Engineering Applications of Artificial Intelligence 120 (2023) 105856

Contents lists available at ScienceDirect

Engineering Applications of Artificial Intelligence

journal homepage: www.elsevier.com/locate/engappai

Artificial intelligence based real-time earthquake prediction

Munish Bhatia a ,∗, Tariq Ahamed Ahanger b , Ankush Manocha c
a
Department of Computer Applications, National Institute of Technology, Kurukshetra, India
b
College of Business Administration, Prince Sattam Bin Abdulaziz University, Al-Kharj, Saudi Arabia
c
Department of Computer Applications, Lovely Professional University, India

ARTICLE INFO ABSTRACT

Keywords: Earthquake prediction is considered a vital endeavour for human safety. Effective earthquake prediction can
Internet of Things drastically reduce human damage, which is of utmost importance to the community and individuals. In the
Edge computing current research world, there is a boom in scientific interest in the prediction of seismic events. With the
ANFIS prediction
technological revolution in data acquisition, communication networks, edge–cloud computing, the Internet of
Smart city
Things (IoT), and big data analysis, it is feasible to develop an intelligent earthquake prediction model for
early warnings at vulnerable locations. Conspicuously, a collaborative IoT–Edge-centered smart earthquake
monitoring and prediction framework using cloud and edge computing are proposed. IoT technology is
utilized to acquire real-time sensor data, which is forwarded to the edge layer for feature classification
utilizing a novel bayesian belief model technique. Furthermore, Adaptive Neuro-Fuzzy Inference System
(ANFIS) mechanism is employed to forecast the magnitude of earthquakes in the cloud layer. Based on
the experimental simulation, enhanced effectiveness is acquired for the presented framework in terms of
classification performance (Precision (92.52%), Sensitivity (91.72%), and Specificity (91.01%)). Additionally,
results show that the utilization of edge computing significantly reduces computational delay (23.06s).
Moreover, enhanced accuracy and throughput are acquired for the presented model in terms of reliability
(95.26%) and stability (92.16%).

1. Introduction the whole earthquake. Each whole number increase in magni-

tude corresponds to a tenfold increase in observed amplitude and
a 32-fold increase in energy output.
Earthquake is among the worst natural calamities. It poses a signif-
2. Time domain:- The term ‘‘time domain’’ refers to the investiga-
icant threat to the economy and human life, as well as to the peaceful
tion of temporal relationships among mathematical functions,
and persistent progress of society (Abbasi et al., 2021; Falanga et al., physical signals, and time series of economic or ecological data.
2022). During the 20th century, earthquake catastrophes claimed the 3. Time–frequency domain:- A time-domain graph depicts the evo-
lives of over 2 million people worldwide1 . Nearly 100,000 of the lution of a signal over time, but a frequency-domain graph
approximately 500,000 earthquakes that occur each year are reported displays the proportion of the signal inside each frequency band
and experienced by common people2 . The death toll caused by the across a range of frequencies.
terrible earthquakes over the last few decades is depicted in Fig. 13 .
To prevent such outcomes, experts from a variety of fields have collab- 1.1. Problem identification
orated to develop a reliable method of earthquake forecasting. During
the analysis of seismic signals, the characteristics can be classified into Seismology continues to face considerable difficulties in predicting
three domains listed as follows: earthquakes, despite its critical importance for human security. Several
studies argue that an earthquake forecast must include the following;
1. Frequency domain:- The frequency domain of a signal is mathe-
matically represented by its spectral density. To get an accurate 1. A specified area or location
number, all frequencies of shaking must be accounted for during 2. A precise period,

∗ Corresponding author.
E-mail addresses: [email protected] (M. Bhatia), [email protected] (T.A. Ahanger), [email protected] (A. Manocha).
1
Source: https://www.statista.com/.
2
Source: https://www.usgs.gov/observatories/.
3
Source: https://www.statista.com/chart/20443/deadliest-earthquakes-since-1900/.

https://doi.org/10.1016/j.engappai.2023.105856
Received 21 September 2022; Received in revised form 27 December 2022; Accepted 11 January 2023
Available online xxxx
0952-1976/© 2023 Elsevier Ltd. All rights reserved.
M. Bhatia, T.A. Ahanger and A. Manocha Engineering Applications of Artificial Intelligence 120 (2023) 105856

Fig. 1. Global frequency of earthquakes.

3. A defined magnitude range et al., 2017). The intelligent identification of real-time earthquake
4. An exact probability of occurrence occurrences using stand-alone technology is difficult due to several
factors, including the availability of redundant data and network ser-
In other words, a prediction of an earthquake must include the time,
vices (Brous et al., 2019; Bhardwaj et al., 2019). Emerging technologies
location, magnitude, probability, and reason for its occurrence. The
of IoT, edge computing, and cloud computing enable dynamic net-
purpose of earthquake forecasts is to aid disaster control organizations
works with effective data analysis to produce enhanced solutions for
in preparing for earthquakes. When a powerful earthquake is forecast,
disaster management. Considering the time-sensitivity of monitoring
disaster control administrators must be alerted to take precaution-
earthquakes, the proposed architecture for real-time data processing
ary measures. In catastrophe preparedness, decisions and activities
incorporates Edge Computing (Cremen et al., 2021; Xu et al., 2021).
are centered on preventing losses. Conspicuously, a variety of earth-
It extends the cloud computing architecture to the network edge,
quake prediction techniques have been implemented to reduce damage.
Using several methods including precursor signals, mathematical com- providing benefits such as decreased delay, quality of service, location-
putation, electromagnetic fields, and anomalous and animal behavior, based services, time-sensitive evaluation, and early alerts (Iwendi et al.,
earthquake scientists have realized the utility of accurate earthquake 2020).
forecasting to society. Unfortunately, because of the complexity and
unpredictability of earthquakes, conventional techniques seldom gen- 1.3. Major contribution
erate effective outcomes (Xin et al., 2022). Henceforth, it presents the
primary motivation for the current research work. The development of Based on the aforementioned aspects, the current architecture in-
monitoring techniques and computing technology acting as a driving tegrates IoT, Edge Computing, cloud computing, and predictive ana-
revolution in computational engineering provides an efficient answer lytic approaches to identify earthquakes in real time and issue pub-
to earthquake prediction. lic safety alerts in advance (Bandyopadhyay and Sen, 2011). Fig. 2
shows the integration of 3 paradigms to realize vital service delivery
1.2. Research motivation aspects (Yu et al., 2017). The integration has enabled real-time accessi-
bility, privacy, latency, and dependability for applications. The primary
The current use of IoT technology has been recognized as an emer- objectives of the proposed research are mentioned ahead;
gent study subject in every domain of science and engineering in
response to earthquake-related losses (Bandyopadhyay and Sen, 2011; 1. Presenting a comprehensive earthquake prediction framework
Gubbi et al., 2013). Small, low-cost sensors are utilized to collect seis- using sophisticated IoT–edge–cloud computing platforms to ex-
mic data for IoT-enabled earthquake monitoring and prediction (Ab- tract optimum features from complicated data.
basi et al., 2021). For early detection, sensor data is evaluated in a 2. Analyzing IoT data and events from IoT sensors and extending
time-sensitive manner (Bassetti and Panizzi, 2022; Tao et al., 2022). to the distant cloud platform, to enhance the data gathering and
Nonetheless, processing the enormous data contents requires a vast analysis processes.
amount of computer resources, which are usually destroyed, unavail- 3. Developing a prototype architecture to collect real-time IoT data
able, or only temporarily available in the disaster-affected area (Lin for effective prediction systems.

2
M. Bhatia, T.A. Ahanger and A. Manocha Engineering Applications of Artificial Intelligence 120 (2023) 105856

Fig. 2. IoT–Edge–Cloud layered architecture.

4. Evaluating the proposed model for performance assessment in use of intelligent sensors and cloud services, advancements have been
terms of several metrics including classification efficacy, predic- achieved in the infrastructure for sensing. Bassetti and Panizzi (2022)
tion efficiency, computational delay, reliability, and stability. created distributed warning framework for the metropolitan city by
combining mesh networks and seismometers. Utilizing a cooperative
Paper organization. Section 2 presents a concise summary of state-
approach for signal processing, the presented system can distinguish
of-the-art works for earthquake prediction. Section 3 describes the
between earthquakes and other forms of ground shaking in a metropo-
architectural layout of the presented model. Section 4 examines the
lis. Roy et al. (2021) used regression techniques and a cloud-based
performance assessment of the presented system. Section 5 concludes
infrastructure to forecast earthquakes in California. Sun et al. (2021)
the paper with research directions for the future.
reviewed techniques for detecting earthquakes based on the study
of electromagnetic waves. Several seismic factors including energy,
2. Literature review
wavelength, frequency, wave magnitude, and zones are abstracted from
Long-term and short-term earthquake forecasting has been a core waves utilizing Fast Fourier Transformation to detect an earthquake.
research domain for several years. Numerous research on occurrences Won et al. (2020) presented an IoT device for intelligent earthquake
and predictions of earthquakes have resulted in dynamic results. The identification. It consists of 3 major components including a high-
seismology community has developed a successful implementation sys- fidelity accelerometer, an inbuilt earthquake identification system, and
tem for early earthquake determination and characterization. Kumar a Bluetooth beacon for alert generation. However, the authors did not
et al. (2018) created a seismic activity identification method using address location awareness or latency-sensitive issues. Moreover, the
the Short-Time-Average/Long-Time-Average (STA/LTA) algorithm. It authors used a seismological software package as a server application
established a collaborative sensing network to monitor seismic activity for data communication. Cremen et al. (2021) presented the early
in a specified region. Specifically, it utilizes an activity identifica- earthquake warning (EEW) system by utilizing a Virtual Seismologist.
tion approach to differentiate between seismic events and local noise. Imeraj et al. (2022) designed and implemented an embedded seis-
Vaezi and Van der Baan (2015) evaluated power spectral density mograph for efficiently and effectively supporting seismological and
techniques for the identification of seismic events. Liao et al. (2022) geophysical data collection networks.
employed a Template-Matching Algorithm to enhance the effectiveness
of earthquake detection. It examined vital characteristics of the earth- 2.1. Research gaps
quake using GPU architecture to expedite computations. The presented
technique is computationally intensive and is limited in recognizing Catastrophe organizations prioritize disaster management in order
earthquakes. In addition, research conducted in recent years has proved to limit the loss of property and lives to the greatest degree feasible.
the applicability of Machine Learning (ML) approaches to seismolog- Several viable decision-making models for forecasting earthquakes in
ical issues. Bhargava and Pasari (2022) proposed Artificial Neural real time have been envisioned using IoT–Edge–Cloud computing. Nev-
Networks for predicting earthquake magnitudes using temporal infor-
ertheless, based on the extensive literature review, Table 1 identifies
mation. Additionally, prediction is performed by employing Seismic
some significant research gaps linked to the current study area. In
Electric Signals as input data to enhance accuracy. Shen and Shen
addition, some of the most critical barriers are as follows:
(2021) presented convolutional neural networks for seismological event
detection. The presented technique depends on the detection of earth- 1. Researchers have not explicitly studied the assessment of earth-
quake signals in the temporal domain and employs the spectral features quake prediction in a real-time scenario.
of phase to differentiate between waves. Wibowo et al. (2021) de- 2. Minimal research effort has been undertaken for earthquake
veloped a P-wave discriminator to alleviate the noise trigger issue real-time detection over a time-scale.
by employing two potent ML algorithms: GANs and Random Forests. 3. The majority of research is restricted to monitoring, and less
Tehseen et al. (2021) introduced an ANFIS-based technique for pre- work is undertaken on the analysis of IoT data for accurately pre-
dicting future earthquakes with a magnitude of at least 5.5. The ANFIS dicting earthquakes in real-time and enhancing public warning
system was modeled using 3 algorithms including Fuzzy C-means, Grid issuance.
Partition, and Subtractive Clustering. Ahamed and Daub (2021) com- 4. The inclusion of ANFIS for accurate earthquake magnitude pre-
pared numerous temporal evaluation techniques such as data network, diction is an additional component that has been lightly investi-
k-nearest neighbors and multi-objective info-fuzzy for predicting the gated in contemporary research.
magnitude of the earthquake event based on sampled data. With the

3
M. Bhatia, T.A. Ahanger and A. Manocha Engineering Applications of Artificial Intelligence 120 (2023) 105856

Fig. 3. Proposed model.

3. Proposed framework data on seismic waves, ground motion, ground-level variations,

vibration noise, and a temporal evaluation of deformations.
Fig. 3 presents a layered framework of the proposed IoT–centric 2. Position Attributes The locational latitudes and longitudes of the
edge–cloud collaborative system for earthquake detection for monitor- afflicted region are utilized to pinpoint the precise geographic
ing seismic signals efficiently. It comprises several layers including the area of the earthquake’s epicenter. It includes latitude, longi-
Data Sensing Layer (DSL), Edge Layer (EL), Data Analysis Layer (DAL), tude, and epicenter.
and Cloud and Information Sharing Layer (CISL). DSL is responsible for
distributing sensors over the geographical region to gather seismic data Table 2 displays the described specific earthquake parameters and
in its raw form. The cumulative data are sent to EL for the classification respective IoT sensors. For earthquake prediction, sensors installed in
of earthquakes in real time. DAL enables efficient extraction of useful the research zone measure the earthquake-inducing characteristics.
data instances based on a temporal manner. CL is responsible for accu-
3.2. Edge layer (EL)
mulating compiled data for forecasting and monitoring results, thereby
assisting disaster response organizations in controlling earthquake loss.
EL layer aligned between IoT devices and DAL. It is comprised of
edge gateways and databases to enable dynamic real-time processing
and storage along with data transmission to the cloud server. Numerous
3.1. Data sensing layer (DSL)
modules are integrated for EL-based computational analysis. It includes
filtering data, extracting characteristics, decreasing sizes, and catego-
DSL is the physical layer for gathering and transforming data into
rizing events for seismic prediction. EL is crucial for preventing data
comprehensible digital signals by utilizing IoT devices. It detects and
transmission to the cloud repository for temporal computations.
transforms seismic characteristics into useful information. An efficient
earthquake detection framework incorporates geological and location
3.2.1. Extracting features
factors.
During seismic data accumulation, it is interrupted by a variety of
1. Geological Attributes: It depicts the geological state in the region distorting sounds of motion artefacts, baseline drift, and instrumenta-
with a substantial effect on earthquake initiation. It comprises tion noise created by the device. For precise and exhaustive earthquake

4
M. Bhatia, T.A. Ahanger and A. Manocha Engineering Applications of Artificial Intelligence 120 (2023) 105856

Table 1
State-of-the-art comparison (A Exist, – Do-not Exist).
Reference Kumar et al. Vaezi and Van der Liao et al. Bhargava and Tehseen Wibowo Shen and This Work
(2018) Baan (2015) (2022) Pasari (2022) et al. (2021) et al. (2021) Shen (2021)
Sensing technology A A A A A A A A
Edge computing – A A – A – – A
Classification – – – – – – – A
Storage A A A A – – A A
Temporal analysis – – – – – A A A
Quantification A A A A A – A A
Delay latency – – – – – – – A
Reliability – – – – – – – A
Predictive behavior – – A – A – – A
Security – – – – – A – A
Stability – – A – – – – A

Table 2
Earthquake attributes.
Attribute Detail IoT-Sensors
Ground motion Motion of the ground due to earthquake Seismic sensors
Seismic noise Ground shaking noise Seismometers
Groundwater levels Earthquake result in the water level rise leading to Tides Water level sensors
Longitude Coordinates for the location of the earthquake Location sensors
Epicenter Focus point of the earthquake Epicenter sensor
Latitude Coordinates for the location of the earthquake Location sensors

Table 3 Table 4
Extracted time features with mathematical description. Extracted frequency features with mathematical description.
Feature domain Definition Mathematical formula Feature domain Definition Mathematical formula
√
1 ∑𝑛 2 Peak frequency 𝑓𝑝 = arg max |𝐹𝑖 (𝑓 )|2
Standard deviation 𝜎x= 𝑛−1 𝑖=1 (𝑥𝑖 − 𝜇𝑥 ) ∑𝑁
( )3 Mean frequency 𝑓𝑚 = ∑𝑖=1
𝑓𝑖 𝑃𝑖
1 ∑𝑛 (𝑥𝑖 −𝜇𝑥 ) 𝑛
𝑃𝑖
Skewness 𝑆𝑘 = 𝑛 𝑖=1 ∑𝑁
𝑖=1
𝜎𝑥 Frequency domain 𝑃𝑖
( ) Mean power 𝑃𝑚 = 𝑖=1

∑𝑛 ( (𝑥𝑖 −𝜇𝑥 ) )4 ∑𝑛𝑁

Kurtosis 𝐾 = 1𝑛 𝑖=1 𝜎𝑥
−3 Spectral moment of order-2 SM = 𝑃𝑖 𝑓𝑖2
𝑖=1
∑𝑛
1 ∑𝑛 2 Total power TP = 𝑃
Time domain Variance 𝑉 = 𝑛−1 𝑖=1
|𝑥𝑖 | 𝑖=1 𝑖
√ ∑
1 𝑛
Root mean square RMS = 𝑛 𝑖=1
|𝑥𝑖 |2
1 ∑𝑛
Mean absolute value MAE = 𝑖=1
|𝑥𝑖 |
𝑛
[ ] square, Mean absolute value, Approximate entropy, Energy entropy,
Approximate entropy AEn(m, r, N) = 𝜙𝑚 (𝑟) − 𝜙𝑚+1 (𝑟)
∑𝑛
and Waveform length. The mathematical description of all the matrices
Energy entropy EEn = − 𝑖=1 𝑝𝑖 log 𝑝𝑖 is presented in Table 3.
∑𝑛−1
Waveform length WL = 𝑖=1 |𝑥𝑖+1 − 𝑥𝑖 | Frequency domain: The Power Spectral Density of a seismic wave
is employed by calculating Peak frequency, Mean frequency, Mean
power Spectral moment of order-2, and Total power as presented in
detection, it is required to filter out high-frequency modules and noise Table 4.
sources. The seismological data were analyzed using a 4th order filter Time-Frequency domain: At each level of DWT decomposition,
to remove the power-line noise. Feature extraction is the procedure of time–frequency characteristics are retrieved by calculating the Stan-
extracting important aspects from unprocessed input values and con- dard deviation of the wavelet, Average wavelet power, Mean of the
densing in reduced features. The success of seismic event categorization absolute value, and Entropy. The mathematical description is provided
is highly dependent on the selection of relevant characteristics. Using in Table 5.
the seismic signal to interpret features, a Discrete Wavelet Transform
(DWT) of 5 levels is used for extracting unique characteristics from 3.2.2. Attribute reduction
DWT coefficients of varying sizes. Due to the non-stationary character After extracting, the attribute dimension is large thereby elevating
of the DWT, it is used for the time–frequency signal analysis. The DWT classifier complexity and lowering error convergence. To reduce com-
employs a large temporal window for reduced frequency and a short- putational complexity and enhance generalization on unlabeled data,
temporal window for higher frequency. It decomposes the initial signal it is essential to reduce the number of features to a minimum. Using a
in a wavelet coefficient that represents the signal vector behavior at feature selection strategy, it is possible to reduce attribute dimension.
multiple frequencies. At the initial level, the seismic signal is assessed The purpose of feature identification is to (i) Dimension reduction,
for high-low pass filtration. This procedure is repeated for each level (ii) Eliminate extraneous characteristics, (iii) Minimize classifier data,
of decomposition. It turns a signal into approximate and detailed and (iv) Enhance classification performance. In the presented research,
coefficients. During the analysis of seismic signals, the characteristics correlation-based Feature Selection (CBFS) is employed to determine
can be classified into three domains listed as follows: usable features. CBFS is a feature identification strategy for determining
the reduced characteristics that are potentially relevant to a given
1. Frequency domain,
task. CBFS selects the vital characteristic for increased precision, while
2. Time domain,
simultaneously lowering the number of features and delays. The impor-
3. Time–frequency domain
tant component of the CBFS technique is a heuristic for evaluating the
Time domain: From seismic datasets, time-domain characteristics value of the attribute subset defined as follows:
are extracted explicitly. Time domain features are extracted by cal- 𝑆𝐷𝐺
MS (Merit) = √
culating Standard deviation, Skewness, Kurtosis, Variance, Root mean 𝐿 + 𝐿(𝑙 − 1)𝑆𝐺𝐺

5
M. Bhatia, T.A. Ahanger and A. Manocha Engineering Applications of Artificial Intelligence 120 (2023) 105856

Table 5
Extracted time–frequency features with mathematical description.
Feature domain Definition Mathematical formula
√ ∑𝑁
1
Standard deviation of wavelet Coefficients in each level: 𝜎 = 𝑁 𝑖=1
(𝑐𝑖 − 𝜇)2
1 ∑𝑁 2
Time-Frequency domain Average wavelet power Coefficients in each level: 𝑃𝑖 = 𝑁 𝑖=0 (𝑐𝑖 )
1 ∑𝑁
Mean absolute value Coefficients in each decomposition: 𝜇𝑖 = 𝑁 𝑖=1 (𝑐𝑖 )
∑𝑁
Entropy 𝐸𝑁 = 𝜎 = 𝑖=1 (𝑐𝑖2 𝑙𝑜𝑔(𝑐𝑖2 ))

𝑗 𝐷
where L is the number of features, SGG represents the mean feature- type 𝐷𝑗 is represented by 𝑃 ( 𝐸 ,𝐸 ,…,𝐸 ). Numerous input parameters
1 2 𝑛
to-feature correlation, and SDG represents the mean feature-to-class are accessible, and a specific metric instance of earthquake detection
correlation. It emphasizes the importance of particular characteristics might have a substantial value; hence, the above-mentioned formula
while class label evaluation. Additionally, it reduces duplicate and may lead to inconsistency in earthquake magnitude detection in real
unnecessary characteristics that are insufficient class predictors. The time. Following this, the revised BBM is stated as
CBFS method is explained as follows:
𝐷𝑗 𝑃 (𝐷𝑗 )𝑃 (𝐸𝑖 ∕𝐷𝑗 )
𝑃( )= .
1. Calculate the correlation between the feature class and the cor- 𝐸𝑖 𝑃 (𝐸𝑖 )
relation between features.
However, the probability of 𝑃 (𝐷𝑗 )𝑃 (𝐸𝑖 ∕𝐷𝑗 ) can be improved based on
2. Evaluate the merit value of each feature.
joint probability function as
Using the outcomes of the previous phase, the subset with the highest => 𝑃 (𝐷𝑗 )𝑃 (𝐸𝑖 ∕𝐸𝑗 ) = 𝑃 (𝐸1 , 𝐸2 , … , 𝐸𝑛 , 𝐷𝑗 )
MS is selected. The selected subset of features and respective labels => 𝑃 (𝐸1 ∕𝐸2 , … , 𝐸𝑛 , 𝐷𝑗 )𝑃 (𝐸2 , … , 𝐸𝑛 , 𝐷𝑗 )
constitute an appropriate feature set. To minimize feature dimension, => 𝑃 (𝐸1 ∕𝐸2 , … , 𝐸𝑛 , 𝐷𝑗 )𝑃 (𝐸2 ∕𝐸3 , … , 𝐸𝑛 , 𝐷𝑗 )𝑃 (𝐸3 , … , 𝐸𝑛 , 𝐷𝑗 )
the Genetic Algorithm (GA) is utilized as a searching approach with => 𝑃 (𝐸1 ∕𝐸2 , … , 𝐸𝑛 , 𝐷𝑗 )𝑃 (𝐸2 ∕𝐸3 , … , 𝐸𝑛 𝐷𝑗 ), … , 𝑃 (𝐸𝑛−1 ∕𝐸𝑛 , … , 𝐸𝑛 𝐷𝑗 ) ∗
CBFS as an evaluating function. The GA is a probabilistic, particu- 𝑃 (𝐸𝑛 ∕𝐷𝑗 )𝑃 (𝐷𝑗 )
lar search strategy capable of identifying vast search areas, which is In addition, it is expected that each characteristic 𝐸𝑖 of the variable
typically required in attribute identification. In the GA mechanism, for disaster identification is independent of any other measure 𝐸𝑗
merit measures are evaluated for nine distinct attribute subsets. This
i.e. 𝑖 ≠ 𝑗. Then 𝑃 (𝐸𝑖 ∕𝐸𝑖+1 , … , 𝐸𝑛 , 𝐷𝑗 ) = 𝑃 (𝐸𝑖 ∕𝐷𝑗 )
process is repeated at different times such that the highest merit value
Therefore, the joint probability is described as follows:
is selected. Ten characteristics (greatest merit value) have been selected 𝑛 𝑃 (𝐷 )𝑃 (𝐸 ∕𝐷 )
=> 𝑃 (𝐷𝑗 ) = 𝛱𝑖=1 𝑗 𝑖 𝑗
for categorization. 𝐷
=> 𝑃 ( 𝑝𝑗 ) = 𝛱𝑖=1
𝑛 𝑃 (𝐷 )𝑃 (𝐸 ∕𝐷 )∕𝑃 (𝐸)
𝑗 𝑖 𝑗
3.2.3. Event classification In the equation described above 𝐷𝑗 denotes the class 1 and class 2
One of the most vital tasks in the presented model is the categoriza- parameters.
tion of the attributes in distinct seismic activities namely earthquakes
and non-earthquake. Various classification techniques of clustering, dis- 3.3. Data analysis layer
criminant analysis, and neural networks, have been evaluated for the
feature classification process. Even though the strategies were helpful
To monitor earthquake data, DAL’s purpose is to extract meaningful
for a variety of classification issues, the Bayesian Belief Model (BBM)
information from the data repository. The proposed model aims to
classifier performed best in the current research. Additionally, the
provide real-time disaster relief services. Earthquake vulnerability is
classification work is essential for reliably forecasting earthquakes with
analyzed in greater detail by abstracting temporal data segments. Using
better precision and reduced classification time to provide real-time
services. The basic layer collects data from numerous IoT devices on Temporal Data Mining, valuable data segments may be retrieved for
catastrophic occurrences. The gathered data is examined to ascertain further analysis. Time-series data may be effectively analyzed using
the probability of a calamity occurring. BBM is an efficient instrument Temporal Data Mining techniques (Xiao et al., 2021; Xing et al., 2022).
for determining the probability of earthquake detection (ED) (Qazi
et al., 2018). BBM uses current research data to establish the Probabil- Definition 2 (Temporal Granule (𝑈 ∗ )). Given the parametric measure in
ity of Earthquake Detection, a probabilistic metric for ED determination the temporal window of 𝛥𝑡, then 𝑈 ∗ is represented as the accumulation
(PoED). of data measures for earthquake-oriented parameters corresponding to
every instance of time 𝑡 ∈ 𝛥𝑡 and is denoted as (𝑡1 , 𝑤1 )(𝑡2 , 𝑤2 ), . . . , (𝑡𝑛 ,
Definition 1 (Probability of Earthquake Detection(PoED)). It is a prob- 𝑤𝑜 ).
abilistic metric to detect the magnitude of an earthquake in real-time
at a particular instance of time 𝛿t. Specifically, the PoED provides a Several data segments that affect the earthquake prediction can be
methodical measurement of an unfavorable environmental occurrence abstracted effectively using the data abstraction of Temporal Granule
(𝑈 ∗ ). Data values may be effectively assessed for further in-depth
Definition 1 determines a quantifiable metric for the detection of the examination using such conceptualization. Disaster relief organizations
earthquake in real time. In other words, PoED evaluates the parameters can use the formulation to give timely and effective care to the people
for the real-time prediction of the earthquake. In the current study, in the event of an earthquake event. Fig. 4 depicts a quick summary of
the BBM method is preferred for classification because of its accuracy, temporal granulation.
feasibility, and efficiency.

3.2.4. BBM-based mathematical analysis 3.4. Cloud layer (CL)

BBM is utilized to categorize the datasets into distinct categories.
As previously stated, two types are characterized using distinct seismic CL is vital in the proposed system. As a result of the restricted
parameters. Let a vector 𝐸𝑖 = (𝐸1 , 𝐸2 , … , 𝐸𝑛 ) represent an instance of storing ability and computation capabilities of the edge layer, seismic
the data, where 𝐸𝑖 represents the earthquake variable i, assuming that event data is kept and processed on Cloud servers as assembled records.
all parameters related to the earthquake are not mutually dependent. It is comprised of two parts: (1) Magnitude forecasting and (2) Cloud
The conditional likelihood that an earthquake will be detected 𝐸𝑖 of storage.

6
M. Bhatia, T.A. Ahanger and A. Manocha Engineering Applications of Artificial Intelligence 120 (2023) 105856

Fig. 4. Temporal analysis.

3.4.1. Magnitude forecasting strength is denoted by 𝑤′𝑗 and is further standardized as follows:
In the current research, the ANFIS paradigm is utilized for the 𝑤𝑗
quantification of identified earthquakes. It is a culmination of ANN 𝑁𝑘3 = 𝑤′𝑘 = 𝑓 𝑜𝑟 𝑘 = 1, 2 (3)
𝑤1 + 𝑤2
and fuzzy theory that integrates the learning powers of a neural system
with the inference capabilities of fuzzy logic and superior knowledge (d) De-fuzzification (Layer 4): This layer has the responsibility
representation to improve prediction capacity. To find the ideal mem- for assessing the contribution to the final output of the 𝑘th rule. The
bership function distribution, the ANFIS technique can generate an corresponding structured consequent variables are classified as the 𝑐𝑗,
input–output mapping based on expert information. In this manner, the 𝑑𝑗 and 𝑟𝑗 parameters. The de-fuzzification function in this layer is as
mathematical formulation of the ANFIS model is depicted ahead. follows:

𝑁𝑘4 = 𝑤′𝑘 𝑘𝑘 = 𝑤′𝑘 (𝑐𝑘 𝑥 + 𝑑𝑘 𝑦 + 𝑟𝑘 ) 𝑓 𝑜𝑟 𝑘 = 1, 2 (4)

3.4.2. Adaptive neuro-Fuzzy inference system (ANFIS)
Numerous academics in various domains employ the ANFIS (e) Output generation (Layer 5): It is the output layer for accu-
paradigm for predictive decision modeling. ANFIS is typically em- mulating all de-fuzzification outputs and computes the finalized result.
ployed to address exceedingly complicated and nonlinear problems. 𝑁𝑘5 as denoted in Eq. (5);
Fig. 5 depicts the 5-layer ANFIS structure with seven input attributes. ∑
Fuzzy logic enables multi-value input from a parental input variable ∑ 𝑤𝑘 𝑔𝑘
𝑁𝑘5 = 𝑤′𝑘 𝑔𝑘 = ∑𝑘 (5)
expressing the connection between a single value set and a collection 𝑘 𝑤𝑘
of additional values. The suggested system uses a fuzzy inference
model for non-linear vector input. Each attribute is evaluated within 3.4.3. Cloud backup
a particular temporal frame using the ANFIS model. As an illustration, This component is responsible for assembling the findings of the
the data measures for evaluating the earthquake in a certain time–space seismic event outbreak analysis including earthquake magnitude. It can
frame are provided to the ANFIS, which then calculates its prediction help seismologists, emergency responders, and disaster management
measure.
agencies to mitigate and handle earthquakes more efficiently.
(a) Fuzzification (Layer 1): The initial layer is the fuzzy layer
which employs membership functions to turn the input into fuzzy sets.
Each adaptive node in this layer is depicted as: 3.5. Information sharing

𝑁𝑘1 = 𝜇𝐻𝑘 (𝑥) (1)

The cloud-based findings are accessible to disaster management
where 𝑁𝑘1is the gaussian function, 𝑥 is the node-specific input value agencies, emergency responders, government organizations, and other
𝑘, and 𝐻𝑘 is the linguistic parameter associated to a node. More- relevant authorities, allowing them to act swiftly to reduce the reper-
over, the first layer is provided with temporal measures for different cussions. A gateway linked to the internet is utilized to transmit the
earthquake-oriented attributes. assembled data to the relevant authorities. From the cloud layer, eval-
(b) Product rule (Layer 2): By implementing the product opera- uated data are transferred directly to the gateway. The gateway can
tion, the nodes pass the input to the next layer and are represented then combine the data and transmit it through the internet to a central
mathematically as; platform maintained by the relevant Emergency Response Teams (ERT).
These data may be accessible via the ERT’s digital website to obtain
𝑁𝑘2 = 𝜇𝐴𝑘 (𝑥) × 𝜇𝐵𝑘 (𝑦), 𝑘 = 1, 2 (2)
the most information possible regarding the catastrophe site and its
where 𝐴𝑘 (𝑥) and 𝐵𝑘 (𝑦) represents the nodes in layer 2. occurrences. Authorities adopt short-term and long-term strategies to
(c) Normalization (Layer 3): Each node computes the ratio of the counteract and alleviate the negative consequences of the earthquake
firing strength rule to the total of all firing strength rules. The firing based on the estimation of the situation’s urgency.

7
M. Bhatia, T.A. Ahanger and A. Manocha Engineering Applications of Artificial Intelligence 120 (2023) 105856

Fig. 5. Generalized ANFIS model.

4. Experimental simulation The current study utilized approximately 150000 waveforms, including
70000 earthquake waveforms and 30000 no-earthquake waveforms ob-
The implementation design and evaluation of the proposed model tained at epicenters ranging from 0.0 to 300.0 km. Each wave consists
are described. The following tasks are computed for the estimation of of 7000 samples with a length of 2 min and a sampling frequency of
the performance of the proposed model. (a) Feature extraction results, 99 hertz. In addition, a location attribute containing data is extracted
(b) ANFIS-based magnitude forecasting results, (c) Feature extraction from (Mousavi et al., 2019).
results, (d) Classification effectiveness at the edge layer, (e) Predic-
Furthermore, a network topology is designed to evaluate the effi-
tion efficiency, (f) Reliability, (g) Stability of the proposed system,
cacy of the proposed solution by following the concept of the IoT–Fog–
(h) Computational complexity.
Cloud tier. Fig. 6 depicts the actual testing setup used to perform the
4.1. Data compilation experiments. Fog cells, fog nodes, sensors, and fog–cloud middleware-
based four types of devices are part of the topology. The middleware
The Stanford Earthquake Dataset (STEAD) database4 contains in- is executing on a system which is connected to the OpenStack cloud
formation on several seismic and geological characteristics. STEAD infrastructure and creates the top tier of the topology. Raspberry Pis
is a massive worldwide dataset comprised of 2 seismogram-labeled are used to arrange the remaining configuration elements. Fog con-
set groups. The first set depicts earthquake waveforms, whereas the trol nodes (FCN) are the root fog coordinator nodes for all nearby
other set depicts waveforms that are not associated with earthquakes. fog devices that handle data from integrated IoT devices. The IoT-
connected devices in this setup are sensor modules that are linked to
4
Source: https://github.com/smousavi05/STEAD. the appropriate Raspberry Pie.

8
M. Bhatia, T.A. Ahanger and A. Manocha Engineering Applications of Artificial Intelligence 120 (2023) 105856

Fig. 6. Designed topology.

Table 6 Table 7
Seismic signal-based 5-level DWT decomposition. CBFS estimation.
Decomposition levels Frequency (Hz) Coefficients S.no Feature Domain Mean
0 26–51 D1 1 Average power of wavelet coefficients Time-Frequency 0.397, 0.020
1 11.802–26 D2 2 Energy entropy Time-Frequency 0.382, 0.027
2 8.062–11.802 D3 3 Mean power Frequency 0.303, 0.026
3 5.273–8.061 D4 4 Entropy Time 0.329, 0.046
4 0–0.5273 A4 5 Root mean square Time 0.476, 0.031
6 Waveform length Time 0.296, 0.09
7 Peak frequency Frequency 0.375, 0.0.35
8 Mean of the absolute value Time-Frequency 0.307, 0.053
Combining and storing datasets on Amazon EC2 to generate the 9 Spectral moment of order-2 Frequency 0.579, 0.029
10 Mean frequency Frequency 0.240, 0.063
finalized data. The stages required in integrating disparate datasets to
generate a single earthquake parameter dataset are depicted in Fig. 7.
Table 8
BBM confusion matrix.
4.2. Feature extraction results
Classification Earthquake No-Earthquake

The Discrete Wavelet Transform (DWT) is utilized for obtaining No-Earthquake 0.27 0.73
Earthquake 0.81 0.7
seismic signal characteristics. It separates the seismic signal into sev-
eral sub-signals with varying frequencies. Identifying wavelet function
and the correct degrees of decomposition is essential when utilizing
DWT to analyze seismic data. In the proposed model, the degrees of data is displayed in Table 7. Based on the highest merit score and
decomposition are set to 4, and the wavelet function evaluated is DB4, standard deviation value, the features for the reduced feature set are
which is appropriate for seismic signal processing. Table 6 depicts 5 determined to be 10.
levels of DWT and decomposition frequency bands. Calculating and
analyzing the DWT coefficient is performed with MATLAB. Fig. 8 4.4. Classification effectiveness: Edge layer
depicts the 5 unique sub-signals of the seismological data, together with
one approximation coefficient A4 and 4 detailed coefficients DI–D4. The suggested approach distinguishes between earthquakes and
Additionally, the original attribute data is created by abstracting the 18 other types of seismological events using a BBM Classifier at the EL
conventional characteristics. Based on the approximation and detailed Layer. The BBM classifier was trained using 5000 seismic data using
coefficients over all subbands, the estimated value is 3.1. WEKA, with 15 iterations used to speed up the training process. In
Table 8, it can be seen that many instances were correctly or incorrectly
4.3. Feature reduction classified as positive or negative. In addition, the collected data is
used to develop statistical measures like accuracy, responsiveness, and
In the current study, implementation is performed for the CBFS selectivity.
with the GA search method. The CBFS method assigns the highest Table 9 shows the outcome of the trained classifier’s classification
measures to feature subsets with the class characteristic but are very of the training dataset. The results show that 97% of all seismic
moderately connected. WEKA tool is used as a searching strategy using events are correctly diagnosed, whereas 3% are incorrectly labeled as
CBFS as a subset assessment mechanism (fitness function). The number something other than earthquakes. Nearly all earthquake data points
of significant features selected by the CBFS–GA for extracted attribute are correctly recognized, with only 7% mislabeled as anything other

9
M. Bhatia, T.A. Ahanger and A. Manocha Engineering Applications of Artificial Intelligence 120 (2023) 105856

earthquake detection. The BBM classifier has a large area under the
ROC curve compared to other methods. Fig. 9 shows the amount of time
it takes classifiers to make their classifications in relation to the total
quantity of seismic samples. The BBM classifier’s classification time
grows more slowly in proportion to the number of samples than that of
other classifiers. The BBM classifier is well-suited for both earthquake
and non-earthquake classification in real-time since its cumulative
classification time remains below normal ranges.

4.5. Prediction analysis

As discussed before, ANFIS is composed of 5 layers, 3 inputs (lat-

itude, longitude, and depth), and 1 output (magnitude forecast). The
ANFIS model was run in MATLAB R2019b for the simulation purpose.
The suggested model’s ANFIS training parameters are described in Ta-
ble 11. Fuzzy networks utilize backpropagation and a hybrid algorithm
to fine-tune the membership function’s variables in order to learn from
data. For training ANFIS models, the Gaussian membership function
is used. The learning procedure modifies the membership function’s
parameters. For a given set of attributes, the accuracy of the output
data is determined by the gradient vector. The error rate is decreased
by adjusting the settings after receiving the gradient vector. It is the
sum of the squared deviations between the expected and actual outputs
that constitute the error. To test how well ANFIS predicted earthquake
magnitudes, a 10 Fold cross-validation technique is employed. Further,
the training method is iterated inside each fold throughout several
epochs until the training error aim is achieved or a maximum number
of iterations is reached. The average percentage of mistakes across all
folds and periods is shown in Fig. 10. The inaccuracy in the forecast
was determined by subtracting the measured Richter magnitude from
the ANFIS-predicted result i.e.

𝐸𝑟𝑟𝑜𝑟 = 𝐴𝑐𝑡𝑢𝑎𝑙 𝑣𝑎𝑙𝑢𝑒 − 𝑃 𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑 𝑣𝑎𝑙𝑢𝑒

Error reduction is analyzed, therefore training is halted at the 55th

epoch. The performance metrics acquired by the ANFIS model for
testing and training data are displayed in Table 12. Coefficient of
Determination (R2), Root Mean Square Error (RMSE), and Mean Ab-
solute Error (MAE) values of 0.908%, 0.915%, and 0.895 respectively,
indicate that the ANFIS model produced acceptable results for fore-
casting earthquake magnitude. Moreover, several statistical parameters
are acquired for determining the prediction efficiency. These include
sensitivity, specificity, accuracy, and precision. For determining the
overall values of the statistical parameters, the proposed model is
compared with state-of-the-art models namely, K-Nearest Neighbor (K-
NN), Artificial Neural Network (ANN), and Support Vector Machine
(SVM). The plot in Fig. 11(a, b, c and d) denotes the accuracy, pre-
cision, sensitivity, and specificity of the proposed model. Computing
Fig. 7. Algorithm 1 Integration of earthquake attribute dataset. over an average of the number of instances, the presented model was
able to attain the highest accuracy value of 95.85%, the precision of
Table 9 96.26%, the sensitivity of 96.89% and specificity of 95.48% which in
BBM classification performance.
comparison to K-NN, cANN, and SVM is better. The findings indicate
S.no Seismic events Sensitivity(%) Specificity(%)
that utilizing the ANFIS model to make predictions is a viable strategy,
1 Non-Earthquake 93 98.51 as the results are significantly more exact and accurate.
2 Earthquake 97 99.54

4.6. Reliability analysis

than earthquakes. In addition, cutting-edge methods like the Support Reliability ensures dependability for the current domain of research.
Vector Machine and the C4.5 decision tree (DT) algorithm have been Decision-making is vital for effectiveness. Henceforth, the performance
compared to the BBM in terms of performance. of reliability analysis must be assessed for better efficacy. Fig. 12 shows
Table 10 displays the sensitivity, specificity, and precision of sev- the results of the reliability assessment simulation. Efficiency levels are
eral classifiers for identifying different types of seismic events. In the registered with greater values approximating 95.26% for the provided
current study, the BBM classifier registered the greatest classification model when the number of datasets for experimental implementation
precision at 92.52%, the highest mean sensitivity at 91.72%, and the increases. The suggested approach appears to be more reliable over
highest mean specificity at 91.01%, making it a strong candidate for large datasets.

10
M. Bhatia, T.A. Ahanger and A. Manocha Engineering Applications of Artificial Intelligence 120 (2023) 105856

Fig. 8. Seismic signals; (a) Earthquake, (b) Non-earthquake.

Fig. 9. Computational delay.

Fig. 10. Computational error.

11
M. Bhatia, T.A. Ahanger and A. Manocha Engineering Applications of Artificial Intelligence 120 (2023) 105856

Table 10
Classification Efficiency; (Precision (Prec), Specificity (Spec), Sensitivity (Sens)).
Model BBM Classifier C4.5 DT Support vector machine
Dataset Prec Spec Sens Prec Spec Sens Prec Spec Sens
15000 95.55% 92.05% 93.92% 92.92% 89.02% 90.77% 92.52% 89.15% 90.11%
30000 92.79% 91.95% 92.32% 91.52% 91.22% 91.72% 91.02% 91.79% 91.05%
45000 93.57% 91.99% 92.52% 91.52% 90.33% 92.35% 90.07% 90.53% 92.09%
60000 93.73% 90.95% 91.32% 92.52% 89.15% 90.59% 92.32% 90.73% 90.59%
75000 91.57% 92.55% 92.55% 91.22% 90.22% 92.09% 91.75% 91.27% 92.39%
90000 92.32% 90.32% 91.32% 92.21% 91.15% 91.07% 92.77% 91.02% 91.93%
105000 93.92% 89.93% 91.31% 91.27% 89.92% 90.25% 91.15% 90.79% 89.17%
120000 92.32% 90.55% 92.72% 92.12% 89.55% 91.15% 92.07% 89.99% 91.35%
135000 91.95% 89.35% 92.52% 90.55% 90.10% 92.05% 89.19% 90.52% 92.15%
150000 92.52% 91.01% 91.72% 91.32% 90.59% 91.10% 91.19% 90.15% 91.70%

Fig. 11. Prediction efficiency.

Fig. 12. Reliability analysis.

12
M. Bhatia, T.A. Ahanger and A. Manocha Engineering Applications of Artificial Intelligence 120 (2023) 105856

Fig. 13. Overall system stability.

Table 11 other hand, combines the procedure of key and signcrypted message
Training parameters of ANFIS. delivery with the expense of consensus.
Structure Value
Inputs 4 5. Conclusion
Fuzzy Rule 9
Membership Function Gaussian Earthquake prediction is a complicated branch of study in which
Number of Layers 6
the intensity of a future catastrophe is forecasted. Conspicuously, a
Output 1
Reaching Epochs Number 49 combination of IoT technology and edge–cloud computing is utilized to
Learning Rules Hybrid monitor and predict earthquakes more efficiently and accurately. The
proposed IoT–edge–cloud-based system provides an effective method
for prediction along with the transmission of data to the remote cloud
server. In the proposed model, seismological data is collected with
4.7. Stability efficacy location information. The edge layer functions as an entry point for
the prediction framework which extracts useful features and classifies
The proposed model is tested for lifespan over time using stability events based on real-time data. Moreover, cloud computing is used to
analysis. In other words, when the system is deployed over large store and compile data for the prediction of earthquake magnitude.
datasets for long-term evaluation, the system’s stability forecasts overall The implementation results demonstrate that the BBM classification
stabilization. The stability of a system is measured using the Mean technique has superior efficiency and precision at the edge layer level.
Absolute Shift (MAS). The MAS rating falls between 0 and 1, with 0 Compared to state-of-the-art classifier models, the BBM classifier’s pre-
indicating minimal stability and 1 indicating greatest stability. Fig. 13 diction accuracy of 92.52% is the maximum. Moreover, the perfor-
depicts the results of the suggested system’s stability analysis. It has mance of the proposed framework is measured in terms of reaction
been determined that the proposed model can register a minimum latency, and computational delay, in which optimal results were regis-
value of 0.58 and a maximum value of 0.81, resulting in an average tered. Conclusively, the presented technique is extremely effective and
value of 0.73. The suggested approach is very stable and suitable for efficient in provisioning an appropriate earthquake prediction system.
earthquake identification. For future research, edge layer considerations including sensor energy
efficiency, and resource usage can be examined. Security is another
4.8. Computational complexity vital aspect of future research exploration.

CRediT authorship contribution statement

It is necessary to calculate the system’s cost to assess the sug-
gested solution’s decision-making efficacy. In this method, two sorts
Munish Bhatia: Conceptualization, Methodology, Software. Tariq
of expenses are estimated: transaction costs and computation costs. Ahamed Ahanger: Data curation, Writing – original draft. Ankush
Computation cost is the time it takes the system to make a judgment Manocha: Visualization, Investigation.
on an event recorded at a certain time instance 𝛿𝑇 . The cost of a
transaction specifies the difficulty associated with the production and Declaration of competing interest
authentication of network blocks. The computed complexity is shown
in Table 13. The authors declare that they have no known competing finan-
Table 13 illustrates the computing cost of the data collection, pre- cial interests or personal relationships that could have appeared to
processing, and event detection processes. The transaction cost, on the influence the work reported in this paper.

Table 12
ANFIS model performance analysis.
Epoch Training dataset Training dataset Training dataset Testing dataset Testing dataset Testing dataset
R2 RMSE MAE R2 RMSE MAE
25 0.911 0.700 0.814 0.725 0.571 0.763
35 0.957 0.801 0.848 0.741 0.630 0.770
45 0.912 0.805 0.772 0.80 0.711 0.776
55 0.901 0.892 0.783 0.807 0.804 0.784

13
M. Bhatia, T.A. Ahanger and A. Manocha Engineering Applications of Artificial Intelligence 120 (2023) 105856

Table 13 Liao, Wu-Yu, Lee, En-Jui, Chen, Da-Yi, Chen, Po, Mu, Dawei, Wu, Yih-Min, 2022. RED-
Computational complexity. PAN: Real-time earthquake detection and phase-picking with multi-task attention
Sr. No. Cost type Complexity network. IEEE Trans. Geosci. Remote Sens..
1. Computation cost O(N log𝑛 ) Lin, Jie, Yu, Wei, Zhang, Nan, Yang, Xinyu, Zhang, Hanlin, Zhao, Wei, 2017. A survey
2. Transactional cost O(N-1 log𝑛 ) on internet of things: Architecture, enabling technologies, security and privacy, and
applications. IEEE Internet Things J. 4 (5), 1125–1142.
Mousavi, S Mostafa, Sheng, Yixiao, Zhu, Weiqiang, Beroza, Gregory C, 2019. Stanford
earthquake dataset (STEAD): A global data set of seismic signals for AI. IEEE Access
7, 179464–179476.
Data availability
Qazi, Abroon, Dickson, Alex, Quigley, John, Gaudenzi, Barbara, 2018. Supply chain
risk network management: A Bayesian belief network and expected utility based
The authors do not have permission to share data. approach for managing supply chain risks. Int. J. Prod. Econ. 196, 24–42.
Roy, Tamal Joyti, Mahmood, Md Ashiq, Roy, Diti, 2021. A machine learning model
References to predict earthquake utilizing neural network. In: 2021 International Conference
on Computer, Communication, Chemical, Materials and Electronic Engineering,
Abbasi, Mahdi, Mohammadi-Pasand, E., Khosravi, Mohammad Reza, 2021. Intelligent IC4ME2. IEEE, pp. 1–4.
workload allocation in IoT–Fog–cloud architecture towards mobile edge computing. Shen, Heather, Shen, Yang, 2021. Array-based convolutional neural networks for
Comput. Commun. 169, 71–80. automatic detection and 4D localization of earthquakes in Hawai ‘i. Seismol. Res.
Ahamed, Sabber, Daub, Eric G., 2021. Application of machine learning techniques to Lett. 92 (5), 2961–2971.
predict rupture propagation and arrest in 2-D dynamic earthquake simulations. Sun, Kai, Cui, Weicheng, Chen, Chi, 2021. Review of underwater sensing technologies
Geophys. J. Int. 224 (3), 1918–1929. and applications. Sensors 21 (23), 7849.
Bandyopadhyay, Debasis, Sen, Jaydip, 2011. Internet of things: Applications and Tao, Hongfeng, Cheng, Long, Qiu, Jier, Stojanovic, Vladimir, 2022. Few shot cross
challenges in technology and standardization. Wirel. Pers. Commun. 58 (1), 49–69. equipment fault diagnosis method based on parameter optimization and feature
Bassetti, Enrico, Panizzi, Emanuele, 2022. Earthquake detection at the edge: IoT mertic. Meas. Sci. Technol. 33 (11), 115005.
crowdsensing network. Information 13 (4), 195. Tehseen, Rabia, Farooq, Muhammad Shoaib, Abid, Adnan, 2021. A framework for the
Bhardwaj, Kartik Krishna, Khanna, Anirudh, Sharma, Deepak Kumar, Chhabra, An- prediction of earthquake using federated learning. PeerJ Comput. Sci. 7, e540.
shuman, 2019. Designing energy-efficient IoT-based intelligent transport system: Vaezi, Yoones, Van der Baan, Mirko, 2015. Comparison of the STA/LTA and power
Need, architecture, characteristics, challenges, and applications. Energy Conserv.
spectral density methods for microseismic event detection. Geophys. Suppl. Mon.
IoT Devices 209–233.
Not. R. Astron. Soc. 203 (3), 1896–1908.
Bhargava, Bharat, Pasari, Sumanta, 2022. Earthquake prediction using deep neural
Wibowo, Adi, Pratama, C, Sahara, David P, Heliani, LS, Rasyid, S, Akbar, Zharfan,
networks. In: 2022 8th International Conference on Advanced Computing and
Muttaqy, Faiz, Sudrajat, Ajat, 2021. Earthquake early warning system using ncheck
Communication Systems, vol. 1. ICACCS, IEEE, pp. 476–479.
and hard-shared orthogonal multitarget regression on deep learning. IEEE Geosci.
Brous, Paul, Janssen, Marijn, Herder, Paulien, 2019. The dual effects of the Internet
Remote Sens. Lett. 19, 1–5.
of Things (IoT): A systematic review of the benefits and risks of IoT adoption by
organizations. Int. J. Inf. Manage. 1–17. Won, Jongbin, Park, Junyoung, Park, Jong-Woong, Kim, In-Ho, 2020. Bleseis: Low-cost
Cremen, Gemma, Zuccolo, Elisa, Galasso, Carmine, 2021. Accuracy and uncertainty IoT sensor for smart earthquake detection and notification. Sensors 20 (10), 2963.
analysis of selected methodological approaches to earthquake early warning in Xiao, Zhiwen, Xu, Xin, Xing, Huanlai, Luo, Shouxi, Dai, Penglin, Zhan, Dawei, 2021.
Europe. Seismol. Res. Lett. 92 (4), 2321–2332. RTFN: A robust temporal feature network for time series classification. Inform. Sci.
Falanga, Mariarosaria, De Lauro, Enza, Petrosino, Simona, Rincon-Yanez, Diego, Sen- 571, 65–86.
atore, Sabrina, 2022. Semantically enhanced IoT-oriented seismic event detection: Xin, Xilin, Tu, Yidong, Stojanovic, Vladimir, Wang, Hai, Shi, Kaibo, He, Shuping,
An application to colima and vesuvius volcanoes. IEEE Internet Things J. 9 (12), Pan, Tianhong, 2022. Online reinforcement learning multiplayer non-zero sum
9789–9803. games of continuous-time Markov jump linear systems. Appl. Math. Comput. 412,
Gubbi, Jayavardhana, Buyya, Rajkumar, Marusic, Slaven, Palaniswami, Marimuthu, 126537.
2013. Internet of Things (IoT): A vision, architectural elements, and future Xing, Huanlai, Xiao, Zhiwen, Qu, Rong, Zhu, Zonghai, Zhao, Bowen, 2022. An efficient
directions. Future Gener. Comput. Syst. 29 (7), 1645–1660. federated distillation learning system for multitask time series classification. IEEE
Imeraj, Petrit, Ali, Maaruf, Imeraj, Gent, 2022. Geotechnical management of isolated Trans. Instrum. Meas. 71, 1–12.
sustainable alpine communities. Ann. Emerg. Technol. Comput. 6 (1), 60–73. Xu, Zhilu, Li, Xiaodi, Stojanovic, Vladimir, 2021. Exponential stability of nonlinear
Iwendi, Celestine, Maddikunta, Praveen Kumar Reddy, Gadekallu, Thippa Reddy, state-dependent delayed impulsive systems with applications. Nonlinear Anal.
Lakshmanna, Kuruva, Bashir, Ali Kashif, Piran, Md Jalil, 2020. A metaheuristic Hybrid Syst. 42, 101088.
optimization approach for energy efficiency in the IoT networks. Softw. - Pract. Yu, Wei, Liang, Fan, He, Xiaofei, Hatcher, William Grant, Lu, Chao, Lin, Jie,
Exp. 1–14.
Yang, Xinyu, 2017. A survey on the edge computing for the Internet of Things.
Kumar, Satish, Vig, Renu, Kapur, Pawan, 2018. Development of earthquake event
IEEE Access 6, 6900–6919.
detection technique based on STA/LTA algorithm for seismic alert system. J. Geol.
Soc. India 92 (6), 679–686.

14