s00521-020-04874-y

Neural Computing and Applications (2020) 32:16205–16233
https://doi.org/10.1007/s00521-020-04874-y
(0123456789().,-volV)
(0123456789().,-volV)
S.I. : APPLYING ARTIFICIAL INTELLIGENCE TO THE INTERNET OF THINGS
Machine learning and data analytics for the IoT

Erwin Adi1 • Adnan Anwar2 • Zubair Baig2 • Sherali Zeadally3
Received: 17 January 2020 / Accepted: 20 March 2020 / Published online: 11 May 2020
Springer-Verlag London Ltd., part of Springer Nature 2020
Abstract
The Internet of Things (IoT) applications have grown in exorbitant numbers, generating a large amount of data required for
intelligent data processing. However, the varying IoT infrastructures (i.e., cloud, edge, fog) and the limitations of the IoT
application layer protocols in transmitting/receiving messages become the barriers in creating intelligent IoT applications.
These barriers prevent current intelligent IoT applications to adaptively learn from other IoT applications. In this paper, we
critically review how IoT-generated data are processed for machine learning analysis and highlight the current challenges
in furthering intelligent solutions in the IoT environment. Furthermore, we propose a framework to enable IoT applications
to adaptively learn from other IoT applications and present a case study in how the framework can be applied to the real
studies in the literature. Finally, we discuss the key factors that have an impact on future intelligent applications for the IoT.
Keywords Cybersecurity Internet of Things Intelligent systems Machine learning
1 Introduction Play, smart vehicles, building automation systems com-

prising smart elevators and temperature control systems,
The Internet of Things (IoT) paradigm is both revolution- and unmanned aerial vehicles such as drones for environ-
ary as well as an enabler of automated and convenient mental monitoring and leisure. The large-scale prolifera-
lifestyles for modern day humans. The evolution of the IoT tion of IoT devices stretch beyond these devices to within
can be attributed to a confluence in advances that took the storage centers such as back-end cloud facilities which
place over the past decade in computing, communication, are geographically dispersed. As a result, a large volume of
and application design. The resulting sphere of influence of data is generated by IoT devices and their supporting
IoT has expanded rapidly to cover the whole human race. platforms, for transfer and subsequent storage and pro-
IoT devices in common use to facilitate our daily activities cessing at back-end cloud storage centers. IoT devices
include the smart phones, home assistants such as Google generate a constant stream of raw data, which cannot be
discerned for meaningful knowledge unless the data are
processed through application of techniques such as
& Zubair Baig knowledge discovery and machine intelligence. The
zubair.baig@deakin.edu.au heterogeneity of the data generated from various IoT
Erwin Adi deployments is dependent on the application domain,
e.adi@unsw.edu.au comprising, smart healthcare, social media, e-agriculture,
Adnan Anwar e-health, smart electricity grids, and smart vehicles. IoT
adnan.anwar@deakin.edu.au devices are designed with custom protocols that consider
Sherali Zeadally the resource constrained nature of these devices, in order to
szeadally@uky.edu preserve power usage associated with device operations.
1
UNSW Canberra Cyber, The University of New South Wales The most common IoT application-layer protocols are
Canberra at ADFA, Canberra, Australia Constrained Application Protocol (CoAP), Message
2
School of Information Technology, Deakin University, Queuing Telemetry Transfer (MQTT), Advanced Message
Geelong, Australia Queuing Protocol (AMQP), and HyperText Transfer Pro-
3
College of Communication and Information, University of tocol (HTTP) [76].
Kentucky, Lexington, USA
123
16206 Neural Computing and Applications (2020) 32:16205–16233
The MQTT and AMQP protocols are deployed on IoT Sect. 3, we define the application of machine learning for
devices with access to a continuous power supply, or to a processing and analysis of IoT data. We propose a novel
renewable power source. Such protocols facilitate transfer framework for data analytics on IoT data emerging from
of longer messages and are thus more power-hungry. On various heterogeneous sources, in Sect. 4. In Sect. 5, we
the other hand, the CoAP protocol is light-weighted and is provide future directions of research work in the discipline
designed for highly resource-constrained IoT devices; of machine learning applications to IoT, and we finally
where resource-constrained devices are those that have conclude the paper in Sect. 6.
very limited computing resources and network bandwidth.
HTTP protocol is the most resource-intensive IoT com-
munication protocol and is best suited for higher end IoT 2 The convergence of machine learning
devices, possessing higher communication, computation, and IoT
and storage capabilities. IoT devices produce a large vol-
ume of data that is locally processed, in limited manner, The convergence of machine learning and IoT paves the
and transferred to a centralized computing node or a cloud way for a prospective advancement in efficiency, accuracy,
storage facility, where it can be further processed or ana- productivity, and overall cost-savings for resource-con-
lyzed to produce knowledge. Machine learning is defined strained IoT devices. When machine learning algorithms
as a family of techniques for analyzing data, wherein the and IoT work together, we can achieve improved perfor-
process of model building on training data is automated, mance for communication and computation, better con-
i.e., requires little to no human intervention. Consequently, trollability, and improved decision making. Due to
the process of categorizing data into various classes is fully advanced monitoring from thousands to billions of ubiq-
automated. The role of data analytics for IoT data pro- uitous sensing devices and improved communication
cessing cannot be understated, and machine learning is a capabilities, IoT has enormous potential to improve the
very strong contributor to facilitate quick processing of quality of human life and potential applications for indus-
large volume data emerging from IoT devices, for gener- trial growth (toward Industry 4.0). IoT’s potential has
ating patterns of interest to analysts of the data. significantly improved with the convergence of machine
Contemporary computation and storage paradigms learning and artificial intelligence. Advanced machine
comprise cloud, fog, and edge computing. Through inte- intelligence techniques have made it possible to mine the
gration of these paradigms with the IoT, a robust data huge volume of IoT sensory data to have better insights
collection, storage, processing, and analytics framework into a range of real-world problems, as well as the ability to
emerge. Such a framework has the ability to provide real- make critical operational decisions. Therefore, to solve
time insights into data patterns and also facilitates the real-world complex problems and to meet the computation
application of machine learning techniques for realizing and communication requirements successfully, IoT and
intelligent data analytics for the IoT. The cloud paradigm is machine learning must complement each other. In recent
a centralized model of data storage that provisions various years, IoT data analytics has gained significant importance
services such as Software as a Service (SaaS), Platform as a and attention because of the following reasons:
Service (PaaS), and Infrastructure as a Service (IaaS), to
High volume of data generated from distributed IoT devi-
enable processing and analysis of IoT data, centrally [22].
ces According to the mobility report by Ericsson, the
Edge computing on the other hand enables the processing
forecasting shows that there will be 18 billion connected
and analysis of IoT data closer to the IoT network, on
IoT devices globally by 2022 [54]. This number will keep
localized computing nodes such as base stations. As a
increasing over time due to the wide adoption of IoT
result, the cost associated with transfer of IoT data to
devices in a wide range of critical applications. Intelligent
centralized nodes is avoided. Fog computing is a middle
data analytics will play an important role to identify and
ground solution between cloud and edge computing para-
predict the future states of any process or system by mining
digms, wherein the processing and analysis of IoT data
this huge amount of data efficiently and intelligently.
need not be located at the edge of the network or even in a
centralized storage facility. Rather, the fog paradigm brings High variability of data types from heterogeneous data
forth the concept of a virtual platform, not exclusively sources Due to a wide range of applications and require-
located at the edge of the network, for processing and ments, a large variety of IoT devices exist, which include
analysis of IoT data. mobile phones, PC/Laptop, tablets to short-range and wide-
In this paper, we study and analyze the role of machine area IoT devices. Due to the heterogeneity of the data, the
learning to facilitate data analytics for the IoT paradigm. features, formats, and attributes of the data are different.
We present a thorough analysis of the integration of Also, based on different IoT application domains, the data
machine learning with the IoT paradigm in Sect. 2. In sources also vary. For example, the IoT devices used for
123
Neural Computing and Applications (2020) 32:16205–16233 16207
medical applications will be different from a smart home 2.1.1 Descriptive analytics
IoT. Moreover, the quality, processing, and storage of data
have also become a challenging task because of its IoT systems can gather data from a few to thousands of
heterogeneity. In [95], the authors highlight some key smart devices and transmit them to a cloud environment.
questions arising due to heterogeneity of data sources. It Based on historical data, it is always possible to generate
includes answers to critical questions such as: how to deal detailed insights into past events by using advanced
with the sampling procedure of the high-frequency machine learning techniques. These groups of machine
streaming data, noise cancellation and filtering of the data, learning-based algorithms that process and summarize the
gathering and merging of the data from heterogeneous data raw data and provide actionable insights basically comprise
sources, data interpretation and interoperability, reasoning, the field of study called descriptive analytics. Data aggre-
situation awareness and knowledge creation from the data, gation, data summarization, mathematical logical opera-
gathering and storing data from heterogeneous data sources tions, data mining (e.g., clustering algorithms), and so on
to meet application’s constraints [95]. are some examples of descriptive analytics. Descriptive
analytics requires a high-volume data. Recent technologi-
Uncertainty in the IoT data streams Uncertainty is very
cal advances have demonstrated that cloud storage is cap-
common in practical data analysis [8]. It may arise in the
able of storing huge volumes of IoT data, and cloud servers
IoT data stream due to the failure of any IoT device or
can process complex tasks using high-performance com-
communication channel during data transfer. Gross errors
puters, and by applying IoT cloud analytics.
and missing data are omnipresent in IoT data streams,
which require advanced analytics to preprocess the data.
2.1.2 Predictive analytics for IoT
Even cyber intrusion could be a valid reason for uncer-
tainty in data [93]. In order to enhance the accuracy during
Predictive analytics rely on historical data and utilize
decision making, it is critical to ensure the proper assess-
advanced statistical or machine learning techniques to
ment, propagation, and representation of uncertainties and
model the behavior or pattern so that it is possible to pre-
develop models and solutions that can deal with these
dict the likelihood of possible future trends or patterns in
factors [95].
data. To summarize, it predicts what will happen in the
Balancing scalability with the efficiency Most of the IoT future by learning the historical patterns and data correla-
data analytics are performed in the cloud. Transferring data tions of existing data. Predictive analytics have been
from the IoT device to the cloud is expensive (in terms of widely used for different applications including predictive
delay), which may be challenging for time-critical appli- maintenance, prediction of price, supply-demand trend, or
cations especially when the number of IoT devices is high. prediction of likelihood of any outcome. According to
For example, in a connected vehicle environment, a large SAS, which is one of the top leading companies in ana-
pool of cars may be required to make decisions in real-time lytics, there are two types of predictive models—(i) clas-
or near real-time. Here, it is important to balance the speed sification-based models that conduct the prediction
and accuracy of the analysis when the number of vehicles analyses by class membership, and (ii) regression-based
increases. models that predict a number based on the historical
observations and likelihood [80]. State-of-the-art predictive
2.1 Classification of various analytics techniques modeling techniques include statistical regression-based
for IoT models, decision trees, and neural network or deep neural
network-based models. Some other widely used algorithms
In this section, we discuss different types of data analytics are based on Bayesian analysis, Gradient boosting,
techniques. Figure 1 shows the major component of the Ensemble model-based analysis, and so on. These predic-
analytical classes. tive analytics techniques are reliant on data for decision
making. The IoT paradigm can help facilitate the data
gathering process from smart IoT devices and can provide
an analytical framework using the cloud or the edge of the
network.
2.1.3 Prescriptive analytics for IoT
Prescriptive analytics suggests how to respond to any

future events based on data analysis. This class of analysis
Fig. 1 Major analytical classes not only predicts the future states but also provides
123
recommendations behind the adoption of the outcome. It is the use case scenario, e.g., public, private, hybrid or
more like a future scenario analysis technique that lever- community-based architecture [89].
ages the benefits of both descriptive and predictive ana-
lytics. While predictive analytics suggests what and when 2.2.2 Edge computing
the event will occur based on future predictions, prescrip-
tive analytics extends the capability by providing insights Edge computing simply refers to the analytical capability
of the future predictions with impact analyses. Prescriptive near the edge of the network of the IoT devices. The
analytics is widely used to optimize the business outcome. importance of edge computing can be better understood by
Prescriptive analytics is suitable in an Industrial IoT (IIoT) realizing the challenges and limitations of cloud comput-
setup where business intelligence-based decisions are made ing. Cloud computing is a centralized way of data pro-
by using the capabilities of cloud/edge computing, big data cessing and may increase the overhead of processing
analytics and machine learning. Services within an IoT- massive data [100]. Sending the raw IoT data streams to
cloud platform can help to make optimal decisions through the cloud would increase the cost and will require high
the deployment of business intelligence tools and through communication bandwidth and power. Instead, edge com-
analytics. puting enables the computing within the ‘local’ network of
the source of the data. Therefore, it solves the bandwidth
2.1.4 Adaptive analytics for IoT problem of sending bulk amounts of data to the central
server. Moreover, edge computing reduces the probability
During actual implementation, the outcome obtained from of a single point of failure and increases analytical effi-
the predictive analytics needs to adjust with real-time data. ciency. Compared with the huge and centralized servers as
For this purpose, adaptive analytics are used to adjust or found in cloud computing, edge computing has smaller
optimize the process outcome based on the recent history edge servers that are distributed. In [89], it has been
of the data and by looking at their correlations. This type of highlighted that cloud computing is suitable for delay-tol-
analysis helps to improve model performance and reduces erant and complex data analysis, whereas edge computing
errors. The advantage of adaptive analytics is that it can is suitable for low-latency real-time operations. Besides,
adjust the outcome of the solution when a new set of input cloud computing requires fairly complex deployments
data is received. Especially in an IoT environment, adap- when compared to edge computing. Figure 3 shows the
tive analytics are a good fit for real-time stream data pro- architecture of a collaborative edge-cloud model for IoT
cessing. Real-time assessment of evolving data streams, networks. From the figure, the end-devices and the core
such as those found in malware [101], can also be subject network cloud are connected through an IoT gateway
to adaptive analytics, to carry out data analytics. where the edge processing is performed. Typically, the
deployment of edge computing requires minimal planning
2.2 Classification of IoT data analytics based and in most cases can be ad hoc in nature [89]. As a result,
on Technological Infrastructure it is predicted that around 45% of the IoT data will use
edge architectures in the near future [90].
2.2.1 Cloud computing
2.2.3 Fog computing
In a cloud computing paradigm, remotely located com-
puting facilities (servers) are utilized using the Internet to Conceptually, fog computing lies somewhere in between
store, gather, manage, and process the data. In the past cloud and edge computing and it acts as a bridge between
decades, especially more recently, cloud computing has the cloud and edge resources [22]. The idea of fog com-
gained much attention because of the new infrastructure puting was first proposed by Cisco [17], to address the
and processing architectures that it provides to support issues associated with latency-sensitive applications.
different services including Software as a Service (SaaS), According to [17], unlike edge computing where comput-
Platform as a Service (PaaS), and Infrastructure as a Ser- ing happens at the edge of the network, fog computing
vice (IaaS) [89]. The cloud provides a variety of services provides networking services between the cloud and end-
delivered through the Internet, the concept of everything as devices as well as provides computing and storage facilities
a Service (XaaS) has emerged. In an IoT-based environ- through a virtual platform. As highlighted in the Cisco
ment, cloud computing can provide several advantages. whitepaper [41], the fog nodes have time-sensitive data
IoT interconnects smart devices. Within an IoT para- analytics capabilities and are a good fit for applications or
digm, a huge volume of generated data needs to be stored services with response times ranging from milliseconds to
and analyzed. A cloud architecture may vary depending on minutes. However, the data storing capabilities of the fog
123
nodes are very limited. On the other hand, cloud tech- blockchains through a blockchain-based network architec-
nologies can store data for up to months or even years [41]. ture. One significant difference between connected vehic-
Figure 2 shows various components of IoT and how ular networks and other IoT application areas is that
analytics relate with these components. Because of the high vehicles have high-speed mobility. Due to the environ-
computational resources and processing capabilities, most mental impact and the nature of the open wireless medium,
of the analytical approaches are suitable in the cloud server high-speed mobility could be a reason for vehicle data
of an IoT setup. On the other hand, lightweight and online faults [113]. This area requires significant research atten-
algorithms are better fit for the edge computing (Fig. 3). tion. Haibin et al. proposed the threshold-based fault
detection and repairing scheme using a dynamic Bayesian
2.3 Applications of IoT data analytics network (DBN) model in [113]. The model considers both
spatial and temporal correlation of the connected vehicle
2.3.1 IoT data analytics for smart vehicles data. This work shows that their proposed analytical
framework can detect faults with good accuracy and a low
IoT has enormous potential in the connected vehicle false alarm rate.
environment, especially for efficient and accurate decision He et al. proposed two advanced analytical frameworks
making using advanced data analytics. Within an Internet for vehicle warranty analysis in the connected vehicle
of vehicle (IoV) paradigm [25], a large number of broad- environment [48]. Two improved data mining algorithms
casting messages are frequently generated with very high based on the Bayes model and a Logistic Regression model
granularity and volume. Therefore, one of the biggest are proposed. These machine learning algorithms are used
challenges is storing and intelligent management of the for effective data mining service development using cloud
huge amount of data [61]. Another key issue is related to data analytics. The authors also demonstrate the intelligent
ensuring the security of the data, as any cyber-related parking cloud service [48]. Typically, it is very challenging
anomalies or cyber intrusion attempts will jeopardize the to find an available parking lot, especially in urban area.
system and may cause fatalities. Hence, it is important to Moreover, it may lead to traffic congestion, stress and
look for intelligent solutions that can deal with such cyber- increase the probability of accidents. To solve this chal-
related incidents. Moreover, a centralized solution lenge, a ‘birth-death’ model is proposed in [48]. The
approach may not be feasible as it is more prone to a single stochastic ‘birth-death’ is a special case of continuous-time
point of failure. In [61], the authors utilize a blockchain- Markov process. This model is utilized in order to predict
based approach for distributed and secure storage of the occupancy probability of any parking space.
vehicular data. The authors also define how the nodes in An autonomous and real-time method has been pro-
the connected vehicle environment could be integrated into posed for crime detection on public bus services in [77].
Fig. 2 Convergence of IoT

architecture and data analytics:
an illustration
123
Fig. 3 A generalized system

model for collaborative edge-
cloud processing in
heterogeneous IoT networks
[89]
The authors highlight the advantages of IoT technologies vehicular environment using data analysis and effective
and demonstrate the advantage of data analytics within a communication. In the first layer (Perception layer), data
fog computing environment. A framework has been are gathered from various IoT sensors and devices. For
developed for intelligent public safety in a connected example, a passenger’s smart cell phone can provide the
vehicle environment. Advanced data analytical capability locational information of the passenger, whereas a smart
for detection, prediction, and prevention of crime has been travel card may contain information related to the pas-
the key focus of this study. The proposed framework shows senger’s identification and financial transactions [72].
significant improvement in performance and device sur- Other smart IoT devices of the environment are smart
vivability over the traditional smart transportation safety terminal boards, Automatic Vehicle Location (AVL),
use cases [77]. Automatic Passengers Counting (APC), digital video
Other than crime detection, IoT connected vehicular cameras, and devices installed for road safety like smart
data analytics can be of use for other critical applications. road cameras, etc. [72]. The second layer (Network layer)
For example, Luo et al proposed a three-tier framework of in the framework is responsible for transferring the source
the connected public transportation system for effective data to the application layer (final layer) using both wired
traffic management [72]. Here, the authors propose inno- and wireless communication. Standard communication
vative solutions for effective scheduling of subway, bus, technologies are used, e.g., Mobile Communication
and shared taxis in an IoT connected environment. A (GSM), General Packet Radio Service (GPRS), Code
periodic pattern mining-based algorithm is deployed for Division Multiple Access (CDMA), the third-Generation
determining the passenger and road flows. Finally, the communication (3G), and the fourth-Generation commu-
capability of evolutionary-based algorithms is used to nication (4G), etc., for public wireless communication and
develop a computational model for dynamic bus schedul- Zigbee, Wi-Fi and Bluetooth for private wireless commu-
ing [72]. Figure 4 shows a three-tier architecture of the nication [72]. In the Application layer, IoT data are used by
IoT-based system for several applications in a connected advanced analytics. Here, optimization and machine
123
Fig. 4 A three-tier architecture of an intelligent transportation system based on IoT [72]
learning algorithms are used for decision making of various prime focus of their work. The authors in [81] show that
applications deployed for subway, bus and shared taxi machine learning algorithms like linear regression and
networks. logistics regression analysis can provide an accept-
Intelligent parking space allocation has been the focus of able outcome to identify these critical parameters.
research presented in [64]. Kong et al. developed a sys- Swarm intelligence-based decision-making approach
tematic approach using IoT and cloud computing to solve has been proposed in [19]. Considering an IoT-connected
the parking allocation problem. The scheme utilizes auc- vehicle environment, this work shows that smart vehicles
tion-based mechanisms to solve the parking allocation can make dynamic and adaptive decisions. Here, the
problem. In [81], Priyashman et al. explored the capability authors show how IoT-based communication is effectively
of two machine learning algorithms for vehicular access utilized for traffic intensity calculations and subsequent
control and identification. Identification of two key critical utilization of IoT data for dynamic decision making using
factors—signal strength and tag settings—has been the swarm intelligence [19].
123
2.3.2 IoT data analytics for smart healthcare learning models. This computational model utilizes the
smart IoT framework for monitoring and communication.
The rapid growth of IoT has benefited different application Early detection of Dyslexia, which is a cognitive dis-
areas including healthcare [109]. IoT technology along ability that affects the regular activities of an individual, is
with data analytics can help to achieve more accurate and possible by using a ledger-based technology within an IoT
improved health diagnoses [37]. One critical requirement is framework [82]. The proposed approach gathers data from
to gather data, predict and make decisions in real time. In smart IoT devices such as smartphones, tabs or laptops and
[37], the authors highlight the importance of the real-time stores the data using blockchain-based distributed ledger
pattern recognition technique deployment for the con- technology. Due to the use of blockchains, it is available to
struction of genomics-based patient models. A dynamic the medical practitioner for evaluation irrespective of the
and adaptive computational model with intelligence needs location, and it also provides security of the information.
to be developed. These innovative models should be able to An agent-based IoT simulation testbed has been devel-
capture the data produced by thousands of IoT-connected oped in [42] to investigate a patient’s sleeping behavior. In
nodes within the smart healthcare paradigm. Advanced the context of a smart bed, the simulator analyzes the
analytics and communication within a connected health- sensor information to detect the posture during sleep time.
care system can ensure lots of benefits including improved An emotion detection model is proposed in [51]. The
resiliency, seamless fusion with different technologies, big proposed solution receives the speech and image signals
data processing and analytics, personalized forecasting of from IoT devices. These signals are analyzed on the edge
patient condition, lifetime monitoring of patient health, cloud and remote cloud to make the decisions. Fourier
ease of use of wearable devices, overall medical health cost transform, different filtering algorithms and support vector
reduction, physician oversight with real-time patient data, machines (SVMs) are used for the detection. The proposed
availability and accessibility of the doctors through framework is suitable to deploy in a big data environment.
advanced communication and efficient healthcare man-
agement [37]. It is also highlighted that the model-based 2.3.3 IoT data analytics in agriculture
approach, which is widely used in the industry, may not be
a suitable fit for the health domain because medical data IoT-based frameworks can be adopted to improve the
changes continuously and is highly prone to uncertainty. operational efficiency and accuracy by using advanced
Therefore, instead of a model-based approach, a data-dri- analytics of the data generated from smart end devices.
ven approach could be a better solution [37]. Figure 5 Several challenges and benefits of IoT and data analytics-
shows a multilayer architecture of the eHealth cloud [37]. enabled smart agriculture frameworks are discussed in
While the data generation and communication take place in [33]. In the recent past, the prospects, opportunities, and
Zone 1 (see Fig. 5), the main computational process is feasibility of Wireless Sensor Networks (WSN) in agri-
conducted in Zone 2; real-time and streaming analytics are culture have been widely studied [75]. The application of
located here. Tools such as Hive, Spark [26, 32], MapRe- WSN in agriculture includes environmental monitoring,
duce [32], HDFS, YARN are used for this purpose in Zone precision, agriculture, machine and process control
2 [37]. Finally, outcomes are displayed or actuated based automation and traceability [33, 57]. Recently, the impor-
on the requirement of the application or services in Zone 2. tance of IoT has been highlighted because of its versatility
A smart sleep monitoring system is proposed in [108]. of the adoption of different wired and wireless technologies
The authors propose a fog computing-based preprocessing as well as the capability to integrate with advanced data
approach implemented on the smart device. Next, the analysis mechanisms. IoT in agriculture empowers the
capability of big data on the cloud is utilized by performing farmers with advanced and improved automation and
batch data processing. Predictive analysis is performed to decision-making processes [33]. It also enables seamlessly
identify the air quality for the treatment of Obtrusive sleep integration among agricultural products, services and
apnoea (OSA) [108]. The outcome of the analysis is dis- knowledge to improve the quality and productivity [33].
played using web user interface to benefit the health pro- IoT is deployed for agricultural asset monitoring and
fessionals in real-time irrespective of their location of management. RFID-based tags are also used for tracking
operation. and tracing of supply chain products [105]. GPS can also
An IoT-enabled automated nutrition monitoring system be used for the same purpose within an IoT framework
is proposed in [96]. The authors propose a meal-prediction along with edge/cloud data analytics [33]. Typically, the
algorithm using a detailed analysis of Bayesian networks. humidity, wind speed, temperature, cloud transients, rain-
To investigate the nutrition balance, another algorithm is fall and weather parameters are very critical for effective
proposed that uses perceptron neural network-based deep farming. IoT with machine intelligence can also help to
predict the localized weather of the agricultural firms.
123
Fig. 5 Multilayer architecture of the eHealth cloud [37]
Moreover, the efficiency of agricultural waste and water irrigation management, greenhouse gas monitoring, and so
management can be improved with the smart use of IoT on [36].
and the big data paradigm. By using intelligent scheduling
algorithms, the challenge of storage management can be 2.3.4 IoT data analytics in energy systems
resolved. IoT devices can be placed strategically to monitor
the storage facilities and then cloud technologies can In recent years, the energy grid has transformed signifi-
deploy machine learning and optimization algorithms to cantly due to the integration of photovoltaics solar cells,
make the decisions [33]. electric vehicles, and storage with the low-voltage distri-
Figure 6 shows the major component of IoT-based smart bution network. These distributed energy resources need to
farming. Out of the four major components, three are be effectively coordinated and controlled. One viable
related to the data (highlighted in the shaded box). The solution is to use the capabilities of smart meter. Within an
physical structure controls the sensors, actuators and IoT environment, smart meters are interconnected and
devices and is responsible for the overall precision in data linked with each other. Advanced analytical decisions are
processing [36]. Data acquisition and data processing are necessary for improving the efficiency and reliability of
similar to those found in standard IoT applications, based energy operations. Therefore, in [4], the authors pointed
upon state-of-the-art IoT protocols. For the data analytics out that the combination of IoT and Big Data can be used
component, various algorithms and models are used for for effective energy management. The authors develop an
data analytics. Some IoT and data analytic-based smart energy management system that can be used in a smart
agriculture applications include livestock health monitor- home. Based on the unique ID address for each IoT object,
ing, stress level monitoring, physical gesture recognition, a System-On-Chip (SoC) module is deployed for data
Rumination, heart rate tracking, livestock location moni- acquisition. A centralized server is used within a Home
toring, climate condition monitoring, pest identification, Area Network, where data are collected, stored and pro-
cessed. One advantage is that the proposed solution is
123
Fig. 6 Major components of

IoT-based smart farming [36]
capable of using off-the-shelf business intelligence tools. within an IoT environment [3]. The combination of CEP
The solution architecture can significantly improve the and historical data has been used to develop a prediction
efficiency of high energy consumption applications such as algorithm using machine learning algorithms in [3]. Here,
air conditioning. This has been validated in a laboratory the authors extend the work for a dynamic IoT data envi-
environment. Considering a smart home architecture, IoT ronment by proposing an adaptive predictive model using
and sensor data optimization-based peer-to-peer energy the moving window regression technique. To determine the
trading architecture is proposed in [55, 85] optimal window size, authors use a spectral component
Complex Event Processing (CEP) is used for reactive analysis of the time series data.
applications to provide a distributed but scalable solution
123
In [6], the authors surveyed the possibilities and SVM-based IDS outperforms the state-of-the-art tech-
opportunities of big data and cloud analytics for the smart niques in accuracy.
grid. Once data are collected from the smart meters, tools
in a cloud platform can perform advanced analytics. IBM
Coremetrics [24] and Google BigQuery [16] are some 3 A critical review on data processing
examples of cloud-based software solutions for data and knowledge discovery for IoT
analytics.
Predictive analysis for decision support systems of a 3.1 Transforming data to knowledge
smart meter IoT network has been presented in [91]. Here,
the authors use a machine learning-based approach for cost IoT devices must process data to produce meaningful
prediction and to improve the performance of the network knowledge. How data becomes meaningful to its consumer
operation. In this work, a Bayesian network model is is illustrated through the architecture shown in Fig. 7. In
compared against three other classifiers including Naı̈ve this architecture, the consumer can be a machine or human.
Bayes, random forest, and decision trees. The proposed When the consumers expect to receive meaningful infor-
method in [91] is validated using the network coverage mation, the data processing unit consists of AI algorithms
data collected from commercial settings. such as machine learning techniques. In this article,
The IoT-connected smart grid integrated with machine meaningful processed data are also referred as
learning can provide several benefits including predictions ‘‘knowledge.’’
of consumption, price, power generation, future optimum Prescriptive, predictive, and adaptive learning in the IoT
schedule, fault detection, adaptive control, sizing, and can be best described through how the applications process
intrusion or anomaly detection [50]. For such applications, data to create knowledge. The above architecture is used
connectivity and data exchange play a significant role. IoT throughout this section to organize the evaluation on how
is being widely deployed for this purpose. In [50], the current studies implement AI in the IoT paradigm. This
authors also discussed the security issues in IoT data ana- section highlights the implementation challenges, discusses
lytics. Lots of research [7, 9, 46, 56] has been done to how studies address them, and identify future research
investigate security issues of traditional energy grid and opportunities.
Supervisory Control And Data Acquisition (SCADA) sys-
tems using various attack detection algorithms [49, 103], 3.2 The application-layer protocol
there is a need to analyze the impact of cybersecurity on
IoT-connected smart grid. In [50], the authors also high- In machine-to-machine communications, an application-
lighted that machine learning can play an important role in layer protocol is required to send data in the format that can
cyber-attack detection in smart grid. be encapsulated by the lower layer protocols. As mentioned
A machine learning-based multi-stage theft detection previously, the widely used IoT application-layer protocols
system is proposed in [69]. In the first stage, multi-modal are Constrained Application Protocol (CoAP), Message
forecasting is used for predicting the consumption behav- Queuing Telemetry Transfer (MQTT), Advanced Message
ior. Multilayer perceptron (MLP), recurrent neural network Queuing Protocol (AMQP), and HyperText Transfer Pro-
(RNN), long short-term memory (LSTM), and gated tocol (HTTP). All these protocols allow for abstraction
recurrent unit (GRU)-based machine learning models are from the lower layers.
used in this stage. In the second stage, decisions are made. Constrained devices have limited computing, commu-
Here, a moving average model is used to identify the nication, and storage resources. Since different IoT devices
anomalies. In the final stage, the historical energy usages have varying resource requirements, the protocols differ in
are used to decide whether the anomaly based on stage 2 is their capabilities to meet these requirements. Resource
an energy theft or not [69]. requirements were derived from the message sizes, over-
A lightweight machine learning-based intrusion detec- head, and latency. As illustrated in Fig. 8, CoAP is the most
tion system (IDS) for IoT is developed in [58]. The pro- lightweight, while HTTP is the most resource intensive.
posed IDS is capable of detecting Denial of Service (DoS) CoAP is used in applications where IoT data are collected
attacks successfully. The authors use SVM-based classifi- from constrained devices at the edge. These devices are
cation techniques to develop the IDS. Unlike other works, typically light in weight, and they include wearable devices
the proposed work considers only one attribute to model or the remote, low maintenance sensors. They have very
the classifier. This assumption made the model computa- limited computing resources, power and capability to
tionally efficient and helped achieve faster solutions. The transmit data wirelessly. Such applications therefore do not
comparative results presented in this paper show that the send knowledge back to the IoT devices. Hence, devices
123
Fig. 7 The architecture of

knowledge creation in the IoT,
adapted from [98]
processed either on the data processor in the cloud or on the

IoT device.
In the third model, smart devices have enough com-
puting resources to turn data to knowledge locally or at the
Fig. 8 IoT application protocols, i.e., Constrained Application
edge. This is suitable for delay-sensitive applications.
Protocol (CoAP), Message Queuing Telemetry Transfer (MQTT),
Advanced Message Queuing Protocol (AMQP), and HyperText Adaptive learning is therefore processed at the edge, either
Transfer Protocol (HTTP). Adapted from [76] on the data processing machine or on the IoT device.
Table 1 illustrates how application protocols were
employing CoAP are suitable for prescriptive and predic- employed in the current literature. The table shows the
tive AI applications, but not adaptive. applications of the studies and the type of data collected
MQTT and AMQP are employed when the applications from the sensors to illustrate the suitability of protocols to
can afford higher-end IoT devices. The IoT devices are transfer the data. The table also shows the type of model
accessible for power recharge or are directly connected to (according to Fig. 7) that the studies adopt. The table pre-
the power supply. In these applications, data can be for- sents the knowledge to illustrate the results from the data
warded to other machines or consumers across the Internet, processing done by the machine learning techniques
including back to the IoT devices. Hence, this scenario can adopted in the studies. The three models show that
be used for prescriptive, predictive, and adaptive AI although application protocols address how IoT applica-
applications. tions can adopt AI to some extent, there are some limita-
On the one hand, the application protocols have tions on how IoT devices adaptively learn.
addressed how constrained devices can send data. On the A recent study [107] showed how a combination of
other hand, machine learning analysis requires high com- these application-layer protocols can be employed to
puting resources and a large amount of data to process the exchange knowledge in the IoT environment. Different
collected data to convert it to meaningful knowledge. protocols were selected according to their degree of con-
Therefore, IoT applications vary the architecture of straint. CoAP was used to allow local communication
knowledge creation (Fig. 7) to transform data to knowl- between a wearable device and an IoT gateway; HTTP was
edge based on the framework and the network bandwidth used to exchange knowledge between the IoT gateway and
available for data transmission. Figure 9 illustrates three the cloud solution; and MQTT was used to send alert from
models on how the architecture of knowledge discovery the IoT gateway to the subscribers such as emergency
can be applied. services and family members. As Table 1 shows, com-
The first model is when a constrained device sends data bining protocols are commonly adopted to support the
to a remote system in a simplex manner. This is the case in varying resource requirements for different types of
data mining, where the data are used or shared with dif- applications.
ferent applications. Each remote application can have its Current studies employing IoT data for machine learn-
own method to transform data to knowledge. Adaptive ing focus on how the consumers can benefit from the IoT
learning takes place in the cloud, since feedbacks cannot be technologies by proposing where the data should be pro-
sent back to the IoT device. In other words, the IoT device cessed, and how an application layer protocol selection can
is unable to adaptively learn. be justified to fit its purpose. The literature reviews [5, 76]
In the second model, constrained devices can send/re- compared how the application-layer protocols exchange
ceive data in duplex communications. The remote system, data based on their client-server architecture, resource
e.g., the cloud, processes the data and brings knowledge discovery framework such as Representational State
back to the device. This model can be seen in many Transfer (REST) and Service-oriented Architecture (SOA),
smartphone applications. Adaptive learning may be the transport protocol such as Transport Control Protocol
123
Fig. 9 Where data are

transformed to knowledge
(TCP), and User Datagram Protocol (UDP), data query preprocessing the data is needed for its use in supervised
method such as Structured Query Language (SQL) and machine learning techniques. In this study [47], data from
publish/subscribe, security protocol, and packet header/ the old IoT devices were adapted to handle the data from
message size. However, few studies have compared how the new devices through a gateway (Fig. 10). Clinical
the current application layer protocols adaptively learn records from doctors were used to train machine learning
knowledge in machine-to-machine communications. techniques, to produce data that can synchronize the old
Although Table 1 shows that knowledge is created in these and the new wearable devices. Here, the gateway serves as
studies, machine-to-machine autonomous learning was not a data adaptor.
fully explored. Creating knowledge from data is considered Data from IoT devices may be represented in raw for-
as the function of a multipurpose box wherein all required mat. Hence, gateways are devised to transform raw data to
data are assumed to have been available before data pro- a usable format, such as in the form of features that can be
cessing employing AI or machine learning techniques. employed for machine learning analysis [47, 67, 107]. The
transformed data may be encapsulated in a network packet
3.3 The gateway with a header and other meta-information such as sensor
messages and addresses, for transmission to other devices
Simply put, a gateway is a multipurpose box at the edge to [60].
perform the necessary action [59]. A gateway converts, As such, gateways can act as an adaptor that convert one
routes, or even processes data. It may be used as a filter that protocol to another. At the application-layer, they adapt
selects data of interest and drops the noise [13, 59], or as based on the protocol used at the edge to the one used by
hub that aggregates data from several IoT devices [68]. the cloud. For example, from constrained IoT devices
While aggregated data at the edge may be too big to be employing CoAP to a cloud service that communicates
transmitted over the Internet, gateways can be designed as through MQTT or AMQP [44, 60, 107]. They may also
the data processor to produce and transmit low bandwidth convert the lower-layer wireless protocols of the local,
messages [68, 107]. Depending on the system specification, constrained devices for enabling Internet-based transmis-
a gateway may not be needed, or may be integrated with sion to the cloud [27]. At the network layer, gateways may
the devices. be used to optimize routing, for example, by selectively
The problem in processing data from the IoT sensors is prioritizing packets in networks connecting vehicles, which
due to the heterogeneity of the data. Consider a health are bursty in nature. In this case, a gateway is deployed to
monitoring system that employs wearable devices to gen- minimize the number of delayed packets [106].
erate data for medical devices [47]. Data heterogeneity Since gateways perform many tasks, it is challenging to
issue will arise when the number of new wearable IoT determine the computing resources required to deploy one.
sensors grows at a much faster pace than how the medical Therefore, studies show how their resource consumption
devices could adapt to the variety of the generated data. can be measured [65]. Relevant evaluation for deployment
The data are unlabeled and varies in dimension. Therefore, includes how throughput and latency perform under
123
Table 1 Knowledge created from the data produced by different applications

Application (from the title of the Protocol Model IoT data Machine Knowledge
publication) learning
Predicting energy consumption MQTT, 1 Energy consumption, temperature, Random Predicted energy
through machine learning using a CoAP humidity, precipitation, wind forest consumption
smart-metering architecture [27] speed, time direction; all at
15-min interval
Smart city traffic monitoring system MQTT 1 Magnetic, light, temperature, fuel Azure Optimum traffic route
based on 5G cellular network, RFID consumption, time to reach machine
and machine learning [79] destination, vehicle weight, road learning
slope
Namatad: inferring occupancy from MQTT 1 Temperature, CO2, air volume, Random Predicted occupancy level
building sensors using machine HVAC temperature forest
learning [30]
Lameness detection as a service: MQTT 1 Step count, lying time, swaps per Random Lameness level (active,
application of machine learning to an hour forest, k-NN normal, dormant)
internet of Cattle [21]
Case study: integrating IoT, streaming AMQP 1 Blood sugar reading Azure Predicted sugar level
analytics and machine learning to machine
improve intelligent diabetes learning
management system [10]
An IoT system to estimate personal AMQP 1 Heart rate, skin temperature, room Support Thermal comfort level
thermal comfort [66] temperature, humidity, air speed vector
machines
Cloud computing-based smart garbage MQTT 1 Level of garbage in a bin, Decision Predicted bin filling
monitoring system [62] timestamp forest pattern
regression
An IoT-based solution for intelligent MQTT, 1 Sheep posture: neck inclination Rule engine Feed on vines or weeds
farming [78] AMQP and distance to the ground
IoT and distributed machine learning MQTT 2 Location, acceleration, heart beat Kalman filter Product recommendation
powered optimal state recommender
solution [86]
Quantifying colorimetric tests using a AMQP 2 Average red, green, blue values Support Quantified peroxide
smartphone app based on machine vector content
learning classifiers [92] machines
On delay-sensitive healthcare data CoAP 3 Pulse rate, respiratory rate, oxygen Deep learning Anomalous health risk
analytics at the network edge based level, sleep condition, fall
on deep learning [35] detection, gait tracking,
washroom visit frequency
Fall detection system for elderly CoAP, 3 Accelerometer data, gyroscope Ensemble Fall (forward, backward,
people using IoT and ensemble MQTT, data random lateral) and normal
machine learning algorithm [107] HTTP forest activities (walking, stairs
climbing, sitting)
Vibration condition monitoring using CoAP 3 Motor vibration level in mV Neural Motor condition (normal,
machine learning [111] network unbalanced)
Early detection system for gas leakage MQTT 3 Temperature, humidity, gas, Classification Disaster risk levels
and fire in smart home using machine smoke, CO, flame, CO2 and
learning [84] regression
trees
varying loads, such as increasing the clients when running heterogeneity from varying sensors, pre-process the data
a combination of services. and synchronize communications with the cloud. However,
Despite having gateways perform the necessary actions, the solution was not directly adopted to adapt heteroge-
they were designed for and evaluated under each specific neous data for different studies. In other words, the current
case study only. The current studies in employing IoT data challenges include not only adapting data from new devi-
for machine learning emphasize on how to adapt data ces, but also from new knowledge, i.e., meaningful data
123
Fig. 10 Gateways adapt

heterogeneous data from
different devices
generated from AI machines. While a gateway is a device level events into management-level events [71]. This
to perform the necessary action, there is a need to integrate management-level information is the knowledge gained.
it with the necessary actions required for discovering Hence, in CEP, knowledge is implemented as maps
knowledge. between event layers. For example, a global pandemic
watch system (as knowledge) was implemented as corre-
3.4 Deriving knowledge from data lations between various event layers such as news feeds
around the globe, short message service (SMS) messages
IoT sensors generate a stream of raw data, much of which from trained agents in some designated area, and electronic
is meaningless to the consumer. Raw data are a blend of reports from health authorities [71].
insight and noises. It does not show a change in events, When correlating events, a knowledgebase is needed to
quantization per time, or meaningful attributes and rela- define what the correlations are. While the relationship
tionships used by the consumers. The challenge to derive definitions are traditionally provided by experts, advance-
meaningful data from sensors is apparent in the IoT when ments in machine learning techniques allow self-assigned
the pervasive sensor devices deliver varying context and rule creation from input streams. In this case, the knowl-
media. In this case, data processing in the IoT follows the edgebase is created as rules generated from a set of
method described in the Complex Event Processing (CEP) supervised rule-learning algorithms [74]. The problem with
approach, which takes events from input streams to process supervised machine learning techniques is that the labeling
data, regardless of the technology or protocol employed process in creating training datasets still requires expert
(Fig. 11). interventions rather than reliance on machine-to-machine
Capturing events implies creating processable data out communications.
of the raw data. Data are sent as input streams; events are Rather than manual intervention, ontologies have been
extracted from some patterns such as a change of state or used to address how machines label data. An ontology is a
time in the data. A series of states can create a pattern of network of objects describing their properties and rela-
interest. However, meaningful patterns can become stale tionships. Given an ontology of a domain, labels can reveal
when a system fails to process them within a specific time. their meaning within the domain. One method of describ-
Thus, systems that derive meaning from input patterns ing meaning is by showing the object property in a label-
implement event listeners. value pair. For example, in the domain of Vehicles, events
Event listeners are implemented in many programming such as speed can be labeled with low-, medium-, and high-
languages, with Java presenting an elegant implementation speed, each with a range of values. A recent study showed
for parallel processing; and as one that can integrate low- how machines can self-assign their labels, given speed
Fig. 11 Complex Event

Processing (CEP) engines listen
to events from input streams to
process data, adapted from [71]
123
values at the input event stream [83]. Another method to The Semantic Web employs HTTP to exchange the
describe meaning is by expressing the object relationship in information describing graphs. This is illustrated in
a subject-predicate-object form. For example, given a Fig. 12b. Each node and edge have a unique Uniform
knowledgebase with an entry ’’access point X is in location Resource Identifier (URI), which is manifested in the form
Y’’ and ’’A has phone B’’, one may derive knowledge that of a web address. Hence, the graph allows machines to
’’A is in location Y’’ upon an event ’’B is connected to follow the ‘‘scent’’ to find information about a specific
access point X’’ [34]. In both above scenarios, knowledge object, such as who Eric Miller is and what his email
is derived through machine-to-machine communications. address is. This demonstrates how an application-layer
While ontologies address how data can be labeled and protocol, HTTP, provides a method for knowledge
associated, the current work still requires a holistic system exchange in a machine-to-machine communication.
that can share outputs between different machines [71]. Locating object properties and relationships is not confined
However, the current reviews on CEP are concerned with through URI as described by the Semantic Web. Compet-
how the approach is implemented and how the various ing approaches include service-oriented knowledge dis-
implementations affect the overall throughput, CPU time covery, which will be demonstrated in the next section.
cost, and communications cost [38]. Consequently, more In its deployment, a gateway can be installed as a hub at
recent studies have proposed different optimization strate- the edge of network to exchange ontology information with
gies such as query reordering, memory management, and the cloud using HTTP, while facilitating local communi-
parallelization, with each strategy compared in terms of cations with the IoT devices under CoAP [107]. The
time complexities. Studies in CEP work are concerned with deployment can be an implementation of a study to alert
CEP performance rather than how machines can exchange remote emergency services when elderly people are in
meaningful data. need, such as when a fall is detected. The method used to
recognize a ‘‘fall’’ requires knowledge about how an
3.5 Machine-to-machine knowledge exchange anomalous event can be detected by device sensors. In this
study, knowledge was exchanged between the gateway and
The Semantic Web was proposed to explain how machines a remote database to describe what normal/anomalous
can semantically communicate with other machines [15]. daily activity features are. The database is equipped with
One fundamental component in Semantic Web is the an ontology that ‘‘a Device is carried by a User’’ and that
ontologies concept. As previously described, an ontology is ‘‘a Device has Events’’. Upon receiving a meaningful input
a taxonomy that defines a set of objects and their rela- event, the gateway (rather than the IoT devices) can dis-
tionships. It can therefore represent a set of inference rules tinguish anomalous from normal daily events.
in AI. The above study demonstrates that the problem with
Figure 12a shows an ontology of a Person in terms of a adopting Semantic Web in the IoT is that HTTP is not
graph (adapted from [87]). The nodes represent objects, suitable for constrained devices (as was previously shown
and the edges represent relationships. A member of a in Fig. 10). HTTP requires higher computing and network
Person object is Me, which has a name, email, and a title resources compared to other IoT application-layer proto-
attribute. These relationships can derive an inference rule cols such as CoAP. The header and message size are lim-
‘‘if an object has an email, it may be a person.’’ ited to a maximum of 2 bytes and 256 MB, respectively, in
CoAP, compared to the unspecified length in HTTP [76].
Fig. 12 Ontologies, adapted from [87]. a The ontology of a Person. b Employing an application-layer protocol to describe an ontology
123
These fundamental differences affect the way they can be described by a knowledgebase [40], where a task consists
adopted in an IoT environment. That is, adopting Semantic of a range of events, and relationships express inference
Web to exchange knowledge degrades the system perfor- rules. This is illustrated in Fig. 13. The decision from
mance when CoAP is employed. Therefore, the current processing these events and rules represents the services
challenge in machine-to-machine knowledge sharing is that a smart object can advertise to other objects. Newly
how constrained IoT devices can exchange semantic mes- discovered services can be updated in the knowledgebase,
sages and can globally find the ontology that the messages extending the ontology, and in return creating new
belong to. The literature has compared how machines knowledge.
exchange IoT data but not knowledge. The current work in Despite smart objects having shown the ability to
intelligent IoT applications requires a review not only in exchange and share knowledge, the use of the constructed
communication protocols and their performance, but also ontology is not widely deployed. The current work is
in how physical/digital modules can smartly communicate. focused on the development and evaluation of ontologies
rather than the usage of the constructed ontologies [11].
3.6 Smart objects Although ontologies allow objects to define and share
properties and relationships of a domain, the ability to
Physical or software objects are regarded as smart when semantically identify services disintegrates when the
they autonomously exchange knowledge to solve prob- ontology is used in other domains. In the various BIM
lems. Such smart objects are aware of and respond to the ontologies for example, different situations cause dividing
changes in their environment and recognize some oppor- interpretations into whether certain data are mandatory or
tunities to proactively make a decision. They must show optional [31]. In addition, rule inference tasks have shown
sociability with other smart objects or people to accomplish to be difficult to adapt given the varying BIM ontologies. A
a goal, and support other applications [104]. survey indicated that two-third of the ontologies reviewed
Most smart object implementations make use of were not reused, while there was no indication whether the
ontologies, or some approach that involves a knowledge- remaining one-third reused ontologies could be applied to a
base system consisting of objects, their properties and different domain [28].
relationships [39]. For example, in a task analysis domain,
such a knowledgebase can consist of a hierarchy of task 3.7 Summary
decomposition starting from the goal, a breakdown of the
state changes, to the low-level task definitions [20]. Message exchange between IoT devices is constrained by
Many implementations have local rather than shared their computing power and network bandwidth. On the
knowledgebases. An example of a rich knowledgebase other hand, smart applications require big data and massive
digital system is a Building Information Model (BIM), computing resources for machine learning-based analysis.
which is a digital platform originally employed for Therefore, current smart applications adapt to such limi-
managing building constructions for planning and cost tations by proposing machine learning analysis at the
effectiveness. The model breaks down the information of a gateway or in the cloud. These applications generate pre-
building, from room layouts to construction materials scriptive and predictive analysis, but implementing adap-
required, forming a hierarchy of properties and relation- tive applications has shown to be challenging in a
ships. With the introduction of IoT, BIM applications constrained computing environment. Adaptive applications
advanced to build green buildings by predicting and sim- require feedback from either human or other smart
ulating energy models [18] and promote safety by detecting machines, which require a duplex communication channel.
hazardous gases on the location where the building workers In addition, feedbacks are processed by algorithms that are
were situated [23]. Despite having state-of-the-art tools for evolutionary, which require more resources than what can
designing knowledgebases, BIM applications do not have a be achieved by constrained devices. Hence, IoT devices are
universally defined ontology to eliminate data interoper- limited in their ability to autonomously learn, build, and
ability issues [12]. share knowledge.
Challenges in implementing smart objects include how To address its limitations, IoT applications employ
to discover and locate shared knowledgebases. Hence, ontologies, i.e., databases describing a network of objects
design frameworks proposed in the literature present a through their properties and relationships. This allows IoT
discovery module to find how a system relates to other solutions to locate inference rules or knowledge that is built
smart object systems and their ontologies [39]. Such a in other machines or the cloud, creating intelligent systems
module is analogous to a gateway. In discovering how for an application. Despite ontologies being widely applied
other systems accomplish a goal, the module allows smart to build smart applications, adopting one ontology to a
objects to hierarchically break down how a task is different domain has shown to be difficult. The ability of an
123
Fig. 13 Smart Objects share

their services, enabling them to
learn the breakdown of rules to
a set of tasks and events
IoT application to semantically identify the meaning of humidity [27], resulting in a predicted energy consumption
messages weakens when applying an ontology from a with 1.7 kWh difference between the actual and predicted
different domain. consumption in a 30-day window.
In the production-farming domain, in [21], the authors
proposed to detect lameness symptoms in dairy cows.
4 A framework for machine learning Diseased cows affect milk productions, and their symptoms
and knowledge discovery for IoT are usually displayed through inactive behaviors such as
lying down for a duration of time. Thus, IoT sensors were
Despite the pervasiveness of IoT devices, the inadequacy placed on cows to collect their activity data such as lying
of machines to adaptively learn becomes the barrier in the time, swaps per hour, and step count. Machine learning
proliferation of intelligent IoT systems. Little has been techniques revealed three kinds of behaviors such as active,
studied in how machines can autonomously share knowl- normal, and dormant in cows. Random forest gave an
edge when the data come from different domains or case accuracy of 91 percent, detected 1 day before some visual
studies. Thus, this section proposes a framework on how signs can be observed; and k-nearest neighbors gave an
machines can autonomously exchange knowledge, create accuracy of 81 percent, detected 3 days before their
new knowledge, and adaptively learn from the knowledge symptoms were noticeable through the visual signs.
so that they can become applicable across different In the production-manufacturing domain, the mainte-
domains or case studies. The framework is adopted from nance of machinery is essential to minimize production
how humans learn and classify knowledge. To begin with, interruptions due to failed motors. Thus, a study [111]
this section reviews the various case studies associated placed 3-axes accelerometers on factory motors to collect
with intelligent big data analytics for IoT. vibration data such as the amplitude and frequency of the
motors. With neural networks, the study was able to
4.1 A review of intelligent big data analytics identify faults from normal motor conditions with 100
for IoT percent accuracy, with confidence level between 80 percent
and 99 percent.
Table 1 shows the various case studies that have leveraged Building management may be seen as a domain that
big data analytics to address IoT challenges by employing connects the smart city and manufacturing domain, because
machine learning techniques. The studies can be grouped their solutions may be applied to both domains. A study
into four domains: smart city, production (i.e., farming & [30] predicted the occupancy level of a building by
manufacturing), building management, and health. The observing the temperature, CO2, air volume, and air con-
following discussions describe the studies given in Table 1. ditioning data. Random forest was employed to classify the
The smart city domain is concerned with bringing effi- data and yielded 95 percent accuracy in predicting the
cient and enjoyable daily living activities. Some examples occupancy level of rooms within the building. Another
include providing optimum traffic route [79], predicting study [84] collected building sensor data such as temper-
garbage bin filling pattern for collection [62], recom- ature, humidity, gas, smoke, CO, flame, and CO2 to equip
mending a product based on one’s location [86], and pre- the building with a early detection system against gas
dicting energy consumption in smart meters [27]. In leakage and fire hazard. The study arranged four risk
predicting energy [27], support vector machines were used levels, from no risk to high risk; the regression trees
to analyze the past energy consumption and the environ- algorithm achieved 99.93 percent accuracy in predicting
ment data of a building such as the temperature and the risk levels.
123
The health domain may be the most pervasive applica- 4.3 Epistemology
tions because the IoT devices are attached to individuals.
The applications predict sugar level for diabetes manage- The philosophy of knowledge, or epistemology, provides
ment [10], estimate thermal comfort level at a workplace the breakdown of how a human gains knowledge. Funda-
[66], and detect when a person falls at home [107]. A study mental in this area is ‘‘a priori’’ and ‘‘a posteriori’’
[35] employed deep learning to detect ambulation events knowledge. A priori is gained through the definitions, such
such as abnormal walking pattern, sleeping habits, and as the classification; apples are fruits. A posteriori is gained
washroom visits. The study provides a real-time solution to through experience and observations, such as apples are
detect anomalous health risks from wearable device data red.
collecting heart rate, respiration rate, and so on, and yiel- Variations from the above definition are also used to
ded a 94 percent accuracy. Such a solution may be viewed describe ideas and human understanding. This is illustrated
as a smart home application because they also bring effi- in Table 2. Kant’s [63] a priori and a posteriori description
ciencies in daily activities (e.g., providing a cyber guardian of knowledge are parallel to Locke’s [70] primary and
for elderly people). secondary quality in human understanding. Primary quality
includes the properties of an object independent of an
4.2 Learning from human-computer interaction observer, e.g., an apple has weight, size, and color. Sec-
ondary quality refers to some object properties according to
How a human can interact with websites can teach how an observer, e.g., apples are red. Descartes [29] used an
machines should interact with other machines. The most additional term, i.e., invented, which denotes ideas gained
relevant property of a website is that its presentation can be from imagination, such as the mermaids and the unicorns.
understood by humans. Users can semantically navigate Henceforth, in this work we adopt the word Primary,
websites to accomplish the goals required of a task. Secondary, and Invented to classify knowledge.
Websites consist of words in the language that is under-
• Primary knowledge is built into ontologies, such as the
stood by the human users. In addition, website layouts are
definition of ‘‘If the soil is wet then crops flourish’’.
arranged in a meaningful structure exhibiting their ‘‘in-
• Secondary knowledge is gained from the IoT sensors,
formation scent’’ so that the users can naturally navigate to
such as ‘‘The soil is dry’’.
find the information being searched.
• Invented knowledge is created after abduction infer-
There are two characteristics of knowledge sharing that
ences, speculating what else serve as the indicators to
can be drawn from human–computer interaction. Thus, we
have the crops flourished such as the sunlight and
propose the following theorems.
fertilizers.
Theorem 1 Naming an entity (i.e. object, properties,
Secondary knowledge can become Primary knowledge
relationship and services) must use words that humans
when the quality it describes can be generalized. This
understand.
agrees with Berkeley [14] who contemplated that primary
Theorem 2 Information must be arranged semantically in knowledge comes from something that is perceived. Sim-
a way that it allows human to derive meaning by following ilarly, when Invented knowledge (such as a hypothesis) has
its structure. been scientifically proven, the parameters it describes
become Secondary knowledge. This agrees with the veri-
Naming an entity must use words in human language
ficationism theory [73], which serves as the groundwork
rather than abbreviation, codes, or binary mapping that
for the proposed framework of machine knowledge in our
only machines can interpret, despite the latter being tech-
paper explained below.
nically more efficient in terms of storage space or network
bandwidth.
4.4 A framework of machine knowledge
In arranging information semantically, the use of
ontologies is the model itself. It allows humans to retrieve
The framework is adapted from the verificationism theory
information based on semantic queries. In this work, we
[73] which provides a breakdown of scientific methods. It
propose the use of ontologies for machines to autono-
is a school of thought where knowledge is gained from
mously share knowledge and adaptively learn from other
experimentally verified observations. As the society has
machines. How this can be articulated may be derived from
developed through this scientific pathway, its framework in
how human classify knowledge.
gaining knowledge through observations can be borrowed
for the design of autonomous learning machines. The
123
Table 2 The categories of

Knowledge gained through Kant [63] Locke [70] Descartes [29]
knowledge
Definition A priori Primary Innate
Experience A posteriori Secondary Adventitious
Imagination – – Invented
framework of machine knowledge proposed in our paper is Secondary knowledge can become Primary knowledge
shown in Fig. 14. after the rules in the database have been verified through
The framework describes how Smart Objects (SOs) can some inductive learning. A family of machine learning
autonomously exchange knowledge from other SOs. It techniques such as rule learning, classification, and Baye-
consists of three databases, namely the Ontology, Param- sian inference can be employed for inductive learning.
eters, and Hypotheses database. When sending data from The Hypotheses database contains the non-verified
the databases, SOs tag the services to signify the Primary, inference rules. It takes rules either directly from other SOs
Secondary, or Invented level of knowledge, accordingly. as the Invented knowledge, or from any ontology defini-
SOs exchange knowledge by advertising and searching for tion, or from its ontology database. Invented knowledge
services with other SOs as what previously is shown in can become Secondary knowledge after some observations
Fig. 13. that the values have converged to a distribution or a pat-
The Ontology database contains inference rules. It takes tern. Statistics can be employed to observe the distribution
rules either directly from other SOs as the Primary of data.
knowledge, or from a well-established ontology definition, The Hypotheses database also learns from the Ontology
or from its Parameters database. database. Primary knowledge from the Ontology database
The Parameters database contains the name-value pairs can become Invented knowledge after speculative rules are
of data. It takes values either from the IoT devices as the created through abduction techniques. For example, given
Secondary knowledge, or from its Hypotheses database. a rule ‘‘apples are red,’’ the speculative rule seeks to
Fig. 14 Smart Objects share

their services, enabling them to
learn the breakdown of rules to
a set of tasks and events
123
conclude if the object is an ‘‘apple’’ when ‘‘red’’ is parameters database is populated with the accelerometer
observed. In this framework, abduction is triggered when values representing the forward/backward/lateral fall and
an observed value (i.e., red) fails to conclude the class (i.e., the name-value pairs representing normal events such as
apple). walking and sitting. Its hypotheses database is still empty.
In creating speculative rules, the framework implements Another study proposed to detect lameness symptoms in
two theorems previously discussed. First, it employs Nat- dairy cows in the farming domain [21] because such a
ural Language Processing (NLP) techniques to understand condition is related to milk production’s effectiveness. It
human language. The NLP can find the synonyms, homo- places IoT devices on cows to detect lying time, swaps per
nyms, and categories of a word. In ‘‘apples are red,’’ the hour, and step count to derive whether the cows were
NLP finds that ‘‘red’’ is a color, and that ‘‘apple’’ is a noun. normal, dormant, or active. Adapting this case to the
It therefore seeks to find other nouns, hence creating a rule framework of machine knowledge, the ontology database
such as ‘‘[noun] are red.’’ The use of NLP in this process would have the relationships as shown in Fig. 16. The
implements Theorem 1. parameters database would contain the name-value pairs of
Second, the framework exchanges ontologies with other the lying time, swaps per hour, and step count. The
SOs. Ontologies are structured in the way that the inference hypotheses database would be empty. The system would be
rules are understandable by humans. Although SOs are named as SO2 in this case.
machines, the structure of the ontologies must convey
meaning. The Semantic Web ontology illustrated in Fig. 12 4.5.2 Exchange primary knowledge with other SO
serves this purpose, as the properties it describes reflect the
model in the real world. Hence, it implements Theorem 2. SO1 learned to expand its knowledge from SO2. It sends a
broadcast query to indicate whether other SOs had any
4.5 Case study Primary knowledge of Event and Sensor or not. SO2
replies by describing its Primary knowledge, i.e., {Normal,
This section provides an example how the framework can Active, Dormant} element Event; and {Lying time, Swaps/
be implemented. It takes the ontology and parameters from hour, Step count} element Sensor. At this point, SO1
published articles and adapts these so that SOs can Ontology database is illustrated as shown in Fig. 13, having
autonomously build knowledge using the proposed both the shaded and non-shaded entities together.
framework. Since SO1 adopted such Primary knowledge, SO1
queried SO2 to send its Secondary knowledge. Hence, SO2
4.5.1 Initialization sends the name-value pairs from the Parameters database.
However, the values learned might not be instantly
A recent study proposed to detect an event when a person adaptable to SO1 because these values were learned from
falls, for the purpose of sending alerts for support in a smart cows. In contrast, human lying time and step count would
home domain [107]. The study classifies normal daily have different values to reflect their normal, active, or
activities from abnormal activities such as when the person dormant condition. Similarly, swaps/hour might be a rele-
falls. The ontology was described by the non-shaded enti- vant observation for cows but not human.
ties in Fig. 15. In this case study, a Smart Object hereby
would be named SO1, and its ontology database is popu-
lated accordingly from the fall detection study. Its
Fig. 15 Illustration of entity

relationship network of SO1
Ontology database. The shaded
ones are learned from SO2
123
knowledge from other SOs, and obtained the relationships

as illustrated in Fig. 17 (the shaded entity was adapted
from a study to monitor traffic in the smart city domain
[43] and the network (formed by the non-shaded entities)
was adapted from a study in the health domain [112]).
Thus, the Ontology database had the following rela-
tionships: a Device is carried by a User; a Device is carried
by a Patient; a Driver is a Patient; a Driver is a Person; a
User is synonymous to a Person. Through abduction
inference, SO1 speculated that SO1 User is a Person; and a
Person is an element of a Role. This relationship was
committed in the Hypotheses database.
Fig. 16 Illustration of entity relationship network of SO2 Ontology Hence, SO1 knowledge would be described as shown in
database
Fig. 18. The dashed entities were added through the above
procedure; and the dashed arrows represent the abduction
inferences in the Hypotheses database. There would be
4.5.3 From secondary knowledge becoming primary more inferences drawn which would create a meshed net-
knowledge work, but Figure illustrates only what is relevant for the
following discussion.
The framework guides how to accept or reject the knowl-
edge learned. When Secondary knowledge is generalized 4.5.5 From invented knowledge becoming secondary
through induction, it becomes Primary knowledge. In this knowledge
case, if it could be verified that SO1 lying time values
reflects an SO1 Event (normal, active, dormant, fall), then In scientific methods, knowledge is gained through obser-
lying time becomes accepted as an element of the SO1 vations. When statistics show an acceptable confidence
Sensor. Suppose that after some time, SO1 has collected interval from the samples then the hypothesis can be
enough data and updated its lying time values from accepted. Therefore, to trigger the statistical analysis, SO1
humans, it will find a range of values that were used to must have enough data samples. To have enough samples,
classify some events. Therefore, these lying time values SO1 should be adopted by the various IoT applications to
will be accepted as the Secondary knowledge and lying collect relevant values.
time as an element of Sensor is reflected in the Ontology Suppose that SO1 was adopted by a study [66] to predict
database. In contrast, swaps per hour are rejected from the personal thermal comfort based on heart rate and body
Ontology database because its values failed to classify temperatures obtained from IoT sensors. The human pro-
Event. grammer adopted the SO1 ontology described in Fig. 18
and created a relationship that ‘‘a Device is attached to a
4.5.4 From primary knowledge becoming invented Pedestrian’’ and ‘‘Comfortable is an Event.’’ Upon
knowledge deployment, data samples were collected from the person’s
heart rate and body temperature. This showed that the
Rejecting an entity would trigger the framework to perform following path had been activated (only one path shown as
abduction inferences. It exhaustively tests all rules in the an example): Heart rate is a unit of a Sensor; a Sensor is a
Ontology database. One path is given here as an example: unit of a Device; a Device is carried by a Person; and a
swaps per hour was measured by a sensor; a sensor was a Pedestrian is a Person. From this application, there was a
unit of a device; a device was carried by a user. Through sample indicating that ‘‘a Pedestrian is a Person.’’
abductive reasoning, SO1 speculates that its ‘‘user’’ was a Interestingly, heart rate, location, and acceleration val-
‘‘person’’ (rather than a cow). ues were the components used by another study [86] to
To have the above reasoning outcome, SO1 went send tailored advertising messages. The human program-
through two processes. First, it employed NLP to find the mer then adopted the solution from the study and added a
synonyms of the word ‘‘user,’’ which resulted in many new relationship that ‘‘Location is a unit of a Sensor.’’ Hence,
words such as ‘‘customer,’’ ‘‘client,’’ ‘‘patron,’’ ‘‘prospect,’’ this application activated two different paths to name its
and ‘‘patient.’’ user:
Second, SO1 exchanged Primary knowledge with other
– Heart rate ! Sensor ! Device ! Person (learned from
SOs to get the ontology structure from the new words. It
the thermal comfort study [66]).
performed the steps as what previously described to expand
123
Fig. 17 The ontology learned to

know what relationships a
‘‘patient’’ has
Fig. 18 From the attributes of

the Sensor, SO1 knows that a
User is like a Person
– Accelerometer ! Sensor ! Device ! User (learned knowledge, can be implemented in the cloud (as in Model
from the smart home study [107]). 2) or use a high-performance computing solution at the
edge (as in Model 3).
Once SO1 had collected data samples whose values
explained the other relevant relationships (e.g., fall, nor-
4.6.1 Latency
mal, comfort), then the relationships that were activated
were asserted. In this example, the fact that the application
By adopting Model 2 and 3 as in Fig. 9, the framework can
indifferently named ‘‘Person’’ and ‘‘User’’ to accomplish a
use ontologies in integrating IoT devices with intelligent
task asserts what earlier was speculative that its User was a
systems. Selecting either model would involve assessing
Person.
how critical the application is. For example, in considering
the proper model for a remote, mission-critical application,
4.6 Discussion
one can adopt Model 2 (i.e., having the resources in the
cloud) provided that the network latency and network
This section demonstrates how the proposed intelligent
availability requirements have been optimized for the
framework can be applied to various IoT domains (i.e.,
remote application. The advantage of this model is that the
smart home, farming, smart city, and health) to autono-
SO in the cloud can learn from remote SOs in the back-
mously exchange and build knowledge. However, the
ground, regardless of the network latency connecting to the
proposed framework does not address how IoT devices can
remote application. The disadvantage is that the network
adaptively learn in constrained environments where com-
latency can become the bottleneck when the need for real-
puting and network resources are very limited (as described
time decisions increases. As another consideration, one can
by Model 1 in Fig. 9). How better computing and network
adopt Model 3 (i.e., having all resources close to the
performance can be achieved is not the focus of our study;
consumer). In this model, the SO can deliver real-time
rather, we showed how adaptive learning can be achieved
knowledge to the consumer, but it relies on the capacity
within the current IoT models. Hence, the framework
and availability of the communications network to initiate
leverages the fact that learning, i.e., transforming data to
knowledge exchange with remote SOs.
123
4.6.2 Natural language processing for machine-to-machine operators have better monitoring and control of their sys-
communications tems and processes, and business intelligence people have
better insights in understanding their challenges and mak-
The framework addresses the current gap in reusing ing informed decisions. Although the convergence of
ontologies in various domains by mandating the use of machine intelligence and IoT has opened up many oppor-
human language in naming the entities and relationships. tunities, there are a few challenges which have constrained
The branch of AI, Natural Language Processing (NLP), is their growth. These future research opportunities will
proposed to intelligently adapt different meaning repre- enable the seamless integration of IoT with data analytics.
sented by the ontologies. Hence, in addition to the above- The following section addresses some of the possible future
mentioned, the contribution of this section is to expand the directions.
use of NLP in machine-to-machine communications.
Perhaps the closest parallel with our work is the one 5.1 Improved cybersecurity
described in [94], wherein machine databases gain
knowledge from human through NLP parsing, converting For improved decision making, machine learning algo-
semantic representations into logical syntax. The work rithms heavily rely on IoT data generated and transmitted
described the consumers as both human and machines that from the IoT devices. Within an IoT framework, different
build knowledge to and retrieve information from the IoT layers, e.g., perception layer, transportation layer, and
knowledgebase. Similarly, the work in [97] proposed NLP- application layer, are vulnerable to cyber-attacks. For
parsed messages to build a knowledgebase, allowing example, malicious code injection, node tampering,
machines to build knowledge from human language. Our impersonation, Denial of Service Attacks, Routing Attacks,
work differs from these works in that it proposed three and Data Transit Attack (man-in-the-middle attack, sniff-
databases in the system, i.e., the Ontology, Parameters, and ing) are some of the examples of cyber-attacks in an IoT
Hypotheses database that allow machine-to-machine com- system model [88]. To secure the IoT system against these
munications. Furthermore, our work demonstrated a novel cyber threats, it is important to ensure a proper trust
case where machines can autonomously expand their management framework. IoT devices themselves require
knowledge through the scientific, abduction method. proper attention because most of the devices lack proper
security mechanisms. As highlighted in [88], the chal-
4.6.3 Drawbacks lenges of ensuring IoT security are different compared with
the traditional IT security, as discussed below.
Although the framework addresses the current challenge of First, users can add security solutions in a traditional IT
reusing ontologies, it adopts a weak definition of intelligent scenario any time; however, most of the IoT devices lack
systems. The Smart Objects given in the case study (i.e., security solutions and others have built-in security solu-
SO1, SO2) are aware of and respond to the changes in their tions, and most of the devices do not support additional
environment, proactively make a decision and show security patches, solutions or updates at a later point of
sociability with other smart objects as well as help others. time, post-device production.
In addition to these, strong AI would involve displaying Second, because of the low memory and processing
emotion and desire, forming a personal character when constraints, only lightweight algorithms are used. As
exposed to a range of situations. However, there is little pointed out in [88], it is always challenging to strike the
discussion on whether IoT systems should display a strong proper balance between lower capability and higher
AI character, or whether it may be counterproductive. security.
Hence, the framework is sound in proposing new insights Third, in an IoT environment, a wide variety of devices
in intelligent IoT systems. are used. Due to this heterogeneous nature, the security risk
is increased due to the integration of devices from different
types, technologies, and vendors.
5 Future research opportunities Fourth, the application layer of IoT suffers from privacy
issues. Data leak and eavesdropping could be potential
The convergence of IoT and large-scale data analytics has consequences. Fifth, IoT communication protocols are also
created enormous opportunities. Machine intelligence vulnerable to cyber-attacks and threats including data
based on IoT data has merged the cyber and physical transit attacks, routing and DoS Attacks, issues in key
worlds together and has improved significantly for inves- management, high computational cost of TLS, and lack of
tigating the real-world challenges from a cyber-physical user control for privacy.
perspective. The efficiency and reliability of the processes Our proposed framework for machine learning and
and systems have improved significantly. Now, the system knowledge discovery for IoT is the sixth dimension to
123
address cybersecurity challenges. Intelligent machines shopping carts; urban cameras can be deployed to detect a
communicating with human language may wrongly inter- missing child; self-driving vehicles and air planes generate
pret messages such as, for example, when interpreting a large amount of data and require fast decision. Interest-
homonyms (i.e., the same words that have different ingly, the authors of [90] pointed out that, in such situation,
meanings such as a ‘‘file’’ which can mean either data or a collaborative learning between the edge devices will
queue). The current active research in this area is lexical- increase because ad hoc machine-to-machine communica-
semantic analysis [52] to tell that ‘‘april in paris lyrics’’ tions do not depend on the cloud data sharing policies.
refers to the name of a song, while ‘‘vacation april in paris’’ The challenge against this trend is that edge devices
refers to holiday information of a city. However, in the (e.g., mobile devices) tend to have lower computing
context of collaborating IoT devices sharing knowledge, a capabilities than a cloud data center. On the other hand,
malicious node or smart object can intentionally spread machine learning data analytics require high processing
fake news. This is possible because homonyms and the power and storage. In this context, our proposed framework
varying lexical-semantics can manipulate the content of a for machine learning in the IoT environment can address
message to deceive others [99]. Thus, cybersecurity con- the challenge. Its future work may comprise the study of a
cerns expand to include machine-to-machine deception federation of edge devices sharing their knowledge. An
techniques. edge device that is equipped with the highest computing
Therefore, future IoT devices, communication protocols, performance can perform the data processing task. An edge
technologies, and the entire platform require improved device that has the highest bandwidth can communicate
security solutions by design. Device manufacturers must with the cloud to offload their computing tasks [1]. In [53],
provide robust in-built security solutions within the devi- a community of IoT devices with enough resources were
ces. Cryptographic algorithms need to be improved, and tasked with making decision on whether an access per-
key management must be made effective. More research is mission can be granted, on behalf of a low-resource device.
needed on authentication schemes and on Public Key In other words, we can further examine a hybrid model
Infrastructure (PKI), focusing on the various IoT tech- comprising Models 2 and 3 (as Fig. 9 illustrates) where a
nologies and applications. Lightweight but robust intrusion group of edge devices communicate to vote which device is
detection techniques need to be developed at the device responsible to transform data into knowledge.
level, edge or at the cloud of the infrastructure.
5.3 Scalability
5.2 Machine learning at the edge
IoT technologies have enabled the platform to communi-
IoT applications will require faster processing and decision cate amongst a large number of connected peers. When
making, which trend would move data processing closer to critical applications and end users require a big number of
the consumer [45]. Sending data to the cloud requires time interconnections, scalability becomes an issue which needs
and high bandwidth. Therefore, analytics at the edge node to be addressed. For example, large amounts of data need
of the IoT has opened the door for future opportunities. to be distributed to several end devices, where the devices
One drive behind for this trend is that the use of sensors simultaneously solve computing challenges. Hence, dis-
proliferates in many areas of life and businesses, such as tributed machine learning algorithms with edge computing
vehicles, manufacturing and health, creating continuous are a potential solution. It enables computing decisions to
streams of data [45, 102]. Such large amount of data take place at the edge, which is closer to the IoT devices.
becomes the raw material for businesses and governments One work in this context is context-aware processing [2],
to gain insights and new knowledge by employing machine where scalability was achieved by exchanging knowledge
learning techniques. The motivations include competition rather than some changes in threshold values (as discussed
or creating better policies set out when data are abundant, in Sect. 3.4, traditional IoT devices exchange information
engendering the urgency to draw knowledge from the data as a change of state or time). In context-aware processing,
[102]. information is exchanged in terms of context, allowing the
Another drive is the convergence between the IoT and term ‘‘morning’’ to be flexibly defined as ‘‘8 am to 12 pm,’’
critical infrastructure due to their demand for mission- or when the road traffic is seen as ‘‘130 vehicles/hour.’’
critical applications [110]. As discussed in Sect. 4.6.1, data This method enables a scalable solution for a network of
are processed closer to the consumer when low latency is collaborating devices because only the required informa-
critical. The current trend is that trivial applications tend to tion according to the context is exchanged, rather than
become critical, or can be repurposed to support critical communicating a number of possible values.
applications. The authors of [90] pointed out that e-com- In regard to the framework for machine learning pro-
merce users need faster time in listing the content of their posed in our work, we note that advancements in NLP
123
research will foster the performance of machine learning way for a future that comprises rapid IoT proliferation,
tasks in a distributed system. NLP puts context behind data application deployment, and strong analytics of high vol-
values, allowing connected IoT devices to collaboratively ume IoT data, we have argued that integrating intelligent
assign labels to their data, learn knowledge from other solutions from different domains have been proven to be
devices, thereby representing distributed machine learning difficult. In this paper, we have justified that our proposed
to solve computing challenges. framework for machine learning and knowledge discovery
for IoT pave ways to integrate adaptive learning techniques
5.4 Hyperconvergence locally, at the edge, through a fog or in the cloud. Conse-
quently, the power of machine learning can be fully har-
Within a hyperconvergence setting, storage facilities are nessed to provide value and benefits to the consumers of
shared amongst a large number of distributed nodes and IoT technology, in the larger context of things.
their combined performance helps to recover from the Our proposed framework opens up some new challenges
resource sharing problem. This paradigm has been shown for future work in machine-to-machine communications. In
as the desired architecture for an IoT data analytics cybersecurity, research can include the study of influence
framework [100]. Our framework for machine learning in of malicious machines on other devices that may lead to
the IoT environment opens the door to this pathway. The system compromise. In distributed systems, machines need
databases shown in Fig. 14 (i.e., the Ontology, Parameters, to be configured to seamlessly distribute resources (hy-
and Hypotheses database) are designed to share their data perconvergence) and allow the scalability of connected IoT
with other IoT devices in their community network, and devices. When a federation of edge devices converge, the
their combined knowledge helps to address cross-domain concurrent advances in machine learning technology are
convergence as discussed in Sect. 4.5. However, as pointed thus realizable in resource constrained IoT systems, and the
in [100], the challenge is that current network systems are future of IoT-enabled human lives is thus realized.
designed for specific applications. Nodes in a community
network are connected with various physical layer tech- Acknowledgements We thank the anonymous reviewers for their
valuable comments which helped us improve the quality, organization
nologies and bandwidths, with some nodes being admin- and presentation of this paper.
istratively more powerful (e.g., routing information, access
control) than others. How a new node can be seamlessly
added to an existing community network in the presence of References
some faults has not been fully addressed. Furthermore,
there is a need to examine how a large (and increasing) 1. Aazam M, Zeadally S, Harras KA (2018) Offloading in fog
amount of data can be shared across large-storage nodes computing for iot: review, enabling technologies, and research
opportunities. Future Generat Comput Syst 87:278–289
such as data centers. Thus, there is a further need to design 2. Akbar A, Carrez F, Moessner K, Sancho J, Rico J (2015)
intelligent solutions for the hyperconvergence setting. Context-aware stream processing for distributed IoT applica-
Additionally, the convergence of intelligent IoT devices tions. In: 2015 IEEE 2nd World Forum on Internet of Things
should not only solve the technical challenges but also (WF-IoT), IEEE, pp 663–668
3. Akbar A, Khan A, Carrez F, Moessner K (2017) Predictive
address regulatory issues. It is essential to justify how analytics for complex iot data streams. IEEE Internet of Things J
intelligent devices can be beneficial in serving humanity 4(5):1571–1582. https://doi.org/10.1109/JIOT.2017.2712672
rather than only satisfying breakthrough enthusiasts. An 4. Al-Ali AR, Zualkernan IA, Rashid M, Gupta R, Alikarar M
example of a current issue is that distributed system solu- (2017) A smart home energy management system using IoT and
big data analytics approach. IEEE Trans Consum Electron
tions consume a large amount of power, which starts to 63(4):426–434. https://doi.org/10.1109/TCE.2017.015014
compete with human basic needs [110]. Thus, future 5. Al-Fuqaha A, Guizani M, Mohammadi M, Aledhari M, Ayyash
directions include more studies on regulating the conver- M (2015) Internet of things: a survey on enabling technologies,
gence of intelligent machines. protocols, and applications. IEEE Commun Surv Tutor
17(4):2347–2376
6. Alahakoon D, Yu X (2016) Smart electricity meter data intel-
ligence for future energy systems: a survey. IEEE Trans Ind
6 Conclusion Inform 12(1):425–436. https://doi.org/10.1109/TII.2015.
2414355
7. Anwar A, Mahmood AN, Tari Z (2015) Identification of vul-
The IoT paradigm has become an integral part of our daily nerable node clusters against false data injection attack in an ami
lives. However, IoT devices are constrained in computation based smart grid. Inf Syst 53:201–212
and communication resources, which are the bottlenecks in 8. Anwar A, Mahmood AN, Pickering M (2017) Modeling and
the development of adaptive, intelligent solutions performance evaluation of stealthy false data injection attacks
on smart grid in the presence of corrupted measurements.
employing machine learning techniques. Although advan-
ces in technologies and platform enhancements pave the
123
J Comput Syst Sci 83(1):58–72. https://doi.org/10.1016/j.jcss. 27. Deligiannis P, Koutroubinas S, Koronias G (2019) Predicting
2016.04.005 energy consumption through machine learning using a smart-
9. Anwar A, Mahmood AN, Tari Z (2017) Ensuring data integrity metering architecture. IEEE Potentials 38(2):29–34
of opf module and energy database by detecting changes in 28. Dermeval D, Vilela J, Bittencourt II, Castro J, Isotani S, Brito P,
power flow patterns in smart grids. IEEE Trans Ind Inform Silva A (2016) Applications of ontologies in requirements
13(6):3299–3311 engineering: a systematic review of the literature. Requir Eng
10. Ara A, Ara A (2017) Case study: integrating iot, streaming 21(4):405–437
analytics and machine learning to improve intelligent diabetes 29. Descartes R (2013) René Descartes: meditations on first phi-
management system. In: 2017 International conference on losophy: with selections from the objections and replies. Cam-
energy. Communication, data analytics and soft computing bridge University Press, Cambridge
(ICECDS), IEEE, pp 3179–3182 30. Dey A, Ling X, Syed A, Zheng Y, Landowski B, Anderson D,
11. Ashraf J, Hussain OK, Hussain FK, Chang EJ (2018) Ontology Stuart K, Tolentino ME (2016) Namatad: Inferring occupancy
usage analysis framework (ousaf). Measuring and analysing the from building sensors using machine learning. In: 2016 IEEE
use of ontologies. Springer, Berlin, pp 49–62 3rd World Forum on Internet of Things (WF-IoT), IEEE,
12. Azhar S (2011) Building information modeling (bim): trends, pp 478–483
benefits, risks, and challenges for the AEC industry. Leadersh 31. Djuedja JFT, Karray MH, Foguem BK, Magniont C, Abanda FH
Manag Eng 11(3):241–252 (2019) Interoperability challenges in building information
13. Bajer M (2017) Building an iot data hub with elasticsearch, modelling (bim). In: Enterprise Interoperability VIII, Springer,
logstash and kibana. In: 2017 5th international conference on pp 275–282
future internet of things and cloud workshops (FiCloudW), 32. Ed-daoudy A, Maalmi K (2019) A new internet of things
IEEE, pp 63–68 architecture for real-time prediction of various diseases using
14. Berkeley G (1881) A treatise concerning the principles of machine learning on big data environment. J Big Data 6(1):104
human knowledge. JB Lippincott & Company, Philadelphia 33. Elijah O, Rahman TA, Orikumhi I, Leow CY, Hindia MN
15. Berners-Lee T, Hendler J, Lassila O et al (2001) The semantic (2018) An overview of internet of things (iot) and data analytics
web. Sci Am 284(5):28–37 in agriculture: benefits and challenges. IEEE Internet of Things J
16. Bigquery: Cloud data warehouse. Google Cloud (2020). https:// 5(5):3758–3773. https://doi.org/10.1109/JIOT.2018.2844296
cloud.google.com/bigquery/. Accessed 18 Oct 2019 34. Endler M, Briot JP, e Silva FS, de Almeida VP, Haeusler EH
17. Bonomi F, Milito R, Zhu J, Addepalli S (2012) Fog computing (2017) Towards stream-based reasoning and machine learning
and its role in the internet of things. In: Proceedings of the First for iot applications. In: 2017 Intelligent Systems Conference
Edition of the MCC Workshop on Mobile Cloud Computing, (IntelliSys), IEEE, pp 202–209
ACM, New York, NY, USA, MCC ’12, pp 13–16 35. Fadlullah ZM, Pathan ASK, Gacanin H (2018) On delay-sen-
18. Bottaccioli L, Aliberti A, Ugliotti F, Patti E, Osello A, Macii E, sitive healthcare data analytics at the network edge based on
Acquaviva A (2017) Building energy modelling and monitoring deep learning. In: 2018 14th international wireless communi-
by integration of IoT devices and building information models. cations & mobile computing conference (IWCMC), IEEE,
In: 2017 IEEE 41st annual computer software and applications pp 388–393
conference (COMPSAC), IEEE, vol 1, pp 914–922 36. Farooq MS, Riaz S, Abid A, Abid K, Naeem MA (2019) A
19. Bui KN, Jung JJ (2019) Aco-based dynamic decision making for survey on the role of iot in agriculture for the implementation of
connected vehicles in iot system. IEEE Trans Ind Inform smart farming. IEEE Access 7:156237–156271
15(10):5648–5655. https://doi.org/10.1109/TII.2019.2906886 37. Firouzi F, Farahani B, Ibrahim M, Chakrabarty K (2018) Key-
20. Burmeister D, Schrader A (2018) Runtime generation and note paper: from eda to iot ehealth: promises, challenges, and
delivery of guidance for smart object ensembles. In: Interna- solutions. IEEE Trans Comput Aided Des Integr Circuits Syst
tional conference on applied human factors and ergonomics. 37(12):2965–2978. https://doi.org/10.1109/TCAD.2018.
Springer, New York, pp 287–296 2801227
21. Byabazaire J, Olariu C, Taneja M, Davy A (2019) Lameness 38. Flouris I, Giatrakos N, Deligiannakis A, Garofalakis M, Kamp
detection as a service: application of machine learning to an M, Mock M (2017) Issues in complex event processing: status
internet of cattle. In: 2019 16th IEEE annual consumer com- and prospects in the big data era. J Syst Softw 127:217–236
munications & networking conference (CCNC), IEEE, pp 1–6 39. Fortino G, Guerrieri A, Russo W (2012) Agent-oriented smart
22. Cao H, Wachowicz M, Renso C, Carlini E (2019) Analytics objects development. In: Proceedings of the 2012 IEEE 16th
everywhere: generating insights from the internet of things. international conference on computer supported cooperative
IEEE Access 7:71749–71769. https://doi.org/10.1109/ACCESS. work in design (CSCWD), IEEE, pp 907–912
2019.2919514 40. Fortino G, Guerrieri A, Russo W, Savaglio C (2015) Towards a
23. Cheung WF, Lin TH, Lin YC (2018) A real-time construction development methodology for smart object-oriented IoT sys-
safety monitoring system for hazardous gas integrating wireless tems: a metamodel approach. In: 2015 IEEE international con-
sensor network and building information modeling technologies. ference on systems, man, and cybernetics, IEEE, pp 1297–1302
Sensors 18(2):436 41. Fog computing and the internet of things: extend the cloud to
24. Coremetrics is part of ibm. IBM (2020). https://www.ibm.com/ where the things are (2015). https://www.cisco.com/c/dam/en_
au-en/digital-marketing/coremetrics-software. Accessed 18 Oct us/solutions/trends/iot/docs/computing-overview.pdf. Accessed
2019 18 Oct 2019
25. Contreras-Castillo J, Zeadally S, Guerrero-Ibañez JA (2018) 42. Garcı́a-Magariño I, Lacuesta R, Lloret J (2017) Agent-based
Internet of vehicles: architecture, protocols, and security. IEEE simulation of smart beds with internet-of-things for exploring
Internet of Things J 5(5):3701–3709 big data analytics. IEEE Access 6:366–379
26. Dash S, Shakyawar SK, Sharma M et al (2019) Big data in 43. Gonzalez-Mendoza M, Velasco-Bermeo N, Orozco OJL (2017)
healthcare: management, analysis and future prospects. J Big The traffic status and pollutant status ontologies for the smart
Data 6:54 city domain. In: Mexican international conference on artificial
intelligence, Springer, pp 95–101
123
44. Granados J, Chu H, Zou Z, Zheng LR (2019) Towards work- 63. Kant I (1781) Critique of pure reason. Modern Classical
load-balanced, live deep learning analytics for confidentiality- Philosophers. Houghton Mifflin, Cambridge, pp 370–456
aware IoT medical platforms. In: 2019 IEEE international 64. Kong XTR, Xu SX, Cheng M, Huang GQ (2018) Iot-enabled
conference on artificial intelligence circuits and systems parking space sharing and allocation mechanisms. IEEE Trans
(AICAS), IEEE, pp 62–66 Autom Sci Eng 15(4):1654–1664. https://doi.org/10.1109/
45. Griffiths F, Ooi M (2018) The fourth industrial revolution-in- TASE.2017.2785241
dustry 4.0 and IoT [trends in future i&m]. IEEE Instrum Meas 65. Krylovskiy A (2015) Internet of things gateways meet linux
Mag 21(6):29–43 containers: performance evaluation and discussion. In: 2015
46. Gunduz MZ, Das R (2020) Cyber-security on smart grid: threats IEEE 2nd world forum on internet of things (WF-IoT), IEEE,
and potential solutions. Comput Netw 169:107094 pp 222–227
47. Hassanalieragh M, Page A, Soyata T, Sharma G, Aktas M, 66. Laftchiev E, Nikovski D (2016) An iot system to estimate per-
Mateos G, Kantarci B, Andreescu S (2015) Health monitoring sonal thermal comfort. In: 2016 IEEE 3rd world forum on
and management using internet-of-things (IoT) sensing with internet of things (WF-IoT), IEEE, pp 672–677
cloud-based processing: opportunities and challenges. In: 2015 67. Lavassani M, Forsström S, Jennehag U, Zhang T (2018) Com-
IEEE international conference on services computing, IEEE, bining fog computing with sensor mote machine learning for
pp 285–292 industrial iot. Sensors 18(5):1532
48. He W, Yan G, Xu LD (2014) Developing vehicular data cloud 68. Li H, Ota K, Dong M (2018) Learning iot in edge: deep learning
services in the iot environment. IEEE Trans Ind Inform for the internet of things with edge computing. IEEE Netw
10(2):1587–1595. https://doi.org/10.1109/TII.2014.2299233 32(1):96–101
49. He Y, Mendis GJ, Wei J (2017) Real-time detection of false data 69. Li W, Logenthiran T, Phan V, Woo WL (2019) A novel smart
injection attacks in smart grid: a deep learning-based intelligent energy theft system (sets) for iot-based smart home. IEEE
mechanism. IEEE Trans Smart Grid 8(5):2505–2516 Internet of Things J 6(3):5531–5539. https://doi.org/10.1109/
50. Hossain E, Khan I, Un-Noor F, Sikander SS, Sunny MSH (2019) JIOT.2019.2903281
Application of big data and machine learning in smart grid, and 70. Locke J, Yolton JW (1993) An Essay Concerning Human
associated security concerns: A review. IEEE Access Understanding. Dent, London
7:13960–13988. https://doi.org/10.1109/ACCESS.2019. 71. Luckham DC (2011) Event processing for business: organizing
2894819 the real-time enterprise. Wiley, New York
51. Hossain MS, Muhammad G (2018) Emotion-aware connected 72. Luo X, Zhang H, Zhang Z, Yu Y, Li K (2019) A new framework
healthcare big data towards 5G. IEEE Internet of Things J of intelligent public transportation system based on the internet
5(4):2399–2406. https://doi.org/10.1109/JIOT.2017.2772959 of things. IEEE Access 7:55290–55304. https://doi.org/10.1109/
52. Hua W, Wang Z, Wang H, Zheng K, Zhou X (2015) Short text ACCESS.2019.2913288
understanding through lexical-semantic analysis. In: 2015 IEEE 73. Matthews D (2006) Epistemic Humility. In: van Gigch JP (ed)
31st international conference on data engineering, IEEE, Wisdom, knowledge, and management, vol 2. C. West
pp 495–506 Churchman and related works series. Springer, New York.
53. Hussein D, Bertin E, Frey V (2017) A community-driven access https://doi.org/10.1007/978-0-387-36506-0_7
control approach in distributed IoT environments. IEEE Com- 74. Mehdiyev N, Krumeich J, Enke D, Werth D, Loos P (2015)
mun Mag 55(3):146–153 Determination of rule patterns in complex event processing
54. Internet of things forecast mobility report (2019). https://www. using machine learning techniques. Proc Comput Sci
ericsson.com/en/mobility-report/internet-of-things-forecast. 61:395–401
Accessed 18 Oct 2019 75. Merrill W (2010) Where is the return on investment in wireless
55. Islam SN (2019) A new pricing scheme for intra-microgrid and sensor networks? IEEE Wirel Commun 17(1):4–6. https://doi.
inter-microgrid local energy trading. Electronics 8(8):898 org/10.1109/MWC.2010.5416341
56. Islam SN, Baig Z, Zeadally S (2019) Physical layer security for 76. Naik N (2017) Choice of effective messaging protocols for IoT
the smart grid: vulnerabilities, threats, and countermeasures. systems: MQTT, CoAP, AMQP and HTTP. In: 2017 IEEE
IEEE Trans Ind Inform 15(12):6522–6530 international systems engineering symposium (ISSE), IEEE,
57. Ivanov S, Bhargava K, Donnelly W (2015) Precision farming: pp 1–7
sensor analytics. IEEE Intell Syst 30(4):76–80. https://doi.org/ 77. Neto AJV, Zhao Z, Rodrigues JJPC, Camboim HB, Braun T
10.1109/MIS.2015.67 (2018) Fog-based crime-assistance in smart iot transportation
58. Jan SU, Ahmed S, Shakhov V, Koo I (2019) Toward a light- system. IEEE Access 6:11101–11111. https://doi.org/10.1109/
weight intrusion detection system for the internet of things. ACCESS.2018.2803439
IEEE Access 7:42450–42471. https://doi.org/10.1109/ACCESS. 78. Nóbrega L, Gonçalves P, Pedreiras P, Pereira J (2019) An iot-
2019.2907965 based solution for intelligent farming. Sensors 19(3):603
59. Janjua ZH, Vecchio M, Antonini M, Antonelli F (2019) Irese: an 79. Pawłowicz B, Salach M, Trybus B (2018) Smart city traffic
intelligent rare-event detection system using unsupervised monitoring system based on 5g cellular network, rfid and
learning on the IoT edge. Eng Appl Artif Intell 84:41–50 machine learning. In: KKIO software engineering conference,
60. Jeong Y, Son S, Lee B (2019) The lightweight autonomous Springer, pp 151–165
vehicle self-diagnosis (lavs) using machine learning based on 80. Predictive analytics history & current advances. SAS. https://
sensors and multi-protocol iot gateway. Sensors 19(11):2534 www.sas.com/en_au/insights/analytics/predictive-analytics.
61. Jiang T, Fang H, Wang H (2019) Blockchain-based internet of html. Accessed 15 Jan 2020
vehicles: distributed network architecture and performance 81. Priyashman V, Ismail W (2019) Signal strength and read rate
analysis. IEEE Internet of Things J 6(3):4640–4649. https://doi. prediction modeling using machine learning algorithms for
org/10.1109/JIOT.2018.2874398 vehicular access control and identification. IEEE Sensors J
62. Joshi J, Reddy J, Reddy P, Agarwal A, Agarwal R, Bagga A, 19(4):1400–1411. https://doi.org/10.1109/JSEN.2018.2880736
Bhargava A (2016) Cloud computing based smart garbage 82. Rahman MA, Hassanain E, Rashid MM, Barnes SJ, Hossain MS
monitoring system. In: 2016 3rd international conference on (2018) Spatial blockchain-based secure mass screening
electronic design (ICED), IEEE, pp 70–75
123
framework for children with dyslexia. IEEE Access 97. Terry GA, Harriger JD, Koepf W, Jonnalagadda SR, Webb-
6:61876–61885. https://doi.org/10.1109/ACCESS.2018. Purkis WD, Gainor MS, Griffin PD (2019) Systems and methods
2875242 for enhanced natural language processing for machine learning
83. Ruta M, Scioscia F, Loseto G, Pinto A, Di Sciascio E (2019) conversations. US Patent App. 16/365,668
Machine learning in the internet of things: a semantic-enhanced 98. Tsai CW, Lai CF, Chiang MC, Yang LT (2013) Data mining for
approach. Semantic Web 10:183–204 internet of things: a survey. IEEE Commun Surv Tutor
84. Salhi L, Silverston T, Yamazaki T, Miyoshi T (2019) Early 16(1):77–97
detection system for gas leakage and fire in smart home using 99. Tsikerdekis M, Zeadally S (2014) Online deception in social
machine learning. In: 2019 IEEE international conference on media. Commun ACM 57(9):72–80
consumer electronics (ICCE), IEEE, pp 1–6 100. Verma S, Kawamoto Y, Fadlullah ZM, Nishiyama H, Kato N
85. Sebastian AJ, Islam SN, Mahmud A, Oo AMT (2019) Optimum (2017) A survey on network methodologies for real-time ana-
local energy trading considering priorities in a microgrid. In: lytics of massive iot data and open research issues. IEEE
2019 IEEE international conference on communications, con- Commun Surv Tutor 19(3):1457–1477. https://doi.org/10.1109/
trol, and computing technologies for smart grids COMST.2017.2694469
(SmartGridComm) 101. Vu DL, Nguyen TK, Nguyen TV, Nguyen TN, Massacci F,
86. Sewak M, Singh S (2016) Iot and distributed machine learning Phung PH (2019) Hit4mal: hybrid image transformation for
powered optimal state recommender solution. In: 2016 interna- malware classification. Trans Emerg Telecommun Technol.
tional conference on internet of things and applications (IOTA), https://doi.org/10.1002/ett.3789
IEEE, pp 101–106 102. Vyas DA, Bhatt D, Jha D (2016) IoT: trends, challenges and
87. Shadbolt N, Berners-Lee T, Hall W (2006) The semantic web future scope. Int J Comput Sci Commun 7(1):186–197
revisited. IEEE Intell Syst 21(3):96–101 103. Wang D, Wang X, Zhang Y, Jin L (2019) Detection of power
88. Shah SA, Seker DZ, Rathore MM, Hameed S, Ben Yahia S, grid disturbances and cyber-attacks based on machine learning.
Draheim D (2019) Towards disaster resilient smart cities: Can J Inf Secur Appl 46:42–52
internet of things and big data analytics be the game changers? 104. Wooldridge M, Jennings NR (1995) Intelligent agents: Theory
IEEE Access 7:91885–91903 and practice. Knowl Eng Rev 10(2):115–152
89. Sharma SK, Wang X (2017) Live data analytics with collabo- 105. Wu Y, Ranasinghe DC, Sheng QZ, Zeadally S, Yu J (2011) Rfid
rative edge and cloud processing in wireless iot networks. IEEE enabled traceability networks: a survey. Distrib Parallel Data-
Access 5:4621–4635. https://doi.org/10.1109/ACCESS.2017. bases 29:397–443
2682640 106. Xie G, Zeng G, Kurachi R, Takada H, Li Z, Li R, Li K (2017)
90. Shi W, Cao J, Zhang Q, Li Y, Xu L (2016) Edge computing: Wcrt analysis of can messages in gateway-integrated in-vehicle
vision and challenges. IEEE Internet of Things J 3(5):637–646. networks. IEEE Trans Veh Technol 66(11):9623–9637
https://doi.org/10.1109/JIOT.2016.2579198 107. Yacchirema D, de Puga JS, Palau C, Esteve M (2019) Fall
91. Siryani J, Tanju B, Eveleigh TJ (2017) A machine learning detection system for elderly people using iot and ensemble
decision-support system improves the internet of things’ smart machine learning algorithm. In: Personal and Ubiquitous Com-
meter operations. IEEE Internet of Things J 4(4):1056–1066. puting, pp 1–17
https://doi.org/10.1109/JIOT.2017.2722358 108. Yacchirema DC, Sarabia-Jácome D, Palau CE, Esteve M (2018)
92. Solmaz ME, Mutlu AY, Alankus G, Kılıç V, Bayram A, Horzum A smart system for sleep monitoring by integrating iot with big
N (2018) Quantifying colorimetric tests using a smartphone app data analytics. IEEE Access 6:35988–36001
based on machine learning classifiers. Sens Actuators B Chem 109. Zeadally S, Bello O (2019) Harnessing the power of internet of
255:1967–1973 things based connectivity to improve healthcare. Internet of
93. Stellios I, Kotzanikolaou P, Psarakis M, Alcaraz C, Lopez J Things, Article ID: 100074
(2018) A survey of iot-enabled cyberattacks: assessing attack 110. Zeadally S, Adi E, Baig Z, Khan I (2020) Harnessing artificial
paths to critical infrastructures and services. IEEE Commun intelligence capabilities to improve cybersecurity. IEEE Access
Surv Tutor 20(4):3453–3495. https://doi.org/10.1109/COMST. 8:23817–23837
2018.2855563 111. Zekveld M, Hancke GP (2018) Vibration condition monitoring
94. Suenbuel A, Waldinger R, Sikka V, Richardson K (2019) Sys- using machine learning. In: IECON 2018-44th annual confer-
tems and methods for natural language processing using ence of the IEEE industrial electronics society, IEEE,
machine-oriented inference rules. US Patent 10,515,154 pp 4742–4747
95. Sun Y, Song H, Jara AJ, Bie R (2016) Internet of things and big 112. Zeshan F, Mohamad R (2012) Medical ontology in the dynamic
data analytics for smart and connected communities. IEEE healthcare environment. Proc Comput Sci 10:340–348
Access 4:766–773. https://doi.org/10.1109/ACCESS.2016. 113. Zhang H, Zhang Q, Liu J, Guo H (2018) Fault detection and
2529723 repairing for intelligent connected vehicles based on dynamic
96. Sundaravadivel P, Kesavan K, Kesavan L, Mohanty SP, Kou- bayesian network model. IEEE Internet of Things J
gianos E (2018) Smart-log: a deep-learning based automated 5(4):2431–2440. https://doi.org/10.1109/JIOT.2018.2844287
nutrition monitoring system in the iot. IEEE Trans Consum
Electron 64(3):390–398. https://doi.org/10.1109/TCE.2018. Publisher’s Note Springer Nature remains neutral with regard to
2867802 jurisdictional claims in published maps and institutional affiliations.
123

s00521-020-04874-y

Uploaded by

Copyright:

Available Formats

s00521-020-04874-y

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

s00521-020-04874-y

Uploaded by

Copyright:

Available Formats

Neural Computing and Applications (2020) 32:16205–16233

S.I. : APPLYING ARTIFICIAL INTELLIGENCE TO THE INTERNET OF THINGS

Machine learning and data analytics for the IoT

Keywords Cybersecurity Internet of Things Intelligent systems Machine learning

1 Introduction Play, smart vehicles, building automation systems com-

2.1.3 Prescriptive analytics for IoT

Prescriptive analytics suggests how to respond to any

Fig. 2 Convergence of IoT

Fig. 3 A generalized system

Fig. 4 A three-tier architecture of an intelligent transportation system based on IoT [72]

Fig. 5 Multilayer architecture of the eHealth cloud [37]

Fig. 6 Major components of

Fig. 7 The architecture of

processed either on the data processor in the cloud or on the

Fig. 9 Where data are

Table 1 Knowledge created from the data produced by different applications

Fig. 10 Gateways adapt

Fig. 11 Complex Event

Fig. 13 Smart Objects share

Table 2 The categories of

Fig. 14 Smart Objects share

Fig. 15 Illustration of entity

knowledge from other SOs, and obtained the relationships

Fig. 17 The ontology learned to

Fig. 18 From the attributes of

You might also like

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.