12
IoT-based systems can provide an efficient and smart solution whichcouldbechallengingforcomplexAGImodels.Besides,
to detect plant stress[270], [271]. To detect the rice nitrogen the diverse range of IoT devices used in agriculture may
stress, Zhu et al. [272] developed a de-striping convolution causecompatibilityissues.Seamlessintegrationofthedevices
neural network to remove strip noise in hyperspectral image fromdifferentmanufacturersandstandardscanbeachallenge.
and designed a nitrogen diagnosis CNN to detect nitrogen Finally, protecting the vast data from cyberattacks is essential
stress for rice leaves. In addition, Elsherbiny et al. [273] to maintaining farm security and animal welfare.
combined CNN and LSTM to fuse the RGB images, the
weather-related factors and soil moisture collected by IoT
F. Smart Transportation
sensors to detect the water stress. The model trained by
multimodal data performed better than that by only RGB 1) Intelligent Motor Drives: The majority of recently pro-
images. posed detection methods can be classified as either physics-
5) Livestock monitoring: Precision livestock farming in- based methods or data-driven methods. Physics-based meth-
volves automated remote detection and monitoring of indi- ods commonly detect cyber-attacks by analyzing pre-defined
viduals for animal health and welfare through the analysis of system performance metrics or residuals between predicted
images, sounds, locations, weight and body condition [274]. system variables and corresponding true measurements [288].
It can detect issues in time and even predict potential issues However, most physics-based methods rely on accurate phys-
based on historical data [275]. Many measurements related to ical models of the target systems, which are unavailable
the health of animals are based on physiological responses, for most cyber-attack scenarios. In real-world applications,
such as body temperature, heart rate and respiration [276]. cyber-attacks are highly unpredictable, and their analytical
To measure heart rate, breathing rate, and oxygen satura- impact models heavily depend on specific attack policies.
tion of dairy cow, Salzer et al. [277] designed a nose ring These factors render the performance of most physics-based
sensor by integration of thermal and photoplethysmography methods unreliable. Recent research has begun to harness the
sensors. Besides, some studies mounted accelerometer, IMU powerofdata-drivenmethodstodevelopmodel-freedetection
and GNSS devices to monitor the behavior of animals[278], methods in power electronics systems, reducing dependency
[279], [280]. Arablouei et al. [281] fused the accelerometry on physical models [176] adopted a specific type of RNN,
and GNSS data to classify animal behavior. The performance namelyanonlinearauto-regressiveexogenousmodel,todetect
of posterior probability fusion is preferable to that based false data injection attacks in microgrids. [289] proposed an
on feature concatenation. However, many sensors are contact attack detection method by combining deep neural networks
or invasive, which may cause potential pressure on animals. and wavelet singular value decomposition. [290] employed
Therefore, many studies are focused on contactless methods, multi-class support vector machines to detect and localize
such as computer vision [282], [283], [284], [285]. Guo et false-data-injection and denial-of-service attacks in inverter-
al. [286] employed object detection methods to obtain the eye based systems. [291] proposed a detection and diagnosis
temperaturefromthermalimageoftheeyesocketin3seconds. method targeting data integrity attacks in solar farms using a
AGI can be a powerful and efficient solution for IoT multilayer long short-term memory network. [178] examined
applications in agriculture. AGI-driven IoT system can adap- theeffectivenessofvariousstandarddata-drivenmethodswith
tively regulate irrigation by analyzing multimodal data, such micro-PMUdataindetectingcyber-attacksinPVfarms.[292],
as the soil moisture, environment data and crop needs to [293] developed anomaly detection methods for electric ve-
optimize water usage and reduce water waste [287]. AGI can hicle traction motor drives using a combination of support
evaluate the soil and crop conditions as well as weed and vectormachines,randomforests,k-nearest-neighborhood,and
pestthreatsthroughIoT-sensorstorecommendprecisedosages logistic regression. [294] employed supervised classification
of fertilizers and pesticides, which can minimize overuse, methods to differentiate cyber-attacks and physical faults in
maximizecrophealth,andreduceenvironmentalimpact.AGI- manufacturing motor drives. Despite the advantages of re-
poweredmicroclimatemanagementsystemscancreateoptimal cently developed data-driven approaches, a significant chal-
conditions for crops in greenhouse by controlling factors like lenge in deep learning-based cyber-attack detection in power
temperature, humidity, light, and gas to improve the plant electronics systems is the requirement for large-volume train-
growth and productivity. In plant pressure detection, AGI ingdatasets.Themodelcanonlylearnfeaturesincorporatedin
can process various datasets from IoT sensors to detect early thetrainingdata,andthealgorithmmayfailwhentestingdata
signals of plant stress caused by diseases and pests. Growers containsdifferentfeatures[120],[117].Toaddressthisissue,a
can enable proactive interventions to prevent crop losses and large-scaletrainingdatasetisnecessarytoincludesimilardata
financialcosts.Inprecisionlivestockfarming,AGIcananalyze to the testing data. However, the computational cost due to
data from wearable IoT devices and contactless sensors to the large volume of training data hinders deep learning model
monitor the health, behavior, and productivity of individual performance.Transferlearningtechniqueshavebeenproposed
animals. This enables early disease detection and optimized to enable machine learning models to leverage knowledge
animalwelfare.However,therearealsosomechallengeswhen from one domain to another [295], thus reducing the amount
using AGI in IoT for agriculture. The agricultural environ- ofrequiredtrainingdata[296].Deeptransferlearningmethods
ment is complex, a robust and reliable AGI-driven system is havebeenutilizedincyber-attackdetectionandfaultdiagnosis
necessary. Furthermore, some agricultural applications, such in intelligent machine systems. Some methods [119], [120],
as spraying and irrigation, require real time data processing, [121] employ deep adversarial models to achieve transfer