safety, inand
in situ sensing, situreal-time
sensing, and
de- real-time detection,
tection,Inoptical optical
biosensors biosensors
have several have several
advantages. advantages.
Its in
widespreadIts widespread
applicability applicability
is con- is constrained
terms of the
bydetection
the high time, sensitivity,
cost [38]. safety,
Electrochemical situ sensing,
biosensors and
are real-time
advantageous de- due to their ability
strainedoptical
tection, by the biosensors
high cost [38].
haveElectrochemical biosensors
several advantages. are advantageous
Its sensitive,
widespread due tois their
to be made smaller, cheaper, more andapplicability
unaffected by con-
solution turbidity [40].
ability to be made smaller,
strained by the highHowever, cheaper,
cost [38].the more
Electrochemicalsensitive, and
biosensors unaffected
are which by solution
advantageous due turbidity
to theirand are significantly
sensor contains electrolytes, have a low stability
[40]. However,
ability to be madethe smaller,
sensor contains
cheaper,electrolytes, which
more sensitive, andhave a low stability
unaffected and are
by solution signifi-
turbidity
cantly
[40]. impactedthe
However, bysensor
external factors,electrolytes,
contains including temperature,
which have ahumidity, and and
low stability pressure. It has
are signifi-
cantly impacted by external factors, including temperature, humidity, and pressure. Itsitu-
some restrictions, because it needs to be calibrated separately for various application has
ations.restrictions,
some Microelectromechanical
because it needs systems
to be (MEMS)
calibratedtypically
separately usefor
beams [41],application
various discs [42], rings
situ-
[43], interdigitated
ations. fingers [44],systems
Microelectromechanical and other resonant
(MEMS) structures
typically to provide
use beams mass-sensitive
[41], discs [42], rings
Processes 2022, 10, 1658 12 of 25

impacted by external factors, including temperature, humidity, and pressure. It has some
restrictions, because it needs to be calibrated separately for various application situations.
Microelectromechanical systems (MEMS) typically use beams [41], discs [42], rings [43],
interdigitated fingers [44], and other resonant structures to provide mass-sensitive sensors
Processes 2022, 10, x FOR PEER REVIEW shown in Figure 11). The microbeam structure has the advantages of small size,
(as 13 ofa simple
26
structure, high sensitivity, easy arraying, and integrationand has become the most potential
device in the field of the simultaneous detection of multiple markers [45].

Figure 11. Mass sensing MEMS biosensors with different structures.

Figure 11. Mass sensing MEMS biosensors with different structures.

There are two operating modes for MEMS microcantilevers used as biosensors: static
There are two operating modes for MEMS microcantilevers used as biosensors: static
mode and dynamic mode. The object that needs to be detected is precisely adsorbed
mode and dynamic mode. The object that needs to be detected is precisely adsorbed on
on the biofunctionalized microcantilever, which changes the microcantilever’s surface
the biofunctionalized microcantilever, which changes the microcantilever’s surface stress
stress
state andstatedeflection
and deflection of the amount
of the amount of analyteof analyte on the microcantilever’s
on the microcantilever’s surface. The surface.
reso- The
resonance mode is another name for the micro-dynamic
nance mode is another name for the micro-dynamic cantilever’s operating mode. Accord-cantilever’s operating mode.
According to the detection
ing to the detection concept, ifconcept,
the chemicalif thebeing
chemical
testedbeing
sticks totested sticks on
or remains to the
or remains
micro- on
the microcantilever
cantilever surface,
surface, it will change it will change
the mass thedevice,
of the mass ofwhich
the device, which
will change thewill change the
resonance
resonance
frequency. frequency. Using the
Using the frequency frequency
shift as a gauge,shift
it as a gauge,
might indicate it might
how much indicate
analytehow themuch
analyte the microcantilever
microcantilever has adsorbedhas adsorbed [46].
[46].
The initial
initial report
reporton onthe
thehybridization
hybridization andand detection
detection of DNAof DNA molecular
molecular fragments
fragments by by
aastatic
static mode
mode microbeam biosensorwas
microbeam biosensor wasmade
madeininthe AmericanJournal
theAmerican Journal of of Science
Science byby Fritz
Fritz et et al.
of
al.IBM
of IBM Zurich
ZurichLaboratory
Laboratory [47]. This
[47]. Thisresearch
researchadvances
advancesthe thestatic
static mode
mode microbeam
microbeam bi- biosen-
sor technology
osensor whilewhile
technology demonstrating
demonstrating howhow the microbeam
the microbeam biosensor
biosensor cancanbe used
be usedto quickly
to
quickly
align DNA.alignTheDNA. The
static static
mode mode microbeam
microbeam biosensor, biosensor,
which has whichsome hassensitivity
some sensitivity
restrictions,
restrictions, accomplishes
accomplishes the detection thegoal
detection goal by deforming
by deforming the cantilever
the cantilever beam throughbeam through
a change a in
change
the in the
surface surfaceThe
tension. tension.
resonantThe resonant
microbeam microbeam
biosensor biosensor
has been has been employed
employed ex- in
extensively
tensively
recent in recent
years, because years, because
in the dynamicin theoperating
dynamic operating
mode, themode, qualitythefactor
qualityoffactor of the
the microbeam
microbeam
can approach can
104approach 104 and
and beyond, andbeyond, and it can
it can achieve achieve
trace detectiontraceindetection
the orderinofthe −18 g [48],
10order
of 10extraordinarily
with
−18 g [48], with extraordinarily
high sensitivity high sensitivity
and more and and morefocus.
more and more focus. According
According to Lange et al.,
to Lange
metal et al.,
oxide metal oxide microcantilevers
microcantilevers can detect volatilecan detect volatile
organic organic
chemical chemical concen-
concentrations with an
trations with an octane detection limit of 0.8 ppm and a mass resolution
octane detection limit of 0.8 ppm and a mass resolution of roughly 2 pg [49]. The Caltech of roughly 2 pg
[49]. The Caltech research team led by M. Roukes discovered
research team led by M. Roukes discovered that, in vacuum, the mass resolution of a that, in vacuum, the mass
resolution of a nanometer-sized double-ended beam sensor can reach 4 kDa (1 Da =1.66 ×
nanometer-sized double-ended beam sensor can reach 4 kDa (1 Da =1.66 × 10−21 g) [50].
10 g) [50]. The dynamic mode microbeam sensor’s drawback, on the other hand, is that
−21

the environment has a significant impact on the resonant frequency and quality factor,
and the high sensitivity of mass detection can only be demonstrated in a vacuum environ-
ment. The resonance frequency, quality factor, and detection sensitivity of the resonant
beam sensor can all be significantly decreased by environmental damping, particularly
Processes 2022, 10, 1658 13 of 25

The dynamic mode microbeam sensor’s drawback, on the other hand, is that the envi-
ronment has a significant impact on the resonant frequency and quality factor, and the
high sensitivity of mass detection can only be demonstrated in a vacuum environment.
The resonance frequency, quality factor, and detection sensitivity of the resonant beam
sensor can all be significantly decreased by environmental damping, particularly damping
caused by liquid environments. Additionally, the microbeam’s stiffness coefficient was
altered, lowering its sensitivity. Based on this investigation, D. Ramos et al. of the Spanish
National Microelectronics Center examined the resonance frequency change after printing
the bacterial solution on the cantilever beam at various positions and discovered that the
single-end clamped beam with only the free end printing the bacterial solution had the
best quality and sensitivity to detection [51]. Nugaeva et al. of the University of Basel
used single-end fixed beam arrays to explore the process of yeast and Aspergillus niger
hyphae growth, as well as the method of the particular immobilization of fungi in the field
of real-time monitoring of microbeam biosensors [52]. By monitoring the change of the
dynamic resonance frequency of the microbeam, the concentration of fungi was detected in
the range of 103–106 CFU/mL, realizing the real-time monitoring of the colony growth.
With the rapid advancement of microfluidic technology following the concept of the
flow cell, polydimethylsiloxane material (PDMS) was used as the encapsulating shell of
the microfluidic channel in the microbeam biosensors reported in recent years to replace
the flow cell [53]. High binding efficiency and high sensitivity detection are successfully
achieved through the creative combination of microcantilevers.

4. The Application Prospect of Neural Network in MEMS Microcantilever Biosensor

A type of sensor with significant current development potential is the MEMS micro-
cantilever biosensor. Due to its benefits in miniaturization, integration, intelligence, low
cost, and mass production, it has been extensively employed in wireless communication,
biomedicine, military defense, consumer electronics, and other disciplines. It is now pos-
sible to use high tech in microcantilever biosensors and spur their development in the
present world, where computer science and technology are advancing one after the other.
Neural networks have advanced significantly in models, learning techniques, and
ap plications in recent years. Due to its traits of self-adaptation, generalization, nonlin-
ear mapping, and extremely parallel processing, it has been widely used in the field of
smart sensors. The application increases the sensor’s intelligence and raises its level of
intelligence.
A neural network is a network made up of certain basic (often adaptive) components
and their massively parallel connections that are hierarchically organized and have a high
capacity for nonlinear description. The neural network continually trains the network in
practical applications based on the input and output data of the object and accomplishes
nonlinear mapping from the input parameters to the output parameters in order to learn
adaptively. Neural networks come in a variety of forms, each with unique advantages,
and they are frequently employed in conjunction with sensors in nonlinear correction,
classification, diagnosis, identification, and other applications.

4.1. Nonlinear Self-Calibration

The majority of a sensor’s input and output properties are now nonlinear. It can
be challenging to adjust for some nonlinear sensor systems using hardware or standard
software. Due to their advantages in parallel processing, fault tolerance, self-adaptation,
and powerful self-learning, neural networks are frequently utilized in the nonlinear self-
correction of sensor systems. Currently, neural networks are gradually being employed in
new domains, including micromechanical systems, in addition to large-scale mechanical
control fields. Steele created a new sensor system with a closed-loop topology with the
goal of simplifying the design and fabrication while addressing the unique nonlinear and
sophisticated electrical circuits of micromechanical sensors [54,55]. Two different types
of neural networks are used in this sensor system to build the mathematical model of
Processes 2022, 10, 1658 14 of 25

Processes 2022, 10, x FOR PEER REVIEW 15 of 26

the micro-sensing system. A compensation neural network completes the input’s static
mapping, while a feedback neural network is utilized to demodulate and linearize the
feedback
feedbacksignal.
signal.The
Thesimulation
simulationresults
resultsdemonstrate
demonstratethethesystem’s
system’sexcellent
excellentlinearity,
linearity,wide
wide
dynamic
dynamicrange,
range,and
andstrong
strong vibration
vibration resistance.
resistance. The
Theclosed-loop
closed-loopdetection
detectionsystem
systemof ofthe
the
microcantilever
microcantileverbeam
beamsensor
sensorcancanuseusethe
thedata
datamodel
modelcreated
createdby
by the
the compensation
compensationneuralneural
network
networkand
andfeedback
feedbackneural
neuralnetwork.
network.SinceSincethe
thesystem
systemuses
usesFPGA
FPGAfor forlogic
logicarchitecture,
architecture,
as
as indicated in the picture, it is incredibly dependable, compact, and affordable. At
indicated in the picture, it is incredibly dependable, compact, and affordable. At the
the
same
sametime,
time,ititaddresses
addressesthethe issues
issues with
with the
the analog
analog phase-locked
phase-lockedloop’s
loop’szero
zerodrift
driftand
andthethe
ease
easewith
withwhich
whichchanges
changesininthe
thepower
powersupply
supplyandandoutside
outsidetemperature
temperaturehavehavean animpact
impacton on
thephase-locking
the phase-locking performance，as
performance, as shown shownin inFigure
Figure12.
12.

Figure 12. The closed-loop detection circuit of the microcantilever sensor.

Figure 12. The closed-loop detection circuit of the microcantilever sensor.
Environmental factors affect sensor properties in a nonlinear way. The system’s
microcantilever
Environmental biosensors can function
factors affect in a variety
sensor properties inof conditions,
a nonlinear including
way. those with
The system’s mi-
varying pressures,
crocantilever humidity
biosensors levels, magnetic
can function in a variety fields, and temperatures.
of conditions, including In thiswith
those instance,
var-
the sensor’s
ying pressures,reaction
humidityis dependent nonlinearly
levels, magnetic onand
fields, bothtemperatures.
the environmental In thisconditions
instance, andthe
the measured
sensor’s reactionvalue, depending
is dependent on the materials
nonlinearly on both and
themanufacturing
environmentalprocesses
conditions utilized
and the to
create the sensor.
measured As a result,on
value, depending getting accurate and
the materials sensor readings under
manufacturing variousutilized
processes environmental
to cre-
circumstances
ate the sensor. As is aachallenging
result, getting problem.
accurate There
sensor have been aunder
readings number of digital
various and analog
environmental
interface circuits described in the past to account for nonlinear
circumstances is a challenging problem. There have been a number of digital and analog response characteristics
and achieve
interface correct
circuits readings
described [56,57].
in the past Environmental factors affect
to account for nonlinear sensorcharacteristics
response properties in
a nonlinear way. The system’s microcantilever biosensors can
and achieve correct readings [56,57]. Environmental factors affect sensor properties function in a variety in of
a
conditions,
nonlinear including
way. those with
The system’s varying pressures,
microcantilever humidity
biosensors levels, magnetic
can function in a varietyfields, and
of con-
temperatures.
ditions, including In this
thoseinstance, the sensor’s
with varying reaction
pressures, is dependent
humidity nonlinearly
levels, magnetic on both
fields, and
temperatures. In this instance, the sensor’s reaction is dependent nonlinearly on both and
the environmental conditions and the measured values, depending on the materials the
manufacturingconditions
environmental processes and utilized to create the
the measured sensor.
values, As a result,
depending getting
on the accurate
materials andsensor
man-
readings under
ufacturing various
processes environmental
utilized circumstances
to create the sensor. As ais a challenging
result, problem.
getting accurate Thereread-
sensor have
been a number of digital and analog interface circuits described
ings under various environmental circumstances is a challenging problem. There have in the past to account
for nonlinear
been a numberresponse
of digitalcharacteristics and achieve
and analog interface circuitscorrect readings
described in the[57,58].
past toIterative
account andfor
noniterative
nonlinear algorithms
response are among
characteristics andthese
achievemethods. Patra suggested
correct readings a neuraland
[57,58]. Iterative network
non-
among them
iterative that automatically
algorithms are among these corrects for the
methods. nonlinear
Patra suggestedimpacts
a neuralof the surrounding
network among
temperature [58], as shown in Figure 13. With temperatures ranging from 0 to 250 ◦ C,
them that automatically corrects for the nonlinear impacts of the surrounding temperature
this method is used to model capacitive pressure sensors that operate under challenging
[58], as shown in Figure 13. With temperatures ranging from 0 to 250 °C, this method is
conditions. The usefulness of the model in predicting the nonlinear effects of various types
used to model capacitive pressure sensors that operate under challenging conditions. The
of ambient temperatures on the properties of the pressure sensor is confirmed by computer
usefulness of the model in predicting the nonlinear effects of various types of ambient
simulation trials. Iterative and noniterative algorithms are among these methods. Patra
temperatures on the properties of the pressure sensor is confirmed by computer simula-
suggested a neural network among them that automatically corrects for the nonlinear
tion trials. Iterative and noniterative algorithms are among these methods. Patra sug-
impacts of the surrounding temperature. With temperatures ranging from 0 to 250 ◦ C,
gested a neural network among them that automatically corrects for the nonlinear impacts
of the surrounding temperature. With temperatures ranging from 0 to 250 °C, this method
Processes
Processes 2022,
2022, 10,
10, x1658
FOR PEER REVIEW 1615ofof26
25

isthis
used to model
method capacitive
is used pressure
to model sensors
capacitive that sensors
pressure operate that
under challenging
operate conditions.
under challenging
The usefulness of the model in predicting the nonlinear effects of various types
conditions. The usefulness of the model in predicting the nonlinear effects of various of ambient
types
temperatures on the properties
of ambient temperatures of the pressure
on the properties of thesensor is confirmed
pressure by computer
sensor is confirmed simula-
by computer
tion trials. trials.
simulation

Figure13.
Figure Linearand
13. Linear andnonlinear
nonlinearrelationship
relationshipbetween
betweentrue
truepressure
pressureand
andpressure
pressureestimated
estimatedby
bythe
the
neuralnetwork
neural networkmodel
modelat
atdifferent
different temperature
temperature values.
values.

Artificial neural
Artificial neural networks
networkshave have become
become aa potent
potent learning
learning method
method for for carrying
carrying outout
difficulttasks
difficult tasksininchanging
changing contexts.
contexts. TheThe effects
effects of these
of these methods
methods on thisoncomplicated
this complicated
prob-
problem
lem are constrained
are constrained under under the supposition
the supposition thatthat environmental
environmental parameter
parameter fluctuation
fluctuation is
modest and that the environmental parameter influence on sensor characteristics isisnot
is modest and that the environmental parameter influence on sensor characteristics not
linearlysignificant.
linearly significant.
To actualize
To actualize thethe sensor
sensor characteristics,
characteristics, the
the linear
linear equation
equation of
of the
the linear
linear segment
segment isis
produced using the best fitting approach, and the inverse function of
produced using the best fitting approach, and the inverse function of the nonlinearthe nonlinear segment
seg-
of the sensor characteristics is mapped using an upgraded BP neural network
ment of the sensor characteristics is mapped using an upgraded BP neural network as a as a corrective
link. The simulation
corrective test resultstest
link. The simulation forresults
linearization suggest that
for linearization this method
suggest that thismay almost
method ten
may
times more effectively minimize the sensor’s nonlinear error [59–61]. The
almost ten times more effectively minimize the sensor’s nonlinear error [59–61]. The ma- majority of these
techniques
jority transfer
of these the inverse
techniques transferfunction of the
the inverse sensingofproperties
function the sensingasproperties
a correction as link in
a cor-
order to accomplish linearization. To correct the sensor’s nonlinear error, Zhang employed
rection link in order to accomplish linearization. To correct the sensor’s nonlinear error,
a single-input, single-output cerebellar neural network (SISO-CMAC) inverse operation.
Zhang employed a single-input, single-output cerebellar neural network (SISO-CMAC)
He also examined and demonstrated the SISO-CMAC algorithm’s convergence using the
inverse operation. He also examined and demonstrated the SISO-CMAC algorithm’s con-
incremental learning approach [62], as shown in Figure 14.
vergence using the incremental learning approach [62], as shown in Figure 14.
The measurement and control system’s performance are directly impacted by the
sensor’s error, which is mostly a nonlinear error. As a result, contemporary research has
begun to focus heavily on the linearization of the microcantilever biosensor’s properties.
 Processes 2022, 10, x FOR PEER REVIEW 17 of 26
Processes 2022, 10, 1658 16 of 25

Figure 14. SISO-CMAC structure.

Figure 14. SISO-CMAC structure.
4.2. Fault Self-Diagnosis
Theneural
The measurement andexcellent
network has control system’s
nonlinearperformance are directly
fitting capabilities impacted
and does by the
not require
sensor’s error, which is mostly a nonlinear error. As a result, contemporary research
the mathematical model of the known system; rather, it simply adapts to the properties has
begun to focus heavily on the linearization of the microcantilever biosensor’s properties.
of the targeted item for defect diagnostics in which the model is elusive. As a result,
neural networks have gradually been used in recent years in study areas including defect
4.2. Faultand
detection Self-Diagnosis
monitoring. Reference [63] used a multi-layer forward feedback neural
network structure network
The neural based on has excellent
the B-P method nonlinear
to monitorfitting
andcapabilities
categorize theandoperational
does not require
state
ofthe
themathematical model of after
processing equipment the known
guiding system; rather, ittraining.
the network’s simply adapts to thetrained
The model properties
withof
thefrequency
the targeted domain
item forsignal
defectisdiagnostics
found to beinmorewhich the model
effective thanistheelusive.
modelAs a result,
trained withneural
the
networks
time domain have gradually
signal, beentoused
according in recent years
the simulation in study
results in thisareas
study, including defect detec-
which demonstrate
that
tionthis
andstrategy can produce
monitoring. Referenceoptimal results.
[63] used Neural networks
a multi-layer forward are used to
feedback identify
neural the
network
chemical
structure and gas components
based on the B-P methodof substances,
to monitor in addition to gas sensor
and categorize fault detection.
the operational The
state of the
literature suggests using neural network models to enhance the optical
processing equipment after guiding the network’s training. The model trained with the sensors based on
the idea thatdomain
frequency two-color mixes
signal is on triacyl
found cellulose
to be sheets have
more effective thanfixed
the features. The findings
model trained with the
demonstrate
time domain that utilizing
signal, a neural
according to network model results
the simulation to process the study,
in this detection value
which increases
demonstrate
the sensor’s
that measuring
this strategy can range
produce while simultaneously
optimal results. Neuralincreasing
networksthe precision
are used and accuracy,
to identify the
with an error
chemical andrate
gasofcomponents
no more than of 2%.
substances, in addition to gas sensor fault detection. The
The learning
literature suggestsobject of the
using neural
neural network
network model
models to is chosenthe
enhance to be a typical
optical sensor
sensors withon
based
good performance, and the sensor output signal data is fed into the neural
the idea that two-color mixes on triacyl cellulose sheets have fixed features. The findings network model
for learning. The
demonstrate neural
that network’s
utilizing a neuralweight
networkis initially
modelset to a random
to process value. A value
the detection signalincreases
output
estimate is created
the sensor’s when range
measuring the test datasimultaneously
while from the gas sensor entersthe
increasing theprecision
neural network. The
and accuracy,
update
with an method of the
error rate weights
of no more thanuses 2%.
the difference between this estimate and the actual
outputThesignal of theobject
learning gas sensor. The neural
of the neural networknetwork
model method
is chosenwilltocontinually
be a typicalupdate
sensor the
with
connection weight of the learning network as long as the output
good performance, and the sensor output signal data is fed into the neural network data from the gas sensor
model
continues to enter
for learning. Thethe network
neural for theweight
network’s operation.
is initially set to a random value. A signal out-
put estimate is created when the test data from the gas sensor enters the neural network.
wij (n + 1) = wij (n) + ηδj Zj (1)
The update method of the weights uses the difference between this estimate and the actual
output
where wijsignal
is the of the gas
neuron sensor.
weight, Zj The
is theneural
neuronnetwork
output,method will continually
η is the learning update
coefficient, and δthe
j
connection weight
is the neuron difference.of the learning network as long as the output data from the gas sensor
continues to enter
The learning the network
process can befor the operation.
stopped when the mean square error (MSE) achieves a
minimal value. Such learning is carried
𝑤 (𝑛 +out
1)by
= several
𝑤 (𝑛) +regression
𝜂𝛿 𝑍 processes. The artificial(1)
neural network model has now finished learning and evolved into the nonlinear output
 where 𝑤 is the neuron weight, 𝑍 is the neuron output, 𝜂 is the learning coefficient,
and 𝛿 is the neuron difference.
Processes 2022, 10, 1658 The learning process can be stopped when the mean square error (MSE) achieves 17 a
of 25
minimal value. Such learning is carried out by several regression processes. The artificial
neural network model has now finished learning and evolved into the nonlinear output
model
modelofofthe
thegas
gassensor
sensoratatthe
thesame
same time
time as the network
network weight
weight reaches
reachesaastable
stablecoefficient.
coeffi-
As can
cient. As be
canseen, the correctness
be seen, of theofdynamic
the correctness nonlinear
the dynamic model
nonlinear of theofgas
model thesensor may be
gas sensor
ensured
may as longasaslong
be ensured sufficient numbers
as sufficient of accurate
numbers training
of accurate samples
training are given.
samples are given.
Thefault
The faultdiagnosis
diagnosisofofthethegas
gassensor
sensorcan
canbebe done
done online
online thanks
thanks toto the
the precise
precise dynamic
dynamic
output model of the gas sensor, as shown in
output model of the gas sensor, as shown in Figure 15. Figure 15.

Figure15.
Figure Neural
15.Neural network
network predictionoutput
prediction outputand
andoutput
outputofofa afaulty
faultygas
gassensor.
sensor.

The graphic shows that, when the gas sensor operates normally, the actual output
The graphic shows that, when the gas sensor operates normally, the actual output of
of the sensor and the anticipated output of the neural network are both consistent. The
the sensor and the anticipated output of the neural network are both consistent. The actual
actual sensor output varies from what the neural network predicts when the gas sensor
sensor output varies
malfunctions. from whatvalue
The estimated the neural network
is compared predicts
with when the gasthreshold
the predetermined sensor malfunc-
value by
tions. The estimated value is compared with the predetermined
measurements in the detection method, making it simple to diagnose sensor threshold value by meas-
faults.
urements
Theinuse theofdetection
artificialmethod, making it simple
neural networks provides to diagnose sensor faults.
greater benefits than conventional
The use of artificial neural networks provides greater
techniques for the dynamic nonlinear modeling of gas sensors to perform benefits than conventional
online fault
techniques for the dynamic nonlinear modeling of gas sensors to perform
testing. The implementation of artificial neural networks in microcantilever biosensors online fault test-is
ing. The implementation of artificial neural networks in microcantilever
made possible by the fact that they are appropriate for the modeling and fault diagnostics biosensors is
made possible by the fact that
of other dynamic nonlinear systems. they are appropriate for the modeling and fault diagnostics
of other dynamicto
According nonlinear
the actualsystems.
system and the output residuals of the neural network observer,
According to the
Reference [64] developed actual system and control
a fault-tolerant the output
systemresiduals of thenetwork
and a neural neural network ob-
state observer
server, Reference
that can diagnose [64]sensor
developed a fault-tolerant
defects. control system
A neural network and apredictor-based
time series neural network sensor
state
observer that can diagnose
defect diagnosis method was sensor
showndefects. A neural[65].
in Reference network time series
The difference predictor-based
between the sensor
sensor
outputdefect
valuediagnosis method
and the actual was output
sensor shown in Referenceusing
is predicted [65]. The difference
a neural networkbetween the
prediction
sensor output value and the actual sensor output is predicted using a neural
model, which is first utilized to create a prediction model for the sensor output sequence network pre-
diction model, the
and whether which is first
sensor will utilized
be faulty.toThe
create a prediction
method modelfault
of the sensor for diagnosis
the sensorcombines
output
sequence and whether
the optimization the sensor
(HPSO) will be
algorithm andfaulty.
neuralThe methodusing
network, of the the
sensor
HPSO fault diagnosisto
algorithm
combines
optimizethetheoptimization
parameters of(HPSO) algorithm
the neural network and neuraland
model, network, usinga prediction
establishes the HPSO algo-
model
rithm to optimize
for sensor the parameters of the neural network model, and establishes a predic-
fault diagnosis.
tion model for sensor fault diagnosis.
4.3. Data Fusion
4.3. Data
TheFusion
precision of the measurement and control system can be significantly increased
by using multi-sensor information fusion to provide rich and comprehensive information
about the system. It represents a significant step in the direction of microcantilever biosensor
development. The microcantilever biosensor’s many beams collect information, which
is analyzed and combined by multi-sensor information fusion technology to create a
thorough and full description of the information about the measured object. Neural
network technology’s nonlinear capability has developed into a useful tool for fusing data
from several sensors effectively and quickly. To increase the precision and accuracy of multi-
Processes 2022, 10, 1658 18 of 25

sensor data processing, References [66,67] made use of rough sets and backpropagation
networks. Reference [68] suggested a multi-sensor data fusion approach based on belief
entropy and belief-based divergence measures. It resolves the issue where combining
extremely incongruent data could lead to outcomes that are contrary to conventional
wisdom. To reduce the uncertainty of sensor data, Reference [69] used fuzzy sets to
combine data from several sensors. Reference [70] was based on the continuous Markov
model. The motion control and somatosensory sensor data fusion algorithm was created.
The outcomes of the experiments demonstrated that the algorithm may successfully lower
the degree of uncertainty in sensor data. A framework for sign language recognition was
suggested on the basis of this. Random finite sets were used in Reference [71], so the
subsequent sensor data was anticipated, as well as fused with the initial sensor data. The
single data fusion techniques can only partially accomplish the goal of fusion due to their
weak resilience and inaccurate and unstable fusion outputs.
In the traditional sense, the realization of neural network multi-sensor integration and
fusion can be divided into three important steps:
(1) According to the requirements of the intelligent system and the form of sensor
integration and fusion, the topology of the neural network is selected.
(2) Comprehensively process the input information of each sensor into an overall
input function and define this function mapping as the mapping function of the relevant
unit, which reflects the statistical law of the environment to the network itself through the
interaction between the neural network and the environment. in the structure.
(3) Learn and understand the sensor output information to determine the distribution
of the values, and complete the knowledge acquisition and information fusion.
Due to traditional multi-sensor information fusion, Chen Ying improved the data
preprocessing and fusion process of the fusion method, as shown in Figure 16. The fusion
process steps of the improved fusion method are as follows:
(1) Preprocess the data collected by the sensor through the Laida criterion, including
data cleaning, replacing outliers, etc.
(2) Denoising the preprocessed data by using the wavelet threshold of the high-
frequency coefficients after the layers are subjected to threshold processing and the obtained
low-frequency coefficients are optimized by the neural network.
(3) Neural network optimization is performed on the transformed and reduced data.
(4) Data fusion is performed through the reliability of the sensor, and the final fusion
result is obtained.
A large-scale, continuously adapting, nonlinear information processing system is a
neural network. It uses a vast network of extensive connections of a significant number
of neurons for information processing and simulates the information processing function
of the human brain. It offers powerful parallel processing, information storage, and self-
learning capabilities. When a neural network is used to combine data from several sensors,
it is first important to choose an appropriate neural network model based on the system’s
needs and the sensor’s properties. To learn it, set the weight distribution and finish the
network training so that the trained neural network may take part in the actual fusion
process; the system decisions and existing sensor data are employed.
Processes
Processes2022,
2022,10,
10,x1658
FOR PEER REVIEW 2019ofof26
25

Figure16.
Figure Thedata
16.The datapreprocessing
preprocessingand
andfusion
fusionprocesses
processesof
ofthe
theimproved
improved fusion
fusion method.
method.

5. Application of Principal Component Analysis in Biosensors

A large-scale, continuously adapting, nonlinear information processing system is a
neuralThe principal
network. component
It uses analysisof(PCA)
a vast network has been
extensive employed
connections of for a long time
a significant as a fault
number of
detection technique to extract pertinent data from multivariate
neurons for information processing and simulates the information processing function sensors. The PCA has beenof
used
the to analyze
human brain.multivariate
It offers powerful data for a number
parallel of multivariate
processing, informationdata storage,
analysis and technolo-
self-
gies, including sensor process monitoring, quality control, and problem diagnosis. The
learning capabilities. When a neural network is used to combine data from several sen-
multi-scale principal component analysis (MSPCA) method, which combines the principal
sors, it is first important to choose an appropriate neural network model based on the
component analysis with wavelet analysis, was first proposed by Bakshi in 1998 [72]. In
system’s needs and the sensor’s properties. To learn it, set the weight distribution and
Reference [73], the authors described a method for identifying sensor failures that uses an
finish the network training so that the trained neural network may take part in the actual
MSPCA model to track the statistical variable T2 in the principal component space, squared
fusion process; the system decisions and existing sensor data are employed.
prediction error (SPE) to identify the sensor drift error, and SPE contributions from each
sensor to identify the faulty sensors [74]. Secondly, the principal component analysis, as an
5. Application of Principal Component Analysis in Biosensors
effective tool for multi-sensor modeling, is often used in sensor data reconstruction [75–78].
The The
PCAprincipal component
is a modeling method analysis
that is (PCA) has been
independent employed
of the for a of
knowledge long thetime as a fault
mathematical
detection technique to extract pertinent data from multivariate
model of the system, and the output of the sensor can build a statistical model for sensors. The PCA hassensor
been
used
faultto analyze multivariate
diagnosis [79]. Under normal data forcircumstances,
a number of multivariate
the sensor data
databaseanalysis Xm×technologies,
n is collected,
including
where m is sensor processofmonitoring,
the number samples andqualityn is thecontrol,
numberand problemThen,
of sensors. diagnosis. The multi-
it is standardized
scale principal component analysis (MSPCA)
to remove the effects of different units of the variables: method, which combines the principal com-
ponent analysis with wavelet analysis, was first proposed by Bakshi in 1998 [72]. In Ref-
erence [73], the authors described aX method = Dσ−1 [ X for−identifying
E( x )] sensor failures that uses an (2)
MSPCA model to track the statistical variable T2 in the principal component space,
= [ x1 , x2 , ·error
where Xprediction
squared ]T ∈ <
· · , xn(SPE) n is the data vector of the specified point in X
to identify the sensor drift error, and SPE contributions m×n , E(X)
= [µ1,
from µ3, µn]toTidentify
qµ2, sensor
each is the average
the faulty vector of X[74].
sensors Dσ = diag
m×n .Secondly, σ1 , σ2 , · · · component
the(principal , σn ), where
analysis, as an 2
effective tool for multi-sensor modeling, is often used in sensor datadatabase
recon-
σi = E( xi − µi ) is the ith standard deviation of X m×n . In this way, the standard
struction [75–78]. The PCA is a modeling method that is independent
X m×n is realized. Calculate the correlation matrix with X m×n and then perform singular of the knowledge of
the mathematicalFinally,
decomposition. modelproject of the system,
X m×n intoandthethe output component
principal of the sensor can X̂
space build
and athestatistical
residual
model for sensor
space X̂, namely: fault diagnosis [79]. Under normal circumstances, the sensor data-
base 𝑋 × is collected, where → m is the number of samples and n is the number of sensors.
X m×n = X̂ + X e = ĈX m×n + CX e m×n (3)
Then, it is standardized to remove the effects of different units of the variables:
where the projection matrix Ĉ = PPT . To detect sensor failures, statistical variables in the
𝑋 = 𝐷 [𝑿 − 𝐸(𝑥)] (2)
subspace need to be specified. Reference [80] presented a comparison between some statis-
where 𝑿 = [𝑥 , proving
tical variables, 𝑥 , ⋯ , 𝑥 ]that ∈ the
ℜ isSPE theindata vector ofspace
the residual the specified
can betterpoint
reflect 𝑋 correlation
in the × , E(X) =
[μ1, μ2, μ3,
changes μn] Tsensor
between is theoutputs
averagethan vector of in
the T2 𝑋 the
× . 𝐷 =
principaldiag (𝜎 , 𝜎
component , ⋯ , 𝜎 ),
space.where 𝜎 =
Therefore,
𝐸(𝑥 − 𝜇 ) is the ith standard deviation of 𝑋 × . In this way, the standard database
SPE can detect sensor failures in real time. The definition of SPE is as follows:
𝑋 × is realized. Calculate the correlation matrix with 𝑋 × and then perform singular
SPE =k Cx e k2 6 δ2 (4)
SPE
 where the projection matrix 𝐶 = 𝑃𝑃 . To detect sensor failures, statistical variables in the
subspace need to be specified. Reference [80] presented a comparison between some sta-
tistical variables, proving that the SPE in the residual space can better reflect the correla-
tion changes between sensor outputs than the T2 in the principal component space. There-
fore, SPE can detect sensor failures in real time. The definition of SPE is as follows:
Processes 2022, 10, 1658 20 of 25
~
SPE =∥ 𝐶 𝑥 ∥ ⩽ 𝛿 (4)

where 𝛿 is the threshold of SPE, which can be calculated from the sample distribution
where SPE is the threshold of SPE, which can be calculated from the sample distribution of
of SPEδ[81].
SPE [81].
Anurag placed detection samples in a 1650 × 6 matrix in the form of datasets, where
Anurag
the rows placed the
represent detection samples
data vectors ininthe
a 1650 × 6 matrix in
microcantilever the form
sensor of datasets,
space. where
It is frequently
the rows represent the data vectors in the microcantilever sensor space. It is frequently
impossible to discriminate between raw data vectors linked to distinct vapor classes be-
impossible to discriminate between raw data vectors linked to distinct vapor classes because
cause of overlap or ambiguity in the chemical interactions between water vapor and pol-
of overlap or ambiguity in the chemical interactions between water vapor and polymers.
ymers. According to the mutually orthogonal directions with the greatest variance, the
According to the mutually orthogonal directions with the greatest variance, the principal
principal component analysis (PCA) can convert the data into a six-dimensional principle
component analysis (PCA) can convert the data into a six-dimensional principle component
component space. The primary components are listed in decreasing order of variance,
space. The primary components are listed in decreasing order of variance, with the variance
with the variance serving as a measure of the information content (as shown in Figure 17).
serving as a measure of the information content (as shown in Figure 17).

Figure 17. Distribution of the principal components.

Figure 17. Distribution of the principal components.
The principal component distribution maps of four sensor arrays constructed utilizing
staticThe
andprincipal
dynamiccomponent
mode sensing, distribution
as well asmaps
FCMof four
and FSCsensor arrays
polymer constructed
options, utiliz-
respectively,
ing static and dynamic mode sensing, as well as FCM and FSC polymer
are shown in the principal component analysis of the response of the MEMS cantilever options, respec-
tively, are
sensor arrayshown in the principal
on respiratory component
volatiles. analysis sensors
Microcantilever of the response
have theofpotential
the MEMSfor canti-
mass
lever sensor array on respiratory volatiles. Microcantilever sensors have the
production, they have the benefits of convenience, cheap cost, and high sensitivity, and potential for
mass production, they have the benefits
they have the possibility to add more features. of convenience, cheap cost, and high sensitivity,
and they
In thehave the PCA
future, possibility
will betorequired
add more features. microcantilever sensors with high tech,
to integrate
since it improves the characteristics of data extraction from these sensors.

6. Challenges and Prospects of AI Applied to Microcantilever Sensors

The development and use of monitoring, optimization, and control technologies have
recently been aided by artificial intelligence, data-driven technologies, cyber–physical sys-
tems, cloud computing, and cognitive computing [82]. Nonlinear algorithms, for instance,
can be used to monitor dynamic systems and provide observer gains [83].
In the progress of the optimization of microcantilever biosensor systems, the lack of a
significant application field for human–computer interactions has prevented a trend of full
integration between microcantilever sensors and high-tech systems such as the Internet of
Things, as shown in Figure 18. As illustrated in the image, there are adaptable applications
and user-friendly human–machine interface applications in various sensor areas, including
wearable sensor fields and human body sensor fields.
 can be used to monitor dynamic systems and provide observer gains [83].
In the progress of the optimization of microcantilever biosensor systems, the lack of
a significant application field for human–computer interactions has prevented a trend of
full integration between microcantilever sensors and high-tech systems such as the Inter-
net of Things, as shown in Figure 18. As illustrated in the image, there are adaptable ap-
Processes 2022, 10, 1658 21 of 25
plications and user-friendly human–machine interface applications in various sensor ar-
eas, including wearable sensor fields and human body sensor fields.

Figure
Figure 18.
18. Human–computer
Human–computer interaction
interaction applications
applications of
of wearable
wearable sensors.
sensors.

In
In addition,
addition, there
there are also some restrictions with the artificial neural algorithm algorithm and and
microcantilever
microcantilever combo. The The study
study of of artificial
artificial neural
neural networks
networks is is aa cutting-edge
cutting-edge field
field that
that
has advanced
has advanced significantly
significantly in
in recent
recent decades.
decades. ItIt can
can get
get over
over the
the current
current bottleneck
bottleneck and and
conduct in-depth research
conduct in-depth research on nonlinear phenomena because of its capacity
nonlinear phenomena because of its capacity for large-scalefor large-scale
parallel processing,
processing, high-fault
high-faulttolerance,
tolerance,highhighself-organization,
self-organization,adaptability,
adaptability, andandability to
ability
approximate
to approximate anyany
nonlinear function.
nonlinear However,
function. the artificial
However, neuralneural
the artificial network still encounters
network still en-
some challenges
counters when the
some challenges training
when samplessamples
the training are excessively vast and
are excessively vastgreat
and precision
great preci- is
needed.
sion The classic
is needed. backpropagation
The classic backpropagation (BP) (BP)
neural network,
neural as shown
network, in Figure
as shown in Figure19 for
19
instance,
for
Processes 2022, 10, x FOR PEER REVIEW hashas
instance, drawbacks,
drawbacks, such
such asas
a ahigh
highoutput
outputmean
meansquare
squareerror
error and
and poor diagnostic
poor diagnostic
23 of 26
precision. Artificial intelligence algorithms used in the sensor industry have thisthis
precision. Artificial intelligence algorithms used in the sensor industry have samesameis-
issue.
sue.

Figure 19.
Figure 19. BP neural network algorithm
algorithm flow.
flow.

Future
Future MEMS
MEMS microcantilever
microcantilever biosensors
biosensors should
should bebe able
able to
to connect
connect toto the
the Internet
Internet
and
and engage in a larger variety of human–computer interactions and interconnections. The
engage in a larger variety of human–computer interactions and interconnections. The
application
application of
of AI
AI in
in biosensors
biosensors can
can benefit
benefit from
from deep-level
deep-level data
data processing,
processing, such
such asas data
data
processing
processing and
and data
data fusion
fusion with
with the
the aid
aid of
of neural
neural algorithms. Today, with
algorithms. Today, with the
the booming
booming
development
development of the Internet, cases of biosensors for human–computer interaction, such
of the Internet, cases of biosensors for human–computer interaction, such as
as
smart bracelets, are seen everywhere. We can envision MEMS microcantilever biosensors
based on neural algorithms and artificial intelligence technologies being used in a range
of applications in the near future, including medical treatments, industrial production,
environmental protection, and others.
Processes 2022, 10, 1658 22 of 25

smart bracelets, are seen everywhere. We can envision MEMS microcantilever biosensors
based on neural algorithms and artificial intelligence technologies being used in a range
of applications in the near future, including medical treatments, industrial production,
environmental protection, and others.

7. Conclusions
Future sensors and microelectromechanical systems (MEMS) will play an increasingly
significant part in our daily lives as we enter the new era of intelligence and experience the
rapid growth of technology. As a result, MEMS sensor systems integrate neural networks
and artificial intelligence (AI), and the following iteration of sensors will have a distinct
development trajectory. This essay explores the possibility of fusing MEMS microcantilever
biosensors with neural algorithms, as well as the specific case of fusing AI with MEMS
biosensors. As a sensor type with several benefits, including portability, affordability,
and high sensitivity, its integration with the Internet should go beyond the use of neural
networks and instead help people interact with other objects in a variety of scenario
applications.

Author Contributions: Conceptualization, J.W. and B.X.; methodology, B.X.; software, B.X.; vali-
dation, B.X.; formal analysis, J.W. and B.X.; investigation, J.W.; resources, B.X.; data curation, B.X.;
writing—original draft preparation B.X.; writing—review and editing, J.W., B.X., L.S., L.Z. and X.W.;
visualization, J.W. and B.X.; supervision, J.W., project administration, J.W.; funding acquisition, J.W.
All authors have read and agreed to the published version of the manuscript.
Funding: This research was funded by [the National Science Foundation of China and Tianjin
Municipal Science and Technology Bureau], grant number [(61804107) and (20JCQNJC00180)]. And
the APC was funded by [the National Science Foundation of China (61804107) and Tianjin Municipal
Science and Technology Bureau (20JCQNJC00180)].
Informed Consent Statement: Written informed consent has been obtained from the patients to
publish this paper.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Binnig, G.; Quate, C.F.; Gerber, C. Atomic Force Microscope. J. Mater. Eng. 1986, 56, 930–933. [CrossRef] [PubMed]
2. Marauska, S.; Claus, M.; Lisec, T.; Wagner, B. Low temperature transient liquid phase bonding of Au/Sn and Cu/Sn electroplated
material systems for MEMS wafer-level packaging. Microsyst. Technol. 2013, 19, 1119–1130. [CrossRef]
3. Xie, J.; Lee, C.; Wang, M.; Liu, Y.; Feng, H. Characterization of heavily doped polysilicon films forCMOS-MEMS thermoelectric
power generators. J. Micromech. Microeng. 2009, 19, 125029. [CrossRef]
4. Yang, B.; Lee, C.; Kotlanka, R.; Xie, J.; Lim, S. A MEMS rotary comb mechanism for harvesting kinetic energyof planar vibrations.
J. Micromech. Microeng. 2010, 20, 065017. [CrossRef]
5. Lee, C.; Yu, A.; Yan, L.; Wang, H.; He, J.; Zhang, Q.; Lau, J. Characterization of intermediate In/Ag layers of low temperature
fluxless solder based wafer bonding for MEMS packaging. Sens. Actuators A Phys. 2009, 154, 85–91. [CrossRef]
6. Lee, C.; Lim, Y.; Yang, B.; Kotlanka, R.; Heng, C.; He, J.; Tang, M.; Xie, J.; Feng, H. Theoretical comparison of theenergy harvesting
capability among various electrostatic mechanisms from structure aspect. Sens. Actuators A Phys. 2009, 156, 208–216. [CrossRef]
7. Koh, K.; Lee, C.; Kobayashi, T. A piezoelectric-driven three-dimensional MEMS VOA using attenuation mechanism with
combination of rotational and translation Effects. IEEE/ASME J. Microelectromech. Syst. 2010, 19, 1370–1379. [CrossRef]
8. Wang, N.; Tsai, J.; Hsiao, F.; Soon, B.; Kwong, D.; Palaniapan, M.; Lee, C. Experimental investigation of a cavity-mode resonator
using a micromachined two-dimensional silicon phononic crystal in a square lattice. IEEE Electron Device Lett. 2011, 32, 821–823.
[CrossRef]
9. Zhu, J.; Liu, X.; Shi, Q.; He, T.; Sun, Z.; Guo, X.; Liu, W.; Sulaiman, O.B.; Dong, B.; Lee, C. Development Trends and Perspectives
of Future Sensors and MEMS/NEMS. Micromachines 2019, 11, 7. [CrossRef]
10. Suh, J.-H.; Cho, I.; Kang, K.; Kweon, S.-J.; Lee, M.; Yoo, H.-J.; Park, I. Fully integrated and portable semiconductor-type multi-gas
sensing module for IoT applications. Sens. Actuators B Chem. 2018, 265, 660–667. [CrossRef]
11. Han, J.H.; Bae, K.M.; Hong, S.K.; Park, H.; Kwak, J.-H.; Wang, H.S.; Joe, D.J.; Park, J.H.; Jung, Y.H.; Hur, S.; et al. Machine
learning-based self-powered acoustic sensor for speaker recognition. Nano Energy 2018, 53, 658–665. [CrossRef]
12. Jung, Y.; Hong, S.K.; Wang, H.S.; Han, J.H.; Pham, T.X.; Park, H.; Kim, J.; Kang, S.; Yoo, C.D.; Lee, K.J. Flexible Piezoelectric
Acoustic Sensors and Machine Learning for Speech Processing. Adv. Mater. 2020, 32, 35. [CrossRef] [PubMed]
Processes 2022, 10, 1658 23 of 25

13. Wang, X.; Gu, Y.; Xiong, Z.; Cui, Z.; Zhang, T. Silk-molded flexible, ultrasensitive, and highly stable electronic skin for monitoring
human physiological signals. Adv. Mater. 2014, 26, 1336–1342. [CrossRef] [PubMed]
14. Park, J.; Lee, Y.; Hong, J.; Lee, Y.; Ha, M.; Jung, Y.; Lim, H.; Kim, S.Y.; Ko, H. Tactile-direction-sensitive and stretchable electronic
skins based on human-skin-inspired interlocked microstructures. ACS Nano 2014, 8, 12020–12029. [CrossRef]
15. Jian, M.; Xia, K.; Wang, Q.; Yin, Z.; Wang, H.; Wang, C.; Wang, H.; Xie, M.; Zhang, Y. Flexible and highly sensitive pressure sensors
based on bionic hierarchical structures. Adv. Funct. Mater. 2017, 27, 1606066. [CrossRef]
16. Fernández-Caramés, T.M.; Fraga-Lamas, P. Electronics towards the internet of smart clothing: A review on iot wearables and
garments for creating intelligent connected e-textiles. Electronics 2019, 7, 405. [CrossRef]
17. Yang, T.; Xie, D.; Li, Z.; Zhu, H. Recent advances in wearable tactile sensors: Materials, sensing mechanisms, and device
performance. Mater. Sci. Eng. R Rep. 2017, 115, 1–37. [CrossRef]
18. Allison, L.; Hoxie, S.; Andrew, T.L. Towards seamlessly integrated textile electronics: Methods to coat fabrics and fibers with
conducting polymers for electronic applications. Chem. Commun. 2017, 53, 7182–7193. [CrossRef]
19. Fernández-Caramés, T.M.; Fraga-Lamas, P. A Review on the Use of Blockchain for the Internet of Things. IEEE Access 2018, 6,
32979–33001. [CrossRef]
20. Akpakwu, G.A.; Silva, B.J.; Hancke, G.P.; Abu-Mahfouz, A.M. A Survey on 5G Networks for the Internet of Things: Communica-
tion Technologies and Challenges. IEEE Access 2018, 6, 3619–3647. [CrossRef]
21. Blanco-Novoa, Ó.; Fernández-Caramés, T.M.; Fraga-Lamas, P.; Vilar-Montesinos, M.A. A Practical Evaluation of Commercial
Industrial Augmented Reality Systems in an Industry 4.0 Shipyard. IEEE Access 2018, 6, 8201–8218. [CrossRef]
22. Fraga-Lamas, P.; Fernández-Caramés, T.M.; Blanco-Novoa, Ó.; Vilar-Montesinos, M.A. A Review on Industrial Augmented
Reality Systems for the Industry 4.0 Shipyard. IEEE Access 2018, 6, 13358–13375. [CrossRef]
23. Fraga-Lamas, P. Enabling Technologies and Cyber-Physical Systems for Mission-Critical Scenarios. Ph.D. Thesis, University of A
Coruña, A Coruña, Spain, 2017.
24. Javaid, N.; Sher, A.; Nasir, H.; Guizani, N. Intelligence in IoT-Based 5G Networks: Opportunities and Challenges. IEEE Commun.
Mag. 2018, 56, 94–100. [CrossRef]
25. Loss, C.; Gonçalves, R.; Lopes, C.; Pinho, P.; Salvado, R. Smart Coat with a Fully Embedded Textile Antenna for IoT Applications.
Sensors 2016, 16, 938. [CrossRef]
26. Mzu, A.; Mmhb, C.; Aa, B.; Savaglio, C. A body sensor data fusion and deep recurrent neural network-based behavior recognition
approach for robust healthcare-ScienceDirect. Inf. Fusion 2020, 55, 105–115.
27. Sak, H.; Senior, A.; Beaufays, F. Long short-term memory recurrent neural network architectures for large scale acoustic modeling.
Interspeech 2014, 338–342. [CrossRef]
28. Williams, R.J.; Peng, J. An Efficient Gradient-Based Algorithm for On-Line Training of Recurrent Network Trajectories. Neural
Comput. 1990, 2, 490–501. [CrossRef]
29. Fortino, G.; Galzarano, S.; Gravina, R.; Li, W. A framework for collaborative computing and multi-sensor data fusion in body
sensor networks. Inf. Fusion 2015, 22, 50–70. [CrossRef]
30. Fortino, G.; Guerrieri, A.; Bellifemine, F.L.; Giannantonio, R. SPINE2: Developing BSN Applications on Heterogeneous Sensor
Nodes. In Proceedings of the 2009 IEEE International Symposium on Industrial Embedded Systems, Lausanne, Switzerland, 8–10
July 2009; IEEE: Torino, Italy; pp. 128–131.
31. Gravina, R.; Fortino, G. Automatic methods for the detection of accelerative cardiac defense response. IEEE Trans. Affect. Comput.
2016, 7, 286–298. [CrossRef]
32. Rebeiz, G.M. RF MEMS: Theory, Design and Technology; John Wiley & Sons: Hoboken, NJ, USA, 2004; pp. 87–120.
33. Kim, Y.J.; Hwang, E.S.; Kim, Y.H. MEMS-Based Biosensor; Springer: New York, NY, USA, 2008.
34. You, Z.; Li, B.; Yu, S. Applications of MEMS devices in nanosatellite. In Proceedings of the 2nd International Conference on
Recent Advances in Space Technologies, Istanbul, Turkey, 9–11 June 2005; pp. 240–243.
35. Brigante, C.M.N.; Abbate, N.; Basile, A.; Faulisi, A.C.; Sessa, S. Towards Miniaturization of a MEMS-Based Wearable Motion
Capture System. IEEE Trans. Ind. Electron. 2011, 58, 3234–3241. [CrossRef]
36. Akiko, Y. A new technique for direct measurement of the shear force necessary to detach a cell from a material. Biomaterials 1998,
19, 871–879.
37. Wu, G. Bioassay of prostate-specific antigen (PSA) using microcantilevers. Nat. Biotechnol. 2001, 19, 856–860. [CrossRef] [PubMed]
38. Mehrotra, P. Biosensors and their applications-A review. J. Oral Biol. Craniofacial Res. 2016, 6, 153. [CrossRef] [PubMed]
39. Zhao, C.; Thuo, M.M.; Liu, X. A microfluidic paper-based electrochemical biosensor array for multiplexed detection of metabolic
biomarkers. Sci. Technol. Adv. Mater. 2013, 14, 4402. [CrossRef] [PubMed]
40. Boisen, A.; Dohn, S.; Keller, S.S.; Schmid, S.; Tenje, M. Cantilever-like micromechanical sensors. Rep. Prog. Phys. 2011, 74, 036101.
[CrossRef]
41. Okan, M.; Duman, M. Functional polymeric nanoparticle decorated microcantilever sensor for specific detection of erythromycin.
Sens. Actuators B-Chem. 2018, 256, 325–333. [CrossRef]
42. Malmir, K.; Habibiyan, H.; Ghafoorifard, H. An ultrasensitive optical label-free polymeric biosensor based on concentric triple
microring resonators with a central microdisk resonator. Opt. Commun. 2016, 365, 150–156. [CrossRef]
43. Camli, B.; Kusakci, E.; Lafci, B.; Salman, S.; Torun, H.; Yalcinkaya, A.D. Cost-effective, Microstrip Antenna Driven Ring Resonator
Microwave Biosensor for Biospecific Detection of Glucose. IEEE J. Sel. Top. Quantum Electron. 2017, 23, 6900706. [CrossRef]
Processes 2022, 10, 1658 24 of 25

44. Partel, S.; Kasemann, S.; Matylitskaya, V.; Thanner, C.; Dincer, C.; Urban, G. A simple fabrication process for disposable
interdigitated electrode arrays with nanogaps for lab-on-a-chip applications. Microelectron. Eng. 2017, 173, 27–32. [CrossRef]
45. Leahy, S.; Lai, Y. A cantilever biosensor exploiting electrokinetic capture to detect Escherichia coli in real time. Sens. Actuators B
Chem. 2017, 238, 292–297. [CrossRef]
46. Hansen, K.M.; Thundat, T. Microcantilever biosensors. Methods 2005, 37, 57–64. [CrossRef] [PubMed]
47. Dammer, U.; Hegner, M.; Anselmetti, D.; Wagner, P.; Dreier, M.; Huber, W.; Güntherodt, H.J. Specific antigen/antibody interactions
measured by force microscopy. Biophys. J. 1996, 70, 2437–2441. [CrossRef]
48. Fritz, J.; Baller, M.K.; Lang, H.P.; Rothuizen, H.; Vettiger, P.; Meyer, E.; Ntherodttch, G.; Gerberand, J.; Gimzewski, K. Translating
Biomolecular Recognition into Nanomechanics. Science 2000, 288, 316–318. [CrossRef] [PubMed]
49. Wang, D.F.; Du, X.; Wang, X.; Ikehara, T.; Maeda, R. Improving picogram mass sensitivity via frequency doubling in coupled
silicon microcantilevers. J. Micromech. Microeng. 2016, 26, 015006. [CrossRef]
50. Thundat, T.; Wachter, E.A.; Sharp, S.L.; Warmack, R.J. Detection of mercury vapor using resonating microcantilevers. Appl. Phys.
Lett. 1995, 66, 1695–1697. [CrossRef]
51. Ramos, D.; Tamayo, J.; Mertens, J.; Calleja, M.; Zaballos, A. Origin of the response of nanomechanical resonators to bacteria
adsorption. J. Appl. Phys. 2006, 100, 061101. [CrossRef]
52. Nugaeva, N.; Gfeller, K.Y.; Backmann, N.; Lang, H.P.; Düggelin, M.; Hegner, M. Micromechanical cantilever array sensors for
selective fungal immobilization and fast growth detection. Biosens. Bioelectron. 2005, 21, 849–856. [CrossRef]
53. Moudgil, A.; Singh, K.K.; Swaminathan, S. MEMS based design and analysis of a biosensor for detection of hepatitis virus. In
Proceedings of the IEEE International Conference on Nanotechnology, Rome, Italy, 27–30 July 2015; pp. 805–808.
54. Kartanas, T.; Ostanin, V.; Challa, P.K.; Daly, R.; Charmet, J.; Knowles, T.P.J. Enhanced Quality Factor Label-free Biosensing with
Microcantilevers Integrated into Microfluidic Systems. Anal. Chem. 2017, 89, 11929–11936. [CrossRef]
55. Fortino, G.; Giannantonio, R.; Gravina, R.; Kuryloski, P.; Jafari, R. Enabling Effective Programming and Flexible Management of
Efficient Body Sensor Network Applications. IEEE Trans. Hum. Mach. Syst. 2013, 43, 115–133. [CrossRef]
56. Patra, J.C.; Kot, A.C.; Panda, G. An intelligent pressure sensor using neural networks. IEEE Trans. Instrum. Meas. 2000, 49,
829–834. [CrossRef]
57. Patra, J.C.; van den Bos, A. Auto-calibration and -compensation of a capacitive pressure sensor using multilayer perceptrons. ISA
Trans. 2000, 39, 175–190. [CrossRef]
58. Patra, J.C.; Ang, E.L.; Das, A.; Chaudhari, N.S. Auto-compensation of nonlinear influence of environmental parameters on the
sensor characteristics using neural networks. ISA Trans. 2005, 44, 165–176. [CrossRef]
59. Daponte, P.; Grimaldi, D. Artificial neural networks in measurements. Measurement 1998, 23, 93–115. [CrossRef]
60. Dias Pereira, J.M.; Silva Girao, P.M.B.; Postolache, O. Fitting transducer characteristics to measured data. IEEE Instrum. Meas.
Mag. 2001, 4, 26–39. [CrossRef]
61. Dias Pereira, J.M.; Postolache, O.; Silva Girao, P.M.B. A temperature compensated system for magnetic field measurements based
on artificial neural networks. IEEE Trans. Instrum. Meas. 1998, 47, 494–498. [CrossRef]
62. Zhang, X.; Liu, Q. Sensor Nonlinear Error Compensation Based on Neural Network Inverse operation. J. Nanjing Norm. Univ.
Nat. Sci. Ed. 2010, 33, 5.
63. Zhuang. Fault Diagnosis of Gas Sensor based on Neural Network. Ind. Instrum. Autom. 2002, 60, 54–55.
64. Shan, F.; Luo, J.; Wu, W.; Shen, X. Optimal Wireless Power Transfer Scheduling for Delay Minimization. In Proceedings of the
IEEE INFOCOM, San Francisco, CA, USA, 10–14 April 2016.
65. Kim, M.U.; Yang, H.J. Min-SINR Maximization with DL SWIPT and UL WPCN in Multicell Multi-Antenna Networks. IEEE Wirel.
Commun. Lett. 2017, 6, 318–321. [CrossRef]
66. Li, K.; Wang, Q. Study on signal recognition and diagnosis for spacecraft based on deep learning method. In Proceedings of the
Prognostics and System Health Management Conference (PHM), Beijing, China, 21–23 October 2015; IEEE: Piscataway, NJ, USA,
2015; pp. 1–5.
67. Guo, L.; Gao, H.; Huang, H.; He, X.; Li, S. Multifeatures Fusion and Nonlinear Dimension Reduction for Intelligent Bearing
Condition Monitoring. Shock Vib. 2016, 2016, 1–10. [CrossRef]
68. Verma, N.K.; Gupta, V.K.; Sharma, M.; Sevakula, R.K. Intelligent Condition Based Monitoring of Rotating Machines Using Sparse
Auto-Encoders. In Proceedings of the 2013 IEEE Conference on Prognostics and Health Management (PHM), Gaithersburg, MD,
USA, 24–27 June 2013; pp. 1–7.
69. Sohaib, M.; Kim, C.-H.; Kim, J.-M. A hybrid feature model and deep-learning-based bearing fault diagnosis. Sensors 2017, 17,
2876. [CrossRef]
70. Cheng, F.; Wang, J.; Qu, L.; Qiao, W. Rotor current-based fault diagnosis for DFIG wind turbine drivetrain gearboxes using
frequency analysis and a deep classifier. IEEE Trans. Ind. Appl. 2017, 54, 1062–1071. [CrossRef]
71. Chen, R.; Chen, S.; He, M.; He, D.; Tang, B. Rolling bearing fault severity identification using deep sparse auto-encoder network
with noise added sample expansion. Proc. Inst. Mech. Eng. Part O J. Risk Reliab. 2017, 231, 666–679. [CrossRef]
72. Bakshi, R.B. Multiscale PCA with application to multivariate statistical process monitoring. AKE J. 1998, 44, 1596–1610. [CrossRef]
73. Luo, R.; Misra, M.; Himmelblau, D.M. Sensor Fault Detection via Multiscale Analysis and Dynamic PCA. Ind. Eng. Chem. Res.
1999, 38, 1489–1495. [CrossRef]
Processes 2022, 10, 1658 25 of 25

74. Hui, D.; Liu, J.H.; Shen, Z. Drift reduction of gas sensor by wavelet and principal component analysis. Sens. Actuators 2003, 96,
354–363. [CrossRef]
75. Mallat, S. A theory for multiresolution signal decomposition: The wavelet representation. IIPEE Trans. Pattern Anal. Mach. Intell.
1989, 11, 674–693. [CrossRef]
76. Dunia, R.; Qin, S.J.; Edgar, T.F.; McAvoy, T.J. Identification of faulty sensors using principal component analysis. AKlhE J. 1996, 42,
2797–2812. [CrossRef]
77. Barker, M.; Hengst, M.; Schmid, J.; Buers, H.-J.; Mittermaier, B.; Klemp, D.; Koppmann, R. Volatile organic compounds in the
exhaled breath of young patients with cystic fibrosis. Eur. Respir. J. 2006, 27, 929–936. [CrossRef]
78. Buszewski, B.; Ligor, T.; Jezierski, T.; W-Piesik, A.; Walczak, M.; Rudnicka, J. Identification of volatile lung cancer markers by gas
chromatography–mass spectrometry: Comparison with discrimination by canines Analytical and Bioanalytical Chemistry. Anal.
Bioanal. Chem. 2012, 404, 141–146. [CrossRef]
79. Buszewski, B.; Kesy, M.; Ligor, T.; Amann, A. Human exhaled air analytics: Biomarkers of diseases. Biomed. Chromatogr. 2007, 21,
553–566. [CrossRef]
80. Amann, A.; Miekisch, W.; Schubert, J.; Buszewski, B.; Ligor, T.; Jezierski, T.; Pleil, J.; Risby, T. Analysis of Exhaled Breath for
Disease Detection. Annu. Rev. Anal. Chem. 2014, 7, 455–482. [CrossRef] [PubMed]
81. Dunia, R.; Qin, S.J.; Edgar, T.F.; McAvoy, T.J. Use of principal component analysis for sensor fault identification. Comput. Chem.
1996, 20, S713–S718. [CrossRef]
82. Gao, Z.; Chen, M.Z.Q.; Zhang, D. Special Issue on “Advances in Condition Monitoring, Optimization and Control for Complex
Industrial Processes”. Processes 2021, 9, 664. [CrossRef]
83. Pen, C.-L.; Chang, W.-J.; Lin, Y.-H. Development of Fuzzy Observer Gain Design Algorithm for Ship Path Estimation Based on
AIS Data. Processes 2022, 10, 33. [CrossRef]